Labor Representation in Governance as an Insurance Mechanism

Labor Representation in Governance as an Insurance Mechanism Abstract We hypothesize that labor participation in governance helps improve risk sharing between employees and employers. It provides an ex post mechanism to enforce implicit insurance contracts protecting employees against adverse shocks. Results based on German establishment-level data show that skilled employees of firms with 50% labor representation on boards are protected against layoffs during adverse industry shocks. They pay an insurance premium of 3.3% in the form of lower wages. Unskilled blue-collar workers are unprotected against shocks. Our evidence suggests that workers capture all the gains from improved risk sharing, whereas shareholders are no better or worse off than without codetermination. 1. Introduction The analysis of risk, hedging, and risk-sharing has been at the core of finance research since the beginnings of modern finance, leading to a remarkable level of sophistication in the analyses of asset allocation, hedging of risky positions, and the pricing of risky securities. In contrast, the risk associated with human capital—a significant portion of wealth and arguably the largest source of risk, and most difficult to diversify for average workers—has received limited attention.1 More recently, however, a new stream of research on finance and labor has demonstrated a critical role of human capital and employment risk in understanding risk-return trade-offs, indirect costs of financial distress, optimal capital structure, and restructuring.2 In this paper, we focus on sharing human capital risk between employers and employees. Our point of departure is implicit contract theory, which holds that the risk-neutral principals of the firm provide human capital protection to risk-averse employees against adverse shocks. Employees accept lower wages in return for insuring their employment and wages (Baily, 1974; Gordon, 1974; Azariadis, 1975; Rudanko, 2011). For employees to enter such implicit insurance contracts requires a means to monitor and enforce contracts when they need protection. Enforceability is essential for implicit insurance contracts to work. If workers accept lower wages today, how can shareholders commit not to lay them off in the future when it is in their best interest to do so? The literature has appealed to arguments based on reputation and investigated the worker-friendly preferences of family owners (e.g., Sraer and Thesmar, 2007; Ellul, Pagano, and Schivardi, 2017).3 To our knowledge, ours is the first study of a contractual mechanism to enforce implicit employment contracts.4 Specifically, we hypothesize that labor representation on corporate boards is such a mechanism to ensure employment guarantees will be honored. In the 1970s, Germany led a trend toward more labor representation, and most countries now grant workers some degree of representation, at least in the form of work councils at the firm or establishment level. The world map in Figure 1 shows countries with labor representation at the firm level and/or at board level as of 2015.5 By 2015, only two countries among the OECD countries—the USA and Singapore—did not grant workers any representation. Figure 1. View largeDownload slide World map of labor representation. Notes: This figure plots labor representation in works councils and corporate boards around the world. We categorize countries in three groups (1) no labor representation (light gray), (2) only work councils (gray), and (3) work councils and board representation (dark gray). This classification is based on variables 30 and 31 in the CBR-LRI data set provided by the Centre for Business Research at the University of Cambridge, UK. [See Adams et al. (2017) and https://www.repository.cam.ac.uk/handle/1810/262929 for a documentation.] For each country, we use the latest information available (for most countries that is 2015). The figure does not display more granular information about the power of work councils and the degree of board representation available in the data set. Figure 1. View largeDownload slide World map of labor representation. Notes: This figure plots labor representation in works councils and corporate boards around the world. We categorize countries in three groups (1) no labor representation (light gray), (2) only work councils (gray), and (3) work councils and board representation (dark gray). This classification is based on variables 30 and 31 in the CBR-LRI data set provided by the Centre for Business Research at the University of Cambridge, UK. [See Adams et al. (2017) and https://www.repository.cam.ac.uk/handle/1810/262929 for a documentation.] For each country, we use the latest information available (for most countries that is 2015). The figure does not display more granular information about the power of work councils and the degree of board representation available in the data set. To test our hypothesis on labor representation on corporate boards, we examine the German system, which requires 50% employee representation on supervisory boards—hereafter referred to as parity-codetermination—when firms have more than 2,000 employees working in Germany. Germany offers a laboratory in which otherwise similar companies have different degrees of labor representation. In addition, the Institute of Employment Research (IAB) in Germany provides detailed, high quality panel data on employment and wages for all establishments located in Germany over our sample period 1990–2008. Establishment level data allow cleaner identification devoid of false detection of insurance that can occur with firm level data. For example, if a firm fires workers in a plant—no employment protection—and adds the same number of different workers at another plant in a different industry, aggregate employment at the firm will be unchanged, making it appear as if the firm provided full employment protection. Establishment data also allow us to construct adverse industry shocks using the establishments of other, non-sample firms in the same industry.6 Our initial identification strategy is a difference-in-differences approach. We are mindful of the fact that parity co-determination is a discrete function of the number of employees in Germany, which is loosely related to firm size. We employ three different strategies to separate the parity-codetermination effect from the effect of employment level in Germany: (1) the number of employees in Germany interacted with the shock indicator; (2) placebo tests using false thresholds different from the parity threshold of 2,000; and (3) regression discontinuity (RD) analyses around the 2,000-employee threshold for parity codetermination. All results support the same conclusions. We also check for the potential endogeneity of parity status and find no evidence that either workers or shareholders successfully manipulate the number of employees in Germany to maintain or avoid the parity status. It appears that the parity status is largely a result of employment decisions driven by business and operational considerations. We find white-collar and skilled blue-collar workers of parity-codetermined firms are protected against layoffs during adverse industry shocks, while those working for non-parity firms are not. Surprisingly, unskilled blue-collar workers of the same parity firms are unprotected from layoffs during industry shocks. We attribute this observation to the absence of unskilled blue-collar worker representation on supervisory boards. We could not identify a single unskilled blue-collar worker among labor representatives, which underscores the importance of worker participation in governance for an effective implementation of implicit employment insurance contracts. From the evidence on employment protection alone, we cannot infer that firms honor implicit insurance contracts against adverse industry shocks. Parity codetermination may result in employee entrenchment sufficient to prevent layoffs (e.g., Pagano and Volpin, 2005; Atanassov and Kim, 2009); employees may not need to pay an insurance premium in the form of lower wages for the increase in job security. There are no theoretical reasons to believe that codetermination makes employees just strong enough to enforce an implicit insurance contract, without causing employee entrenchment. Therefore, whether codetermination leads to an insurance relationship or employee entrenchment is an empirical question. We find that workers with vocational and higher educational qualifications, two categories that cover most skilled blue-collar and white-collar workers, receive significantly lower wages at parity-codetermined firms. Their average wage concession is about 3.3%, supporting the notion of a quid pro quo in exchange for employment insurance. Are the wage differentials between parity-codetermined firms and non-parity firms sufficient to compensate shareholders for risks they bear in providing insurance? Providing insurance is costly, because it increases operating leverage. We find parity firms’ operating profits and valuation suffer more during shock periods compared with non-parity firms. While we observe that employees make concessions, this concession may still be too low to compensate shareholders for the additional risk. Short of a structural model of the costs of reduced flexibility to firms, we cannot directly infer whether the wage concessions are adequate.7 We investigate this issue indirectly by relating the parity status to firm performance as measured through the cycle over the non-shock and shock periods. We find no significant relation between parity codetermination and operating performance or shareholder value, which is in line with the overall conclusions reached in the literature on the effect of German codetermination on firm performance.8 On average, parity firms perform no worse or better than non-parity firms. Hence, whatever efficiency gains arise from improved risk sharing, workers seem to capture all these gains, making the minimum wage concessions that allow shareholders to be no worse off, but also no better off than without codetermination. The lack of shareholder gains may explain why firms do not voluntarily adopt codetermination, an observation often used to support arguments against worker participation in governance (Jensen and Meckling, 1979). The fact that shareholders do not lose from parity codetermination may also be the reason why we do not find evidence for shareholders manipulating the number of employees in Germany so as to evade parity status. Parity codetermination seems to be, if anything, a second-order consideration from the point of view of shareholder value, so that other, first-order business and operating considerations drive firms’ employment decisions in Germany. Our paper makes multiple contributions. We contribute to the literature on implicit insurance contracts by focusing on how such contracts are enforced.9 This literature has relied mostly on reputation as an enforcement mechanism, which is fragile when companies undergo ownership changes (Shleifer and Summers, 1988). To our knowledge, we are first to think of codetermination as an enforcement mechanism in an implicit insurance relationship. The closest theoretical contribution in this context is Malcomson (1983), who argues that unions may act as an enforcement mechanism for implicit wage contracts. An earlier literature on life-long employment in Japanese firms views the security of long-term employment relationships as mechanisms to facilitate greater investments in firm-specific human capital (Hashimoto and Raisian, 1985; Abraham and Houseman, 1989). However, these papers do not specify how firms commit to protecting long-term employment relationships. Faleye, Mehrotra, and Morck (2006) and Kim and Ouimet (2014) analyze the implications of employee stock ownership in the USA, but do not explore the possibility that employee equity ownership might help enforce implicit contracts. A strand of the family firm literature argues that it is easier for family firms to commit to implicit contracts because their managers have a longer time horizon.10Sraer and Thesmar (2007) show that family firms in France insure workers against employment shocks; however, they recognize that even with the long-term perspective provided by dynastic management, family-owners face incentives to renege on prior promises to insure workers during downturns. To explain their finding, they rely on preferences “irrationally tying top management to employees” (p. 729). Our argument is based on an explicit mechanism and self-interest and not on the preferences of decision-makers. Ellul, Pagano, and Schivardi (2017) find evidence of employment insurance in family firms for a cross-section of countries. However, insurance by family firms seems only partial; workers accept higher fluctuations in wages and are not covered against shocks originating from banking crises, when access to capital dries up. Similarly, D’Aurizio and Romano (2013) show that family firms in Italy do not insure workers employed in locations further away from company headquarters. Ours is also the first to provide a quantitative estimate for the insurance premium based on establishment-level data within the context of the quid pro quo of the implicit insurance relationship.11Guiso, Pistaferri, and Schivardi (2005) examine a matched employee-firm panel of Italian firms and show that firms protect workers against wage shocks; by contrast, our study analyzes employment shocks. In addition, Guiso, Pistaferri, and Schivardi (2005) lack a benchmark and a control group of firms that provide less or no insurance. Similarly, Beaudry and DiNardo (1991) provide evidence for the implications of implicit-contracting models for wage dynamics over the business cycle, but without looking at enforcement mechanisms or insurance premiums. In a broader context, our paper provides evidence that balances critical views on labor representation in governance. The critics argue that it is detrimental to the interests of investors, because sharing control rights with labor through representation on the board can have a harmful impact on firm performance. We find German codetermination improves risk sharing without hurting firm performance, suggesting a gain to employee welfare without damaging shareholders. 2. Theoretical Considerations and Hypothesis Development In the simplest version of the insurance argument, diversified, risk-neutral investors (firms and entrepreneurs) insure risk-averse workers against firm-level shocks by promising them a wage that does not vary with their productivity from period to period. In most models, insurance affects wages as well as the employment status of workers (e.g., Baily, 1974; Azariadis, 1975; Holmstrom, 1983; Gamber, 1988). Workers give up a portion of their wages in return for protection against adverse shocks to wages and employment and receive wages that are sometimes above and sometimes below their marginal product. The insurance provided to workers shifts employment risk from workers to investors, but an effective risk transfer requires a commitment device. Workers first give up a portion of their wages. Later they have to count on firms’ honoring contracts in the event of adverse shocks. The theoretical literature on implicit contracts simply assumes that firms have the ability to commit to long-term contracts. However, commitment is crucial. Workers often have to locate close to the firm or make investments in firm-specific human capital well before the firm has to honor its side of the bargain, which makes them vulnerable to breaches of implicit contracts (Shleifer and Summers, 1988). We argue that parity-codetermination serves as a commitment device by allowing workers to influence employment decisions. Hypothesis 1: Parity-codetermination provides an ex-post enforcement mechanism, which ensures workers receive protection against adverse shocks to employment and wages. This hypothesis explicitly incorporates employment guarantees, which imply that firms do not fire workers even when layoffs are ex post efficient. If workers and firms could engage in frictionless bargaining, they would always agree to sever the employment relationship ex post by negotiating suitable transfers, which makes ex post inefficient employment of workers unsustainable. Models with employment insurance implicitly rule out such frictionless bargaining between firms and workers.12 Specifically, our hypothesis relies labor market frictions such as mobility costs (Baily, 1974; Beaudry and DiNardo, 1991) or search frictions (Rudanko, 2011), so that firms do not have to pay the competitive market wage in every period. It also requires the assumption that firms have better access to capital markets than workers and therefore enjoy a privileged position to insure workers; otherwise, workers could insure themselves directly through capital markets (e.g., Berk and Walden, 2013).13 When firms act as insurers to workers, they enter a quid-pro-quo relationship, whereby workers receive job guarantees in return for an insurance premium. The standard form of compensating firms in the implicit contracting literature is through lower wages. We hypothesize that only parity-codetermined firms provide insurance to workers, so only parity-codetermined firms receive insurance payments from workers in the form of lower wages: Hypothesis 2: Firms with parity-codetermination pay on average lower wages than non-parity firms. Lower wages are not necessarily the only way to compensate shareholders, but in all likelihood, lower wages are the most efficient way to provide an insurance premium because they do not distort the allocation of resources. 3. Institutional Background, Data, and Empirical Design 3.1 The German Governance System and Wage-Setting Process Germany has a two-tier board system. The management board (Vorstand) manages day-to-day operations and the supervisory board (Aufsichtsrat) monitors the management board; approves key strategic decisions; and appoints and dismisses management board members, including the CEO, and decides their compensation. The two boards are strictly separated and no member of one board can be a member of the other for the same company at the same time. The structure, size, and composition of the supervisory board are regulated by the German stock corporation act (Aktiengesetz) and the codetermination act (Mitbestimmungsgesetz) as well as other laws. The regulation requires that half of the supervisory board members are worker representatives for corporations with more than 2,000 employees working in Germany. For corporations with more than 500 up to 2,000 employees in Germany, one-third of the members of the supervisory board have to represent workers. Worker representatives are elected by the company’s workers. Depending on the size of the supervisory board, two or three seats of the worker representatives are reserved for union representatives. One seat is always reserved for a representative from middle management (leitende Angestellte).14 The chairman of the board casts the deciding vote in case of a tie. Wages in German firms are often set through collective bargaining agreements between trade unions and employers’ associations. Some firms are covered by sectoral agreements, whereas others negotiate firm-level agreements. Agreements are binding on all workers as German law forbids discriminatory wage policies that disadvantage non-union members. However, firms still have discretion over wages through firm-level wage agreements and the way in which they categorize workers, resulting in significant variation in wages across firms (Kohn and Lembcke, 2007). Unionization in Germany is generally not high and is also falling. Based on survey data, Schnabel and Wagner (2007) estimate it to be about one-third in 1992, declining to around 20% in 2004. Coverage by unionized wage agreements also has declined (Hassel, 1999). See Institutional Details in the Online Appendix for more information on German codetermination and collective bargaining. 3.2 Data Employment and wage data at the establishment level are obtained from the IAB. The IAB is the research organization of the German employment agency, the Bundesagentur für Arbeit (BA). The BA collects worker and employer contributions to unemployment insurance and distributes unemployment benefits. All German businesses are required to report detailed information on employment and wages to the BA. The individual-level data are aggregated at the establishment level, made anonymous, and offered for scientific use by the IAB. An establishment is any facility having a separate physical address, such as a factory, service station, restaurant, or office building. The IAB owns detailed establishment-level data on industry, location, employment, employee education, age, nationality, and wages. The data are in the form of establishment-level statistics, such as sums, medians, quartiles, and averages on wages and employment according to different classifications and breakdowns. We use the first three-digits of the NACE code, which identifies 224 separate economic sub-sectors (groups).15 The NACE (Rev. 1.1) classification is available from the IAB database only for 2003 and afterward. We assign an establishment’s NACE (Rev. 1.1) classification in 2003 to all its prior sample years. Some establishments may have changed their industry classification prior to 2003, in which case they would receive new establishment IDs. To avoid assigning incorrect industry codes, we drop all establishments changing industry classifications over time in the entire IAB database, as well as establishment-year observations with missing information on industry classification. These screens yield approximately 33.4 million establishment-year observations on approximately 3.5 million establishments for the sample period 1990–2008. Our sample firms are drawn from all companies included in the two main German stock market indices, DAX and MDAX, at any point over the 19-year period from 1990 to 2008. There are 184 such firms, for which we hand-collect data on the composition of the supervisory board from annual reports and Hoppenstedt company profiles. We do not include non-listed firms, because information on worker representation on their boards is usually unavailable. Stock market data come from Datastream, balance sheet and accounting data from Worldscope. At our request, the IAB matched our sample of listed firms with their establishment-level database using an automatic procedure, based on company name and address information (city, zip code, street, and house number). All cases not unambiguously matched by the automatic matching procedure are checked by hand to avoid mismatching. This procedure results in 284,538 establishment-years matched to 2,168 firm-years for 142 of the 184 firms. Of these, 1,461 firm-years (67.4%) are subject to parity codetermination, 442 (20.4%) subject to one-third codetermination, and the remaining 265 firm-years are not covered by any codetermination arrangement. The matching was performed for 2004, 2005, and 2006. Firms are dropped if they do not exist during the period 2004–2006, because we cannot match them to the IAB data. IAB does not have a firm identifier, which is why manual matching is necessary. All establishments are matched only once to our sample firms and, if establishments were sold prior to 2004, they do not enter our sample because IAB cannot match them. This matching procedure does not allow us to identify changes in establishment ownership after 2006. (At the time of matching establishments to firms, establishment data were not available for 2007 or 2008.) Thus, if an establishment belonging to a parity (non-parity) firm is sold to a non-parity (parity) firm in 2007 or 2008, it will be treated as if it still belonged to a parity (non-parity) firm after the sale. This will blur the distinction between parity and non-parity status of the establishment and potentially lead to attenuation bias working against finding significant results. The IAB distinguishes employees by occupational status. The three most important groups are unskilled blue-collar workers, skilled blue-collar workers, and white-collar employees.16 The IAB also reports three different qualification levels at each establishment by educational and vocational qualifications: (1) low-qualified employees possess neither an upper secondary school graduation certificate as their highest school qualification nor a vocational qualification. (2) Qualified employees have either an upper secondary school graduation certificate as their highest school qualification or a vocational qualification. (3) Highly qualified employees have a degree from a specialized college of higher education or a university degree.17 Unfortunately, over our sample period an increasing number of firms stopped reporting information on qualifications, leading to a steady increase in the number of employees with unknown qualifications. If firms’ decision not to report their employees’ qualification is random, the increasing trend in the number of employees with unknown qualifications should not bias our results. We include only establishments with more than 50 employees. These establishments almost always have works councils, which have significant information rights as well as the right to demand compensation for dismissals and may exert some influence at the establishment level (see Wiedemann, 1980). Nonetheless, our results are robust to an alternative sample construction with a cutoff point at ten employees. While fiscal years of German firms are mostly from January to December, establishment years for IAB data are from July to June. We therefore lag all variables from Worldscope by 6 months relative to IAB years. Effectively, we assign year-end values from Worldscope to June 30 information on employment and wages of the next year. Table I provides summary statistics. Monetary units are normalized to 2005 euros. Panels A and B provide statistics at the establishment and firm level, respectively. All accounting and market variables are taken from Worldscope and Datastream. None of the firm-level variables in Panel B are from IAB. Variable definitions and data sources are listed in Table II. Our sample selection criteria tilt the sample toward large, mature firms (mean firm age: 85 years) with high book-value leverage and low betas. About two-thirds of firm-year observations are for parity-codetermined firms. The distributions of sales, market capitalization, and the number of establishments are highly skewed; the mean exceeds the 75th percentile for all three size measures. Table I. Descriptive statistics This table presents descriptive statistics for all key variables used in this paper. Panel A reports summary statistics at the establishment level. N is the number of establishment-years the respective variable is available. Only establishments with more than 50 employees are used. DailyWageP50LQ is the median daily gross wage for low-qualified employees. DailyWageP50Q is the median daily gross wage for qualified employees. DailyWageP50HQ is the median daily gross wage for highly qualified employees. Monetary units are normalized to 2005 euros. Panel B reports summary statistics at the firm level. N is the number of firm-years the respective variable is available. Panel A Variable Mean Median Std Min P25 P75 Max N Number of employees 517.47 148 2,099.29 51 81 346 61,380 54,042 Number of unskilled 97.14 5 700.35 0 0 31 32,733 54,042 Number of skilled 103.32 10 584.98 0 0 49 19,658 54,042 Number of white collar 223.80 64 894.00 0 31 148 29,084 54,042 DailyWageP25 81.73 76.66 27.98 1.02 61.20 97.99 214.42 53,956 DailyWageP50 94.23 88.38 32.59 7.66 69.56 113.53 228.92 53,956 DailyWageP75 108.76 104.68 34.87 7.66 81.01 132.69 228.92 53,956 DailyWageP50LQ 82.50 77.52 29.07 1.87 61.99 99.04 781.59 44,783 DailyWageP50Q 93.11 88.53 30.25 7.66 70.37 110.98 199.33 53,811 DailyWageP50HQ 124.56 126.03 34.84 0.60 99.96 150.47 335.52 40,459 EstAge 15.64 16 9.880 0 6 24 33 54,042 MedianEmplAge 38.84 39 4.973 17 36 42 60 54,042 RatioWhiteCollar 0.477 0.446 0.297 0 0.228 0.746 1 54,042 Shock 0.058 0 0.233 0 0 0 1 52,756 HHI 0.405 0.304 0.314 0.007 0.145 0.563 1 54,042 Panel B Variable Mean Median Std Min P25 P75 Max N Beta 0.678 0.620 0.467 −3.198 0.324 0.997 3.002 1,832 Firm age 84.5 86 53.3 0 36 124 259 1,989 Leverage 0.392 0.358 0.273 0 0.169 0.582 0.996 2,052 MCap (bn €) 5.382 0.976 12.327 0.003 0.297 4.141 198.186 2,002 NetPPE (bn €) 2.451 0.288 7.337 0 0.092 1.349 74.003 2,057 Parity 0.674 1 0.469 0 0 1 1 2,168 ROA 0.074 0.069 0.096 −1.152 0.031 0.110 0.671 1,926 Sales (bn €) 8.663 1.793 17.950 0.005 0.591 7.694 157.5 2,064 Tobin’s Q 1.546 1.224 1.010 0.454 1.054 1.602 12.529 1,991 Firm shock 0.048 0 0.177 0 0 0 1 2,126 Number of establishments 26.045 6 56.827 1 3 20 497 2,075 Panel A Variable Mean Median Std Min P25 P75 Max N Number of employees 517.47 148 2,099.29 51 81 346 61,380 54,042 Number of unskilled 97.14 5 700.35 0 0 31 32,733 54,042 Number of skilled 103.32 10 584.98 0 0 49 19,658 54,042 Number of white collar 223.80 64 894.00 0 31 148 29,084 54,042 DailyWageP25 81.73 76.66 27.98 1.02 61.20 97.99 214.42 53,956 DailyWageP50 94.23 88.38 32.59 7.66 69.56 113.53 228.92 53,956 DailyWageP75 108.76 104.68 34.87 7.66 81.01 132.69 228.92 53,956 DailyWageP50LQ 82.50 77.52 29.07 1.87 61.99 99.04 781.59 44,783 DailyWageP50Q 93.11 88.53 30.25 7.66 70.37 110.98 199.33 53,811 DailyWageP50HQ 124.56 126.03 34.84 0.60 99.96 150.47 335.52 40,459 EstAge 15.64 16 9.880 0 6 24 33 54,042 MedianEmplAge 38.84 39 4.973 17 36 42 60 54,042 RatioWhiteCollar 0.477 0.446 0.297 0 0.228 0.746 1 54,042 Shock 0.058 0 0.233 0 0 0 1 52,756 HHI 0.405 0.304 0.314 0.007 0.145 0.563 1 54,042 Panel B Variable Mean Median Std Min P25 P75 Max N Beta 0.678 0.620 0.467 −3.198 0.324 0.997 3.002 1,832 Firm age 84.5 86 53.3 0 36 124 259 1,989 Leverage 0.392 0.358 0.273 0 0.169 0.582 0.996 2,052 MCap (bn €) 5.382 0.976 12.327 0.003 0.297 4.141 198.186 2,002 NetPPE (bn €) 2.451 0.288 7.337 0 0.092 1.349 74.003 2,057 Parity 0.674 1 0.469 0 0 1 1 2,168 ROA 0.074 0.069 0.096 −1.152 0.031 0.110 0.671 1,926 Sales (bn €) 8.663 1.793 17.950 0.005 0.591 7.694 157.5 2,064 Tobin’s Q 1.546 1.224 1.010 0.454 1.054 1.602 12.529 1,991 Firm shock 0.048 0 0.177 0 0 0 1 2,126 Number of establishments 26.045 6 56.827 1 3 20 497 2,075 Table I. Descriptive statistics This table presents descriptive statistics for all key variables used in this paper. Panel A reports summary statistics at the establishment level. N is the number of establishment-years the respective variable is available. Only establishments with more than 50 employees are used. DailyWageP50LQ is the median daily gross wage for low-qualified employees. DailyWageP50Q is the median daily gross wage for qualified employees. DailyWageP50HQ is the median daily gross wage for highly qualified employees. Monetary units are normalized to 2005 euros. Panel B reports summary statistics at the firm level. N is the number of firm-years the respective variable is available. Panel A Variable Mean Median Std Min P25 P75 Max N Number of employees 517.47 148 2,099.29 51 81 346 61,380 54,042 Number of unskilled 97.14 5 700.35 0 0 31 32,733 54,042 Number of skilled 103.32 10 584.98 0 0 49 19,658 54,042 Number of white collar 223.80 64 894.00 0 31 148 29,084 54,042 DailyWageP25 81.73 76.66 27.98 1.02 61.20 97.99 214.42 53,956 DailyWageP50 94.23 88.38 32.59 7.66 69.56 113.53 228.92 53,956 DailyWageP75 108.76 104.68 34.87 7.66 81.01 132.69 228.92 53,956 DailyWageP50LQ 82.50 77.52 29.07 1.87 61.99 99.04 781.59 44,783 DailyWageP50Q 93.11 88.53 30.25 7.66 70.37 110.98 199.33 53,811 DailyWageP50HQ 124.56 126.03 34.84 0.60 99.96 150.47 335.52 40,459 EstAge 15.64 16 9.880 0 6 24 33 54,042 MedianEmplAge 38.84 39 4.973 17 36 42 60 54,042 RatioWhiteCollar 0.477 0.446 0.297 0 0.228 0.746 1 54,042 Shock 0.058 0 0.233 0 0 0 1 52,756 HHI 0.405 0.304 0.314 0.007 0.145 0.563 1 54,042 Panel B Variable Mean Median Std Min P25 P75 Max N Beta 0.678 0.620 0.467 −3.198 0.324 0.997 3.002 1,832 Firm age 84.5 86 53.3 0 36 124 259 1,989 Leverage 0.392 0.358 0.273 0 0.169 0.582 0.996 2,052 MCap (bn €) 5.382 0.976 12.327 0.003 0.297 4.141 198.186 2,002 NetPPE (bn €) 2.451 0.288 7.337 0 0.092 1.349 74.003 2,057 Parity 0.674 1 0.469 0 0 1 1 2,168 ROA 0.074 0.069 0.096 −1.152 0.031 0.110 0.671 1,926 Sales (bn €) 8.663 1.793 17.950 0.005 0.591 7.694 157.5 2,064 Tobin’s Q 1.546 1.224 1.010 0.454 1.054 1.602 12.529 1,991 Firm shock 0.048 0 0.177 0 0 0 1 2,126 Number of establishments 26.045 6 56.827 1 3 20 497 2,075 Panel A Variable Mean Median Std Min P25 P75 Max N Number of employees 517.47 148 2,099.29 51 81 346 61,380 54,042 Number of unskilled 97.14 5 700.35 0 0 31 32,733 54,042 Number of skilled 103.32 10 584.98 0 0 49 19,658 54,042 Number of white collar 223.80 64 894.00 0 31 148 29,084 54,042 DailyWageP25 81.73 76.66 27.98 1.02 61.20 97.99 214.42 53,956 DailyWageP50 94.23 88.38 32.59 7.66 69.56 113.53 228.92 53,956 DailyWageP75 108.76 104.68 34.87 7.66 81.01 132.69 228.92 53,956 DailyWageP50LQ 82.50 77.52 29.07 1.87 61.99 99.04 781.59 44,783 DailyWageP50Q 93.11 88.53 30.25 7.66 70.37 110.98 199.33 53,811 DailyWageP50HQ 124.56 126.03 34.84 0.60 99.96 150.47 335.52 40,459 EstAge 15.64 16 9.880 0 6 24 33 54,042 MedianEmplAge 38.84 39 4.973 17 36 42 60 54,042 RatioWhiteCollar 0.477 0.446 0.297 0 0.228 0.746 1 54,042 Shock 0.058 0 0.233 0 0 0 1 52,756 HHI 0.405 0.304 0.314 0.007 0.145 0.563 1 54,042 Panel B Variable Mean Median Std Min P25 P75 Max N Beta 0.678 0.620 0.467 −3.198 0.324 0.997 3.002 1,832 Firm age 84.5 86 53.3 0 36 124 259 1,989 Leverage 0.392 0.358 0.273 0 0.169 0.582 0.996 2,052 MCap (bn €) 5.382 0.976 12.327 0.003 0.297 4.141 198.186 2,002 NetPPE (bn €) 2.451 0.288 7.337 0 0.092 1.349 74.003 2,057 Parity 0.674 1 0.469 0 0 1 1 2,168 ROA 0.074 0.069 0.096 −1.152 0.031 0.110 0.671 1,926 Sales (bn €) 8.663 1.793 17.950 0.005 0.591 7.694 157.5 2,064 Tobin’s Q 1.546 1.224 1.010 0.454 1.054 1.602 12.529 1,991 Firm shock 0.048 0 0.177 0 0 0 1 2,126 Number of establishments 26.045 6 56.827 1 3 20 497 2,075 Table II. Variable definitions This table defines all variables used in this paper. Board data are taken from Hoppenstedt company profiles and annual reports. Employment and wage data are from the IAB Establishment History Panel. Accounting data are taken from Worldscope and market data from Datastream. The numbers in brackets refer to Worldscope items, taken from the Worldscope Data Definition Guide. Variable Description Source Number of employees Total number of employees in the establishment IAB Number of skilled Number of skilled (blue-collar) employees (at least vocational training) IAB Number of unskilled Number of unskilled (blue-collar) employees (no formal qualification) IAB Number of white collar Number of white-collar employees (at least vocational training) IAB Beta CAPM beta estimated over the prior calendar year using daily returns Datastream EstAge Age of the establishment in year IAB Firm employees Sum of all employees across all establishments of the firm in Germany IAB Firm age Age of the firm in year Worldscope HHI Herfindahl–Hirschman index calculated using number of employees of all establishments with more than 50 employees in the same county and industry (three-digit NACE code). IAB Leverage =Total debt [03255]/(total debt+common equity [03501]) Worldscope MCap Market capitalization [08001] Worldscope MedianEmplAge Median age of all employees in the establishment IAB NetPPE Net property, plant, and equipment [02501] Worldscope Parity =1 if 50% of all members of the company’s supervisory board are classified as employee representatives Hoppenstedt, annual reports PPE dummy =1 if NetPPE [02501] declines by more than 15% Worldscope RatioWhiteCollar =#WhiteCollar/#Employees IAB ROA =EBITDAt [18198]/((total assetst [02999]+total assetst–1)/2) Worldscope Sales =Net sales or revenues [01001] in 2005 euros Worldscope Shock =1 if employment in non-sample establishments in the same industry (three-digit NACE-code) as the establishment decreases by more than 5% with no increase in employment in the following year. A detailed description of the definition is provided in Section 3.3.2. IAB DailyWageP25 1st quartile of gross average daily wage for all full-time employees in 2005 euros IAB DailyWageP50 Median of gross average daily wage for all full-time employees in 2005 euros IAB DailyWageP75 Third quartile of gross average daily wage for all full-time employees in 2005 euros IAB Tobin’s Q = (market capitalization [08001]+total assets [02999] – common equity [03501])/total assets Worldscope Variable Description Source Number of employees Total number of employees in the establishment IAB Number of skilled Number of skilled (blue-collar) employees (at least vocational training) IAB Number of unskilled Number of unskilled (blue-collar) employees (no formal qualification) IAB Number of white collar Number of white-collar employees (at least vocational training) IAB Beta CAPM beta estimated over the prior calendar year using daily returns Datastream EstAge Age of the establishment in year IAB Firm employees Sum of all employees across all establishments of the firm in Germany IAB Firm age Age of the firm in year Worldscope HHI Herfindahl–Hirschman index calculated using number of employees of all establishments with more than 50 employees in the same county and industry (three-digit NACE code). IAB Leverage =Total debt [03255]/(total debt+common equity [03501]) Worldscope MCap Market capitalization [08001] Worldscope MedianEmplAge Median age of all employees in the establishment IAB NetPPE Net property, plant, and equipment [02501] Worldscope Parity =1 if 50% of all members of the company’s supervisory board are classified as employee representatives Hoppenstedt, annual reports PPE dummy =1 if NetPPE [02501] declines by more than 15% Worldscope RatioWhiteCollar =#WhiteCollar/#Employees IAB ROA =EBITDAt [18198]/((total assetst [02999]+total assetst–1)/2) Worldscope Sales =Net sales or revenues [01001] in 2005 euros Worldscope Shock =1 if employment in non-sample establishments in the same industry (three-digit NACE-code) as the establishment decreases by more than 5% with no increase in employment in the following year. A detailed description of the definition is provided in Section 3.3.2. IAB DailyWageP25 1st quartile of gross average daily wage for all full-time employees in 2005 euros IAB DailyWageP50 Median of gross average daily wage for all full-time employees in 2005 euros IAB DailyWageP75 Third quartile of gross average daily wage for all full-time employees in 2005 euros IAB Tobin’s Q = (market capitalization [08001]+total assets [02999] – common equity [03501])/total assets Worldscope Table II. Variable definitions This table defines all variables used in this paper. Board data are taken from Hoppenstedt company profiles and annual reports. Employment and wage data are from the IAB Establishment History Panel. Accounting data are taken from Worldscope and market data from Datastream. The numbers in brackets refer to Worldscope items, taken from the Worldscope Data Definition Guide. Variable Description Source Number of employees Total number of employees in the establishment IAB Number of skilled Number of skilled (blue-collar) employees (at least vocational training) IAB Number of unskilled Number of unskilled (blue-collar) employees (no formal qualification) IAB Number of white collar Number of white-collar employees (at least vocational training) IAB Beta CAPM beta estimated over the prior calendar year using daily returns Datastream EstAge Age of the establishment in year IAB Firm employees Sum of all employees across all establishments of the firm in Germany IAB Firm age Age of the firm in year Worldscope HHI Herfindahl–Hirschman index calculated using number of employees of all establishments with more than 50 employees in the same county and industry (three-digit NACE code). IAB Leverage =Total debt [03255]/(total debt+common equity [03501]) Worldscope MCap Market capitalization [08001] Worldscope MedianEmplAge Median age of all employees in the establishment IAB NetPPE Net property, plant, and equipment [02501] Worldscope Parity =1 if 50% of all members of the company’s supervisory board are classified as employee representatives Hoppenstedt, annual reports PPE dummy =1 if NetPPE [02501] declines by more than 15% Worldscope RatioWhiteCollar =#WhiteCollar/#Employees IAB ROA =EBITDAt [18198]/((total assetst [02999]+total assetst–1)/2) Worldscope Sales =Net sales or revenues [01001] in 2005 euros Worldscope Shock =1 if employment in non-sample establishments in the same industry (three-digit NACE-code) as the establishment decreases by more than 5% with no increase in employment in the following year. A detailed description of the definition is provided in Section 3.3.2. IAB DailyWageP25 1st quartile of gross average daily wage for all full-time employees in 2005 euros IAB DailyWageP50 Median of gross average daily wage for all full-time employees in 2005 euros IAB DailyWageP75 Third quartile of gross average daily wage for all full-time employees in 2005 euros IAB Tobin’s Q = (market capitalization [08001]+total assets [02999] – common equity [03501])/total assets Worldscope Variable Description Source Number of employees Total number of employees in the establishment IAB Number of skilled Number of skilled (blue-collar) employees (at least vocational training) IAB Number of unskilled Number of unskilled (blue-collar) employees (no formal qualification) IAB Number of white collar Number of white-collar employees (at least vocational training) IAB Beta CAPM beta estimated over the prior calendar year using daily returns Datastream EstAge Age of the establishment in year IAB Firm employees Sum of all employees across all establishments of the firm in Germany IAB Firm age Age of the firm in year Worldscope HHI Herfindahl–Hirschman index calculated using number of employees of all establishments with more than 50 employees in the same county and industry (three-digit NACE code). IAB Leverage =Total debt [03255]/(total debt+common equity [03501]) Worldscope MCap Market capitalization [08001] Worldscope MedianEmplAge Median age of all employees in the establishment IAB NetPPE Net property, plant, and equipment [02501] Worldscope Parity =1 if 50% of all members of the company’s supervisory board are classified as employee representatives Hoppenstedt, annual reports PPE dummy =1 if NetPPE [02501] declines by more than 15% Worldscope RatioWhiteCollar =#WhiteCollar/#Employees IAB ROA =EBITDAt [18198]/((total assetst [02999]+total assetst–1)/2) Worldscope Sales =Net sales or revenues [01001] in 2005 euros Worldscope Shock =1 if employment in non-sample establishments in the same industry (three-digit NACE-code) as the establishment decreases by more than 5% with no increase in employment in the following year. A detailed description of the definition is provided in Section 3.3.2. IAB DailyWageP25 1st quartile of gross average daily wage for all full-time employees in 2005 euros IAB DailyWageP50 Median of gross average daily wage for all full-time employees in 2005 euros IAB DailyWageP75 Third quartile of gross average daily wage for all full-time employees in 2005 euros IAB Tobin’s Q = (market capitalization [08001]+total assets [02999] – common equity [03501])/total assets Worldscope 3.3 Research Design Our empirical strategy is to compare how a negative shock affects employee layoffs by establishments owned by parity-codetermined firms differently from those owned by firms with one-third or no labor representation on the supervisory board. Then we compare establishment-level wages between parity and non-parity firms through the cycle of the shock and non-shock period. To cross-check robustness of results based on a particular empirical design, we employ three different strategies: difference-in-differences estimations, placebo tests, and RD analyses. 3.3.a. Specification The difference-in-differences analysis relies on exogenous, negative industry shocks to employment and on the dummy variable Parity, which is one in any establishment-year when the firm has 50% worker representation on the supervisory board, and zero otherwise. Our definition of Parity is based on actual board data and is not inferred from employment data. The baseline specification is: yijkt=αt+αi+γXijkt+δParityjt+θShockkt+βParityjt×Shockkt+ϵijkt. (1) The dependent variable, yijkt, is the log change in employment or the log of the median daily wage, where i indexes establishments, j indexes firms, k indexes industries, and t indexes time. Parityjt is the parity dummy, Shockkt is the shock dummy defined in the next section, and ϵijkt is an error term. We control for year-fixed effects, αt, and establishment-fixed effects, αi. Establishment-fixed effects absorb firm-fixed effects, and control for unobservable time-invariant firm and establishment characteristics. The vector of time-varying control variables, Xijkt, includes the logarithm of the number of employees working in Germany, the logarithm of total sales, firm leverage, and establishment age. We cluster standard errors at the firm level. Abadie et al. (2017) argue that the requirement to cluster follows from the sampling strategy and the experimental design. We follow a two-stage sampling process, in which we identify first firms and then match establishments to firms. The argument of Abadie et al. (2017) suggests clustering at the first-stage variable for such a sampling process; in our case, the firm level. A different approach would cluster observations at the industry-year level, the level at which we define shocks. We present our results with standard errors clustered at the firm level and provide a robustness check with clustering at the industry-year level in Tables A-5 (Panel A) and A-6 (Panel A) in the Online Appendix. Standard errors and significance levels are not much affected by our choice of clustering. We shall refer to establishment-year observations in which the firm has 50% worker representation as parity observations and to all others, including those with one-third representation, as non-parity observations.18 For brevity and ease of exposition, we will mostly, and somewhat imprecisely, distinguish between “parity firms” and “non-parity firms,” to avoid the more cumbersome expression “establishment-year observations.” This focus on parity establishments helps to preserve the sample size of non-parity establishments, which is smaller than that of parity establishments. Table I, Panel B shows 67.4% of our sample establishment-years are owned by parity firms. For employment analysis, the coefficient of main interest is the β on the interaction of Parity and Shock. It measures the differential impact industry shocks have on employment of parity and non-parity establishments. Our hypothesis predicts β>0; parity establishments maintain higher levels of employment after an industry-wide shock than non-parity establishments. For wages, the regression does not contain Shock and δ is of main interest. Our hypothesis predicts δ<0; employees of parity firms pay insurance premiums in the form of lower wages. 3.3.b. Definition of shocks We identify adverse employment shocks using a drop in employment by non-sample German firms and foreign firms with establishments in Germany. The main requirement is that shocks are exogenous and sufficiently correlated with the economic environment of our sample firms so that they warrant major adjustments. We do not require that shocks cannot be anticipated. A suitable definition of shocks requires that Shock has a significant impact in Equation (1). If the estimate of θ is not significant, then either the definition of Shock is unsuitable and the assumed shock has no impact on non-parity establishments, or workers at non-parity establishments are also insured. In either case we cannot infer whether workers at parity establishments receive more insurance from estimates of β. We do not define shocks based on employment changes in other European countries because Germany follows a different business cycle and the correlations between employment at our sample firms and other European firms are too weak. For example, during 2011–2012, the German economy was booming while most other European countries were in, or on the verge of, a recession. Employment shocks are defined at the industry level. An industry is subject to a shock if non-sample firms’ establishments located in Germany belonging to the same three-digit NACE-code industry as a whole suffer a decrease of at least 5% in employment. These establishments may belong to either German or foreign firms. Our sample firms are likely to be under economic pressure to decrease their payroll when other non-sample firms in the same industry lay off 5% or more of their workers. We use the 5% threshold to ensure that shocks are strong enough to have a material effect and frequent enough to permit identification. Results based on a lower threshold of 2.5% are similar.19 Notice that these non-sample German and foreign firms are used only to define shocks, not as controls. The non-sample firms include many large non-listed, family owned, or foreign firms with establishments located in Germany.20 The mean (median) total sales and the number of employees of the largest 100 non-sample firms used to identify shocks are €10.2 bn (€7.0 bn) and 33,500 (19,700) in 2006, respectively. These numbers are quite comparable to the corresponding numbers for our sample firms in 2006, which are €11.7 bn (€2.0 bn) and 38,700 (9,200), respectively. Ultimately, if the non-sample firms were unsuitable to define industry shocks, attenuation bias would move the coefficient of Shock in Equation (1) toward zero and render it insignificant. In defining industry shocks, we exclude 5% declines in non-sample firms’ employment followed by an immediate recovery in the subsequent year, because other firms in the industry do not react to such shocks. Shocks followed by immediate recovery render the coefficient θ in Equation (1) insignificant (see Panel B of Table A-5 in the Online Appendix). One scenario in which this may happen has some non-sample firms lose market share or exit the market, and other firms in the industry subsequently increase employment to take the space left by declining or exiting firms. We exclude these very short-lived shocks, because they provide no identification. On the other hand, sufficiently severe and long-lasting shocks will require all firms, including parity-firms, to adjust their workforce to avoid bankruptcy (Gamber, 1988; Guiso, Pistaferri, and Schivardi, 2005). We therefore end a shock period after 4 years, again, because permanent shocks cannot provide identification. Thus, a shock period ends either after a resumption of growth, or after 4 consecutive years of non-positive growth, whichever occurs first. We also define a shock over a 2-year interval. Table A-1 in the Online Appendix illustrates how Shock is defined for four possible sequences of employment growth over 5 years. To get a feel for how our definition identifies employment shocks during our sample period, we estimate OLS regressions of the shock dummy as the dependent variable using the 4-year or 2-year definitions of shock. The independent variables are year dummies, and the regression does not have an intercept. Adjusted R2s are around 18%, indicating that much of the variation in shocks is industry-specific and is not driven by the business cycle. Figure A-1 in the Online Appendix plots the year dummy coefficients in the regressions and German unemployment rate by year. The two shock definitions are highly correlated with each other and move with the unemployment rate. We report estimation results based on the 4-year interval. With this definition, 5.8% of establishment-years of the sample firms are in shock periods. Baseline results using the 2-year interval are similar. 4. Empirical Results We begin with an investigation of how layoffs at establishments owned by parity firms differ from those owned by non-parity firms when the industry suffers a negative shock to employment. We then compare wages between parity and non-parity firms. 4.1 Employment Table III reports estimation results with different combinations of control variables. The results are consistent with the insurance hypothesis. The coefficients on Shock × Parity in the first two columns imply parity-codetermined firms retain about 8% more employees in comparison to non-parity firms during shock periods. The majority of non-parity firms in our sample have one-third worker representation. Hence, the coefficients reflect, to a large extent, the incremental impact of moving from one-third-codetermination to parity-codetermination, and to a lesser extent the impact of moving from no employee representation to parity codetermination. Table III. Employment changes—all employees This table presents OLS estimation results with the log change in employment as the dependent variable. Only establishments with more than 50 employees are included. Column (4) includes six additional variables not shown on the table: LogFirmEmployees × Parity, LogFirmEmployees × (1 – Parity), LogFirmEmployees2 × Parity, LogFirmEmployees2 × (1 – Parity), LogFirmEmployees³ × Parity, and LogFirmEmployees³ × (1 – Parity). T-statistics are reported in parentheses. Standard errors are clustered at the firm level. The table also reports the p-value for the t-test that Shock + Shock × Parity = 0, except in Column (3), which includes interaction terms of Shock with LogFirmEmployees and LogFirmEmployees2. Dependent variable Log change in number of employees (1) (2) (3) (4) Shock×Parity 0.082 0.081 0.064 0.083 (2.37) (2.22) (1.80) (2.46) Shock×LogFirmEmployees 0.264 (1.86) Shock×LogFirmEmployees2 –0.013 (–1.79) Shock –0.096 –0.095 –1.323 –0.097 (–3.30) (–3.05) (–2.05) (–3.44) Parity –0.044 –0.042 –0.039 8.651 (–1.39) (–1.39) (–1.27) (3.34) LogEstAge –0.594 –0.594 –0.594 –0.594 (–10.8) (–10.7) (–10.8) (–10.8) LogSales 0.005 –0.012 –0.014 0.006 (0.22) (–0.08) (–0.10) (0.30) Leverage –0.016 –0.018 –0.020 –0.017 (–0.34) (–0.44) (–0.48) (–0.36) LogFirmEmployees 0.054 0.147 0.137 (1.46) (0.81) (0.75) LogSales2 0.000 0.000 (0.10) (0.13) LogFirmEmployees2 –0.005 –0.005 (–0.47) (–0.42) Adjusted R2 0.174 0.174 0.175 0.175 Observations 48,896 48,896 48,896 48,896 t-test: Shock×Parity+Shock=0 0.502 0.508 0.507 Year F.E. Yes Yes Yes Yes Add. employment polynomials No No No Yes Establishment F.E. Yes Yes Yes Yes Dependent variable Log change in number of employees (1) (2) (3) (4) Shock×Parity 0.082 0.081 0.064 0.083 (2.37) (2.22) (1.80) (2.46) Shock×LogFirmEmployees 0.264 (1.86) Shock×LogFirmEmployees2 –0.013 (–1.79) Shock –0.096 –0.095 –1.323 –0.097 (–3.30) (–3.05) (–2.05) (–3.44) Parity –0.044 –0.042 –0.039 8.651 (–1.39) (–1.39) (–1.27) (3.34) LogEstAge –0.594 –0.594 –0.594 –0.594 (–10.8) (–10.7) (–10.8) (–10.8) LogSales 0.005 –0.012 –0.014 0.006 (0.22) (–0.08) (–0.10) (0.30) Leverage –0.016 –0.018 –0.020 –0.017 (–0.34) (–0.44) (–0.48) (–0.36) LogFirmEmployees 0.054 0.147 0.137 (1.46) (0.81) (0.75) LogSales2 0.000 0.000 (0.10) (0.13) LogFirmEmployees2 –0.005 –0.005 (–0.47) (–0.42) Adjusted R2 0.174 0.174 0.175 0.175 Observations 48,896 48,896 48,896 48,896 t-test: Shock×Parity+Shock=0 0.502 0.508 0.507 Year F.E. Yes Yes Yes Yes Add. employment polynomials No No No Yes Establishment F.E. Yes Yes Yes Yes Table III. Employment changes—all employees This table presents OLS estimation results with the log change in employment as the dependent variable. Only establishments with more than 50 employees are included. Column (4) includes six additional variables not shown on the table: LogFirmEmployees × Parity, LogFirmEmployees × (1 – Parity), LogFirmEmployees2 × Parity, LogFirmEmployees2 × (1 – Parity), LogFirmEmployees³ × Parity, and LogFirmEmployees³ × (1 – Parity). T-statistics are reported in parentheses. Standard errors are clustered at the firm level. The table also reports the p-value for the t-test that Shock + Shock × Parity = 0, except in Column (3), which includes interaction terms of Shock with LogFirmEmployees and LogFirmEmployees2. Dependent variable Log change in number of employees (1) (2) (3) (4) Shock×Parity 0.082 0.081 0.064 0.083 (2.37) (2.22) (1.80) (2.46) Shock×LogFirmEmployees 0.264 (1.86) Shock×LogFirmEmployees2 –0.013 (–1.79) Shock –0.096 –0.095 –1.323 –0.097 (–3.30) (–3.05) (–2.05) (–3.44) Parity –0.044 –0.042 –0.039 8.651 (–1.39) (–1.39) (–1.27) (3.34) LogEstAge –0.594 –0.594 –0.594 –0.594 (–10.8) (–10.7) (–10.8) (–10.8) LogSales 0.005 –0.012 –0.014 0.006 (0.22) (–0.08) (–0.10) (0.30) Leverage –0.016 –0.018 –0.020 –0.017 (–0.34) (–0.44) (–0.48) (–0.36) LogFirmEmployees 0.054 0.147 0.137 (1.46) (0.81) (0.75) LogSales2 0.000 0.000 (0.10) (0.13) LogFirmEmployees2 –0.005 –0.005 (–0.47) (–0.42) Adjusted R2 0.174 0.174 0.175 0.175 Observations 48,896 48,896 48,896 48,896 t-test: Shock×Parity+Shock=0 0.502 0.508 0.507 Year F.E. Yes Yes Yes Yes Add. employment polynomials No No No Yes Establishment F.E. Yes Yes Yes Yes Dependent variable Log change in number of employees (1) (2) (3) (4) Shock×Parity 0.082 0.081 0.064 0.083 (2.37) (2.22) (1.80) (2.46) Shock×LogFirmEmployees 0.264 (1.86) Shock×LogFirmEmployees2 –0.013 (–1.79) Shock –0.096 –0.095 –1.323 –0.097 (–3.30) (–3.05) (–2.05) (–3.44) Parity –0.044 –0.042 –0.039 8.651 (–1.39) (–1.39) (–1.27) (3.34) LogEstAge –0.594 –0.594 –0.594 –0.594 (–10.8) (–10.7) (–10.8) (–10.8) LogSales 0.005 –0.012 –0.014 0.006 (0.22) (–0.08) (–0.10) (0.30) Leverage –0.016 –0.018 –0.020 –0.017 (–0.34) (–0.44) (–0.48) (–0.36) LogFirmEmployees 0.054 0.147 0.137 (1.46) (0.81) (0.75) LogSales2 0.000 0.000 (0.10) (0.13) LogFirmEmployees2 –0.005 –0.005 (–0.47) (–0.42) Adjusted R2 0.174 0.174 0.175 0.175 Observations 48,896 48,896 48,896 48,896 t-test: Shock×Parity+Shock=0 0.502 0.508 0.507 Year F.E. Yes Yes Yes Yes Add. employment polynomials No No No Yes Establishment F.E. Yes Yes Yes Yes As expected, Shock has a significantly negative coefficient, which implies non-parity firms suffer a sharp decline in employment. We performed a standard t-test for the restriction that the coefficients on Shock and Shock × Parity add up to zero, which would indicate full insurance. In no specification can we reject the null hypothesis that the coefficients on Shock and Shock × Parity have the same magnitude with opposite signs. It appears employees working for parity firms are fully protected against negative industry shocks. 4.1.a. Identification issues To address identification issues, we first check the validity of the parallel trends assumption, the possible confounding effects of the level of employment in Germany (hereafter, the employment level), and whether we find similar results in placebo tests. Then, we examine the potential endogeneity of the parity status and conduct RD analysis. In addition, we examine changes in parity status due to industry shocks, cross-industry effects, and robustness to various alternative specifications. Parallel trends. In our difference-in-differences estimation, the outcome variable of the control group is used to calculate the expected counterfactual, assuming that parity and non-parity firms have the same time trend if there are no industry shocks. Thus, we check whether the log change in the number of employees of parity and non-parity firms were following the same trend before an industry shock. Figure 2 plots log change in the number of employees separately for our sample of parity and non-parity firms from 4 years prior to 3 years after the first shock year, where the first year of the shock is defined as event year 0. We only include establishments in industries without a shock during the 4 years prior to year 0. To ensure that changes in the sample do not affect trends in employment, we require that, during the event period, all establishments have data available for all years and do not change their parity status. Figure 2. View largeDownload slide Pre- and post-shock trends in employment changes of establishments of parity and non-parity firms. Notes: This figure plots trends in average log changes of employment for establishments of parity (black line) and non-parity firms (gray line) in event time. Time=0 is the first year of an industry shock. We require Shock=0 for all years before Time=0. Only establishments with more than 50 employees and data available for the complete event period are included. Establishments switching parity status during the event period are excluded. Figure 2. View largeDownload slide Pre- and post-shock trends in employment changes of establishments of parity and non-parity firms. Notes: This figure plots trends in average log changes of employment for establishments of parity (black line) and non-parity firms (gray line) in event time. Time=0 is the first year of an industry shock. We require Shock=0 for all years before Time=0. Only establishments with more than 50 employees and data available for the complete event period are included. Establishments switching parity status during the event period are excluded. Figure 2 shows remarkably parallel trends in average employment growth of parity and non-parity establishments until year –1, and a sharp divergence in year 0. The sharp drop in employment of non-parity establishments in year 0 vis-à-vis parity establishments, which show no significant decline in employment, reflects what we find in Table III: Incomplete employment protection for non-parity establishments versus full protection for workers of parity establishments.21 Parity and firm-level employment. An important concern for our specification is that Parity is defined with respect to the number of employees in Germany. Note that only the number of employees in Germany is relevant for whether a firm is subject to parity codetermination. Hence, Parity is only loosely related to firm size.22 We separate the Parity effect from the employment level effects in Column (3) of Table III by including interactions of Shock with the log of the total number of employees in Germany, LogFirmEmployees, and LogFirmEmployees2. If Parity has no unique effect other than just being an indicator for having more than 2,000 employees in Germany, then the coefficient on Shock × Parity should become insignificant if we interact Shock and the employment-level variables. However, the coefficient of Shock × Parity remains positive and significant, albeit at a lower statistical significance. The parity status per se has a significant positive effect on employment during shock periods.23 We also allow the impact of LogFirmEmployees and its higher-order terms to be different for parity and non-parity firms in Column (4) by interacting these terms with Parity and (1 – Parity). None of the coefficients of employment-level variables is significant. (For brevity, the coefficients are not reported.) The coefficient on the interaction of Shock and Parity remains positive and significant. Placebo tests. Another way to test the unique effect of parity-codetermination during industry shocks is to conduct placebo tests using false parity dummies. We reestimate Columns (1) and (2) in Table III, separately for parity and non-parity firms, with Placebo replacing Parity in the regressions. The thresholds for the placebo dummy, Placebo, equal the median number of employees in Germany, which is 1,318 for non-parity firms and 10,458 for parity firms. The results are presented in Table A-2 in the Online Appendix. Coefficients of both, Placebo and Shock × Placebo, are statistically zero for both subsamples, demonstrating that the 2,000-threshold is unique and the treatment effect is real. We take this placebo test idea further by reestimating Specification (2) in Table III, with Placebo equal to one if employment exceeds a threshold that varies from 1,500 to 2,500 in increments of 50, that is, 1,500, 1,550, 1,600, …, 2,500. Our hypothesis predicts that the 2,000 threshold is uniquely relevant. The more we move the cut-off above (below) 2,000, the more firm-years we misclassify as non-parity (parity) firms, which creates attenuation bias and reduces the coefficients on Parity and Shock × Parity. For example, if Placebo is based on 2,500 employees, all establishments belonging to firms with a number of employees in Germany between 2,000 and 2,500 are classified as non-parity, even though they are parity codetermined. Hence, if the parity effect is unique, we should find both coefficient estimates and significance levels peak at 2,000, and decline as we move the cut-off further away from 2,000. Figure 3 presents the results for the estimated coefficients on the interaction term (right-hand axis, gray line) and the corresponding t-statistics (left-hand axis, black line). As expected, coefficients and significance levels both peak around the 2,000 threshold. Notice that if our results were driven by the employment-level effect alone, without a separate effect of Parity, no noise would have been added at Placebo thresholds greater than 2,000, and the coefficient would have increased (or at least plateaued out) as Placebo increases further away from 2,000. Figure 3. View largeDownload slide Employment—placebo tests. Notes: This figure shows the coefficients on the interaction term Placebo×Shock (right-hand axis, gray line) and the corresponding t-statistics (left-hand axis, black line) from estimating Column (2) in Table III with Placebo instead of Parity. We vary the placebo threshold in steps of 50 from 1,500 to 2,500 and let Placebo be equal to one for any establishment of a firm with employment in Germany above the corresponding placebo threshold. Figure 3. View largeDownload slide Employment—placebo tests. Notes: This figure shows the coefficients on the interaction term Placebo×Shock (right-hand axis, gray line) and the corresponding t-statistics (left-hand axis, black line) from estimating Column (2) in Table III with Placebo instead of Parity. We vary the placebo threshold in steps of 50 from 1,500 to 2,500 and let Placebo be equal to one for any establishment of a firm with employment in Germany above the corresponding placebo threshold. Endogeneity of parity status. Parity may be endogenous, as employers and employees may attempt to influence their firm’s parity status through influencing the number of employees in Germany. Workers may want to keep the number above 2,000 to maintain parity status, whereas shareholders may attempt to keep the number below 2,000 to maintain non-parity status.24 Such attempts would lead to an accumulation of firms around the 2,000 threshold. To investigate whether there is an unusual concentration of firms located right below or above the 2,000 employee threshold, we first plot a histogram of the frequency of the distribution of all sample firm-year observations with 500–3,500 workers employed in Germany in Figure A-2 of the Online Appendix. It shows there are more firms with fewer employees. More important, it shows scattered and minor peaks throughout the whole range of 500–3,500, but does not show an unusual concentration around the 2,000 threshold. To conduct a more formal test, we first use the McCrary (2008) test, which uses a smoothed version of the histogram, testing for a discontinuity of the density at the 2,000 threshold. Figure 4 shows the smoothed density estimate, which allows for a discontinuous jump of the density at the threshold. Different estimates are obtained depending on the range of data included; we use ranges from 1,500 to 2,500 employees and expand it symmetrically; the largest interval is from zero to 4,000 employees. In all cases the estimate of the discontinuity is smaller than its standard error; we cannot reject that the density is continuous at 2,000 at any conventional significance level. Figure 4. View largeDownload slide Distribution of firms by number of employees (density plot). Notes: The figure shows the histogram and estimated density of all firm-year observations for which the number of employees in Germany is between 0 and 4,000 with breakpoint at 2,000 employees. The density and the confidence band are estimated using the McCrary (2008) estimator using the default bin size of 59 and bandwidth of 904. Figure 4. View largeDownload slide Distribution of firms by number of employees (density plot). Notes: The figure shows the histogram and estimated density of all firm-year observations for which the number of employees in Germany is between 0 and 4,000 with breakpoint at 2,000 employees. The density and the confidence band are estimated using the McCrary (2008) estimator using the default bin size of 59 and bandwidth of 904. To be cautious, we conduct two additional tests. The first involves a non-parametric regression of employment growth on the employment level. If firms perceive parity codetermination as very costly, they may sacrifice growth if employment is just below the 2,000 threshold. Employees, on the other hand, may resist contraction more if the level of employment is just above the 2,000 threshold. Hence, we expect a negative relation between the number of employees and employment growth below the threshold, and a positive relation above the threshold. Figure 5 plots a non-parametric regression allowing the slope of the regression of employment growth on employment to change arbitrarily with the level of employment. We find no evidence in support of the manipulation hypothesis. Although never significant, the regression is upward sloping just below the threshold, and downward sloping just above the threshold, the opposite of what we would expect if workers or shareholders would influence the number of employees to prevent board composition from changing. Figure 5. View largeDownload slide Employment growth around the 2,000-employee cutoff. Notes: This figure shows the result of a local polynomial regression with local-mean smoothing (and 95% confidence band) of firm employment growth on the number of firm employees between 1,500 and 2,500 in Germany. Epanechnikov kernel with bandwidth 50 are used. Figure 5. View largeDownload slide Employment growth around the 2,000-employee cutoff. Notes: This figure shows the result of a local polynomial regression with local-mean smoothing (and 95% confidence band) of firm employment growth on the number of firm employees between 1,500 and 2,500 in Germany. Epanechnikov kernel with bandwidth 50 are used. For the second test, we tabulate the frequency of firms crossing various employment-level thresholds from below or above by counting how many times our sample firms cross 1,500, 1,600, … up to 2,500 in the increment of 100. If shareholders prevent crossing the 2,000 threshold from below, and workers prevent crossing the 2,000 threshold from above, then we should observe fewer crossings over 2,000 compared with all other thresholds. Table A-3 in the Online Appendix reports the tabulation results, which show more or less even distribution across all thresholds without a noticeable difference for the 2,000 threshold. In sum, all analyses lead to the same conclusion: There is no indication that firms cluster around the 2,000 threshold or manipulate the likelihood of crossing the threshold, either upward or downward. It appears that business and operational considerations are the dominant factor in deciding the employment level. There is no evidence that either workers or shareholders influence employment decisions to suit their interests in terms of board representation. As we shall show later, parity codetermination has no discernible impact on firm profitability or value; hence, firms are unlikely to alter investment decisions just to avoid codetermination. Whether a firm is just above or just below the 2,000-threshold seems rather random. RD analysis. The above conclusion allows us to perform RD analysis. Testing whether our conclusions hold locally around the 2,000-employee cut-off point improves the covariate balance relative to difference-in-differences analysis. However, it comes at the cost of lowering the power of the tests stemming from smaller sample size: RD analysis has to be restricted to industry-shock periods, which drastically reduces our sample size to 5.8% of establishment-years in our sample. We regress employment in shock periods on Parity using a local linear regression: yijkt=f(FirmEmployeesit)+δParityjt+ηjt, (2) where f() is a kernel-weighted local polynomial of degree one. Table IV reports the results and Figure 6 plots local polynomial regressions around the 2,000 threshold with 95% confidence intervals in dashed lines.25 We show results for the optimized bandwidth based on Imbens and Kalyanaraman (2012) as well as results from doubling or halving the optimal bandwidth. We apply a sharp RD design, which is appropriate here, because firms never elect to be parity-codetermined if they are below the 2,000 threshold. Table IV. Employment change—RD analysis This table presents results for a kernel regression using a triangular kernel and the optimized Imbens and Kalyanaraman, 2012 bandwidth (bw = 1). The dependent variable is log change in employment. We modify the optimized bandwidth by factors of 0.5 and 2 to check robustness. A sharp RD design is assumed, where the treatment variable (Parity) changes from one to zero at 2,000 employees (firm level). The z-statistics for the coefficient estimates are reported in parentheses below the estimates. Dependent variable Log change in employment Parity (bw = 1) 0.148 (2.02) Parity (bw = 0.5) 0.092 (1.10) Parity (bw = 2) 0.134 (1.95) Observations 314 Dependent variable Log change in employment Parity (bw = 1) 0.148 (2.02) Parity (bw = 0.5) 0.092 (1.10) Parity (bw = 2) 0.134 (1.95) Observations 314 Table IV. Employment change—RD analysis This table presents results for a kernel regression using a triangular kernel and the optimized Imbens and Kalyanaraman, 2012 bandwidth (bw = 1). The dependent variable is log change in employment. We modify the optimized bandwidth by factors of 0.5 and 2 to check robustness. A sharp RD design is assumed, where the treatment variable (Parity) changes from one to zero at 2,000 employees (firm level). The z-statistics for the coefficient estimates are reported in parentheses below the estimates. Dependent variable Log change in employment Parity (bw = 1) 0.148 (2.02) Parity (bw = 0.5) 0.092 (1.10) Parity (bw = 2) 0.134 (1.95) Observations 314 Dependent variable Log change in employment Parity (bw = 1) 0.148 (2.02) Parity (bw = 0.5) 0.092 (1.10) Parity (bw = 2) 0.134 (1.95) Observations 314 Figure 6. View largeDownload slide Discontinuity at the 2,000-employee cutoff: Employment. Notes: This figure shows the results of two local polynomial regressions for log change in employment during shock periods around the parity cutoff of 2,000 employees. Figure 6. View largeDownload slide Discontinuity at the 2,000-employee cutoff: Employment. Notes: This figure shows the results of two local polynomial regressions for log change in employment during shock periods around the parity cutoff of 2,000 employees. In spite of the big drop in sample size, the results show a statistically significant upward jump of employment residuals above the 2,000 cutoff for the optimal bandwidth, consistent with the shock influencing employment at non-parity firms more strongly than at parity firms. Results remain significant if we increase the bandwidth, but not if we halve it, which is equivalent to a further reduction in sample size that is already very small. Continuous measures of industry shocks. Sraer and Thesmar (2007) and Ellul, Pagano, and Schivardi (2017) study implicit employment contracts by family-firms using continuous measures of shocks. To check the robustness of our conclusions to continuous measures of shocks, we re-estimate our baseline specifications with measures of shocks similar to those used in Sraer and Thesmar (2007) and Ellul, Pagano, and Schivardi (2017). First, we follow Sraer and Thesmar (2007) and define a new variable LogMeanIndEmplit, the logarithm of the industry mean of contemporaneous employment of non-sample firms in the three-digit NACE industry firm i belongs to, excluding the employment of firm i itself.26 We interact LogMeanIndEmplit with the Parity indicator, in keeping with our earlier analysis. Here, Parity takes the place of indicators for family involvement in the family-firms literature.27Table V (Panel A) shows the results. Note that with this definition of shocks, our hypothesis predicts opposite signs: a positive sign on LogMeanIndEmplit and a negative sign on the interaction with Parity. The coefficient on the inaction term is negative and highly significant in all specifications. Specifications controlling for year-fixed effects (Columns 3 and 4) show coefficients of 0.29 and 0.36 on LogMeanIndEmpl, a range of elasticity estimates quantitatively similar to those in Sraer and Thesmar (2007). With this definition, insurance in parity firms manifests in a reduction of the sensitivity to industry-level shocks of about 50% for parity firms. Table V. Employment changes—continuous shock measure This table presents OLS estimation results with (1) log employment as the dependent variable in Panel A, and (2) the log change in employment as the dependent variable in Panel B. Only establishments with more than 50 employees are included. Panel A adapts the methodology of Sraer and Thesmar (2007) [see their equation (4) and Table 7]. We use the log of contemporaneous average employment of non-sample firms’ establishments in the same industry (NACE, LogMeanIndEmpl) instead of our Shock definition. Panel B adapts the methodology of Ellul, Pagano, and Schivardi (2017) (see their Table 2). We use the contemporaneous log change in employment of non-sample firms’ establishments in the same industry (LogChangeIndEmpl) instead of our Shock definition. T-statistics are reported in parentheses. Standard errors are clustered at the firm level. Panel A Dependent variable Log number of employees (1) (2) (3) (4) LogMeanIndEmpl 0.492 0.480 0.286 0.359 (5.68) (5.90) (3.00) (3.74) Parity×LogMeanIndEmpl –0.172 –0.172 –0.141 –0.209 (–2.12) (–2.15) (–1.94) (–2.51) Parity 0.620 0.623 0.527 0.645 (1.09) (1.11) (0.91) (1.12) LogEstAge –0.020 0.112 0.113 (–0.57) (3.58) (3.70) Adjusted R2 0.909 0.909 0.912 0.912 Observations 54,042 54,042 54,042 54,042 t-Test: LogMIE×Parity+LogMIE=0 0.036 0.037 0.295 0.262 Establishment F.E. Yes Yes Yes Yes Year F.E. No No Yes No Year×Parity F.E. No No No Yes Panel B Dependent variable Log change in number of employees (1) (2) (3) (4) LogChangeIndEmpl 0.740 0.598 0.424 0.453 (3.03) (2.52) (2.00) (2.91) Parity×LogChangeIndEmpl –0.681 –0.522 –0.271 –0.295 (–2.38) (–2.04) (–2.04) (–2.71) Parity –0.051 0.070 –0.001 0.406 (–4.06) (1.53) (–0.06) (7.82) LogEstAge –0.344 –0.596 –0.596 (–9.50) (–11.51) (–11.34) Adjusted R2 0.084 0.153 0.179 0.180 Observations 50,134 50,134 50,134 50,134 t-Test: LogCIE×Parity+LogCIE=0 0.549 0.441 0.151 0.147 Establishment F.E. Yes Yes Yes Yes Year F.E. No No Yes No Year×Parity F.E. No No No Yes Panel A Dependent variable Log number of employees (1) (2) (3) (4) LogMeanIndEmpl 0.492 0.480 0.286 0.359 (5.68) (5.90) (3.00) (3.74) Parity×LogMeanIndEmpl –0.172 –0.172 –0.141 –0.209 (–2.12) (–2.15) (–1.94) (–2.51) Parity 0.620 0.623 0.527 0.645 (1.09) (1.11) (0.91) (1.12) LogEstAge –0.020 0.112 0.113 (–0.57) (3.58) (3.70) Adjusted R2 0.909 0.909 0.912 0.912 Observations 54,042 54,042 54,042 54,042 t-Test: LogMIE×Parity+LogMIE=0 0.036 0.037 0.295 0.262 Establishment F.E. Yes Yes Yes Yes Year F.E. No No Yes No Year×Parity F.E. No No No Yes Panel B Dependent variable Log change in number of employees (1) (2) (3) (4) LogChangeIndEmpl 0.740 0.598 0.424 0.453 (3.03) (2.52) (2.00) (2.91) Parity×LogChangeIndEmpl –0.681 –0.522 –0.271 –0.295 (–2.38) (–2.04) (–2.04) (–2.71) Parity –0.051 0.070 –0.001 0.406 (–4.06) (1.53) (–0.06) (7.82) LogEstAge –0.344 –0.596 –0.596 (–9.50) (–11.51) (–11.34) Adjusted R2 0.084 0.153 0.179 0.180 Observations 50,134 50,134 50,134 50,134 t-Test: LogCIE×Parity+LogCIE=0 0.549 0.441 0.151 0.147 Establishment F.E. Yes Yes Yes Yes Year F.E. No No Yes No Year×Parity F.E. No No No Yes Table V. Employment changes—continuous shock measure This table presents OLS estimation results with (1) log employment as the dependent variable in Panel A, and (2) the log change in employment as the dependent variable in Panel B. Only establishments with more than 50 employees are included. Panel A adapts the methodology of Sraer and Thesmar (2007) [see their equation (4) and Table 7]. We use the log of contemporaneous average employment of non-sample firms’ establishments in the same industry (NACE, LogMeanIndEmpl) instead of our Shock definition. Panel B adapts the methodology of Ellul, Pagano, and Schivardi (2017) (see their Table 2). We use the contemporaneous log change in employment of non-sample firms’ establishments in the same industry (LogChangeIndEmpl) instead of our Shock definition. T-statistics are reported in parentheses. Standard errors are clustered at the firm level. Panel A Dependent variable Log number of employees (1) (2) (3) (4) LogMeanIndEmpl 0.492 0.480 0.286 0.359 (5.68) (5.90) (3.00) (3.74) Parity×LogMeanIndEmpl –0.172 –0.172 –0.141 –0.209 (–2.12) (–2.15) (–1.94) (–2.51) Parity 0.620 0.623 0.527 0.645 (1.09) (1.11) (0.91) (1.12) LogEstAge –0.020 0.112 0.113 (–0.57) (3.58) (3.70) Adjusted R2 0.909 0.909 0.912 0.912 Observations 54,042 54,042 54,042 54,042 t-Test: LogMIE×Parity+LogMIE=0 0.036 0.037 0.295 0.262 Establishment F.E. Yes Yes Yes Yes Year F.E. No No Yes No Year×Parity F.E. No No No Yes Panel B Dependent variable Log change in number of employees (1) (2) (3) (4) LogChangeIndEmpl 0.740 0.598 0.424 0.453 (3.03) (2.52) (2.00) (2.91) Parity×LogChangeIndEmpl –0.681 –0.522 –0.271 –0.295 (–2.38) (–2.04) (–2.04) (–2.71) Parity –0.051 0.070 –0.001 0.406 (–4.06) (1.53) (–0.06) (7.82) LogEstAge –0.344 –0.596 –0.596 (–9.50) (–11.51) (–11.34) Adjusted R2 0.084 0.153 0.179 0.180 Observations 50,134 50,134 50,134 50,134 t-Test: LogCIE×Parity+LogCIE=0 0.549 0.441 0.151 0.147 Establishment F.E. Yes Yes Yes Yes Year F.E. No No Yes No Year×Parity F.E. No No No Yes Panel A Dependent variable Log number of employees (1) (2) (3) (4) LogMeanIndEmpl 0.492 0.480 0.286 0.359 (5.68) (5.90) (3.00) (3.74) Parity×LogMeanIndEmpl –0.172 –0.172 –0.141 –0.209 (–2.12) (–2.15) (–1.94) (–2.51) Parity 0.620 0.623 0.527 0.645 (1.09) (1.11) (0.91) (1.12) LogEstAge –0.020 0.112 0.113 (–0.57) (3.58) (3.70) Adjusted R2 0.909 0.909 0.912 0.912 Observations 54,042 54,042 54,042 54,042 t-Test: LogMIE×Parity+LogMIE=0 0.036 0.037 0.295 0.262 Establishment F.E. Yes Yes Yes Yes Year F.E. No No Yes No Year×Parity F.E. No No No Yes Panel B Dependent variable Log change in number of employees (1) (2) (3) (4) LogChangeIndEmpl 0.740 0.598 0.424 0.453 (3.03) (2.52) (2.00) (2.91) Parity×LogChangeIndEmpl –0.681 –0.522 –0.271 –0.295 (–2.38) (–2.04) (–2.04) (–2.71) Parity –0.051 0.070 –0.001 0.406 (–4.06) (1.53) (–0.06) (7.82) LogEstAge –0.344 –0.596 –0.596 (–9.50) (–11.51) (–11.34) Adjusted R2 0.084 0.153 0.179 0.180 Observations 50,134 50,134 50,134 50,134 t-Test: LogCIE×Parity+LogCIE=0 0.549 0.441 0.151 0.147 Establishment F.E. Yes Yes Yes Yes Year F.E. No No Yes No Year×Parity F.E. No No No Yes We also re-estimate our baseline regressions following Ellul, Pagano, and Schivardi (2017), who use first differences and total log changes as “shock” variable, a somewhat different approach from Sraer and Thesmar (2007). Specifically, we regress the log change in employment on LogChangeIndEmpl, the log change in employment of non-sample firms’ establishments in the same industry. Table V (Panel B) reports the results, which imply that parity firms’ response to industry-level employment changes are only about one-third of that of non-parity firms. In sum, our conclusions based on discrete shocks are robust to continuous measures of industry shocks. An alternative sample construction and alternative definitions of shock. We reestimate the regressions in Table III using other alternative specifications and report the results Table A-5 of the Online Appendix. Specifically, in Panel C, we include all establishments with more than 10 employees. In Panel D, we define Shock by a 2.5% drop in industry-wide employment among non-sample firms instead of the 5% drop. In Panel E, we use the 2-year interval in defining Shock instead of the 4-year interval. In Panel F, we control for outliers by truncating establishment-years with employment growth in the 1st and 99th percentile. Results are robust to these alternative specifications. Changes in parity status due to industry shocks. An industry shock may cause a sufficient drop in employment to make a parity firm become a non-parity firm. If both happen simultaneously but we observe the drop in employment after the change in parity status, we will incorrectly attribute the employment change to non-parity firms, underestimating the decline in employment among parity firms. To check this possibility, we first read annual reports of all firms changing their status from parity to non-parity.28 In almost all cases, worker representatives stay on the supervisory board until the annual shareholder meeting the year after the number of employees in Germany drops below 2,000; hence, our findings are highly unlikely to be affected by the simultaneous drop in employment and change in parity status. We are also concerned about establishments switching from parity to non-parity status that were in an industry shock the year of or the 2 years before the change in parity status. A big drop in employment in those establishments could cause the change in parity status. Our sample contains 236 establishments switching from parity to non-parity status. Of the 236 establishments, only 23 were in an industry shock in the year of or the 2 years before the parity status change. We drop all observations (49) belonging to these 23 establishments for 3 years after the switch, and reestimate Columns (1)–(3) of Table III. The reestimation results, presented in Table A-4 of the Online Appendix, are robust. Cross-industry effects. Firms operating in multiple industries may react to a shock in one industry by transferring workers to establishments in non-shock industries (Tate and Yang, 2016). Such transfers would be recorded as employment losses in establishments affected by a shock, lowering the coefficient on Shock × Parity, and as employment gains in establishments unaffected by the shock, not affecting the coefficient on Shock × Parity. Hence, the net effect would be to bias our results to finding less insurance. In spite of this potential bias, we find full insurance for parity firms.29 4.1.b. Local labor market conditions and positive industry shocks In this section, we extend our hypothesis to consider how the insurance effect of parity codetermination depends on local labor market conditions and how parity and non-parity firms differ in their response to positive industry shocks. Local labor market conditions. We expect implicit employment insurance be more prevalent when employment risk is greater. Labor markets are local and migration across local labor market has declined (Molloy, Smith, and Wozniak, 2017). Accordingly, employees become more vulnerable to losing their jobs if employers have more local monopsony power (Benmelech, Bergman, and Kim, 2018). Hence, we hypothesize that if one employer accounts for a larger fraction of employment in an industry in their local labor market, demand for insurance will be greater and codetermination as an enforcement mechanism will have greater impact on preventing layoffs. To test this prediction, we adapt the research design of Benmelech, Bergman, and Kim (2018) and measure employer concentration for each county–industry–year for the 401 counties and 222 NACE industries (three-digit level) by calculating the Hirschman–Herfindahl index (HHI) for the employment shares of local employers. We then define the indicator variable Herfindahl, which is one if the HHI is above the median across all counties and industries, and zero otherwise. We estimate triple-differences using the same specifications as Table III and report the results in Table VI. The variable of main interest here is the triple interaction of Shock, Parity, and Herfindahl. Its coefficient measures whether and how the effect of parity codetermination on insuring employment mechanism depends on employer concentration at the industry level. The effect is economically strong, with coefficient values around 0.05, and statistically significant at the 10% level in all specifications. From the interaction of Shock and Herfindahl, we also observe that higher employer concentration increases the vulnerability of employees to shocks, although this effect is not significant. These results imply that the insurance effect of parity codetermination is greater, the more vulnerable employees are to the market power of local employers, providing further corroborating evidence in support of our hypothesis. Table VI. Employment changes and local labor market conditions This table repeats the analysis of Table III with additional interaction terms using Herfindahl. Herfindahl is a dummy variable equal to one if the employment HHI in the establishment’s industry-county is above the sample median. The employment HHI is calculated using all establishments with more than 50 employees in the same county and industry (three-digit NACE code). The controls are not reported to preserve space. T-statistics are reported in parentheses. Standard errors are clustered at the firm level. Dependent variable Log change in number of employees (1) (2) (3) (4) Shock×Parity 0.052 0.051 0.045 0.048 (1.95) (1.88) (1.79) (1.94) Shock×LogFirmEmployees 0.245 (1.69) Shock×LogFirmEmployees2 –0.012 (–1.60) Shock –0.170 –0.168 –1.338 –0.166 (–2.57) (–2.50) (–2.98) (–2.56) Herfindahl 0.076 0.075 0.075 0.075 (2.91) (2.94) (2.91) (2.91) Parity×Herfindahl –0.001 –0.001 –0.001 –0.002 (–0.19) (–0.19) (–0.20) (–0.22) Shock×Herfindahl 0.103 0.102 0.114 0.099 (1.39) (1.37) (1.60) (1.35) Parity×Shock×Herfindahl 0.047 0.047 0.060 0.044 (1.72) (1.71) (1.98) (1.68) Parity –0.040 –0.039 –0.035 8.626 (–1.30) (–1.30) (–1.17) (3.36) Adjusted R2 0.175 0.175 0.175 0.175 Observations 48,896 48,896 48,896 48,896 Year F.E. Yes Yes Yes Yes Add. employment polynomials No No No Yes Establishment F.E. Yes Yes Yes Yes Dependent variable Log change in number of employees (1) (2) (3) (4) Shock×Parity 0.052 0.051 0.045 0.048 (1.95) (1.88) (1.79) (1.94) Shock×LogFirmEmployees 0.245 (1.69) Shock×LogFirmEmployees2 –0.012 (–1.60) Shock –0.170 –0.168 –1.338 –0.166 (–2.57) (–2.50) (–2.98) (–2.56) Herfindahl 0.076 0.075 0.075 0.075 (2.91) (2.94) (2.91) (2.91) Parity×Herfindahl –0.001 –0.001 –0.001 –0.002 (–0.19) (–0.19) (–0.20) (–0.22) Shock×Herfindahl 0.103 0.102 0.114 0.099 (1.39) (1.37) (1.60) (1.35) Parity×Shock×Herfindahl 0.047 0.047 0.060 0.044 (1.72) (1.71) (1.98) (1.68) Parity –0.040 –0.039 –0.035 8.626 (–1.30) (–1.30) (–1.17) (3.36) Adjusted R2 0.175 0.175 0.175 0.175 Observations 48,896 48,896 48,896 48,896 Year F.E. Yes Yes Yes Yes Add. employment polynomials No No No Yes Establishment F.E. Yes Yes Yes Yes Table VI. Employment changes and local labor market conditions This table repeats the analysis of Table III with additional interaction terms using Herfindahl. Herfindahl is a dummy variable equal to one if the employment HHI in the establishment’s industry-county is above the sample median. The employment HHI is calculated using all establishments with more than 50 employees in the same county and industry (three-digit NACE code). The controls are not reported to preserve space. T-statistics are reported in parentheses. Standard errors are clustered at the firm level. Dependent variable Log change in number of employees (1) (2) (3) (4) Shock×Parity 0.052 0.051 0.045 0.048 (1.95) (1.88) (1.79) (1.94) Shock×LogFirmEmployees 0.245 (1.69) Shock×LogFirmEmployees2 –0.012 (–1.60) Shock –0.170 –0.168 –1.338 –0.166 (–2.57) (–2.50) (–2.98) (–2.56) Herfindahl 0.076 0.075 0.075 0.075 (2.91) (2.94) (2.91) (2.91) Parity×Herfindahl –0.001 –0.001 –0.001 –0.002 (–0.19) (–0.19) (–0.20) (–0.22) Shock×Herfindahl 0.103 0.102 0.114 0.099 (1.39) (1.37) (1.60) (1.35) Parity×Shock×Herfindahl 0.047 0.047 0.060 0.044 (1.72) (1.71) (1.98) (1.68) Parity –0.040 –0.039 –0.035 8.626 (–1.30) (–1.30) (–1.17) (3.36) Adjusted R2 0.175 0.175 0.175 0.175 Observations 48,896 48,896 48,896 48,896 Year F.E. Yes Yes Yes Yes Add. employment polynomials No No No Yes Establishment F.E. Yes Yes Yes Yes Dependent variable Log change in number of employees (1) (2) (3) (4) Shock×Parity 0.052 0.051 0.045 0.048 (1.95) (1.88) (1.79) (1.94) Shock×LogFirmEmployees 0.245 (1.69) Shock×LogFirmEmployees2 –0.012 (–1.60) Shock –0.170 –0.168 –1.338 –0.166 (–2.57) (–2.50) (–2.98) (–2.56) Herfindahl 0.076 0.075 0.075 0.075 (2.91) (2.94) (2.91) (2.91) Parity×Herfindahl –0.001 –0.001 –0.001 –0.002 (–0.19) (–0.19) (–0.20) (–0.22) Shock×Herfindahl 0.103 0.102 0.114 0.099 (1.39) (1.37) (1.60) (1.35) Parity×Shock×Herfindahl 0.047 0.047 0.060 0.044 (1.72) (1.71) (1.98) (1.68) Parity –0.040 –0.039 –0.035 8.626 (–1.30) (–1.30) (–1.17) (3.36) Adjusted R2 0.175 0.175 0.175 0.175 Observations 48,896 48,896 48,896 48,896 Year F.E. Yes Yes Yes Yes Add. employment polynomials No No No Yes Establishment F.E. Yes Yes Yes Yes Positive shocks. Since our hypothesis is about insurance against adverse industry shocks, it does not yield direct predictions about how positive industry shocks affect employment of parity firms differently from non-parity firms. However, it is plausible that when parity firms know it is difficult to lay off workers because of implicit contracts enforceable through parity codetermination, they will be more cautious to add workers when business conditions call for more employees. Similarly, if they have retained workers after negative shocks in the past, they are less likely to add new workers at the same rate as non-parity firms. The provision of job security is tantamount to increased firing costs in a dynamic model with adjustment costs. Based on calibrations of such a model (Bentolila and Bertola, 1990), we expect parity firms to be less responsive to positive shocks than non-parity firms. To test this prediction, we compare employment growth in parity and non-parity establishments when the industry as a whole experiences positive growth in employment. The positive growth indicator, Pos. Growth, is equal to one when non-sample firms’ establishments located in Germany belonging to the same industry as a whole increase their work force. We reestimate Table III with Pos. Growth in the place of Shock and report the results in Panel G of Table A-5 of the Online Appendix. The results are consistent with our conjecture. While non-parity establishments increase the number of employees by about 4%, parity establishments show a significantly smaller increase, roughly half those of non-parity establishments. 4.1.c. Employment by occupational status The estimation results based on all employees may mask important heterogeneity across different types of employees. Table VII therefore re-estimates the employment regressions separately for each skill level. We include the same set of control variables as in Specifications (1) and (2) in Table III. For white-collar workers and skilled blue-collar workers, the coefficient on Shock × Parity is positive, economically large, and statistically significant for both specifications. The results for unskilled blue-collar workers are in sharp contrast; both specifications show an insignificant coefficient on Shock × Parity and the sum of Shock and Shock × Parity is significantly negative at the 10% level. Unlike white-collar and skilled blue-collar workers, there is no evidence these workers are protected against an industry-wide decline in employment. Table VII. Employment change—white-collar, skilled blue-collar, and unskilled blue-collar employees This table presents OLS estimation results with the log change in the number of white-collar, skilled blue-collar, or unskilled blue-collar employees as the dependent variable. Only establishments with more than 50 employees are included. All regressions control for year- and establishment-fixed effects. T-statistics are reported in parentheses. Standard errors are clustered at the firm level. The table also reports the p-value for the t-test that Shock + Shock × Parity = 0. Dependent variable White-collar employees Skilled blue-collar employees Unskilled blue-collar employees (1) (2) (3) (4) (5) (6) Shock×Parity 0.076 0.076 0.137 0.133 0.005 0.006 (2.23) (2.14) (2.79) (2.76) (0.12) (0.14) Shock –0.082 –0.082 –0.112 –0.108 –0.063 –0.064 (–2.63) (–2.50) (–2.63) (–2.57) (–2.19) (–2.29) Parity –0.051 –0.050 –0.108 –0.104 0.030 0.029 (–1.57) (–1.59) (–2.41) (–2.92) (1.06) (1.01) LogEstAge –0.633 –0.633 –0.467 –0.467 –0.503 –0.503 (–10.79) (–10.83) (–8.07) (–8.08) (–12.95) (–12.97) LogSales 0.021 –0.005 0.023 –0.019 –0.001 –0.018 (0.85) (–0.03) (1.15) (–0.14) (–0.06) (–0.15) Leverage 0.077 0.074 –0.035 –0.041 –0.043 –0.043 (0.80) (0.82) (–0.76) (–0.88) (–1.94) (–1.80) LogFirmEmployees 0.062 0.154 0.088 0.355 0.087 0.016 (1.37) (0.81) (2.44) (2.51) (2.15) (0.08) LogSales2 0.001 0.001 0.000 (0.13) (0.31) (0.13) LogFirmEmployees2 –0.005 –0.014 0.004 (–0.43) (–1.77) (0.33) Adjusted R2 0.171 0.171 0.064 0.065 0.076 0.076 Observations 48,896 48,896 48,896 48,896 48,896 48,896 t-test: Shock×Parity+ Shock=0 0.709 0.719 0.406 0.397 0.098 0.091 Dependent variable White-collar employees Skilled blue-collar employees Unskilled blue-collar employees (1) (2) (3) (4) (5) (6) Shock×Parity 0.076 0.076 0.137 0.133 0.005 0.006 (2.23) (2.14) (2.79) (2.76) (0.12) (0.14) Shock –0.082 –0.082 –0.112 –0.108 –0.063 –0.064 (–2.63) (–2.50) (–2.63) (–2.57) (–2.19) (–2.29) Parity –0.051 –0.050 –0.108 –0.104 0.030 0.029 (–1.57) (–1.59) (–2.41) (–2.92) (1.06) (1.01) LogEstAge –0.633 –0.633 –0.467 –0.467 –0.503 –0.503 (–10.79) (–10.83) (–8.07) (–8.08) (–12.95) (–12.97) LogSales 0.021 –0.005 0.023 –0.019 –0.001 –0.018 (0.85) (–0.03) (1.15) (–0.14) (–0.06) (–0.15) Leverage 0.077 0.074 –0.035 –0.041 –0.043 –0.043 (0.80) (0.82) (–0.76) (–0.88) (–1.94) (–1.80) LogFirmEmployees 0.062 0.154 0.088 0.355 0.087 0.016 (1.37) (0.81) (2.44) (2.51) (2.15) (0.08) LogSales2 0.001 0.001 0.000 (0.13) (0.31) (0.13) LogFirmEmployees2 –0.005 –0.014 0.004 (–0.43) (–1.77) (0.33) Adjusted R2 0.171 0.171 0.064 0.065 0.076 0.076 Observations 48,896 48,896 48,896 48,896 48,896 48,896 t-test: Shock×Parity+ Shock=0 0.709 0.719 0.406 0.397 0.098 0.091 Table VII. Employment change—white-collar, skilled blue-collar, and unskilled blue-collar employees This table presents OLS estimation results with the log change in the number of white-collar, skilled blue-collar, or unskilled blue-collar employees as the dependent variable. Only establishments with more than 50 employees are included. All regressions control for year- and establishment-fixed effects. T-statistics are reported in parentheses. Standard errors are clustered at the firm level. The table also reports the p-value for the t-test that Shock + Shock × Parity = 0. Dependent variable White-collar employees Skilled blue-collar employees Unskilled blue-collar employees (1) (2) (3) (4) (5) (6) Shock×Parity 0.076 0.076 0.137 0.133 0.005 0.006 (2.23) (2.14) (2.79) (2.76) (0.12) (0.14) Shock –0.082 –0.082 –0.112 –0.108 –0.063 –0.064 (–2.63) (–2.50) (–2.63) (–2.57) (–2.19) (–2.29) Parity –0.051 –0.050 –0.108 –0.104 0.030 0.029 (–1.57) (–1.59) (–2.41) (–2.92) (1.06) (1.01) LogEstAge –0.633 –0.633 –0.467 –0.467 –0.503 –0.503 (–10.79) (–10.83) (–8.07) (–8.08) (–12.95) (–12.97) LogSales 0.021 –0.005 0.023 –0.019 –0.001 –0.018 (0.85) (–0.03) (1.15) (–0.14) (–0.06) (–0.15) Leverage 0.077 0.074 –0.035 –0.041 –0.043 –0.043 (0.80) (0.82) (–0.76) (–0.88) (–1.94) (–1.80) LogFirmEmployees 0.062 0.154 0.088 0.355 0.087 0.016 (1.37) (0.81) (2.44) (2.51) (2.15) (0.08) LogSales2 0.001 0.001 0.000 (0.13) (0.31) (0.13) LogFirmEmployees2 –0.005 –0.014 0.004 (–0.43) (–1.77) (0.33) Adjusted R2 0.171 0.171 0.064 0.065 0.076 0.076 Observations 48,896 48,896 48,896 48,896 48,896 48,896 t-test: Shock×Parity+ Shock=0 0.709 0.719 0.406 0.397 0.098 0.091 Dependent variable White-collar employees Skilled blue-collar employees Unskilled blue-collar employees (1) (2) (3) (4) (5) (6) Shock×Parity 0.076 0.076 0.137 0.133 0.005 0.006 (2.23) (2.14) (2.79) (2.76) (0.12) (0.14) Shock –0.082 –0.082 –0.112 –0.108 –0.063 –0.064 (–2.63) (–2.50) (–2.63) (–2.57) (–2.19) (–2.29) Parity –0.051 –0.050 –0.108 –0.104 0.030 0.029 (–1.57) (–1.59) (–2.41) (–2.92) (1.06) (1.01) LogEstAge –0.633 –0.633 –0.467 –0.467 –0.503 –0.503 (–10.79) (–10.83) (–8.07) (–8.08) (–12.95) (–12.97) LogSales 0.021 –0.005 0.023 –0.019 –0.001 –0.018 (0.85) (–0.03) (1.15) (–0.14) (–0.06) (–0.15) Leverage 0.077 0.074 –0.035 –0.041 –0.043 –0.043 (0.80) (0.82) (–0.76) (–0.88) (–1.94) (–1.80) LogFirmEmployees 0.062 0.154 0.088 0.355 0.087 0.016 (1.37) (0.81) (2.44) (2.51) (2.15) (0.08) LogSales2 0.001 0.001 0.000 (0.13) (0.31) (0.13) LogFirmEmployees2 –0.005 –0.014 0.004 (–0.43) (–1.77) (0.33) Adjusted R2 0.171 0.171 0.064 0.065 0.076 0.076 Observations 48,896 48,896 48,896 48,896 48,896 48,896 t-test: Shock×Parity+ Shock=0 0.709 0.719 0.406 0.397 0.098 0.091 To explain this surprising finding, we hand collect information on the occupational status and the educational and vocational qualification of labor representatives on supervisory boards for 1990, 1999, and 2008, the beginning, the middle, and the end of our sample period, for all firms which provide the relevant information in their annual reports. Table VIII, Panel A, categorizes labor representatives into unskilled blue-collar, skilled blue-collar, white-collar, and union representatives. The occupational status of union representatives is usually not reported, so they form a separate category in the table. In Panel B, we categorize labor representatives as low-qualified, qualified, and highly qualified. Here, we exclude union representatives, because information on union representatives’ educational/vocational qualifications is usually not reported. Panel C shows the occupational status and qualification categories are closely related. Table VIII. Employee representatives by occupational status and qualification Panel A reports the occupational status and Panel B the educational and vocational qualification of labor representatives on supervisory boards. We hand collected this information for all sample firms existing in 1990, 1999, and 2008, the beginning, the middle, and the end of our sample period. The tabulation is based on 97, 231, and 229 labor representatives of 15, 35, and 48 sample firms, respectively, which report the relevant personal information in annual reports. To follow the structure of the IAB data, we categorized labor representatives in Panel A into (1) unskilled blue collar, (2) skilled blue collar, (3) white collar, and (4) union representatives. The occupational status of union representatives is usually not reported, but their occupational status tends to be white collar. Panel B categorizes labor representatives into (1) low qualified, (2) qualified, and (3) highly qualified. In Panel B, union representatives are excluded because their qualification is usually not reported. Panel C presents how the classification based on occupational status corresponds to the breakdown by educational and vocational qualification. It is based on a random sample of 2% of all employees covered by the IAB database between 1975 and 2008 (“Sample of Integrated Labour Market Biographies”). Panel A Occupational status 1990 1999 2008 Unskilled blue collar (%) 0.0 0.0 0.0 Skilled blue collar (%) 42.3 22.5 22.3 White collar (%) 36.1 54.5 56.3 Union representative (%) 21.6 22.9 21.4 Sum (%) 100.0 100.0 100.0 Number of labor representatives 97 231 229 Number of firms 15 35 48 Panel B Qualification 1990 1999 2008 Low qualified (%) 0.0 0.0 0.0 Qualified (%) 86.8 92.1 59.4 Highly qualified (%) 13.2 7.9 40.6 Sum (%) 100.0 100.0 100.0 Number of labor representatives 76 178 180 Number of firms 15 35 48 Panel A Occupational status 1990 1999 2008 Unskilled blue collar (%) 0.0 0.0 0.0 Skilled blue collar (%) 42.3 22.5 22.3 White collar (%) 36.1 54.5 56.3 Union representative (%) 21.6 22.9 21.4 Sum (%) 100.0 100.0 100.0 Number of labor representatives 97 231 229 Number of firms 15 35 48 Panel B Qualification 1990 1999 2008 Low qualified (%) 0.0 0.0 0.0 Qualified (%) 86.8 92.1 59.4 Highly qualified (%) 13.2 7.9 40.6 Sum (%) 100.0 100.0 100.0 Number of labor representatives 76 178 180 Number of firms 15 35 48 Panel C Unskilled blue collar Skilled blue collar White collar Sum Low qualified (%) 15.5 2.2 2.5 20.2 Qualified (%) 9.8 25.6 36.6 72.0 Highly qualified (%) 0.1 0.1 7.7 7.9 Sum (%) 25.4 27.9 46.8 100.0 Panel C Unskilled blue collar Skilled blue collar White collar Sum Low qualified (%) 15.5 2.2 2.5 20.2 Qualified (%) 9.8 25.6 36.6 72.0 Highly qualified (%) 0.1 0.1 7.7 7.9 Sum (%) 25.4 27.9 46.8 100.0 Table VIII. Employee representatives by occupational status and qualification Panel A reports the occupational status and Panel B the educational and vocational qualification of labor representatives on supervisory boards. We hand collected this information for all sample firms existing in 1990, 1999, and 2008, the beginning, the middle, and the end of our sample period. The tabulation is based on 97, 231, and 229 labor representatives of 15, 35, and 48 sample firms, respectively, which report the relevant personal information in annual reports. To follow the structure of the IAB data, we categorized labor representatives in Panel A into (1) unskilled blue collar, (2) skilled blue collar, (3) white collar, and (4) union representatives. The occupational status of union representatives is usually not reported, but their occupational status tends to be white collar. Panel B categorizes labor representatives into (1) low qualified, (2) qualified, and (3) highly qualified. In Panel B, union representatives are excluded because their qualification is usually not reported. Panel C presents how the classification based on occupational status corresponds to the breakdown by educational and vocational qualification. It is based on a random sample of 2% of all employees covered by the IAB database between 1975 and 2008 (“Sample of Integrated Labour Market Biographies”). Panel A Occupational status 1990 1999 2008 Unskilled blue collar (%) 0.0 0.0 0.0 Skilled blue collar (%) 42.3 22.5 22.3 White collar (%) 36.1 54.5 56.3 Union representative (%) 21.6 22.9 21.4 Sum (%) 100.0 100.0 100.0 Number of labor representatives 97 231 229 Number of firms 15 35 48 Panel B Qualification 1990 1999 2008 Low qualified (%) 0.0 0.0 0.0 Qualified (%) 86.8 92.1 59.4 Highly qualified (%) 13.2 7.9 40.6 Sum (%) 100.0 100.0 100.0 Number of labor representatives 76 178 180 Number of firms 15 35 48 Panel A Occupational status 1990 1999 2008 Unskilled blue collar (%) 0.0 0.0 0.0 Skilled blue collar (%) 42.3 22.5 22.3 White collar (%) 36.1 54.5 56.3 Union representative (%) 21.6 22.9 21.4 Sum (%) 100.0 100.0 100.0 Number of labor representatives 97 231 229 Number of firms 15 35 48 Panel B Qualification 1990 1999 2008 Low qualified (%) 0.0 0.0 0.0 Qualified (%) 86.8 92.1 59.4 Highly qualified (%) 13.2 7.9 40.6 Sum (%) 100.0 100.0 100.0 Number of labor representatives 76 178 180 Number of firms 15 35 48 Panel C Unskilled blue collar Skilled blue collar White collar Sum Low qualified (%) 15.5 2.2 2.5 20.2 Qualified (%) 9.8 25.6 36.6 72.0 Highly qualified (%) 0.1 0.1 7.7 7.9 Sum (%) 25.4 27.9 46.8 100.0 Panel C Unskilled blue collar Skilled blue collar White collar Sum Low qualified (%) 15.5 2.2 2.5 20.2 Qualified (%) 9.8 25.6 36.6 72.0 Highly qualified (%) 0.1 0.1 7.7 7.9 Sum (%) 25.4 27.9 46.8 100.0 These tabulations reveal a striking phenomenon: We do not find a single unskilled blue-collar or low-qualified worker among the labor representatives for any firm in any year. Although companies not included in this tabulation may have unskilled blue-collar or low-qualified workers on their boards, the numbers in Table VIII reveal the lack of real representation of unskilled blue-collar or low-qualified workers.30 The implicit insurance contract works only for those who have their own kind represented on the board, an important enforcement mechanism to ensure employers honor the contract. There appears to be an agency problem between unskilled workers and their representatives on supervisory boards. Why unskilled workers cannot remedy this lack of proper representation is puzzling and warrants further investigation, which would require an in-depth analysis of the election process of worker representatives, an issue beyond the scope of this paper. 4.2 Wages The protection against layoffs may not be the result of implementing implicit insurance contracts. The employment protection could be due to employee entrenchment through worker–management collusion (Pagano and Volpin, 2005; Atanassov and Kim, 2009). Whereas the implicit-insurance hypothesis predicts lower wages (Hypothesis 2), the entrenchment hypothesis predicts, if anything, the opposite. Such entrenched employees are unlikely to accept lower wages. Thus, we relate wages to Parity, the coefficient of which measures the wage difference between parity-codetermined and non-parity firms. We use the median wage at each establishment, because the IAB provides only the first quartile, the median, and the third quartile wages. We use two sets of control variables: (1) the same control variables as in Column (2) in Table III and (2) these variables plus the number of employees in the establishment, the median employee age, and the percentage of white-collar employees. Prior research suggests the additional control variables help explain average employee wages (e.g., Oi and Idson, 1999; Brown and Medoff, 1989). We take logs of all level variables when estimating regressions. 4.2.a. Wages for all employees The first two columns in Table IX report estimation results for all employees. The coefficients on Parity are negative and highly significant. The point estimate in Column (2) indicates employees of parity-codetermined firms receive on average about 3.3% lower wages, consistent with Hypothesis 2 and implicit contract theory. Table IX. Wages—all, highly qualified, qualified, and low-qualified employees This table presents OLS estimation results with median wages of all, highly qualified, qualified, or low-qualified employees as the dependent variable. The wage variables are defined as the log of median gross average daily wage for (1) all full-time employees, (2) with higher educational qualifications, (3) with educational/vocational qualifications, and (4) without educational/vocational qualifications. Only establishments with more than 50 employees are included. All regressions control for year- and establishment-fixed effects. The t-statistics are reported in parentheses. Standard errors are clustered at the firm level. Dependent variable All employees Highly qualified Qualified Low-qualified (1) (2) (3) (4) (5) (6) (7) (8) Parity –0.031 –0.033 –0.031 –0.031 –0.029 –0.032 –0.033 –0.034 (–3.10) (–3.78) (–2.49) (–2.56) (–3.01) (–3.70) (–1.59) (–1.66) LogEstAge 0.049 0.049 0.059 0.060 0.050 0.051 0.031 0.030 (3.55) (3.42) (6.18) (6.06) (3.61) (3.49) (1.90) (1.79) LogSales –0.212 –0.194 –0.022 –0.018 –0.244 –0.229 –0.058 –0.046 (–2.27) (–2.37) (–0.37) (–0.32) (–2.70) (–2.85) (–0.60) (–0.52) Leverage –0.024 –0.023 0.005 0.006 –0.016 –0.015 –0.075 –0.073 (–0.93) (–0.90) (0.23) (0.29) (–0.65) (–0.64) (–2.88) (–2.85) LogFirmEmployees 0.032 0.067 –0.062 –0.053 0.039 0.065 0.016 0.023 (0.34) (0.77) (–1.04) (–0.89) (0.44) (0.78) (0.14) (0.22) LogSales2 0.006 0.005 0.001 0.001 0.006 0.006 0.002 0.002 (2.49) (2.61) (0.73) (0.68) (2.83) (3.00) (0.90) (0.85) LogFirmEmployees2 –0.003 –0.004 0.004 0.003 –0.003 –0.003 –0.002 –0.002 (–0.48) (–0.73) (0.94) (0.88) (–0.49) (–0.64) (–0.27) (–0.28) Log#Employees –0.033 –0.009 –0.032 –0.015 (–4.04) (–1.58) (–3.80) (–1.84) LogMedianEmplAge 0.185 0.071 0.194 0.206 (3.98) (2.45) (5.65) (4.86) RatioWhiteCollar 0.153 0.022 0.072 0.047 (2.89) (1.03) (1.59) (0.75) Adjusted R2 0.942 0.945 0.825 0.826 0.926 0.929 0.800 0.801 Observations 51,205 51,205 38,670 38,670 51,060 51,060 42,336 42,336 Dependent variable All employees Highly qualified Qualified Low-qualified (1) (2) (3) (4) (5) (6) (7) (8) Parity –0.031 –0.033 –0.031 –0.031 –0.029 –0.032 –0.033 –0.034 (–3.10) (–3.78) (–2.49) (–2.56) (–3.01) (–3.70) (–1.59) (–1.66) LogEstAge 0.049 0.049 0.059 0.060 0.050 0.051 0.031 0.030 (3.55) (3.42) (6.18) (6.06) (3.61) (3.49) (1.90) (1.79) LogSales –0.212 –0.194 –0.022 –0.018 –0.244 –0.229 –0.058 –0.046 (–2.27) (–2.37) (–0.37) (–0.32) (–2.70) (–2.85) (–0.60) (–0.52) Leverage –0.024 –0.023 0.005 0.006 –0.016 –0.015 –0.075 –0.073 (–0.93) (–0.90) (0.23) (0.29) (–0.65) (–0.64) (–2.88) (–2.85) LogFirmEmployees 0.032 0.067 –0.062 –0.053 0.039 0.065 0.016 0.023 (0.34) (0.77) (–1.04) (–0.89) (0.44) (0.78) (0.14) (0.22) LogSales2 0.006 0.005 0.001 0.001 0.006 0.006 0.002 0.002 (2.49) (2.61) (0.73) (0.68) (2.83) (3.00) (0.90) (0.85) LogFirmEmployees2 –0.003 –0.004 0.004 0.003 –0.003 –0.003 –0.002 –0.002 (–0.48) (–0.73) (0.94) (0.88) (–0.49) (–0.64) (–0.27) (–0.28) Log#Employees –0.033 –0.009 –0.032 –0.015 (–4.04) (–1.58) (–3.80) (–1.84) LogMedianEmplAge 0.185 0.071 0.194 0.206 (3.98) (2.45) (5.65) (4.86) RatioWhiteCollar 0.153 0.022 0.072 0.047 (2.89) (1.03) (1.59) (0.75) Adjusted R2 0.942 0.945 0.825 0.826 0.926 0.929 0.800 0.801 Observations 51,205 51,205 38,670 38,670 51,060 51,060 42,336 42,336 Table IX. Wages—all, highly qualified, qualified, and low-qualified employees This table presents OLS estimation results with median wages of all, highly qualified, qualified, or low-qualified employees as the dependent variable. The wage variables are defined as the log of median gross average daily wage for (1) all full-time employees, (2) with higher educational qualifications, (3) with educational/vocational qualifications, and (4) without educational/vocational qualifications. Only establishments with more than 50 employees are included. All regressions control for year- and establishment-fixed effects. The t-statistics are reported in parentheses. Standard errors are clustered at the firm level. Dependent variable All employees Highly qualified Qualified Low-qualified (1) (2) (3) (4) (5) (6) (7) (8) Parity –0.031 –0.033 –0.031 –0.031 –0.029 –0.032 –0.033 –0.034 (–3.10) (–3.78) (–2.49) (–2.56) (–3.01) (–3.70) (–1.59) (–1.66) LogEstAge 0.049 0.049 0.059 0.060 0.050 0.051 0.031 0.030 (3.55) (3.42) (6.18) (6.06) (3.61) (3.49) (1.90) (1.79) LogSales –0.212 –0.194 –0.022 –0.018 –0.244 –0.229 –0.058 –0.046 (–2.27) (–2.37) (–0.37) (–0.32) (–2.70) (–2.85) (–0.60) (–0.52) Leverage –0.024 –0.023 0.005 0.006 –0.016 –0.015 –0.075 –0.073 (–0.93) (–0.90) (0.23) (0.29) (–0.65) (–0.64) (–2.88) (–2.85) LogFirmEmployees 0.032 0.067 –0.062 –0.053 0.039 0.065 0.016 0.023 (0.34) (0.77) (–1.04) (–0.89) (0.44) (0.78) (0.14) (0.22) LogSales2 0.006 0.005 0.001 0.001 0.006 0.006 0.002 0.002 (2.49) (2.61) (0.73) (0.68) (2.83) (3.00) (0.90) (0.85) LogFirmEmployees2 –0.003 –0.004 0.004 0.003 –0.003 –0.003 –0.002 –0.002 (–0.48) (–0.73) (0.94) (0.88) (–0.49) (–0.64) (–0.27) (–0.28) Log#Employees –0.033 –0.009 –0.032 –0.015 (–4.04) (–1.58) (–3.80) (–1.84) LogMedianEmplAge 0.185 0.071 0.194 0.206 (3.98) (2.45) (5.65) (4.86) RatioWhiteCollar 0.153 0.022 0.072 0.047 (2.89) (1.03) (1.59) (0.75) Adjusted R2 0.942 0.945 0.825 0.826 0.926 0.929 0.800 0.801 Observations 51,205 51,205 38,670 38,670 51,060 51,060 42,336 42,336 Dependent variable All employees Highly qualified Qualified Low-qualified (1) (2) (3) (4) (5) (6) (7) (8) Parity –0.031 –0.033 –0.031 –0.031 –0.029 –0.032 –0.033 –0.034 (–3.10) (–3.78) (–2.49) (–2.56) (–3.01) (–3.70) (–1.59) (–1.66) LogEstAge 0.049 0.049 0.059 0.060 0.050 0.051 0.031 0.030 (3.55) (3.42) (6.18) (6.06) (3.61) (3.49) (1.90) (1.79) LogSales –0.212 –0.194 –0.022 –0.018 –0.244 –0.229 –0.058 –0.046 (–2.27) (–2.37) (–0.37) (–0.32) (–2.70) (–2.85) (–0.60) (–0.52) Leverage –0.024 –0.023 0.005 0.006 –0.016 –0.015 –0.075 –0.073 (–0.93) (–0.90) (0.23) (0.29) (–0.65) (–0.64) (–2.88) (–2.85) LogFirmEmployees 0.032 0.067 –0.062 –0.053 0.039 0.065 0.016 0.023 (0.34) (0.77) (–1.04) (–0.89) (0.44) (0.78) (0.14) (0.22) LogSales2 0.006 0.005 0.001 0.001 0.006 0.006 0.002 0.002 (2.49) (2.61) (0.73) (0.68) (2.83) (3.00) (0.90) (0.85) LogFirmEmployees2 –0.003 –0.004 0.004 0.003 –0.003 –0.003 –0.002 –0.002 (–0.48) (–0.73) (0.94) (0.88) (–0.49) (–0.64) (–0.27) (–0.28) Log#Employees –0.033 –0.009 –0.032 –0.015 (–4.04) (–1.58) (–3.80) (–1.84) LogMedianEmplAge 0.185 0.071 0.194 0.206 (3.98) (2.45) (5.65) (4.86) RatioWhiteCollar 0.153 0.022 0.072 0.047 (2.89) (1.03) (1.59) (0.75) Adjusted R2 0.942 0.945 0.825 0.826 0.926 0.929 0.800 0.801 Observations 51,205 51,205 38,670 38,670 51,060 51,060 42,336 42,336 We cannot completely rule out that employees offer wage cuts in exchange for other benefits we do not observe; for example, for better working conditions. While possible, we believe our explanation is more parsimonious. There is no reason to believe that other aspects of the labor contract, such as working conditions, safety, etc. change abruptly at the 2,000 threshold of employees in Germany. Working conditions are negotiated with work councils at the establishment level and usually not part of board-level negotiations. The discontinuous change at the 2,000 threshold documented through the placebo tests and RD analysis above speaks more in favor of a quid pro quo in exchange for employment insurance. Estimated coefficients on controls are mostly consistent with intuition. Older employees and employees working in older establishments and establishments with a greater proportion of white collar workers are paid more. However, the number of employees in establishments is associated with lower wages. This is somewhat surprising given the finding in Brown and Medoff (1989) that an increase in the number of employees is associated with higher wages. The difference could be due to differences in sample and specification. Our sample is at the establishment level, heavily skewed toward large firms, and our regression contains a number of other firm size variables, whereas Brown and Medoff (1989) rely on firm-level data, include small businesses, and their specification contains fewer size controls. We do not include Shock and Shock × Parity in the regressions, because we are primarily interested in the difference in average wages between parity and non-parity firms. For consistency with the employment regressions, we include Shock and Shock × Parity in the wage regressions and report results in Table A-6, Panel B, Columns (1) and (2). The results are robust; employees of parity firms receive 3.0–3.1% lower wages during non-shock periods and the difference is significant. The coefficients on Shock are insignificant. The Shock × Parity term shows positive but insignificant coefficients. These results confirm our prior that German wages are sticky and do not respond to industry shocks in significant ways. 4.2.b. Wages by qualification The IAB does not provide the wage data broken down by occupational status. It provides only the breakdown by qualifications. To see how the classification based on qualifications corresponds to occupational status, at our request, IAB cross-tabulated the percentage of employees belonging to each type of occupational status and qualification based on a random sample of 2% of all employees covered by its database between 1975 and 2008 (Sample of Integrated Labour Market Biographies). The tabulation is shown in Table VIII, Panel C. Most highly qualified workers tend to be white-collar workers; most qualified workers are either white collar or skilled blue-collar workers; and most low-qualified workers are unskilled blue-collar workers. Columns (3)–(8) in Table IX report separate estimates for employees grouped by qualification.31 Coefficients on Parity range from –2.9% to –3.4% for all three qualification levels. The coefficients are highly significant for the qualified and highly qualified groups, suggesting that skilled blue-collar and white-collar employees of parity firms receive significantly lower wages. For low-qualified employees, the coefficient on Parity is not significant, even though the size of the coefficient is similar. This group of employees has larger standard errors, probably because roughly one-fourth belongs to skilled blue-collar or white-collar workers (see Table VIII, Panel C). 4.2.c. Identification issues Placebo tests. As in the employment analysis, we re-estimate the wage results while replacing the parity dummy with placebo dummies. The results are reported in Table A-8 of the Online Appendix, separately for parity and non-parity firms. Coefficients on placebo dummies are statistically zero for all four regressions. We also repeat placebo regressions for all employees with Placebo equal to one if employment exceeds a threshold that varies from 1,500 to 2,500 in increments of 50. Figure 7 shows the regression coefficients on Placebo (right axis, gray line) and the t-statistic (left axis, black line). The lowest values are reached around the 2,000 threshold, supporting the conclusion that the relevance of this threshold is genuine.32 Figure 7. View largeDownload slide Wages—placebo tests. Notes: This figure shows the coefficients on Placebo (right-hand axis, gray line) and the corresponding t-statistics (left-hand axis, black line) from estimating Column (2) in Table IX with Placebo instead of Parity. We vary the placebo threshold in steps of 50 from 1,500 to 2,500 and let Placebo be equal to one for any establishment of a firm with employment in Germany above the corresponding placebo threshold. Figure 7. View largeDownload slide Wages—placebo tests. Notes: This figure shows the coefficients on Placebo (right-hand axis, gray line) and the corresponding t-statistics (left-hand axis, black line) from estimating Column (2) in Table IX with Placebo instead of Parity. We vary the placebo threshold in steps of 50 from 1,500 to 2,500 and let Placebo be equal to one for any establishment of a firm with employment in Germany above the corresponding placebo threshold. RD analysis. We also test whether our conclusions regarding wages hold locally around the 2,000-employee cut-off with RD analyses. We follow the same steps as in the employment analysis. Unlike the employment analysis, however, we do not see a drop in sample size because the estimation is done for the entire cycle of shock and non-shock periods. Results are reported in Table X, with Figure 8 plotting local polynomial regressions around the 2,000 threshold with 95% confidence intervals in dashed lines. The drop in wages above the 2,000 threshold is significant for the optimal bandwidth, implying that parity firms’ employees are paid less than those at non-parity firms. The estimate for the insurance premium at the optimal bandwidth is about 8%, much larger than those estimated via the difference-in-differences approach, because the RD analyses do not contain the control variables used in the difference-in-differences approach. Table X. Wages—RD analysis This table presents results for a kernel regression using a triangular kernel and the optimized (Imbens and Kalyanaraman, 2012) bandwidth (bw = 1). The dependent variable is log median daily wage. We modify the optimized bandwidth by factors of 0.5 and 2 to check robustness. A sharp RD design is assumed, where the treatment variable (Parity) changes from zero to one at 2,000 employees (firm level). The z-statistics for the coefficient estimates are reported in parentheses below the estimates. Dependent variable Log median daily wage Parity (bw = 1) –0.079 (–1.95) Parity (bw = 0.5) –0.148 (–2.58) Parity (bw = 2) –0.039 (–1.26) Observations 4,962 Dependent variable Log median daily wage Parity (bw = 1) –0.079 (–1.95) Parity (bw = 0.5) –0.148 (–2.58) Parity (bw = 2) –0.039 (–1.26) Observations 4,962 Table X. Wages—RD analysis This table presents results for a kernel regression using a triangular kernel and the optimized (Imbens and Kalyanaraman, 2012) bandwidth (bw = 1). The dependent variable is log median daily wage. We modify the optimized bandwidth by factors of 0.5 and 2 to check robustness. A sharp RD design is assumed, where the treatment variable (Parity) changes from zero to one at 2,000 employees (firm level). The z-statistics for the coefficient estimates are reported in parentheses below the estimates. Dependent variable Log median daily wage Parity (bw = 1) –0.079 (–1.95) Parity (bw = 0.5) –0.148 (–2.58) Parity (bw = 2) –0.039 (–1.26) Observations 4,962 Dependent variable Log median daily wage Parity (bw = 1) –0.079 (–1.95) Parity (bw = 0.5) –0.148 (–2.58) Parity (bw = 2) –0.039 (–1.26) Observations 4,962 Figure 8. View largeDownload slide Discontinuity at the 2,000-employee cutoff: wages. Notes: This figure shows the results of two local polynomial regressions for log median daily wage around the parity cutoff of 2,000 employees. Figure 8. View largeDownload slide Discontinuity at the 2,000-employee cutoff: wages. Notes: This figure shows the results of two local polynomial regressions for log median daily wage around the parity cutoff of 2,000 employees. Alternative sample constructions. As with the employment regressions, we re-estimate our baseline regressions while (1) lowering the threshold of the minimum number of employees in an establishment for inclusion in our sample to ten employees and (2) truncating the sample at the 1st and 99th percentile of employment growth. The re-estimation results reported in Panels C and D of Table A-6 of the Online Appendix are robust. 5. Operating Risks and Firm Performance In this section, we examine how the employment insurance affects firm risk and performance over the cycle of non-shock and shock periods. These tests require regression estimation at the firm level, so we redefine our shock measure as FirmShock, the proportion of a firm’s employees working in establishments in industries for which Shock = 1. FirmShock is a weighted average of Shock in a given firm-year, ranging between 0 and 1. For example, if 60% of a firm’s employees work in industries in which Shock equals 1, and the remaining 40% work in industries not subject to a shock in that year, then FirmShock equals 0.6. 5.1 Operating Leverage Providing job guarantees limits firms’ ability to reduce payroll in response to changes in technology, consumer taste, or general business conditions. Then wages become more fixed costs rather than variable costs, increasing operating leverage. To test this prediction, we relate Parity to two different measures of operating leverage. First, we follow Mandelker and Rhee (1984) and Chen, Kacperczyk, and Ortiz-Molina (2011) and use the elasticity of a firm’s operating income (after depreciation) with respect to sales, the Mandelker and Rhee degree of operating leverage (MRDOL). We estimate it for firms with positive EBIT using a time-series regression of log (EBIT) on log (Sales) using the five most recent annual observations. Second, we use the Novy-Marx (2011) operating leverage measure, which is cost of goods sold plus selling, general and administrative expenses, scaled by the book value of assets (NMOL). For each measure of operating leverage, we use two different specifications, one with year- and industry-fixed effects; the other with year- and firm-fixed effects. The control variables are similar to Chen, Kacperczyk, and Ortiz-Molina (2011). Results are reported in Table XI, Panel A without control variables. The coefficient on Parity is positive and statistically significant, regardless of the specification. It appears employment insurance through parity codetermination increases operating leverage. Table XI. Firm-level regressions: operating leverage, effects on ROA and Q Panel A presents OLS estimation results with (1) the Mandelker and Rhee (1984) measure of operating leverage (MRDOL) and (2) the Novy-Marx (2011) measure of operating leverage (NMOL) as the dependent variable. Following Mandelker and Rhee (1984), we estimate for firms with positive EBIT the degree of operating leverage using a time-series regression of log(EBIT) on log(Sales) using the five most recent annual observations. Novy-Marx (2011) defines operating leverage as cost of goods sold plus selling, general and administrative expenses, scaled by the book value of assets. Control variables (LogAssets, LogTobinsQ, FixedAssets/Assets, Leverage, LogFirmAge, and FirmEmployees/Assets) are not reported to conserve space. The t-statistics are reported in parentheses. Standard errors are clustered at the firm level. Panel B presents OLS estimation results with (1) ROA and (2) log Tobin’s Q as the dependent variable. The FirmShock variable is defined as the weighted average of Shock across all establishments in a firm-year. All regressions control for year- and firm-fixed effects. The t-statistics are reported in parentheses. Standard errors are clustered at the firm level. Panel A Dependent variable MRDOL NMOL (1) (2) (3) (4) Parity 0.177 0.135 0.297 0.127 (2.16) (1.79) (4.67) (2.60) Adjusted R2 0.146 0.241 0.736 0.913 Observations 1,442 1,442 953 953 Year F.E. Yes Yes Yes Yes Industry F.E. Yes No Yes No Firm F.E. No Yes No Yes Panel B Dependent variable ROA LogTobinsQ (1) (2) (3) (4) FirmShock×Parity –0.036 –0.092 (–2.59) (–1.80) FirmShock –0.030 –0.075 (–2.40) (–1.70) Parity 0.003 –0.003 0.031 0.025 (0.29) (–0.38) (1.58) (1.27) LogFirmAge –0.005 –0.006 –0.037 –0.038 (–0.75) (–0.83) (–2.26) (–2.32) LogSales –0.029 –0.029 –0.747 –0.757 (–0.65) (–0.67) (–8.39) (–8.53) Leverage –0.056 –0.055 –0.249 –0.249 (–5.15) (–5.06) (–9.72) (–9.72) LogFirmEmployees –0.031 –0.030 0.277 0.280 (–1.85) (–1.76) (6.32) (6.39) LogSales2 0.001 0.001 0.018 0.018 (1.36) (1.38) (8.38) (8.53) LogFirmEmployees2 0.002 0.002 –0.019 –0.019 (1.32) (1.25) (–5.95) (–6.03) Adjusted R2 0.545 0.543 0.682 0.681 Observations 1,734 1,734 1,885 1,885 Panel A Dependent variable MRDOL NMOL (1) (2) (3) (4) Parity 0.177 0.135 0.297 0.127 (2.16) (1.79) (4.67) (2.60) Adjusted R2 0.146 0.241 0.736 0.913 Observations 1,442 1,442 953 953 Year F.E. Yes Yes Yes Yes Industry F.E. Yes No Yes No Firm F.E. No Yes No Yes Panel B Dependent variable ROA LogTobinsQ (1) (2) (3) (4) FirmShock×Parity –0.036 –0.092 (–2.59) (–1.80) FirmShock –0.030 –0.075 (–2.40) (–1.70) Parity 0.003 –0.003 0.031 0.025 (0.29) (–0.38) (1.58) (1.27) LogFirmAge –0.005 –0.006 –0.037 –0.038 (–0.75) (–0.83) (–2.26) (–2.32) LogSales –0.029 –0.029 –0.747 –0.757 (–0.65) (–0.67) (–8.39) (–8.53) Leverage –0.056 –0.055 –0.249 –0.249 (–5.15) (–5.06) (–9.72) (–9.72) LogFirmEmployees –0.031 –0.030 0.277 0.280 (–1.85) (–1.76) (6.32) (6.39) LogSales2 0.001 0.001 0.018 0.018 (1.36) (1.38) (8.38) (8.53) LogFirmEmployees2 0.002 0.002 –0.019 –0.019 (1.32) (1.25) (–5.95) (–6.03) Adjusted R2 0.545 0.543 0.682 0.681 Observations 1,734 1,734 1,885 1,885 Table XI. Firm-level regressions: operating leverage, effects on ROA and Q Panel A presents OLS estimation results with (1) the Mandelker and Rhee (1984) measure of operating leverage (MRDOL) and (2) the Novy-Marx (2011) measure of operating leverage (NMOL) as the dependent variable. Following Mandelker and Rhee (1984), we estimate for firms with positive EBIT the degree of operating leverage using a time-series regression of log(EBIT) on log(Sales) using the five most recent annual observations. Novy-Marx (2011) defines operating leverage as cost of goods sold plus selling, general and administrative expenses, scaled by the book value of assets. Control variables (LogAssets, LogTobinsQ, FixedAssets/Assets, Leverage, LogFirmAge, and FirmEmployees/Assets) are not reported to conserve space. The t-statistics are reported in parentheses. Standard errors are clustered at the firm level. Panel B presents OLS estimation results with (1) ROA and (2) log Tobin’s Q as the dependent variable. The FirmShock variable is defined as the weighted average of Shock across all establishments in a firm-year. All regressions control for year- and firm-fixed effects. The t-statistics are reported in parentheses. Standard errors are clustered at the firm level. Panel A Dependent variable MRDOL NMOL (1) (2) (3) (4) Parity 0.177 0.135 0.297 0.127 (2.16) (1.79) (4.67) (2.60) Adjusted R2 0.146 0.241 0.736 0.913 Observations 1,442 1,442 953 953 Year F.E. Yes Yes Yes Yes Industry F.E. Yes No Yes No Firm F.E. No Yes No Yes Panel B Dependent variable ROA LogTobinsQ (1) (2) (3) (4) FirmShock×Parity –0.036 –0.092 (–2.59) (–1.80) FirmShock –0.030 –0.075 (–2.40) (–1.70) Parity 0.003 –0.003 0.031 0.025 (0.29) (–0.38) (1.58) (1.27) LogFirmAge –0.005 –0.006 –0.037 –0.038 (–0.75) (–0.83) (–2.26) (–2.32) LogSales –0.029 –0.029 –0.747 –0.757 (–0.65) (–0.67) (–8.39) (–8.53) Leverage –0.056 –0.055 –0.249 –0.249 (–5.15) (–5.06) (–9.72) (–9.72) LogFirmEmployees –0.031 –0.030 0.277 0.280 (–1.85) (–1.76) (6.32) (6.39) LogSales2 0.001 0.001 0.018 0.018 (1.36) (1.38) (8.38) (8.53) LogFirmEmployees2 0.002 0.002 –0.019 –0.019 (1.32) (1.25) (–5.95) (–6.03) Adjusted R2 0.545 0.543 0.682 0.681 Observations 1,734 1,734 1,885 1,885 Panel A Dependent variable MRDOL NMOL (1) (2) (3) (4) Parity 0.177 0.135 0.297 0.127 (2.16) (1.79) (4.67) (2.60) Adjusted R2 0.146 0.241 0.736 0.913 Observations 1,442 1,442 953 953 Year F.E. Yes Yes Yes Yes Industry F.E. Yes No Yes No Firm F.E. No Yes No Yes Panel B Dependent variable ROA LogTobinsQ (1) (2) (3) (4) FirmShock×Parity –0.036 –0.092 (–2.59) (–1.80) FirmShock –0.030 –0.075 (–2.40) (–1.70) Parity 0.003 –0.003 0.031 0.025 (0.29) (–0.38) (1.58) (1.27) LogFirmAge –0.005 –0.006 –0.037 –0.038 (–0.75) (–0.83) (–2.26) (–2.32) LogSales –0.029 –0.029 –0.747 –0.757 (–0.65) (–0.67) (–8.39) (–8.53) Leverage –0.056 –0.055 –0.249 –0.249 (–5.15) (–5.06) (–9.72) (–9.72) LogFirmEmployees –0.031 –0.030 0.277 0.280 (–1.85) (–1.76) (6.32) (6.39) LogSales2 0.001 0.001 0.018 0.018 (1.36) (1.38) (8.38) (8.53) LogFirmEmployees2 0.002 0.002 –0.019 –0.019 (1.32) (1.25) (–5.95) (–6.03) Adjusted R2 0.545 0.543 0.682 0.681 Observations 1,734 1,734 1,885 1,885 The higher operating leverage implies that parity firms will suffer greater relative reductions in profitability and valuation from industry shocks than non-parity firms. To test these implications, we estimate difference-in-differences in ROA and the logarithm of Tobin’s Q, our measures of profitability and valuation, respectively. ROA is defined as EBITDA/Total Assets to avoid changes in depreciation and amortization methods during our sample period affecting estimation results. Tobin’s Q is defined as the market value of common equity plus total assets minus the book value of common equity divided by total assets. Our main interest is again the coefficient of FirmShock × Parity, which we expect to be negative for ROA and Q. The estimation results are reported in Table XI, Panel B, Columns (1) and (3). The coefficient on FirmShock × Parity is significant and negative in both specifications. The economic magnitude is also large. The estimates for ROA show that profitability of parity-codetermined firms falls by 3.6% more if all employees of a firm are affected by a shock. This number is substantial; Table I shows that the mean (median) ROA of all firm-year observations in the sample is 7.5% (6.9%) and the coefficient on FirmShock for non-parity firms is only –3.0%. The incremental decline in Tobin’s Q for parity firms is 9.2% if all employees are affected by a shock, which is larger than the decline of 7.5% for non-parity firms. These estimates for ROA and Tobin’s Q suggest that parity codetermination more than doubles the negative impact of shocks on profitability and valuation relative to non-parity firms. 5.2 Are Shareholders Properly Compensated for Bearing the Risk? These results imply that the employment insurance increases shareholder risk—the risk of getting hurt more during industry downturns. Is the insurance premium in the form of lower wages sufficient to compensate shareholders for the risk? More generally, are shareholders worse or better off as the result of parity codetermination? We investigate this issue by reestimating the ROA and Q regressions while omitting FirmShock and FirmShock ×Parity and report the re-estimated coefficients of Parity in Columns (2) and (4) of Table XI, Panel B. The coefficient on Parity is insignificant for both ROA and Q, implying operating profits and shareholder value through the cycle are unrelated to parity codetermination. This conclusion is in line with previous findings in the literature on German codetermination.33 It appears workers capture all the efficiency gains from the improved risk-sharing arrangement through the implicit insurance contract with firms; however, they do pay an insurance premium that is on average sufficient to cover the risks of providing the insurance. Put differently, the wage concessions workers make to firms seem to be just sufficient to compensate shareholders for higher operating leverage. 6. Conclusion We find parity-codetermined firms provide skilled employees greater protection against layoffs during adverse industry shocks. There are two alternative explanations. According to the insurance hypothesis, parity-codetermination serves as an enforcement mechanism to ensure firms honor implicit insurance contracts, whereby workers receive protection against adverse shocks in return for accepting lower wages. The alternative explanation, employee entrenchment, suggests the worker control rights bestowed by parity-codetermination allow employees to prevent layoffs when the industry suffers adverse shocks, but without offering adequate concessions in the form of lower wages. We distinguish the two competing explanations by examining the wage differential between parity and non-parity firms and by analyzing firm valuation and performance. We find employees of parity firms receive significantly lower wages relative to those working for non-parity firms. Our firm performance analyses indicate that shareholders do not gain or lose from improved risk sharing and workers reap all the benefits from this arrangement. Both findings favor the insurance hypothesis. However, only skilled blue-collar and white-collar workers benefit from employment insurance, whereas unskilled blue-collar workers do not. We attribute this finding to the lack of real representation of unskilled blue-collar workers on supervisory boards. Supplementary Material Supplementary data are available at Review of Finance online. Footnotes * We are grateful for helpful comments and suggestions made by Daniel Ferreira, Fangjian Fu, Yao Lu, David Matsa, Marco Pagano, Page Quimet, David Sraer, Alex Stomper, Gary Twite, and participants at the Ackerman Conference on Corporate Governance at Bar Ilan University, the Duke-UNC Conference on Corporate Finance, the Singapore International Conference on Finance, the Rothschild Caesarea Center 11th Annual Conference, the CEPR/Study Center Gerzensee European Summer Symposium, the China International Conference in Finance, CSEF Conference on Finance and Labor, 2014 American Finance Association Annual Meetings, and seminar participants at the University of Amsterdam, University of British Columbia, University of Hong Kong, University of Lausanne, University of Mannheim, University of Maryland, University of Michigan, UNSW, and Nanyang Technological University. We are also grateful to Stefan Bender and his team at the Institut für Arbeitsmarkt- und Berufsforschung for providing access to their data. 1 Exceptions are, for example, Fama and Schwert (1977), Campbell (1996), and Jagannathan and Wang (1996). 2 On risk-return trade-offs see Palacios-Huerta (2003) and Lustig and Van Nieuwerburgh (2008). On financial distress, optimal capital structure, and restructuring, see Berk, Stanton, and Zechner (2010); Agrawal and Matsa (2013); Graham et al. (2016); and Agrawal and Tambe (2016). These studies are selected from a much larger literature on finance and labor. We do not provide a systematic and comprehensive survey of this literature here. 3 Azariadis (1975) assumes that firms that do not honor implicit contracts would “suffer a catastrophic loss in reputation” (p. 1187) and Rudanko (2011) invokes a similar assumption with the claim “equilibrium contracts are likely to be self-enforcing for a range of reasonable parameterizations.” (pp. 2823–2824). Thomas and Worrall (1988) assume that agents can always renege on a contract, but at an exogenously imposed cost: they will be excluded from all trading in the future if they renege on a contract. 4 Maybe the closest to our approach is Malcomson (1983), who argues unions enforce implicit contracts. 5 Data are from the CBR-LRI dataset (Adams et al., 2017) and refer to earlier years if data for 2016 were not available. Of the 117 countries covered, 27 grant workers board-level representation, albeit to varying degrees. Between 1970 and 2016, there were 28 changes of these laws in 19 countries, of which 22 led to more, and only six to lower, worker representation. Moreover, 92 countries provide workers with influence at the plant level or firm level through work councils. These laws were amended 78 times in 58 countries since 1970, leading to more influence for workers in 60 cases and less influence in 18 cases. 6 See Section 3.3.2 for a detailed description of how industry shocks are defined and constructed. 7 Guiso, Pistaferri, and Schivardi (2005) provide a calibration based on a parameterization of Lucas’s tree model following Clark, Leslie, and Symons (1994). They estimate that workers would be indifferent between the partial insurance the authors find for Italian firms, and uninsured consumption that is on average 9% higher than insured consumption. 8 Renaud (2007) surveys 13 studies investigating the impact of codetermination on company performance using different methodological approaches, sample constructions, and performance variables. The overall evidence seems inconclusive, suggesting no systematic relation between parity codetermination and firm performance. 9 See Azariadis and Stiglitz (1983) and Rosen (1985) for early surveys of this literature. 10 Bach and Serrano-Velarde (2015) argue that it is family links between CEOs and their successors that enhances firms’ ability to commit to implicit contracts. 11 Ellul, Pagano, and Schivardi (2017) estimate this premium to be 6% based on firm-level data. 12 Without frictions in the labor market, only partial insurance is feasible, because workers always receive pay increases if their marginal product rises above their wage. Harris and Holmstrom (1982) and Thomas and Worrall (1988) discuss contracting problems in such a setting. 13 Berk and Walden (2013) argue that firms insure workers’ human capital risk and investors spread this risk by investing in diversified portfolios. Workers could insure themselves directly through participation in capital markets. However, Berk and Walden (2013) show indirect insurance through firms is sufficiently close to being optimal so that workers prefer it to direct participation in capital markets even if the costs of direct participation are small. 14 For firms with more than 2,000 and up to 10,000 employees, the supervisory board has twelve seats, of which six are shareholder representatives, three are from the firm’s workers, one is from middle management, and two are from unions. For firms with up to 20,000 (over 20,000 employees), the size of the supervisory board increases to 16 (20) and the number of representatives from each constituency becomes: shareholders 8 (10), workers 5 (6), middle management 1 (1), and union representatives 2 (3). 15 The industry classification is based on the Statistical Classification of Economic Activities in the European Community (NACE), a six-digit industry classification. NACE is similar to North American Industry Classification System (NAICS). The first four levels are the same for all European countries. The IAB database contains different versions of the NACE classification. We use NACE Revision 1.1, which is based on the International Standard Industrial Classification (ISIC Rev. 3) of the United Nations. 16 Other groups are employees in vocational training, home workers, master craftsmen, and part-time employees. We do not analyze these groups of employees because they usually form only a small fraction of employees and are present in relatively few establishments. 17 In Germany, a relatively small fraction of students obtains an upper secondary school degree (high school, Abitur), although the fraction rose from 31% in 1992 to 45% in 2008. The typical career path in Germany is to leave school after tenth grade to enter vocational training. In 2009, 57.8% of the German population had such a vocational qualification, while 27.8% had none of these qualifications. IAB classifies the former as qualified employees and the latter as low-qualified employees (see Hethey-Maier and Seth, 2010). 18 Gorton and Schmid (2004) and several of the studies surveyed by Renaud (2007) also use parity codetermination as their main variable for labor representation. 19 We experimented with two other definitions of shocks. The first alternative makes shocks comparable across industries with different cash-flow volatilities by scaling shocks with the standard deviation of the industry-growth rate of employment, so that a lower cut-off applies to more volatile industries. The results are qualitatively similar, but statistically weaker. The second alternative uses sales growth or growth in operating income of firms from other European countries to define industry-level shocks. These analyses mostly yield insignificant estimates on the shock variable [insignificant θ in Equation (1)]. 20 Examples include Bosch, Aldi, Boehringer Ingelheim, Edeka, Rewe Group, Haniel, Shell Germany, BP Germany, Ford, Coca Cola, Procter and Gamble, Dow Chemical, Pfizer, IBM, Hewlett-Packard, ExxonMobil, Vodafone, Gazprom Germania, Sanofi-Aventis Germany, Telefónica Germany, and Fujitsu. 21 If parity firms hold on to more employees during negative shock periods, they may hire fewer new employees in expansion periods. Thus, the trends may not to be strictly parallel. However, Figure 2 suggests that this effect is too subtle to carry to noticeable weight. 22 The correlations between the parity indicator and standard measures of firm size are 0.15, 0.16, and 0.27 for, respectively, firms’ total assets, market capitalization, and sales. 23 The coefficients on the interaction terms of Shock with the employment-level variables suggest employment protection increases with the employment level in Germany at a decreasing rate. 24 Shareholder-directors’ resistance to crossing the 2,000 threshold, if any, cannot be due to their concern about losing their board seats, because when a non-parity firm crosses the threshold board size increases to accommodate the additional employee representatives. In our sample, when a firm becomes parity-codetermined, on average, its board size increases by five seats, leading to one more shareholder and four more employee representatives. 25 We cannot report sample averages for each bin, as is customary for local polynomial plots, because the IAB, our data provider, does not allow us to report moments of samples with fewer than thirty observations. 26 We use employment to construct a continuous measure of shocks instead of sales used in Sraer and Thesmar (2007) because the IAB does not provide data on sales at the establishment level. 27 Our specification replicates equation (4) of Sraer and Thesmar (2007). Their specification in Table VII, Column (3) closely resembles our specification in Table V (Panel A), Column (3), which includes year- and establishment-fixed effects. 28 The loss of employees causing the changes in parity status is mostly due to major asset sales. In most cases, the annual reports mention sales of certain parts of the firm, which are intended to change the focus of operations. 29 Employee transfers between establishments belonging to the firm and same industry (i.e., moving employees from one establishment to another in the same industry) will cancel out and will not be observed. They are not of concern because such transfers are unlikely to be a response to industry shocks. 30 Interestingly, the IAB’s tabulation based on a random sample shows a disproportionately large percentage of foreign workers in the unskilled blue-collar worker category. Whereas Germans represent 93% of skilled blue-collar workers and 96% of white-collar workers, they represent only 80% of unskilled blue-collar workers, with the rest being foreign workers. 31 To see how well the qualification categories correspond to the occupational categories for the regression purpose, we repeat the employment regressions using the breakdown by educational and vocational qualifications and report the estimation results in Table A-7 in the Online Appendix. The results are qualitatively similar to those based on occupational status in Table VII. The statistical significance of the coefficient on Shock × Parity is lower for highly qualified and qualified workers because of the imperfect correlation between occupational status and qualification. 32 The figure does not show a sharp drop at 2,000 because we gradually “misclassify” more and more establishments as we move the placebo threshold further away from the true threshold. 33 Of the four studies surveyed by Renaud (2007) that use either Tobin’s Q or the market-to-book ratio, two find negative effects and the other two find no effect of worker representation. In a study not covered by Renaud (2007), Petry (2018) performs an event study around the transition dates when firms announce changes in their parity status and finds a negative impact of 2.3% of changing from non-parity to parity status. This seems to contradict Baums and Frick (1999), who also conduct an event study based on twenty-eight court cases that addressed the power of unions and workers on codetermined supervisory boards and find no impact on firm value. References Abadie A. , Athey S. , Imbens G. W. , Wooldridge J. ( 2017 ): When should you adjust standard errors for clustering? NBER Working paper 24003. Abraham K. G. , Houseman S. N. ( 1989 ): Job security and work force adjustment: how different are U.S. and Japanese practices? , Journal of the Japanese and International Economies 3 , 500 – 521 . Google Scholar CrossRef Search ADS Adams Z. , Bastani P. , Bishop L. , Deakin S. ( 2017 ): The CBR–LRI dataset: methods, properties & potential of leximetric coding of labour laws, International Journal of Comparative Labour Law and Industrial Relations 33 , 59 – 91 . Addison J. T. , Pahnke A. , Bryson A. , Bellmann L. , Teixeira P. ( 2010 ): The state of collective bargaining and worker representation in Germany: the erosion continues, IZA Discussion Paper 5030. Agrawal A. , Tambe P. ( 2016 ): Private equity and workers’ career paths: the role of technological change , Review of Financial Studies 29 , 2455 – 2489 . Google Scholar CrossRef Search ADS Agrawal A. K. , Matsa D. A. ( 2013 ): Labor unemployment risk and corporate financing decisions , Journal of Financial Economics 108 , 449 – 470 . Google Scholar CrossRef Search ADS Atanassov J. , Kim E. H. ( 2009 ): Labor and corporate governance: international evidence from restructuring , Journal of Finance 64 , 341 – 374 . Google Scholar CrossRef Search ADS Azariadis C. ( 1975 ): Implicit contracts and underemployment equilibria , Journal of Political Economy 83 , 1183 – 1202 . Google Scholar CrossRef Search ADS Azariadis C. , Stiglitz J. E. ( 1983 ): Implicit contracts and fixed price equilibria , The Quarterly Journal of Economics 98 , 2 – 22 . Bach L. , Serrano-Velarde N. ( 2015 ): CEO identity and labor contracts: evidence from CEO transitions , Journal of Corporate Finance 33 , 227 – 242 . Google Scholar CrossRef Search ADS Baily M. N. ( 1974 ): Wages and employment under uncertain demand , Review of Economic Studies 41 , 37 – 50 . Google Scholar CrossRef Search ADS Baums T. , Frick B. ( 1999 ): The market value of the codetermined firm, in: Blair M. M. , Roe M. J. (eds.), Employees and Corporate Governance , Brookings Institution Press , Washington, D.C ., pp. 206 – 234 . Beaudry P. , DiNardo J. ( 1991 ): The effect of implicit contracts on the movement of wages over the business cycle: evidence from micro data , Journal of Political Economy 99 , 665 – 688 . Google Scholar CrossRef Search ADS Benmelech E. , Bergman N. , Kim H. ( 2018 ): Strong employers and weak employees: how does employer concentration affect wages?, National Bureau of Economic Research Working Paper Series 24307. Bentolila S. , Bertola G. ( 1990 ): Firing costs and labour demand: how bad is eurosclerosis? , The Review of Economic Studies 57 , 381 – 402 . Google Scholar CrossRef Search ADS Berk J. B. , Stanton R. , Zechner J. ( 2010 ): Human capital, bankruptcy, and capital structure , Journal of Finance 65 , 891 – 926 . Google Scholar CrossRef Search ADS Berk J. B. , Walden J. ( 2013 ): Limited capital market participation and human capital risk , Review of Asset Pricing Studies 3 , 1 – 37 . Google Scholar CrossRef Search ADS Brown C. , Medoff J. L. ( 1989 ): The employer size–wage effect , Journal of Political Economy 97 , 1027 – 1059 . Google Scholar CrossRef Search ADS Campbell J. Y. ( 1996 ): Understanding risk and return , Journal of Political Economy 104 , 298 – 345 . Google Scholar CrossRef Search ADS Chen H. , Kacperczyk M. T. , Ortiz-Molina H. ( 2011 ): Labor unions, operating flexibility, and the cost of equity , Journal of Financial and Quantitative Analysis 46 , 25 – 58 . Google Scholar CrossRef Search ADS Clark K. , Leslie D. , Symons E. ( 1994 ): The costs of recession , The Economic Journal 104 , 20 – 36 . Google Scholar CrossRef Search ADS D’Aurizio L. , Romano L. ( 2013 ): Family firms and the great recession: out of sight, out of mind? Bank of Italy Working paper 905. Dittmann I. , Maug E. , Schneider C. ( 2010 ): Bankers on the boards of German firms: what they do, what they are worth, and why they are (still) there , Review of Finance 14 , 35 – 71 . Google Scholar CrossRef Search ADS Ellul A. , Pagano M. , Schivardi F. ( 2017 ): Employment and wage insurance within firms: worldwide evidence. ECGI Finance working paper 517/2017. Faleye O. , Mehrotra V. , Morck R. ( 2006 ): When labor has a voice in corporate governance , Journal of Financial and Quantitative Analysis 41 , 489 – 510 . Google Scholar CrossRef Search ADS Fama E. F. , Schwert G. W. ( 1977 ): Human capital and capital market equilibrium , Journal of Financial Economics 4 , 95 – 125 . Google Scholar CrossRef Search ADS Gamber E. N. ( 1988 ): Long-term risk-sharing wage contracts in an economy subject to permanent and temporary shocks , Journal of Labor Economics 6 , 83 – 99 . Google Scholar CrossRef Search ADS Gordon D. F. ( 1974 ): A neo-classical theory of keynesian unemployment , Economic Inquiry 12 , 431 – 459 . Google Scholar CrossRef Search ADS Gorton G. , Schmid F. A. ( 2004 ): Capital, labor, and the firm: a study of German codetermination , Journal of the European Economic Association 2 , 863 – 905 . Google Scholar CrossRef Search ADS Graham J. R. , Kim H. , Li S. , Qiu J. ( 2016 ): Employee costs of corporate bankruptcy. Working paper, Duke University. Guiso L. , Pistaferri L. , Schivardi F. ( 2005 ): Insurance within the firm , Journal of Political Economy 113 , 1054 – 1087 . Google Scholar CrossRef Search ADS Harris M. , Holmstrom B. ( 1982 ): A theory of wage dynamics , Review of Economic Studies 49 , 315 – 333 . Google Scholar CrossRef Search ADS Hashimoto M. , Raisian J. ( 1985 ): Employment tenure and earnings profiles in Japan and the United States , The American Economic Review 75 , 721 – 735 . Hassel A. ( 1999 ): The erosion of the German system of industrial relations , British Journal of Industrial Relations 37 , 483 – 505 . Google Scholar CrossRef Search ADS Hethey-Maier T. , Seth S. ( 2010 ): The establishment history panel (BHP) 1975–2008: handbook version 1.0.2. FDZ Datenreport, Documentation on Labor Market Data 04/2010. Holmstrom B. ( 1983 ): Equilibrium long-term labor contracts , Quarterly Journal of Economics 98, 23 – 54 . Google Scholar CrossRef Search ADS Imbens G. , Kalyanaraman K. ( 2012 ): Optimal bandwidth choice for the regression discontinuity estimator , The Review of Economic Studies 79 , 933 – 959 . Google Scholar CrossRef Search ADS Jagannathan R. , Wang Z. ( 1996 ): The conditional CAPM and the cross-section of expected returns , Journal of Finance 51 , 3 – 53 . Google Scholar CrossRef Search ADS Jensen M. C. , Meckling W. H. ( 1979 ): Rights and production functions: an application to labor-managed firms and codetermination , Journal of Business 52 , 469 – 506 . Google Scholar CrossRef Search ADS Kim E. H. , Ouimet P. ( 2014 ): Broad-based employee stock ownership: motives and outcomes , Journal of Finance 69 , 1273 – 1319 . Google Scholar CrossRef Search ADS Kohn K. , Lembcke A. C. ( 2007 ): Wage distributions by bargaining regime , AStA Wirtschafts– Und Sozialstatistisches Archiv 1 , 247 – 261 . Google Scholar CrossRef Search ADS Lustig H. , Van Nieuwerburgh S. ( 2008 ): The returns on human capital: good news on wall street is bad news on main street , Review of Financial Studies 21 , 2097 – 2137 . Google Scholar CrossRef Search ADS Malcomson J. M. ( 1983 ): Trade unions and economic efficiency , Ecoonomic Journal 93 , 51 – 65 . Mandelker G. N. , Rhee S. G. ( 1984 ): The impact of the degrees of operating and financial leverage on systematic risk of common stock , Journal of Financial and Quantitative Analysis 19 , 45 – 57 . Google Scholar CrossRef Search ADS McCrary J. ( 2008 ): Manipulation of the running variable in the regression discontinuity design: a density test , Journal of Econometrics 142 , 698 – 714 . Google Scholar CrossRef Search ADS Molloy R. , Smith C. L. , Wozniak A. ( 2017 ): Job changing and the decline in long-distance migration in the United States , Demography 54 , 631 – 653 . Google Scholar CrossRef Search ADS PubMed Novy-Marx R. ( 2011 ): Operating leverage , Review of Finance 15 , 103 – 134 . Google Scholar CrossRef Search ADS Oi W. Y. , Idson T. L. ( 1999 ) Chapter 33: Firm size and wages, in: Ashenfelter O. C. , Card D. (eds.), Handbook of Labor Economics , Elsevier , Amsterdam , pp. 2165 – 2214 . Pagano M. , Volpin P. F. ( 2005 ): Managers, workers, and corporate control , Journal of Finance 60 , 841 – 868 . Google Scholar CrossRef Search ADS Palacios-Huerta I. ( 2003 ): An empirical analysis of the risk properties of human capital returns , The American Economic Review 93 , 948 – 964 . Google Scholar CrossRef Search ADS Petry S. ( 2018 ): Mandatory worker representation on the board and its effect on shareholder wealth , Financial Management 47 , 25 – 54 . Google Scholar CrossRef Search ADS Renaud S. ( 2007 ): Dynamic efficiency of supervisory board codetermination in Germany , Labour 21 , 689 – 712 . Google Scholar CrossRef Search ADS Rosen S. ( 1985 ): Implicit contracts: a survey , Journal of Economic Literature 23 , 1144 – 1175 . Rudanko L. ( 2011 ): Aggregate and idiosyncratic risk in a frictional labor market , American Economic Review 101 , 2823 – 2843 . Google Scholar CrossRef Search ADS Schnabel C. , Wagner J. ( 2007 ): The persistent decline in unionization in Western and Eastern Germany, 1980–2004: what can we learn from a decomposition analysis? , Industrielle Beziehungen/the German Journal of Industrial Relations 142 , 118 – 132 . Shleifer A. , Summers L. H. ( 1988 ) Breach of trust in hostile takeovers, in: Auerbach A. J. (ed.), Corporate Takeovers: Causes and Consequences , University of Chicago Press , Chicago , pp. 33 – 56 . Sraer D. , Thesmar D. ( 2007 ): Performance and behavior of family firms: evidence from the french stock market , Journal of the European Economic Association 5 , 709 – 751 . Google Scholar CrossRef Search ADS Tate G. A. , Yang L. ( 2016 ): The human factor in acquisitions: cross-industry labor mobility and corporate diversification. US Census Bureau Center for Economic Studies Paper CES-WP-15-31. Thomas J. , Worrall T. ( 1988 ): Self-enforcing wage contracts , Review of Economic Studies 55 , 541 – 553 . Google Scholar CrossRef Search ADS Wiedemann H. ( 1980 ): Codetermination by workers in German enterprises , American Journal of Comparative Law 28 , 79 – 92 . Google Scholar CrossRef Search ADS © The Author(s) 2018. Published by Oxford University Press on behalf of the European Finance Association. All rights reserved. For Permissions, please email: journals.permissions@oup.com This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/about_us/legal/notices) http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Review of Finance Oxford University Press

Labor Representation in Governance as an Insurance Mechanism

Review of Finance , Volume Advance Article (4) – May 26, 2018

Loading next page...
 
/lp/ou_press/labor-representation-in-governance-as-an-insurance-mechanism-XQPW3PuG0Q
Publisher
Oxford University Press
Copyright
© The Author(s) 2018. Published by Oxford University Press on behalf of the European Finance Association. All rights reserved. For Permissions, please email: journals.permissions@oup.com
ISSN
1572-3097
eISSN
1573-692X
D.O.I.
10.1093/rof/rfy012
Publisher site
See Article on Publisher Site

Abstract

Abstract We hypothesize that labor participation in governance helps improve risk sharing between employees and employers. It provides an ex post mechanism to enforce implicit insurance contracts protecting employees against adverse shocks. Results based on German establishment-level data show that skilled employees of firms with 50% labor representation on boards are protected against layoffs during adverse industry shocks. They pay an insurance premium of 3.3% in the form of lower wages. Unskilled blue-collar workers are unprotected against shocks. Our evidence suggests that workers capture all the gains from improved risk sharing, whereas shareholders are no better or worse off than without codetermination. 1. Introduction The analysis of risk, hedging, and risk-sharing has been at the core of finance research since the beginnings of modern finance, leading to a remarkable level of sophistication in the analyses of asset allocation, hedging of risky positions, and the pricing of risky securities. In contrast, the risk associated with human capital—a significant portion of wealth and arguably the largest source of risk, and most difficult to diversify for average workers—has received limited attention.1 More recently, however, a new stream of research on finance and labor has demonstrated a critical role of human capital and employment risk in understanding risk-return trade-offs, indirect costs of financial distress, optimal capital structure, and restructuring.2 In this paper, we focus on sharing human capital risk between employers and employees. Our point of departure is implicit contract theory, which holds that the risk-neutral principals of the firm provide human capital protection to risk-averse employees against adverse shocks. Employees accept lower wages in return for insuring their employment and wages (Baily, 1974; Gordon, 1974; Azariadis, 1975; Rudanko, 2011). For employees to enter such implicit insurance contracts requires a means to monitor and enforce contracts when they need protection. Enforceability is essential for implicit insurance contracts to work. If workers accept lower wages today, how can shareholders commit not to lay them off in the future when it is in their best interest to do so? The literature has appealed to arguments based on reputation and investigated the worker-friendly preferences of family owners (e.g., Sraer and Thesmar, 2007; Ellul, Pagano, and Schivardi, 2017).3 To our knowledge, ours is the first study of a contractual mechanism to enforce implicit employment contracts.4 Specifically, we hypothesize that labor representation on corporate boards is such a mechanism to ensure employment guarantees will be honored. In the 1970s, Germany led a trend toward more labor representation, and most countries now grant workers some degree of representation, at least in the form of work councils at the firm or establishment level. The world map in Figure 1 shows countries with labor representation at the firm level and/or at board level as of 2015.5 By 2015, only two countries among the OECD countries—the USA and Singapore—did not grant workers any representation. Figure 1. View largeDownload slide World map of labor representation. Notes: This figure plots labor representation in works councils and corporate boards around the world. We categorize countries in three groups (1) no labor representation (light gray), (2) only work councils (gray), and (3) work councils and board representation (dark gray). This classification is based on variables 30 and 31 in the CBR-LRI data set provided by the Centre for Business Research at the University of Cambridge, UK. [See Adams et al. (2017) and https://www.repository.cam.ac.uk/handle/1810/262929 for a documentation.] For each country, we use the latest information available (for most countries that is 2015). The figure does not display more granular information about the power of work councils and the degree of board representation available in the data set. Figure 1. View largeDownload slide World map of labor representation. Notes: This figure plots labor representation in works councils and corporate boards around the world. We categorize countries in three groups (1) no labor representation (light gray), (2) only work councils (gray), and (3) work councils and board representation (dark gray). This classification is based on variables 30 and 31 in the CBR-LRI data set provided by the Centre for Business Research at the University of Cambridge, UK. [See Adams et al. (2017) and https://www.repository.cam.ac.uk/handle/1810/262929 for a documentation.] For each country, we use the latest information available (for most countries that is 2015). The figure does not display more granular information about the power of work councils and the degree of board representation available in the data set. To test our hypothesis on labor representation on corporate boards, we examine the German system, which requires 50% employee representation on supervisory boards—hereafter referred to as parity-codetermination—when firms have more than 2,000 employees working in Germany. Germany offers a laboratory in which otherwise similar companies have different degrees of labor representation. In addition, the Institute of Employment Research (IAB) in Germany provides detailed, high quality panel data on employment and wages for all establishments located in Germany over our sample period 1990–2008. Establishment level data allow cleaner identification devoid of false detection of insurance that can occur with firm level data. For example, if a firm fires workers in a plant—no employment protection—and adds the same number of different workers at another plant in a different industry, aggregate employment at the firm will be unchanged, making it appear as if the firm provided full employment protection. Establishment data also allow us to construct adverse industry shocks using the establishments of other, non-sample firms in the same industry.6 Our initial identification strategy is a difference-in-differences approach. We are mindful of the fact that parity co-determination is a discrete function of the number of employees in Germany, which is loosely related to firm size. We employ three different strategies to separate the parity-codetermination effect from the effect of employment level in Germany: (1) the number of employees in Germany interacted with the shock indicator; (2) placebo tests using false thresholds different from the parity threshold of 2,000; and (3) regression discontinuity (RD) analyses around the 2,000-employee threshold for parity codetermination. All results support the same conclusions. We also check for the potential endogeneity of parity status and find no evidence that either workers or shareholders successfully manipulate the number of employees in Germany to maintain or avoid the parity status. It appears that the parity status is largely a result of employment decisions driven by business and operational considerations. We find white-collar and skilled blue-collar workers of parity-codetermined firms are protected against layoffs during adverse industry shocks, while those working for non-parity firms are not. Surprisingly, unskilled blue-collar workers of the same parity firms are unprotected from layoffs during industry shocks. We attribute this observation to the absence of unskilled blue-collar worker representation on supervisory boards. We could not identify a single unskilled blue-collar worker among labor representatives, which underscores the importance of worker participation in governance for an effective implementation of implicit employment insurance contracts. From the evidence on employment protection alone, we cannot infer that firms honor implicit insurance contracts against adverse industry shocks. Parity codetermination may result in employee entrenchment sufficient to prevent layoffs (e.g., Pagano and Volpin, 2005; Atanassov and Kim, 2009); employees may not need to pay an insurance premium in the form of lower wages for the increase in job security. There are no theoretical reasons to believe that codetermination makes employees just strong enough to enforce an implicit insurance contract, without causing employee entrenchment. Therefore, whether codetermination leads to an insurance relationship or employee entrenchment is an empirical question. We find that workers with vocational and higher educational qualifications, two categories that cover most skilled blue-collar and white-collar workers, receive significantly lower wages at parity-codetermined firms. Their average wage concession is about 3.3%, supporting the notion of a quid pro quo in exchange for employment insurance. Are the wage differentials between parity-codetermined firms and non-parity firms sufficient to compensate shareholders for risks they bear in providing insurance? Providing insurance is costly, because it increases operating leverage. We find parity firms’ operating profits and valuation suffer more during shock periods compared with non-parity firms. While we observe that employees make concessions, this concession may still be too low to compensate shareholders for the additional risk. Short of a structural model of the costs of reduced flexibility to firms, we cannot directly infer whether the wage concessions are adequate.7 We investigate this issue indirectly by relating the parity status to firm performance as measured through the cycle over the non-shock and shock periods. We find no significant relation between parity codetermination and operating performance or shareholder value, which is in line with the overall conclusions reached in the literature on the effect of German codetermination on firm performance.8 On average, parity firms perform no worse or better than non-parity firms. Hence, whatever efficiency gains arise from improved risk sharing, workers seem to capture all these gains, making the minimum wage concessions that allow shareholders to be no worse off, but also no better off than without codetermination. The lack of shareholder gains may explain why firms do not voluntarily adopt codetermination, an observation often used to support arguments against worker participation in governance (Jensen and Meckling, 1979). The fact that shareholders do not lose from parity codetermination may also be the reason why we do not find evidence for shareholders manipulating the number of employees in Germany so as to evade parity status. Parity codetermination seems to be, if anything, a second-order consideration from the point of view of shareholder value, so that other, first-order business and operating considerations drive firms’ employment decisions in Germany. Our paper makes multiple contributions. We contribute to the literature on implicit insurance contracts by focusing on how such contracts are enforced.9 This literature has relied mostly on reputation as an enforcement mechanism, which is fragile when companies undergo ownership changes (Shleifer and Summers, 1988). To our knowledge, we are first to think of codetermination as an enforcement mechanism in an implicit insurance relationship. The closest theoretical contribution in this context is Malcomson (1983), who argues that unions may act as an enforcement mechanism for implicit wage contracts. An earlier literature on life-long employment in Japanese firms views the security of long-term employment relationships as mechanisms to facilitate greater investments in firm-specific human capital (Hashimoto and Raisian, 1985; Abraham and Houseman, 1989). However, these papers do not specify how firms commit to protecting long-term employment relationships. Faleye, Mehrotra, and Morck (2006) and Kim and Ouimet (2014) analyze the implications of employee stock ownership in the USA, but do not explore the possibility that employee equity ownership might help enforce implicit contracts. A strand of the family firm literature argues that it is easier for family firms to commit to implicit contracts because their managers have a longer time horizon.10Sraer and Thesmar (2007) show that family firms in France insure workers against employment shocks; however, they recognize that even with the long-term perspective provided by dynastic management, family-owners face incentives to renege on prior promises to insure workers during downturns. To explain their finding, they rely on preferences “irrationally tying top management to employees” (p. 729). Our argument is based on an explicit mechanism and self-interest and not on the preferences of decision-makers. Ellul, Pagano, and Schivardi (2017) find evidence of employment insurance in family firms for a cross-section of countries. However, insurance by family firms seems only partial; workers accept higher fluctuations in wages and are not covered against shocks originating from banking crises, when access to capital dries up. Similarly, D’Aurizio and Romano (2013) show that family firms in Italy do not insure workers employed in locations further away from company headquarters. Ours is also the first to provide a quantitative estimate for the insurance premium based on establishment-level data within the context of the quid pro quo of the implicit insurance relationship.11Guiso, Pistaferri, and Schivardi (2005) examine a matched employee-firm panel of Italian firms and show that firms protect workers against wage shocks; by contrast, our study analyzes employment shocks. In addition, Guiso, Pistaferri, and Schivardi (2005) lack a benchmark and a control group of firms that provide less or no insurance. Similarly, Beaudry and DiNardo (1991) provide evidence for the implications of implicit-contracting models for wage dynamics over the business cycle, but without looking at enforcement mechanisms or insurance premiums. In a broader context, our paper provides evidence that balances critical views on labor representation in governance. The critics argue that it is detrimental to the interests of investors, because sharing control rights with labor through representation on the board can have a harmful impact on firm performance. We find German codetermination improves risk sharing without hurting firm performance, suggesting a gain to employee welfare without damaging shareholders. 2. Theoretical Considerations and Hypothesis Development In the simplest version of the insurance argument, diversified, risk-neutral investors (firms and entrepreneurs) insure risk-averse workers against firm-level shocks by promising them a wage that does not vary with their productivity from period to period. In most models, insurance affects wages as well as the employment status of workers (e.g., Baily, 1974; Azariadis, 1975; Holmstrom, 1983; Gamber, 1988). Workers give up a portion of their wages in return for protection against adverse shocks to wages and employment and receive wages that are sometimes above and sometimes below their marginal product. The insurance provided to workers shifts employment risk from workers to investors, but an effective risk transfer requires a commitment device. Workers first give up a portion of their wages. Later they have to count on firms’ honoring contracts in the event of adverse shocks. The theoretical literature on implicit contracts simply assumes that firms have the ability to commit to long-term contracts. However, commitment is crucial. Workers often have to locate close to the firm or make investments in firm-specific human capital well before the firm has to honor its side of the bargain, which makes them vulnerable to breaches of implicit contracts (Shleifer and Summers, 1988). We argue that parity-codetermination serves as a commitment device by allowing workers to influence employment decisions. Hypothesis 1: Parity-codetermination provides an ex-post enforcement mechanism, which ensures workers receive protection against adverse shocks to employment and wages. This hypothesis explicitly incorporates employment guarantees, which imply that firms do not fire workers even when layoffs are ex post efficient. If workers and firms could engage in frictionless bargaining, they would always agree to sever the employment relationship ex post by negotiating suitable transfers, which makes ex post inefficient employment of workers unsustainable. Models with employment insurance implicitly rule out such frictionless bargaining between firms and workers.12 Specifically, our hypothesis relies labor market frictions such as mobility costs (Baily, 1974; Beaudry and DiNardo, 1991) or search frictions (Rudanko, 2011), so that firms do not have to pay the competitive market wage in every period. It also requires the assumption that firms have better access to capital markets than workers and therefore enjoy a privileged position to insure workers; otherwise, workers could insure themselves directly through capital markets (e.g., Berk and Walden, 2013).13 When firms act as insurers to workers, they enter a quid-pro-quo relationship, whereby workers receive job guarantees in return for an insurance premium. The standard form of compensating firms in the implicit contracting literature is through lower wages. We hypothesize that only parity-codetermined firms provide insurance to workers, so only parity-codetermined firms receive insurance payments from workers in the form of lower wages: Hypothesis 2: Firms with parity-codetermination pay on average lower wages than non-parity firms. Lower wages are not necessarily the only way to compensate shareholders, but in all likelihood, lower wages are the most efficient way to provide an insurance premium because they do not distort the allocation of resources. 3. Institutional Background, Data, and Empirical Design 3.1 The German Governance System and Wage-Setting Process Germany has a two-tier board system. The management board (Vorstand) manages day-to-day operations and the supervisory board (Aufsichtsrat) monitors the management board; approves key strategic decisions; and appoints and dismisses management board members, including the CEO, and decides their compensation. The two boards are strictly separated and no member of one board can be a member of the other for the same company at the same time. The structure, size, and composition of the supervisory board are regulated by the German stock corporation act (Aktiengesetz) and the codetermination act (Mitbestimmungsgesetz) as well as other laws. The regulation requires that half of the supervisory board members are worker representatives for corporations with more than 2,000 employees working in Germany. For corporations with more than 500 up to 2,000 employees in Germany, one-third of the members of the supervisory board have to represent workers. Worker representatives are elected by the company’s workers. Depending on the size of the supervisory board, two or three seats of the worker representatives are reserved for union representatives. One seat is always reserved for a representative from middle management (leitende Angestellte).14 The chairman of the board casts the deciding vote in case of a tie. Wages in German firms are often set through collective bargaining agreements between trade unions and employers’ associations. Some firms are covered by sectoral agreements, whereas others negotiate firm-level agreements. Agreements are binding on all workers as German law forbids discriminatory wage policies that disadvantage non-union members. However, firms still have discretion over wages through firm-level wage agreements and the way in which they categorize workers, resulting in significant variation in wages across firms (Kohn and Lembcke, 2007). Unionization in Germany is generally not high and is also falling. Based on survey data, Schnabel and Wagner (2007) estimate it to be about one-third in 1992, declining to around 20% in 2004. Coverage by unionized wage agreements also has declined (Hassel, 1999). See Institutional Details in the Online Appendix for more information on German codetermination and collective bargaining. 3.2 Data Employment and wage data at the establishment level are obtained from the IAB. The IAB is the research organization of the German employment agency, the Bundesagentur für Arbeit (BA). The BA collects worker and employer contributions to unemployment insurance and distributes unemployment benefits. All German businesses are required to report detailed information on employment and wages to the BA. The individual-level data are aggregated at the establishment level, made anonymous, and offered for scientific use by the IAB. An establishment is any facility having a separate physical address, such as a factory, service station, restaurant, or office building. The IAB owns detailed establishment-level data on industry, location, employment, employee education, age, nationality, and wages. The data are in the form of establishment-level statistics, such as sums, medians, quartiles, and averages on wages and employment according to different classifications and breakdowns. We use the first three-digits of the NACE code, which identifies 224 separate economic sub-sectors (groups).15 The NACE (Rev. 1.1) classification is available from the IAB database only for 2003 and afterward. We assign an establishment’s NACE (Rev. 1.1) classification in 2003 to all its prior sample years. Some establishments may have changed their industry classification prior to 2003, in which case they would receive new establishment IDs. To avoid assigning incorrect industry codes, we drop all establishments changing industry classifications over time in the entire IAB database, as well as establishment-year observations with missing information on industry classification. These screens yield approximately 33.4 million establishment-year observations on approximately 3.5 million establishments for the sample period 1990–2008. Our sample firms are drawn from all companies included in the two main German stock market indices, DAX and MDAX, at any point over the 19-year period from 1990 to 2008. There are 184 such firms, for which we hand-collect data on the composition of the supervisory board from annual reports and Hoppenstedt company profiles. We do not include non-listed firms, because information on worker representation on their boards is usually unavailable. Stock market data come from Datastream, balance sheet and accounting data from Worldscope. At our request, the IAB matched our sample of listed firms with their establishment-level database using an automatic procedure, based on company name and address information (city, zip code, street, and house number). All cases not unambiguously matched by the automatic matching procedure are checked by hand to avoid mismatching. This procedure results in 284,538 establishment-years matched to 2,168 firm-years for 142 of the 184 firms. Of these, 1,461 firm-years (67.4%) are subject to parity codetermination, 442 (20.4%) subject to one-third codetermination, and the remaining 265 firm-years are not covered by any codetermination arrangement. The matching was performed for 2004, 2005, and 2006. Firms are dropped if they do not exist during the period 2004–2006, because we cannot match them to the IAB data. IAB does not have a firm identifier, which is why manual matching is necessary. All establishments are matched only once to our sample firms and, if establishments were sold prior to 2004, they do not enter our sample because IAB cannot match them. This matching procedure does not allow us to identify changes in establishment ownership after 2006. (At the time of matching establishments to firms, establishment data were not available for 2007 or 2008.) Thus, if an establishment belonging to a parity (non-parity) firm is sold to a non-parity (parity) firm in 2007 or 2008, it will be treated as if it still belonged to a parity (non-parity) firm after the sale. This will blur the distinction between parity and non-parity status of the establishment and potentially lead to attenuation bias working against finding significant results. The IAB distinguishes employees by occupational status. The three most important groups are unskilled blue-collar workers, skilled blue-collar workers, and white-collar employees.16 The IAB also reports three different qualification levels at each establishment by educational and vocational qualifications: (1) low-qualified employees possess neither an upper secondary school graduation certificate as their highest school qualification nor a vocational qualification. (2) Qualified employees have either an upper secondary school graduation certificate as their highest school qualification or a vocational qualification. (3) Highly qualified employees have a degree from a specialized college of higher education or a university degree.17 Unfortunately, over our sample period an increasing number of firms stopped reporting information on qualifications, leading to a steady increase in the number of employees with unknown qualifications. If firms’ decision not to report their employees’ qualification is random, the increasing trend in the number of employees with unknown qualifications should not bias our results. We include only establishments with more than 50 employees. These establishments almost always have works councils, which have significant information rights as well as the right to demand compensation for dismissals and may exert some influence at the establishment level (see Wiedemann, 1980). Nonetheless, our results are robust to an alternative sample construction with a cutoff point at ten employees. While fiscal years of German firms are mostly from January to December, establishment years for IAB data are from July to June. We therefore lag all variables from Worldscope by 6 months relative to IAB years. Effectively, we assign year-end values from Worldscope to June 30 information on employment and wages of the next year. Table I provides summary statistics. Monetary units are normalized to 2005 euros. Panels A and B provide statistics at the establishment and firm level, respectively. All accounting and market variables are taken from Worldscope and Datastream. None of the firm-level variables in Panel B are from IAB. Variable definitions and data sources are listed in Table II. Our sample selection criteria tilt the sample toward large, mature firms (mean firm age: 85 years) with high book-value leverage and low betas. About two-thirds of firm-year observations are for parity-codetermined firms. The distributions of sales, market capitalization, and the number of establishments are highly skewed; the mean exceeds the 75th percentile for all three size measures. Table I. Descriptive statistics This table presents descriptive statistics for all key variables used in this paper. Panel A reports summary statistics at the establishment level. N is the number of establishment-years the respective variable is available. Only establishments with more than 50 employees are used. DailyWageP50LQ is the median daily gross wage for low-qualified employees. DailyWageP50Q is the median daily gross wage for qualified employees. DailyWageP50HQ is the median daily gross wage for highly qualified employees. Monetary units are normalized to 2005 euros. Panel B reports summary statistics at the firm level. N is the number of firm-years the respective variable is available. Panel A Variable Mean Median Std Min P25 P75 Max N Number of employees 517.47 148 2,099.29 51 81 346 61,380 54,042 Number of unskilled 97.14 5 700.35 0 0 31 32,733 54,042 Number of skilled 103.32 10 584.98 0 0 49 19,658 54,042 Number of white collar 223.80 64 894.00 0 31 148 29,084 54,042 DailyWageP25 81.73 76.66 27.98 1.02 61.20 97.99 214.42 53,956 DailyWageP50 94.23 88.38 32.59 7.66 69.56 113.53 228.92 53,956 DailyWageP75 108.76 104.68 34.87 7.66 81.01 132.69 228.92 53,956 DailyWageP50LQ 82.50 77.52 29.07 1.87 61.99 99.04 781.59 44,783 DailyWageP50Q 93.11 88.53 30.25 7.66 70.37 110.98 199.33 53,811 DailyWageP50HQ 124.56 126.03 34.84 0.60 99.96 150.47 335.52 40,459 EstAge 15.64 16 9.880 0 6 24 33 54,042 MedianEmplAge 38.84 39 4.973 17 36 42 60 54,042 RatioWhiteCollar 0.477 0.446 0.297 0 0.228 0.746 1 54,042 Shock 0.058 0 0.233 0 0 0 1 52,756 HHI 0.405 0.304 0.314 0.007 0.145 0.563 1 54,042 Panel B Variable Mean Median Std Min P25 P75 Max N Beta 0.678 0.620 0.467 −3.198 0.324 0.997 3.002 1,832 Firm age 84.5 86 53.3 0 36 124 259 1,989 Leverage 0.392 0.358 0.273 0 0.169 0.582 0.996 2,052 MCap (bn €) 5.382 0.976 12.327 0.003 0.297 4.141 198.186 2,002 NetPPE (bn €) 2.451 0.288 7.337 0 0.092 1.349 74.003 2,057 Parity 0.674 1 0.469 0 0 1 1 2,168 ROA 0.074 0.069 0.096 −1.152 0.031 0.110 0.671 1,926 Sales (bn €) 8.663 1.793 17.950 0.005 0.591 7.694 157.5 2,064 Tobin’s Q 1.546 1.224 1.010 0.454 1.054 1.602 12.529 1,991 Firm shock 0.048 0 0.177 0 0 0 1 2,126 Number of establishments 26.045 6 56.827 1 3 20 497 2,075 Panel A Variable Mean Median Std Min P25 P75 Max N Number of employees 517.47 148 2,099.29 51 81 346 61,380 54,042 Number of unskilled 97.14 5 700.35 0 0 31 32,733 54,042 Number of skilled 103.32 10 584.98 0 0 49 19,658 54,042 Number of white collar 223.80 64 894.00 0 31 148 29,084 54,042 DailyWageP25 81.73 76.66 27.98 1.02 61.20 97.99 214.42 53,956 DailyWageP50 94.23 88.38 32.59 7.66 69.56 113.53 228.92 53,956 DailyWageP75 108.76 104.68 34.87 7.66 81.01 132.69 228.92 53,956 DailyWageP50LQ 82.50 77.52 29.07 1.87 61.99 99.04 781.59 44,783 DailyWageP50Q 93.11 88.53 30.25 7.66 70.37 110.98 199.33 53,811 DailyWageP50HQ 124.56 126.03 34.84 0.60 99.96 150.47 335.52 40,459 EstAge 15.64 16 9.880 0 6 24 33 54,042 MedianEmplAge 38.84 39 4.973 17 36 42 60 54,042 RatioWhiteCollar 0.477 0.446 0.297 0 0.228 0.746 1 54,042 Shock 0.058 0 0.233 0 0 0 1 52,756 HHI 0.405 0.304 0.314 0.007 0.145 0.563 1 54,042 Panel B Variable Mean Median Std Min P25 P75 Max N Beta 0.678 0.620 0.467 −3.198 0.324 0.997 3.002 1,832 Firm age 84.5 86 53.3 0 36 124 259 1,989 Leverage 0.392 0.358 0.273 0 0.169 0.582 0.996 2,052 MCap (bn €) 5.382 0.976 12.327 0.003 0.297 4.141 198.186 2,002 NetPPE (bn €) 2.451 0.288 7.337 0 0.092 1.349 74.003 2,057 Parity 0.674 1 0.469 0 0 1 1 2,168 ROA 0.074 0.069 0.096 −1.152 0.031 0.110 0.671 1,926 Sales (bn €) 8.663 1.793 17.950 0.005 0.591 7.694 157.5 2,064 Tobin’s Q 1.546 1.224 1.010 0.454 1.054 1.602 12.529 1,991 Firm shock 0.048 0 0.177 0 0 0 1 2,126 Number of establishments 26.045 6 56.827 1 3 20 497 2,075 Table I. Descriptive statistics This table presents descriptive statistics for all key variables used in this paper. Panel A reports summary statistics at the establishment level. N is the number of establishment-years the respective variable is available. Only establishments with more than 50 employees are used. DailyWageP50LQ is the median daily gross wage for low-qualified employees. DailyWageP50Q is the median daily gross wage for qualified employees. DailyWageP50HQ is the median daily gross wage for highly qualified employees. Monetary units are normalized to 2005 euros. Panel B reports summary statistics at the firm level. N is the number of firm-years the respective variable is available. Panel A Variable Mean Median Std Min P25 P75 Max N Number of employees 517.47 148 2,099.29 51 81 346 61,380 54,042 Number of unskilled 97.14 5 700.35 0 0 31 32,733 54,042 Number of skilled 103.32 10 584.98 0 0 49 19,658 54,042 Number of white collar 223.80 64 894.00 0 31 148 29,084 54,042 DailyWageP25 81.73 76.66 27.98 1.02 61.20 97.99 214.42 53,956 DailyWageP50 94.23 88.38 32.59 7.66 69.56 113.53 228.92 53,956 DailyWageP75 108.76 104.68 34.87 7.66 81.01 132.69 228.92 53,956 DailyWageP50LQ 82.50 77.52 29.07 1.87 61.99 99.04 781.59 44,783 DailyWageP50Q 93.11 88.53 30.25 7.66 70.37 110.98 199.33 53,811 DailyWageP50HQ 124.56 126.03 34.84 0.60 99.96 150.47 335.52 40,459 EstAge 15.64 16 9.880 0 6 24 33 54,042 MedianEmplAge 38.84 39 4.973 17 36 42 60 54,042 RatioWhiteCollar 0.477 0.446 0.297 0 0.228 0.746 1 54,042 Shock 0.058 0 0.233 0 0 0 1 52,756 HHI 0.405 0.304 0.314 0.007 0.145 0.563 1 54,042 Panel B Variable Mean Median Std Min P25 P75 Max N Beta 0.678 0.620 0.467 −3.198 0.324 0.997 3.002 1,832 Firm age 84.5 86 53.3 0 36 124 259 1,989 Leverage 0.392 0.358 0.273 0 0.169 0.582 0.996 2,052 MCap (bn €) 5.382 0.976 12.327 0.003 0.297 4.141 198.186 2,002 NetPPE (bn €) 2.451 0.288 7.337 0 0.092 1.349 74.003 2,057 Parity 0.674 1 0.469 0 0 1 1 2,168 ROA 0.074 0.069 0.096 −1.152 0.031 0.110 0.671 1,926 Sales (bn €) 8.663 1.793 17.950 0.005 0.591 7.694 157.5 2,064 Tobin’s Q 1.546 1.224 1.010 0.454 1.054 1.602 12.529 1,991 Firm shock 0.048 0 0.177 0 0 0 1 2,126 Number of establishments 26.045 6 56.827 1 3 20 497 2,075 Panel A Variable Mean Median Std Min P25 P75 Max N Number of employees 517.47 148 2,099.29 51 81 346 61,380 54,042 Number of unskilled 97.14 5 700.35 0 0 31 32,733 54,042 Number of skilled 103.32 10 584.98 0 0 49 19,658 54,042 Number of white collar 223.80 64 894.00 0 31 148 29,084 54,042 DailyWageP25 81.73 76.66 27.98 1.02 61.20 97.99 214.42 53,956 DailyWageP50 94.23 88.38 32.59 7.66 69.56 113.53 228.92 53,956 DailyWageP75 108.76 104.68 34.87 7.66 81.01 132.69 228.92 53,956 DailyWageP50LQ 82.50 77.52 29.07 1.87 61.99 99.04 781.59 44,783 DailyWageP50Q 93.11 88.53 30.25 7.66 70.37 110.98 199.33 53,811 DailyWageP50HQ 124.56 126.03 34.84 0.60 99.96 150.47 335.52 40,459 EstAge 15.64 16 9.880 0 6 24 33 54,042 MedianEmplAge 38.84 39 4.973 17 36 42 60 54,042 RatioWhiteCollar 0.477 0.446 0.297 0 0.228 0.746 1 54,042 Shock 0.058 0 0.233 0 0 0 1 52,756 HHI 0.405 0.304 0.314 0.007 0.145 0.563 1 54,042 Panel B Variable Mean Median Std Min P25 P75 Max N Beta 0.678 0.620 0.467 −3.198 0.324 0.997 3.002 1,832 Firm age 84.5 86 53.3 0 36 124 259 1,989 Leverage 0.392 0.358 0.273 0 0.169 0.582 0.996 2,052 MCap (bn €) 5.382 0.976 12.327 0.003 0.297 4.141 198.186 2,002 NetPPE (bn €) 2.451 0.288 7.337 0 0.092 1.349 74.003 2,057 Parity 0.674 1 0.469 0 0 1 1 2,168 ROA 0.074 0.069 0.096 −1.152 0.031 0.110 0.671 1,926 Sales (bn €) 8.663 1.793 17.950 0.005 0.591 7.694 157.5 2,064 Tobin’s Q 1.546 1.224 1.010 0.454 1.054 1.602 12.529 1,991 Firm shock 0.048 0 0.177 0 0 0 1 2,126 Number of establishments 26.045 6 56.827 1 3 20 497 2,075 Table II. Variable definitions This table defines all variables used in this paper. Board data are taken from Hoppenstedt company profiles and annual reports. Employment and wage data are from the IAB Establishment History Panel. Accounting data are taken from Worldscope and market data from Datastream. The numbers in brackets refer to Worldscope items, taken from the Worldscope Data Definition Guide. Variable Description Source Number of employees Total number of employees in the establishment IAB Number of skilled Number of skilled (blue-collar) employees (at least vocational training) IAB Number of unskilled Number of unskilled (blue-collar) employees (no formal qualification) IAB Number of white collar Number of white-collar employees (at least vocational training) IAB Beta CAPM beta estimated over the prior calendar year using daily returns Datastream EstAge Age of the establishment in year IAB Firm employees Sum of all employees across all establishments of the firm in Germany IAB Firm age Age of the firm in year Worldscope HHI Herfindahl–Hirschman index calculated using number of employees of all establishments with more than 50 employees in the same county and industry (three-digit NACE code). IAB Leverage =Total debt [03255]/(total debt+common equity [03501]) Worldscope MCap Market capitalization [08001] Worldscope MedianEmplAge Median age of all employees in the establishment IAB NetPPE Net property, plant, and equipment [02501] Worldscope Parity =1 if 50% of all members of the company’s supervisory board are classified as employee representatives Hoppenstedt, annual reports PPE dummy =1 if NetPPE [02501] declines by more than 15% Worldscope RatioWhiteCollar =#WhiteCollar/#Employees IAB ROA =EBITDAt [18198]/((total assetst [02999]+total assetst–1)/2) Worldscope Sales =Net sales or revenues [01001] in 2005 euros Worldscope Shock =1 if employment in non-sample establishments in the same industry (three-digit NACE-code) as the establishment decreases by more than 5% with no increase in employment in the following year. A detailed description of the definition is provided in Section 3.3.2. IAB DailyWageP25 1st quartile of gross average daily wage for all full-time employees in 2005 euros IAB DailyWageP50 Median of gross average daily wage for all full-time employees in 2005 euros IAB DailyWageP75 Third quartile of gross average daily wage for all full-time employees in 2005 euros IAB Tobin’s Q = (market capitalization [08001]+total assets [02999] – common equity [03501])/total assets Worldscope Variable Description Source Number of employees Total number of employees in the establishment IAB Number of skilled Number of skilled (blue-collar) employees (at least vocational training) IAB Number of unskilled Number of unskilled (blue-collar) employees (no formal qualification) IAB Number of white collar Number of white-collar employees (at least vocational training) IAB Beta CAPM beta estimated over the prior calendar year using daily returns Datastream EstAge Age of the establishment in year IAB Firm employees Sum of all employees across all establishments of the firm in Germany IAB Firm age Age of the firm in year Worldscope HHI Herfindahl–Hirschman index calculated using number of employees of all establishments with more than 50 employees in the same county and industry (three-digit NACE code). IAB Leverage =Total debt [03255]/(total debt+common equity [03501]) Worldscope MCap Market capitalization [08001] Worldscope MedianEmplAge Median age of all employees in the establishment IAB NetPPE Net property, plant, and equipment [02501] Worldscope Parity =1 if 50% of all members of the company’s supervisory board are classified as employee representatives Hoppenstedt, annual reports PPE dummy =1 if NetPPE [02501] declines by more than 15% Worldscope RatioWhiteCollar =#WhiteCollar/#Employees IAB ROA =EBITDAt [18198]/((total assetst [02999]+total assetst–1)/2) Worldscope Sales =Net sales or revenues [01001] in 2005 euros Worldscope Shock =1 if employment in non-sample establishments in the same industry (three-digit NACE-code) as the establishment decreases by more than 5% with no increase in employment in the following year. A detailed description of the definition is provided in Section 3.3.2. IAB DailyWageP25 1st quartile of gross average daily wage for all full-time employees in 2005 euros IAB DailyWageP50 Median of gross average daily wage for all full-time employees in 2005 euros IAB DailyWageP75 Third quartile of gross average daily wage for all full-time employees in 2005 euros IAB Tobin’s Q = (market capitalization [08001]+total assets [02999] – common equity [03501])/total assets Worldscope Table II. Variable definitions This table defines all variables used in this paper. Board data are taken from Hoppenstedt company profiles and annual reports. Employment and wage data are from the IAB Establishment History Panel. Accounting data are taken from Worldscope and market data from Datastream. The numbers in brackets refer to Worldscope items, taken from the Worldscope Data Definition Guide. Variable Description Source Number of employees Total number of employees in the establishment IAB Number of skilled Number of skilled (blue-collar) employees (at least vocational training) IAB Number of unskilled Number of unskilled (blue-collar) employees (no formal qualification) IAB Number of white collar Number of white-collar employees (at least vocational training) IAB Beta CAPM beta estimated over the prior calendar year using daily returns Datastream EstAge Age of the establishment in year IAB Firm employees Sum of all employees across all establishments of the firm in Germany IAB Firm age Age of the firm in year Worldscope HHI Herfindahl–Hirschman index calculated using number of employees of all establishments with more than 50 employees in the same county and industry (three-digit NACE code). IAB Leverage =Total debt [03255]/(total debt+common equity [03501]) Worldscope MCap Market capitalization [08001] Worldscope MedianEmplAge Median age of all employees in the establishment IAB NetPPE Net property, plant, and equipment [02501] Worldscope Parity =1 if 50% of all members of the company’s supervisory board are classified as employee representatives Hoppenstedt, annual reports PPE dummy =1 if NetPPE [02501] declines by more than 15% Worldscope RatioWhiteCollar =#WhiteCollar/#Employees IAB ROA =EBITDAt [18198]/((total assetst [02999]+total assetst–1)/2) Worldscope Sales =Net sales or revenues [01001] in 2005 euros Worldscope Shock =1 if employment in non-sample establishments in the same industry (three-digit NACE-code) as the establishment decreases by more than 5% with no increase in employment in the following year. A detailed description of the definition is provided in Section 3.3.2. IAB DailyWageP25 1st quartile of gross average daily wage for all full-time employees in 2005 euros IAB DailyWageP50 Median of gross average daily wage for all full-time employees in 2005 euros IAB DailyWageP75 Third quartile of gross average daily wage for all full-time employees in 2005 euros IAB Tobin’s Q = (market capitalization [08001]+total assets [02999] – common equity [03501])/total assets Worldscope Variable Description Source Number of employees Total number of employees in the establishment IAB Number of skilled Number of skilled (blue-collar) employees (at least vocational training) IAB Number of unskilled Number of unskilled (blue-collar) employees (no formal qualification) IAB Number of white collar Number of white-collar employees (at least vocational training) IAB Beta CAPM beta estimated over the prior calendar year using daily returns Datastream EstAge Age of the establishment in year IAB Firm employees Sum of all employees across all establishments of the firm in Germany IAB Firm age Age of the firm in year Worldscope HHI Herfindahl–Hirschman index calculated using number of employees of all establishments with more than 50 employees in the same county and industry (three-digit NACE code). IAB Leverage =Total debt [03255]/(total debt+common equity [03501]) Worldscope MCap Market capitalization [08001] Worldscope MedianEmplAge Median age of all employees in the establishment IAB NetPPE Net property, plant, and equipment [02501] Worldscope Parity =1 if 50% of all members of the company’s supervisory board are classified as employee representatives Hoppenstedt, annual reports PPE dummy =1 if NetPPE [02501] declines by more than 15% Worldscope RatioWhiteCollar =#WhiteCollar/#Employees IAB ROA =EBITDAt [18198]/((total assetst [02999]+total assetst–1)/2) Worldscope Sales =Net sales or revenues [01001] in 2005 euros Worldscope Shock =1 if employment in non-sample establishments in the same industry (three-digit NACE-code) as the establishment decreases by more than 5% with no increase in employment in the following year. A detailed description of the definition is provided in Section 3.3.2. IAB DailyWageP25 1st quartile of gross average daily wage for all full-time employees in 2005 euros IAB DailyWageP50 Median of gross average daily wage for all full-time employees in 2005 euros IAB DailyWageP75 Third quartile of gross average daily wage for all full-time employees in 2005 euros IAB Tobin’s Q = (market capitalization [08001]+total assets [02999] – common equity [03501])/total assets Worldscope 3.3 Research Design Our empirical strategy is to compare how a negative shock affects employee layoffs by establishments owned by parity-codetermined firms differently from those owned by firms with one-third or no labor representation on the supervisory board. Then we compare establishment-level wages between parity and non-parity firms through the cycle of the shock and non-shock period. To cross-check robustness of results based on a particular empirical design, we employ three different strategies: difference-in-differences estimations, placebo tests, and RD analyses. 3.3.a. Specification The difference-in-differences analysis relies on exogenous, negative industry shocks to employment and on the dummy variable Parity, which is one in any establishment-year when the firm has 50% worker representation on the supervisory board, and zero otherwise. Our definition of Parity is based on actual board data and is not inferred from employment data. The baseline specification is: yijkt=αt+αi+γXijkt+δParityjt+θShockkt+βParityjt×Shockkt+ϵijkt. (1) The dependent variable, yijkt, is the log change in employment or the log of the median daily wage, where i indexes establishments, j indexes firms, k indexes industries, and t indexes time. Parityjt is the parity dummy, Shockkt is the shock dummy defined in the next section, and ϵijkt is an error term. We control for year-fixed effects, αt, and establishment-fixed effects, αi. Establishment-fixed effects absorb firm-fixed effects, and control for unobservable time-invariant firm and establishment characteristics. The vector of time-varying control variables, Xijkt, includes the logarithm of the number of employees working in Germany, the logarithm of total sales, firm leverage, and establishment age. We cluster standard errors at the firm level. Abadie et al. (2017) argue that the requirement to cluster follows from the sampling strategy and the experimental design. We follow a two-stage sampling process, in which we identify first firms and then match establishments to firms. The argument of Abadie et al. (2017) suggests clustering at the first-stage variable for such a sampling process; in our case, the firm level. A different approach would cluster observations at the industry-year level, the level at which we define shocks. We present our results with standard errors clustered at the firm level and provide a robustness check with clustering at the industry-year level in Tables A-5 (Panel A) and A-6 (Panel A) in the Online Appendix. Standard errors and significance levels are not much affected by our choice of clustering. We shall refer to establishment-year observations in which the firm has 50% worker representation as parity observations and to all others, including those with one-third representation, as non-parity observations.18 For brevity and ease of exposition, we will mostly, and somewhat imprecisely, distinguish between “parity firms” and “non-parity firms,” to avoid the more cumbersome expression “establishment-year observations.” This focus on parity establishments helps to preserve the sample size of non-parity establishments, which is smaller than that of parity establishments. Table I, Panel B shows 67.4% of our sample establishment-years are owned by parity firms. For employment analysis, the coefficient of main interest is the β on the interaction of Parity and Shock. It measures the differential impact industry shocks have on employment of parity and non-parity establishments. Our hypothesis predicts β>0; parity establishments maintain higher levels of employment after an industry-wide shock than non-parity establishments. For wages, the regression does not contain Shock and δ is of main interest. Our hypothesis predicts δ<0; employees of parity firms pay insurance premiums in the form of lower wages. 3.3.b. Definition of shocks We identify adverse employment shocks using a drop in employment by non-sample German firms and foreign firms with establishments in Germany. The main requirement is that shocks are exogenous and sufficiently correlated with the economic environment of our sample firms so that they warrant major adjustments. We do not require that shocks cannot be anticipated. A suitable definition of shocks requires that Shock has a significant impact in Equation (1). If the estimate of θ is not significant, then either the definition of Shock is unsuitable and the assumed shock has no impact on non-parity establishments, or workers at non-parity establishments are also insured. In either case we cannot infer whether workers at parity establishments receive more insurance from estimates of β. We do not define shocks based on employment changes in other European countries because Germany follows a different business cycle and the correlations between employment at our sample firms and other European firms are too weak. For example, during 2011–2012, the German economy was booming while most other European countries were in, or on the verge of, a recession. Employment shocks are defined at the industry level. An industry is subject to a shock if non-sample firms’ establishments located in Germany belonging to the same three-digit NACE-code industry as a whole suffer a decrease of at least 5% in employment. These establishments may belong to either German or foreign firms. Our sample firms are likely to be under economic pressure to decrease their payroll when other non-sample firms in the same industry lay off 5% or more of their workers. We use the 5% threshold to ensure that shocks are strong enough to have a material effect and frequent enough to permit identification. Results based on a lower threshold of 2.5% are similar.19 Notice that these non-sample German and foreign firms are used only to define shocks, not as controls. The non-sample firms include many large non-listed, family owned, or foreign firms with establishments located in Germany.20 The mean (median) total sales and the number of employees of the largest 100 non-sample firms used to identify shocks are €10.2 bn (€7.0 bn) and 33,500 (19,700) in 2006, respectively. These numbers are quite comparable to the corresponding numbers for our sample firms in 2006, which are €11.7 bn (€2.0 bn) and 38,700 (9,200), respectively. Ultimately, if the non-sample firms were unsuitable to define industry shocks, attenuation bias would move the coefficient of Shock in Equation (1) toward zero and render it insignificant. In defining industry shocks, we exclude 5% declines in non-sample firms’ employment followed by an immediate recovery in the subsequent year, because other firms in the industry do not react to such shocks. Shocks followed by immediate recovery render the coefficient θ in Equation (1) insignificant (see Panel B of Table A-5 in the Online Appendix). One scenario in which this may happen has some non-sample firms lose market share or exit the market, and other firms in the industry subsequently increase employment to take the space left by declining or exiting firms. We exclude these very short-lived shocks, because they provide no identification. On the other hand, sufficiently severe and long-lasting shocks will require all firms, including parity-firms, to adjust their workforce to avoid bankruptcy (Gamber, 1988; Guiso, Pistaferri, and Schivardi, 2005). We therefore end a shock period after 4 years, again, because permanent shocks cannot provide identification. Thus, a shock period ends either after a resumption of growth, or after 4 consecutive years of non-positive growth, whichever occurs first. We also define a shock over a 2-year interval. Table A-1 in the Online Appendix illustrates how Shock is defined for four possible sequences of employment growth over 5 years. To get a feel for how our definition identifies employment shocks during our sample period, we estimate OLS regressions of the shock dummy as the dependent variable using the 4-year or 2-year definitions of shock. The independent variables are year dummies, and the regression does not have an intercept. Adjusted R2s are around 18%, indicating that much of the variation in shocks is industry-specific and is not driven by the business cycle. Figure A-1 in the Online Appendix plots the year dummy coefficients in the regressions and German unemployment rate by year. The two shock definitions are highly correlated with each other and move with the unemployment rate. We report estimation results based on the 4-year interval. With this definition, 5.8% of establishment-years of the sample firms are in shock periods. Baseline results using the 2-year interval are similar. 4. Empirical Results We begin with an investigation of how layoffs at establishments owned by parity firms differ from those owned by non-parity firms when the industry suffers a negative shock to employment. We then compare wages between parity and non-parity firms. 4.1 Employment Table III reports estimation results with different combinations of control variables. The results are consistent with the insurance hypothesis. The coefficients on Shock × Parity in the first two columns imply parity-codetermined firms retain about 8% more employees in comparison to non-parity firms during shock periods. The majority of non-parity firms in our sample have one-third worker representation. Hence, the coefficients reflect, to a large extent, the incremental impact of moving from one-third-codetermination to parity-codetermination, and to a lesser extent the impact of moving from no employee representation to parity codetermination. Table III. Employment changes—all employees This table presents OLS estimation results with the log change in employment as the dependent variable. Only establishments with more than 50 employees are included. Column (4) includes six additional variables not shown on the table: LogFirmEmployees × Parity, LogFirmEmployees × (1 – Parity), LogFirmEmployees2 × Parity, LogFirmEmployees2 × (1 – Parity), LogFirmEmployees³ × Parity, and LogFirmEmployees³ × (1 – Parity). T-statistics are reported in parentheses. Standard errors are clustered at the firm level. The table also reports the p-value for the t-test that Shock + Shock × Parity = 0, except in Column (3), which includes interaction terms of Shock with LogFirmEmployees and LogFirmEmployees2. Dependent variable Log change in number of employees (1) (2) (3) (4) Shock×Parity 0.082 0.081 0.064 0.083 (2.37) (2.22) (1.80) (2.46) Shock×LogFirmEmployees 0.264 (1.86) Shock×LogFirmEmployees2 –0.013 (–1.79) Shock –0.096 –0.095 –1.323 –0.097 (–3.30) (–3.05) (–2.05) (–3.44) Parity –0.044 –0.042 –0.039 8.651 (–1.39) (–1.39) (–1.27) (3.34) LogEstAge –0.594 –0.594 –0.594 –0.594 (–10.8) (–10.7) (–10.8) (–10.8) LogSales 0.005 –0.012 –0.014 0.006 (0.22) (–0.08) (–0.10) (0.30) Leverage –0.016 –0.018 –0.020 –0.017 (–0.34) (–0.44) (–0.48) (–0.36) LogFirmEmployees 0.054 0.147 0.137 (1.46) (0.81) (0.75) LogSales2 0.000 0.000 (0.10) (0.13) LogFirmEmployees2 –0.005 –0.005 (–0.47) (–0.42) Adjusted R2 0.174 0.174 0.175 0.175 Observations 48,896 48,896 48,896 48,896 t-test: Shock×Parity+Shock=0 0.502 0.508 0.507 Year F.E. Yes Yes Yes Yes Add. employment polynomials No No No Yes Establishment F.E. Yes Yes Yes Yes Dependent variable Log change in number of employees (1) (2) (3) (4) Shock×Parity 0.082 0.081 0.064 0.083 (2.37) (2.22) (1.80) (2.46) Shock×LogFirmEmployees 0.264 (1.86) Shock×LogFirmEmployees2 –0.013 (–1.79) Shock –0.096 –0.095 –1.323 –0.097 (–3.30) (–3.05) (–2.05) (–3.44) Parity –0.044 –0.042 –0.039 8.651 (–1.39) (–1.39) (–1.27) (3.34) LogEstAge –0.594 –0.594 –0.594 –0.594 (–10.8) (–10.7) (–10.8) (–10.8) LogSales 0.005 –0.012 –0.014 0.006 (0.22) (–0.08) (–0.10) (0.30) Leverage –0.016 –0.018 –0.020 –0.017 (–0.34) (–0.44) (–0.48) (–0.36) LogFirmEmployees 0.054 0.147 0.137 (1.46) (0.81) (0.75) LogSales2 0.000 0.000 (0.10) (0.13) LogFirmEmployees2 –0.005 –0.005 (–0.47) (–0.42) Adjusted R2 0.174 0.174 0.175 0.175 Observations 48,896 48,896 48,896 48,896 t-test: Shock×Parity+Shock=0 0.502 0.508 0.507 Year F.E. Yes Yes Yes Yes Add. employment polynomials No No No Yes Establishment F.E. Yes Yes Yes Yes Table III. Employment changes—all employees This table presents OLS estimation results with the log change in employment as the dependent variable. Only establishments with more than 50 employees are included. Column (4) includes six additional variables not shown on the table: LogFirmEmployees × Parity, LogFirmEmployees × (1 – Parity), LogFirmEmployees2 × Parity, LogFirmEmployees2 × (1 – Parity), LogFirmEmployees³ × Parity, and LogFirmEmployees³ × (1 – Parity). T-statistics are reported in parentheses. Standard errors are clustered at the firm level. The table also reports the p-value for the t-test that Shock + Shock × Parity = 0, except in Column (3), which includes interaction terms of Shock with LogFirmEmployees and LogFirmEmployees2. Dependent variable Log change in number of employees (1) (2) (3) (4) Shock×Parity 0.082 0.081 0.064 0.083 (2.37) (2.22) (1.80) (2.46) Shock×LogFirmEmployees 0.264 (1.86) Shock×LogFirmEmployees2 –0.013 (–1.79) Shock –0.096 –0.095 –1.323 –0.097 (–3.30) (–3.05) (–2.05) (–3.44) Parity –0.044 –0.042 –0.039 8.651 (–1.39) (–1.39) (–1.27) (3.34) LogEstAge –0.594 –0.594 –0.594 –0.594 (–10.8) (–10.7) (–10.8) (–10.8) LogSales 0.005 –0.012 –0.014 0.006 (0.22) (–0.08) (–0.10) (0.30) Leverage –0.016 –0.018 –0.020 –0.017 (–0.34) (–0.44) (–0.48) (–0.36) LogFirmEmployees 0.054 0.147 0.137 (1.46) (0.81) (0.75) LogSales2 0.000 0.000 (0.10) (0.13) LogFirmEmployees2 –0.005 –0.005 (–0.47) (–0.42) Adjusted R2 0.174 0.174 0.175 0.175 Observations 48,896 48,896 48,896 48,896 t-test: Shock×Parity+Shock=0 0.502 0.508 0.507 Year F.E. Yes Yes Yes Yes Add. employment polynomials No No No Yes Establishment F.E. Yes Yes Yes Yes Dependent variable Log change in number of employees (1) (2) (3) (4) Shock×Parity 0.082 0.081 0.064 0.083 (2.37) (2.22) (1.80) (2.46) Shock×LogFirmEmployees 0.264 (1.86) Shock×LogFirmEmployees2 –0.013 (–1.79) Shock –0.096 –0.095 –1.323 –0.097 (–3.30) (–3.05) (–2.05) (–3.44) Parity –0.044 –0.042 –0.039 8.651 (–1.39) (–1.39) (–1.27) (3.34) LogEstAge –0.594 –0.594 –0.594 –0.594 (–10.8) (–10.7) (–10.8) (–10.8) LogSales 0.005 –0.012 –0.014 0.006 (0.22) (–0.08) (–0.10) (0.30) Leverage –0.016 –0.018 –0.020 –0.017 (–0.34) (–0.44) (–0.48) (–0.36) LogFirmEmployees 0.054 0.147 0.137 (1.46) (0.81) (0.75) LogSales2 0.000 0.000 (0.10) (0.13) LogFirmEmployees2 –0.005 –0.005 (–0.47) (–0.42) Adjusted R2 0.174 0.174 0.175 0.175 Observations 48,896 48,896 48,896 48,896 t-test: Shock×Parity+Shock=0 0.502 0.508 0.507 Year F.E. Yes Yes Yes Yes Add. employment polynomials No No No Yes Establishment F.E. Yes Yes Yes Yes As expected, Shock has a significantly negative coefficient, which implies non-parity firms suffer a sharp decline in employment. We performed a standard t-test for the restriction that the coefficients on Shock and Shock × Parity add up to zero, which would indicate full insurance. In no specification can we reject the null hypothesis that the coefficients on Shock and Shock × Parity have the same magnitude with opposite signs. It appears employees working for parity firms are fully protected against negative industry shocks. 4.1.a. Identification issues To address identification issues, we first check the validity of the parallel trends assumption, the possible confounding effects of the level of employment in Germany (hereafter, the employment level), and whether we find similar results in placebo tests. Then, we examine the potential endogeneity of the parity status and conduct RD analysis. In addition, we examine changes in parity status due to industry shocks, cross-industry effects, and robustness to various alternative specifications. Parallel trends. In our difference-in-differences estimation, the outcome variable of the control group is used to calculate the expected counterfactual, assuming that parity and non-parity firms have the same time trend if there are no industry shocks. Thus, we check whether the log change in the number of employees of parity and non-parity firms were following the same trend before an industry shock. Figure 2 plots log change in the number of employees separately for our sample of parity and non-parity firms from 4 years prior to 3 years after the first shock year, where the first year of the shock is defined as event year 0. We only include establishments in industries without a shock during the 4 years prior to year 0. To ensure that changes in the sample do not affect trends in employment, we require that, during the event period, all establishments have data available for all years and do not change their parity status. Figure 2. View largeDownload slide Pre- and post-shock trends in employment changes of establishments of parity and non-parity firms. Notes: This figure plots trends in average log changes of employment for establishments of parity (black line) and non-parity firms (gray line) in event time. Time=0 is the first year of an industry shock. We require Shock=0 for all years before Time=0. Only establishments with more than 50 employees and data available for the complete event period are included. Establishments switching parity status during the event period are excluded. Figure 2. View largeDownload slide Pre- and post-shock trends in employment changes of establishments of parity and non-parity firms. Notes: This figure plots trends in average log changes of employment for establishments of parity (black line) and non-parity firms (gray line) in event time. Time=0 is the first year of an industry shock. We require Shock=0 for all years before Time=0. Only establishments with more than 50 employees and data available for the complete event period are included. Establishments switching parity status during the event period are excluded. Figure 2 shows remarkably parallel trends in average employment growth of parity and non-parity establishments until year –1, and a sharp divergence in year 0. The sharp drop in employment of non-parity establishments in year 0 vis-à-vis parity establishments, which show no significant decline in employment, reflects what we find in Table III: Incomplete employment protection for non-parity establishments versus full protection for workers of parity establishments.21 Parity and firm-level employment. An important concern for our specification is that Parity is defined with respect to the number of employees in Germany. Note that only the number of employees in Germany is relevant for whether a firm is subject to parity codetermination. Hence, Parity is only loosely related to firm size.22 We separate the Parity effect from the employment level effects in Column (3) of Table III by including interactions of Shock with the log of the total number of employees in Germany, LogFirmEmployees, and LogFirmEmployees2. If Parity has no unique effect other than just being an indicator for having more than 2,000 employees in Germany, then the coefficient on Shock × Parity should become insignificant if we interact Shock and the employment-level variables. However, the coefficient of Shock × Parity remains positive and significant, albeit at a lower statistical significance. The parity status per se has a significant positive effect on employment during shock periods.23 We also allow the impact of LogFirmEmployees and its higher-order terms to be different for parity and non-parity firms in Column (4) by interacting these terms with Parity and (1 – Parity). None of the coefficients of employment-level variables is significant. (For brevity, the coefficients are not reported.) The coefficient on the interaction of Shock and Parity remains positive and significant. Placebo tests. Another way to test the unique effect of parity-codetermination during industry shocks is to conduct placebo tests using false parity dummies. We reestimate Columns (1) and (2) in Table III, separately for parity and non-parity firms, with Placebo replacing Parity in the regressions. The thresholds for the placebo dummy, Placebo, equal the median number of employees in Germany, which is 1,318 for non-parity firms and 10,458 for parity firms. The results are presented in Table A-2 in the Online Appendix. Coefficients of both, Placebo and Shock × Placebo, are statistically zero for both subsamples, demonstrating that the 2,000-threshold is unique and the treatment effect is real. We take this placebo test idea further by reestimating Specification (2) in Table III, with Placebo equal to one if employment exceeds a threshold that varies from 1,500 to 2,500 in increments of 50, that is, 1,500, 1,550, 1,600, …, 2,500. Our hypothesis predicts that the 2,000 threshold is uniquely relevant. The more we move the cut-off above (below) 2,000, the more firm-years we misclassify as non-parity (parity) firms, which creates attenuation bias and reduces the coefficients on Parity and Shock × Parity. For example, if Placebo is based on 2,500 employees, all establishments belonging to firms with a number of employees in Germany between 2,000 and 2,500 are classified as non-parity, even though they are parity codetermined. Hence, if the parity effect is unique, we should find both coefficient estimates and significance levels peak at 2,000, and decline as we move the cut-off further away from 2,000. Figure 3 presents the results for the estimated coefficients on the interaction term (right-hand axis, gray line) and the corresponding t-statistics (left-hand axis, black line). As expected, coefficients and significance levels both peak around the 2,000 threshold. Notice that if our results were driven by the employment-level effect alone, without a separate effect of Parity, no noise would have been added at Placebo thresholds greater than 2,000, and the coefficient would have increased (or at least plateaued out) as Placebo increases further away from 2,000. Figure 3. View largeDownload slide Employment—placebo tests. Notes: This figure shows the coefficients on the interaction term Placebo×Shock (right-hand axis, gray line) and the corresponding t-statistics (left-hand axis, black line) from estimating Column (2) in Table III with Placebo instead of Parity. We vary the placebo threshold in steps of 50 from 1,500 to 2,500 and let Placebo be equal to one for any establishment of a firm with employment in Germany above the corresponding placebo threshold. Figure 3. View largeDownload slide Employment—placebo tests. Notes: This figure shows the coefficients on the interaction term Placebo×Shock (right-hand axis, gray line) and the corresponding t-statistics (left-hand axis, black line) from estimating Column (2) in Table III with Placebo instead of Parity. We vary the placebo threshold in steps of 50 from 1,500 to 2,500 and let Placebo be equal to one for any establishment of a firm with employment in Germany above the corresponding placebo threshold. Endogeneity of parity status. Parity may be endogenous, as employers and employees may attempt to influence their firm’s parity status through influencing the number of employees in Germany. Workers may want to keep the number above 2,000 to maintain parity status, whereas shareholders may attempt to keep the number below 2,000 to maintain non-parity status.24 Such attempts would lead to an accumulation of firms around the 2,000 threshold. To investigate whether there is an unusual concentration of firms located right below or above the 2,000 employee threshold, we first plot a histogram of the frequency of the distribution of all sample firm-year observations with 500–3,500 workers employed in Germany in Figure A-2 of the Online Appendix. It shows there are more firms with fewer employees. More important, it shows scattered and minor peaks throughout the whole range of 500–3,500, but does not show an unusual concentration around the 2,000 threshold. To conduct a more formal test, we first use the McCrary (2008) test, which uses a smoothed version of the histogram, testing for a discontinuity of the density at the 2,000 threshold. Figure 4 shows the smoothed density estimate, which allows for a discontinuous jump of the density at the threshold. Different estimates are obtained depending on the range of data included; we use ranges from 1,500 to 2,500 employees and expand it symmetrically; the largest interval is from zero to 4,000 employees. In all cases the estimate of the discontinuity is smaller than its standard error; we cannot reject that the density is continuous at 2,000 at any conventional significance level. Figure 4. View largeDownload slide Distribution of firms by number of employees (density plot). Notes: The figure shows the histogram and estimated density of all firm-year observations for which the number of employees in Germany is between 0 and 4,000 with breakpoint at 2,000 employees. The density and the confidence band are estimated using the McCrary (2008) estimator using the default bin size of 59 and bandwidth of 904. Figure 4. View largeDownload slide Distribution of firms by number of employees (density plot). Notes: The figure shows the histogram and estimated density of all firm-year observations for which the number of employees in Germany is between 0 and 4,000 with breakpoint at 2,000 employees. The density and the confidence band are estimated using the McCrary (2008) estimator using the default bin size of 59 and bandwidth of 904. To be cautious, we conduct two additional tests. The first involves a non-parametric regression of employment growth on the employment level. If firms perceive parity codetermination as very costly, they may sacrifice growth if employment is just below the 2,000 threshold. Employees, on the other hand, may resist contraction more if the level of employment is just above the 2,000 threshold. Hence, we expect a negative relation between the number of employees and employment growth below the threshold, and a positive relation above the threshold. Figure 5 plots a non-parametric regression allowing the slope of the regression of employment growth on employment to change arbitrarily with the level of employment. We find no evidence in support of the manipulation hypothesis. Although never significant, the regression is upward sloping just below the threshold, and downward sloping just above the threshold, the opposite of what we would expect if workers or shareholders would influence the number of employees to prevent board composition from changing. Figure 5. View largeDownload slide Employment growth around the 2,000-employee cutoff. Notes: This figure shows the result of a local polynomial regression with local-mean smoothing (and 95% confidence band) of firm employment growth on the number of firm employees between 1,500 and 2,500 in Germany. Epanechnikov kernel with bandwidth 50 are used. Figure 5. View largeDownload slide Employment growth around the 2,000-employee cutoff. Notes: This figure shows the result of a local polynomial regression with local-mean smoothing (and 95% confidence band) of firm employment growth on the number of firm employees between 1,500 and 2,500 in Germany. Epanechnikov kernel with bandwidth 50 are used. For the second test, we tabulate the frequency of firms crossing various employment-level thresholds from below or above by counting how many times our sample firms cross 1,500, 1,600, … up to 2,500 in the increment of 100. If shareholders prevent crossing the 2,000 threshold from below, and workers prevent crossing the 2,000 threshold from above, then we should observe fewer crossings over 2,000 compared with all other thresholds. Table A-3 in the Online Appendix reports the tabulation results, which show more or less even distribution across all thresholds without a noticeable difference for the 2,000 threshold. In sum, all analyses lead to the same conclusion: There is no indication that firms cluster around the 2,000 threshold or manipulate the likelihood of crossing the threshold, either upward or downward. It appears that business and operational considerations are the dominant factor in deciding the employment level. There is no evidence that either workers or shareholders influence employment decisions to suit their interests in terms of board representation. As we shall show later, parity codetermination has no discernible impact on firm profitability or value; hence, firms are unlikely to alter investment decisions just to avoid codetermination. Whether a firm is just above or just below the 2,000-threshold seems rather random. RD analysis. The above conclusion allows us to perform RD analysis. Testing whether our conclusions hold locally around the 2,000-employee cut-off point improves the covariate balance relative to difference-in-differences analysis. However, it comes at the cost of lowering the power of the tests stemming from smaller sample size: RD analysis has to be restricted to industry-shock periods, which drastically reduces our sample size to 5.8% of establishment-years in our sample. We regress employment in shock periods on Parity using a local linear regression: yijkt=f(FirmEmployeesit)+δParityjt+ηjt, (2) where f() is a kernel-weighted local polynomial of degree one. Table IV reports the results and Figure 6 plots local polynomial regressions around the 2,000 threshold with 95% confidence intervals in dashed lines.25 We show results for the optimized bandwidth based on Imbens and Kalyanaraman (2012) as well as results from doubling or halving the optimal bandwidth. We apply a sharp RD design, which is appropriate here, because firms never elect to be parity-codetermined if they are below the 2,000 threshold. Table IV. Employment change—RD analysis This table presents results for a kernel regression using a triangular kernel and the optimized Imbens and Kalyanaraman, 2012 bandwidth (bw = 1). The dependent variable is log change in employment. We modify the optimized bandwidth by factors of 0.5 and 2 to check robustness. A sharp RD design is assumed, where the treatment variable (Parity) changes from one to zero at 2,000 employees (firm level). The z-statistics for the coefficient estimates are reported in parentheses below the estimates. Dependent variable Log change in employment Parity (bw = 1) 0.148 (2.02) Parity (bw = 0.5) 0.092 (1.10) Parity (bw = 2) 0.134 (1.95) Observations 314 Dependent variable Log change in employment Parity (bw = 1) 0.148 (2.02) Parity (bw = 0.5) 0.092 (1.10) Parity (bw = 2) 0.134 (1.95) Observations 314 Table IV. Employment change—RD analysis This table presents results for a kernel regression using a triangular kernel and the optimized Imbens and Kalyanaraman, 2012 bandwidth (bw = 1). The dependent variable is log change in employment. We modify the optimized bandwidth by factors of 0.5 and 2 to check robustness. A sharp RD design is assumed, where the treatment variable (Parity) changes from one to zero at 2,000 employees (firm level). The z-statistics for the coefficient estimates are reported in parentheses below the estimates. Dependent variable Log change in employment Parity (bw = 1) 0.148 (2.02) Parity (bw = 0.5) 0.092 (1.10) Parity (bw = 2) 0.134 (1.95) Observations 314 Dependent variable Log change in employment Parity (bw = 1) 0.148 (2.02) Parity (bw = 0.5) 0.092 (1.10) Parity (bw = 2) 0.134 (1.95) Observations 314 Figure 6. View largeDownload slide Discontinuity at the 2,000-employee cutoff: Employment. Notes: This figure shows the results of two local polynomial regressions for log change in employment during shock periods around the parity cutoff of 2,000 employees. Figure 6. View largeDownload slide Discontinuity at the 2,000-employee cutoff: Employment. Notes: This figure shows the results of two local polynomial regressions for log change in employment during shock periods around the parity cutoff of 2,000 employees. In spite of the big drop in sample size, the results show a statistically significant upward jump of employment residuals above the 2,000 cutoff for the optimal bandwidth, consistent with the shock influencing employment at non-parity firms more strongly than at parity firms. Results remain significant if we increase the bandwidth, but not if we halve it, which is equivalent to a further reduction in sample size that is already very small. Continuous measures of industry shocks. Sraer and Thesmar (2007) and Ellul, Pagano, and Schivardi (2017) study implicit employment contracts by family-firms using continuous measures of shocks. To check the robustness of our conclusions to continuous measures of shocks, we re-estimate our baseline specifications with measures of shocks similar to those used in Sraer and Thesmar (2007) and Ellul, Pagano, and Schivardi (2017). First, we follow Sraer and Thesmar (2007) and define a new variable LogMeanIndEmplit, the logarithm of the industry mean of contemporaneous employment of non-sample firms in the three-digit NACE industry firm i belongs to, excluding the employment of firm i itself.26 We interact LogMeanIndEmplit with the Parity indicator, in keeping with our earlier analysis. Here, Parity takes the place of indicators for family involvement in the family-firms literature.27Table V (Panel A) shows the results. Note that with this definition of shocks, our hypothesis predicts opposite signs: a positive sign on LogMeanIndEmplit and a negative sign on the interaction with Parity. The coefficient on the inaction term is negative and highly significant in all specifications. Specifications controlling for year-fixed effects (Columns 3 and 4) show coefficients of 0.29 and 0.36 on LogMeanIndEmpl, a range of elasticity estimates quantitatively similar to those in Sraer and Thesmar (2007). With this definition, insurance in parity firms manifests in a reduction of the sensitivity to industry-level shocks of about 50% for parity firms. Table V. Employment changes—continuous shock measure This table presents OLS estimation results with (1) log employment as the dependent variable in Panel A, and (2) the log change in employment as the dependent variable in Panel B. Only establishments with more than 50 employees are included. Panel A adapts the methodology of Sraer and Thesmar (2007) [see their equation (4) and Table 7]. We use the log of contemporaneous average employment of non-sample firms’ establishments in the same industry (NACE, LogMeanIndEmpl) instead of our Shock definition. Panel B adapts the methodology of Ellul, Pagano, and Schivardi (2017) (see their Table 2). We use the contemporaneous log change in employment of non-sample firms’ establishments in the same industry (LogChangeIndEmpl) instead of our Shock definition. T-statistics are reported in parentheses. Standard errors are clustered at the firm level. Panel A Dependent variable Log number of employees (1) (2) (3) (4) LogMeanIndEmpl 0.492 0.480 0.286 0.359 (5.68) (5.90) (3.00) (3.74) Parity×LogMeanIndEmpl –0.172 –0.172 –0.141 –0.209 (–2.12) (–2.15) (–1.94) (–2.51) Parity 0.620 0.623 0.527 0.645 (1.09) (1.11) (0.91) (1.12) LogEstAge –0.020 0.112 0.113 (–0.57) (3.58) (3.70) Adjusted R2 0.909 0.909 0.912 0.912 Observations 54,042 54,042 54,042 54,042 t-Test: LogMIE×Parity+LogMIE=0 0.036 0.037 0.295 0.262 Establishment F.E. Yes Yes Yes Yes Year F.E. No No Yes No Year×Parity F.E. No No No Yes Panel B Dependent variable Log change in number of employees (1) (2) (3) (4) LogChangeIndEmpl 0.740 0.598 0.424 0.453 (3.03) (2.52) (2.00) (2.91) Parity×LogChangeIndEmpl –0.681 –0.522 –0.271 –0.295 (–2.38) (–2.04) (–2.04) (–2.71) Parity –0.051 0.070 –0.001 0.406 (–4.06) (1.53) (–0.06) (7.82) LogEstAge –0.344 –0.596 –0.596 (–9.50) (–11.51) (–11.34) Adjusted R2 0.084 0.153 0.179 0.180 Observations 50,134 50,134 50,134 50,134 t-Test: LogCIE×Parity+LogCIE=0 0.549 0.441 0.151 0.147 Establishment F.E. Yes Yes Yes Yes Year F.E. No No Yes No Year×Parity F.E. No No No Yes Panel A Dependent variable Log number of employees (1) (2) (3) (4) LogMeanIndEmpl 0.492 0.480 0.286 0.359 (5.68) (5.90) (3.00) (3.74) Parity×LogMeanIndEmpl –0.172 –0.172 –0.141 –0.209 (–2.12) (–2.15) (–1.94) (–2.51) Parity 0.620 0.623 0.527 0.645 (1.09) (1.11) (0.91) (1.12) LogEstAge –0.020 0.112 0.113 (–0.57) (3.58) (3.70) Adjusted R2 0.909 0.909 0.912 0.912 Observations 54,042 54,042 54,042 54,042 t-Test: LogMIE×Parity+LogMIE=0 0.036 0.037 0.295 0.262 Establishment F.E. Yes Yes Yes Yes Year F.E. No No Yes No Year×Parity F.E. No No No Yes Panel B Dependent variable Log change in number of employees (1) (2) (3) (4) LogChangeIndEmpl 0.740 0.598 0.424 0.453 (3.03) (2.52) (2.00) (2.91) Parity×LogChangeIndEmpl –0.681 –0.522 –0.271 –0.295 (–2.38) (–2.04) (–2.04) (–2.71) Parity –0.051 0.070 –0.001 0.406 (–4.06) (1.53) (–0.06) (7.82) LogEstAge –0.344 –0.596 –0.596 (–9.50) (–11.51) (–11.34) Adjusted R2 0.084 0.153 0.179 0.180 Observations 50,134 50,134 50,134 50,134 t-Test: LogCIE×Parity+LogCIE=0 0.549 0.441 0.151 0.147 Establishment F.E. Yes Yes Yes Yes Year F.E. No No Yes No Year×Parity F.E. No No No Yes Table V. Employment changes—continuous shock measure This table presents OLS estimation results with (1) log employment as the dependent variable in Panel A, and (2) the log change in employment as the dependent variable in Panel B. Only establishments with more than 50 employees are included. Panel A adapts the methodology of Sraer and Thesmar (2007) [see their equation (4) and Table 7]. We use the log of contemporaneous average employment of non-sample firms’ establishments in the same industry (NACE, LogMeanIndEmpl) instead of our Shock definition. Panel B adapts the methodology of Ellul, Pagano, and Schivardi (2017) (see their Table 2). We use the contemporaneous log change in employment of non-sample firms’ establishments in the same industry (LogChangeIndEmpl) instead of our Shock definition. T-statistics are reported in parentheses. Standard errors are clustered at the firm level. Panel A Dependent variable Log number of employees (1) (2) (3) (4) LogMeanIndEmpl 0.492 0.480 0.286 0.359 (5.68) (5.90) (3.00) (3.74) Parity×LogMeanIndEmpl –0.172 –0.172 –0.141 –0.209 (–2.12) (–2.15) (–1.94) (–2.51) Parity 0.620 0.623 0.527 0.645 (1.09) (1.11) (0.91) (1.12) LogEstAge –0.020 0.112 0.113 (–0.57) (3.58) (3.70) Adjusted R2 0.909 0.909 0.912 0.912 Observations 54,042 54,042 54,042 54,042 t-Test: LogMIE×Parity+LogMIE=0 0.036 0.037 0.295 0.262 Establishment F.E. Yes Yes Yes Yes Year F.E. No No Yes No Year×Parity F.E. No No No Yes Panel B Dependent variable Log change in number of employees (1) (2) (3) (4) LogChangeIndEmpl 0.740 0.598 0.424 0.453 (3.03) (2.52) (2.00) (2.91) Parity×LogChangeIndEmpl –0.681 –0.522 –0.271 –0.295 (–2.38) (–2.04) (–2.04) (–2.71) Parity –0.051 0.070 –0.001 0.406 (–4.06) (1.53) (–0.06) (7.82) LogEstAge –0.344 –0.596 –0.596 (–9.50) (–11.51) (–11.34) Adjusted R2 0.084 0.153 0.179 0.180 Observations 50,134 50,134 50,134 50,134 t-Test: LogCIE×Parity+LogCIE=0 0.549 0.441 0.151 0.147 Establishment F.E. Yes Yes Yes Yes Year F.E. No No Yes No Year×Parity F.E. No No No Yes Panel A Dependent variable Log number of employees (1) (2) (3) (4) LogMeanIndEmpl 0.492 0.480 0.286 0.359 (5.68) (5.90) (3.00) (3.74) Parity×LogMeanIndEmpl –0.172 –0.172 –0.141 –0.209 (–2.12) (–2.15) (–1.94) (–2.51) Parity 0.620 0.623 0.527 0.645 (1.09) (1.11) (0.91) (1.12) LogEstAge –0.020 0.112 0.113 (–0.57) (3.58) (3.70) Adjusted R2 0.909 0.909 0.912 0.912 Observations 54,042 54,042 54,042 54,042 t-Test: LogMIE×Parity+LogMIE=0 0.036 0.037 0.295 0.262 Establishment F.E. Yes Yes Yes Yes Year F.E. No No Yes No Year×Parity F.E. No No No Yes Panel B Dependent variable Log change in number of employees (1) (2) (3) (4) LogChangeIndEmpl 0.740 0.598 0.424 0.453 (3.03) (2.52) (2.00) (2.91) Parity×LogChangeIndEmpl –0.681 –0.522 –0.271 –0.295 (–2.38) (–2.04) (–2.04) (–2.71) Parity –0.051 0.070 –0.001 0.406 (–4.06) (1.53) (–0.06) (7.82) LogEstAge –0.344 –0.596 –0.596 (–9.50) (–11.51) (–11.34) Adjusted R2 0.084 0.153 0.179 0.180 Observations 50,134 50,134 50,134 50,134 t-Test: LogCIE×Parity+LogCIE=0 0.549 0.441 0.151 0.147 Establishment F.E. Yes Yes Yes Yes Year F.E. No No Yes No Year×Parity F.E. No No No Yes We also re-estimate our baseline regressions following Ellul, Pagano, and Schivardi (2017), who use first differences and total log changes as “shock” variable, a somewhat different approach from Sraer and Thesmar (2007). Specifically, we regress the log change in employment on LogChangeIndEmpl, the log change in employment of non-sample firms’ establishments in the same industry. Table V (Panel B) reports the results, which imply that parity firms’ response to industry-level employment changes are only about one-third of that of non-parity firms. In sum, our conclusions based on discrete shocks are robust to continuous measures of industry shocks. An alternative sample construction and alternative definitions of shock. We reestimate the regressions in Table III using other alternative specifications and report the results Table A-5 of the Online Appendix. Specifically, in Panel C, we include all establishments with more than 10 employees. In Panel D, we define Shock by a 2.5% drop in industry-wide employment among non-sample firms instead of the 5% drop. In Panel E, we use the 2-year interval in defining Shock instead of the 4-year interval. In Panel F, we control for outliers by truncating establishment-years with employment growth in the 1st and 99th percentile. Results are robust to these alternative specifications. Changes in parity status due to industry shocks. An industry shock may cause a sufficient drop in employment to make a parity firm become a non-parity firm. If both happen simultaneously but we observe the drop in employment after the change in parity status, we will incorrectly attribute the employment change to non-parity firms, underestimating the decline in employment among parity firms. To check this possibility, we first read annual reports of all firms changing their status from parity to non-parity.28 In almost all cases, worker representatives stay on the supervisory board until the annual shareholder meeting the year after the number of employees in Germany drops below 2,000; hence, our findings are highly unlikely to be affected by the simultaneous drop in employment and change in parity status. We are also concerned about establishments switching from parity to non-parity status that were in an industry shock the year of or the 2 years before the change in parity status. A big drop in employment in those establishments could cause the change in parity status. Our sample contains 236 establishments switching from parity to non-parity status. Of the 236 establishments, only 23 were in an industry shock in the year of or the 2 years before the parity status change. We drop all observations (49) belonging to these 23 establishments for 3 years after the switch, and reestimate Columns (1)–(3) of Table III. The reestimation results, presented in Table A-4 of the Online Appendix, are robust. Cross-industry effects. Firms operating in multiple industries may react to a shock in one industry by transferring workers to establishments in non-shock industries (Tate and Yang, 2016). Such transfers would be recorded as employment losses in establishments affected by a shock, lowering the coefficient on Shock × Parity, and as employment gains in establishments unaffected by the shock, not affecting the coefficient on Shock × Parity. Hence, the net effect would be to bias our results to finding less insurance. In spite of this potential bias, we find full insurance for parity firms.29 4.1.b. Local labor market conditions and positive industry shocks In this section, we extend our hypothesis to consider how the insurance effect of parity codetermination depends on local labor market conditions and how parity and non-parity firms differ in their response to positive industry shocks. Local labor market conditions. We expect implicit employment insurance be more prevalent when employment risk is greater. Labor markets are local and migration across local labor market has declined (Molloy, Smith, and Wozniak, 2017). Accordingly, employees become more vulnerable to losing their jobs if employers have more local monopsony power (Benmelech, Bergman, and Kim, 2018). Hence, we hypothesize that if one employer accounts for a larger fraction of employment in an industry in their local labor market, demand for insurance will be greater and codetermination as an enforcement mechanism will have greater impact on preventing layoffs. To test this prediction, we adapt the research design of Benmelech, Bergman, and Kim (2018) and measure employer concentration for each county–industry–year for the 401 counties and 222 NACE industries (three-digit level) by calculating the Hirschman–Herfindahl index (HHI) for the employment shares of local employers. We then define the indicator variable Herfindahl, which is one if the HHI is above the median across all counties and industries, and zero otherwise. We estimate triple-differences using the same specifications as Table III and report the results in Table VI. The variable of main interest here is the triple interaction of Shock, Parity, and Herfindahl. Its coefficient measures whether and how the effect of parity codetermination on insuring employment mechanism depends on employer concentration at the industry level. The effect is economically strong, with coefficient values around 0.05, and statistically significant at the 10% level in all specifications. From the interaction of Shock and Herfindahl, we also observe that higher employer concentration increases the vulnerability of employees to shocks, although this effect is not significant. These results imply that the insurance effect of parity codetermination is greater, the more vulnerable employees are to the market power of local employers, providing further corroborating evidence in support of our hypothesis. Table VI. Employment changes and local labor market conditions This table repeats the analysis of Table III with additional interaction terms using Herfindahl. Herfindahl is a dummy variable equal to one if the employment HHI in the establishment’s industry-county is above the sample median. The employment HHI is calculated using all establishments with more than 50 employees in the same county and industry (three-digit NACE code). The controls are not reported to preserve space. T-statistics are reported in parentheses. Standard errors are clustered at the firm level. Dependent variable Log change in number of employees (1) (2) (3) (4) Shock×Parity 0.052 0.051 0.045 0.048 (1.95) (1.88) (1.79) (1.94) Shock×LogFirmEmployees 0.245 (1.69) Shock×LogFirmEmployees2 –0.012 (–1.60) Shock –0.170 –0.168 –1.338 –0.166 (–2.57) (–2.50) (–2.98) (–2.56) Herfindahl 0.076 0.075 0.075 0.075 (2.91) (2.94) (2.91) (2.91) Parity×Herfindahl –0.001 –0.001 –0.001 –0.002 (–0.19) (–0.19) (–0.20) (–0.22) Shock×Herfindahl 0.103 0.102 0.114 0.099 (1.39) (1.37) (1.60) (1.35) Parity×Shock×Herfindahl 0.047 0.047 0.060 0.044 (1.72) (1.71) (1.98) (1.68) Parity –0.040 –0.039 –0.035 8.626 (–1.30) (–1.30) (–1.17) (3.36) Adjusted R2 0.175 0.175 0.175 0.175 Observations 48,896 48,896 48,896 48,896 Year F.E. Yes Yes Yes Yes Add. employment polynomials No No No Yes Establishment F.E. Yes Yes Yes Yes Dependent variable Log change in number of employees (1) (2) (3) (4) Shock×Parity 0.052 0.051 0.045 0.048 (1.95) (1.88) (1.79) (1.94) Shock×LogFirmEmployees 0.245 (1.69) Shock×LogFirmEmployees2 –0.012 (–1.60) Shock –0.170 –0.168 –1.338 –0.166 (–2.57) (–2.50) (–2.98) (–2.56) Herfindahl 0.076 0.075 0.075 0.075 (2.91) (2.94) (2.91) (2.91) Parity×Herfindahl –0.001 –0.001 –0.001 –0.002 (–0.19) (–0.19) (–0.20) (–0.22) Shock×Herfindahl 0.103 0.102 0.114 0.099 (1.39) (1.37) (1.60) (1.35) Parity×Shock×Herfindahl 0.047 0.047 0.060 0.044 (1.72) (1.71) (1.98) (1.68) Parity –0.040 –0.039 –0.035 8.626 (–1.30) (–1.30) (–1.17) (3.36) Adjusted R2 0.175 0.175 0.175 0.175 Observations 48,896 48,896 48,896 48,896 Year F.E. Yes Yes Yes Yes Add. employment polynomials No No No Yes Establishment F.E. Yes Yes Yes Yes Table VI. Employment changes and local labor market conditions This table repeats the analysis of Table III with additional interaction terms using Herfindahl. Herfindahl is a dummy variable equal to one if the employment HHI in the establishment’s industry-county is above the sample median. The employment HHI is calculated using all establishments with more than 50 employees in the same county and industry (three-digit NACE code). The controls are not reported to preserve space. T-statistics are reported in parentheses. Standard errors are clustered at the firm level. Dependent variable Log change in number of employees (1) (2) (3) (4) Shock×Parity 0.052 0.051 0.045 0.048 (1.95) (1.88) (1.79) (1.94) Shock×LogFirmEmployees 0.245 (1.69) Shock×LogFirmEmployees2 –0.012 (–1.60) Shock –0.170 –0.168 –1.338 –0.166 (–2.57) (–2.50) (–2.98) (–2.56) Herfindahl 0.076 0.075 0.075 0.075 (2.91) (2.94) (2.91) (2.91) Parity×Herfindahl –0.001 –0.001 –0.001 –0.002 (–0.19) (–0.19) (–0.20) (–0.22) Shock×Herfindahl 0.103 0.102 0.114 0.099 (1.39) (1.37) (1.60) (1.35) Parity×Shock×Herfindahl 0.047 0.047 0.060 0.044 (1.72) (1.71) (1.98) (1.68) Parity –0.040 –0.039 –0.035 8.626 (–1.30) (–1.30) (–1.17) (3.36) Adjusted R2 0.175 0.175 0.175 0.175 Observations 48,896 48,896 48,896 48,896 Year F.E. Yes Yes Yes Yes Add. employment polynomials No No No Yes Establishment F.E. Yes Yes Yes Yes Dependent variable Log change in number of employees (1) (2) (3) (4) Shock×Parity 0.052 0.051 0.045 0.048 (1.95) (1.88) (1.79) (1.94) Shock×LogFirmEmployees 0.245 (1.69) Shock×LogFirmEmployees2 –0.012 (–1.60) Shock –0.170 –0.168 –1.338 –0.166 (–2.57) (–2.50) (–2.98) (–2.56) Herfindahl 0.076 0.075 0.075 0.075 (2.91) (2.94) (2.91) (2.91) Parity×Herfindahl –0.001 –0.001 –0.001 –0.002 (–0.19) (–0.19) (–0.20) (–0.22) Shock×Herfindahl 0.103 0.102 0.114 0.099 (1.39) (1.37) (1.60) (1.35) Parity×Shock×Herfindahl 0.047 0.047 0.060 0.044 (1.72) (1.71) (1.98) (1.68) Parity –0.040 –0.039 –0.035 8.626 (–1.30) (–1.30) (–1.17) (3.36) Adjusted R2 0.175 0.175 0.175 0.175 Observations 48,896 48,896 48,896 48,896 Year F.E. Yes Yes Yes Yes Add. employment polynomials No No No Yes Establishment F.E. Yes Yes Yes Yes Positive shocks. Since our hypothesis is about insurance against adverse industry shocks, it does not yield direct predictions about how positive industry shocks affect employment of parity firms differently from non-parity firms. However, it is plausible that when parity firms know it is difficult to lay off workers because of implicit contracts enforceable through parity codetermination, they will be more cautious to add workers when business conditions call for more employees. Similarly, if they have retained workers after negative shocks in the past, they are less likely to add new workers at the same rate as non-parity firms. The provision of job security is tantamount to increased firing costs in a dynamic model with adjustment costs. Based on calibrations of such a model (Bentolila and Bertola, 1990), we expect parity firms to be less responsive to positive shocks than non-parity firms. To test this prediction, we compare employment growth in parity and non-parity establishments when the industry as a whole experiences positive growth in employment. The positive growth indicator, Pos. Growth, is equal to one when non-sample firms’ establishments located in Germany belonging to the same industry as a whole increase their work force. We reestimate Table III with Pos. Growth in the place of Shock and report the results in Panel G of Table A-5 of the Online Appendix. The results are consistent with our conjecture. While non-parity establishments increase the number of employees by about 4%, parity establishments show a significantly smaller increase, roughly half those of non-parity establishments. 4.1.c. Employment by occupational status The estimation results based on all employees may mask important heterogeneity across different types of employees. Table VII therefore re-estimates the employment regressions separately for each skill level. We include the same set of control variables as in Specifications (1) and (2) in Table III. For white-collar workers and skilled blue-collar workers, the coefficient on Shock × Parity is positive, economically large, and statistically significant for both specifications. The results for unskilled blue-collar workers are in sharp contrast; both specifications show an insignificant coefficient on Shock × Parity and the sum of Shock and Shock × Parity is significantly negative at the 10% level. Unlike white-collar and skilled blue-collar workers, there is no evidence these workers are protected against an industry-wide decline in employment. Table VII. Employment change—white-collar, skilled blue-collar, and unskilled blue-collar employees This table presents OLS estimation results with the log change in the number of white-collar, skilled blue-collar, or unskilled blue-collar employees as the dependent variable. Only establishments with more than 50 employees are included. All regressions control for year- and establishment-fixed effects. T-statistics are reported in parentheses. Standard errors are clustered at the firm level. The table also reports the p-value for the t-test that Shock + Shock × Parity = 0. Dependent variable White-collar employees Skilled blue-collar employees Unskilled blue-collar employees (1) (2) (3) (4) (5) (6) Shock×Parity 0.076 0.076 0.137 0.133 0.005 0.006 (2.23) (2.14) (2.79) (2.76) (0.12) (0.14) Shock –0.082 –0.082 –0.112 –0.108 –0.063 –0.064 (–2.63) (–2.50) (–2.63) (–2.57) (–2.19) (–2.29) Parity –0.051 –0.050 –0.108 –0.104 0.030 0.029 (–1.57) (–1.59) (–2.41) (–2.92) (1.06) (1.01) LogEstAge –0.633 –0.633 –0.467 –0.467 –0.503 –0.503 (–10.79) (–10.83) (–8.07) (–8.08) (–12.95) (–12.97) LogSales 0.021 –0.005 0.023 –0.019 –0.001 –0.018 (0.85) (–0.03) (1.15) (–0.14) (–0.06) (–0.15) Leverage 0.077 0.074 –0.035 –0.041 –0.043 –0.043 (0.80) (0.82) (–0.76) (–0.88) (–1.94) (–1.80) LogFirmEmployees 0.062 0.154 0.088 0.355 0.087 0.016 (1.37) (0.81) (2.44) (2.51) (2.15) (0.08) LogSales2 0.001 0.001 0.000 (0.13) (0.31) (0.13) LogFirmEmployees2 –0.005 –0.014 0.004 (–0.43) (–1.77) (0.33) Adjusted R2 0.171 0.171 0.064 0.065 0.076 0.076 Observations 48,896 48,896 48,896 48,896 48,896 48,896 t-test: Shock×Parity+ Shock=0 0.709 0.719 0.406 0.397 0.098 0.091 Dependent variable White-collar employees Skilled blue-collar employees Unskilled blue-collar employees (1) (2) (3) (4) (5) (6) Shock×Parity 0.076 0.076 0.137 0.133 0.005 0.006 (2.23) (2.14) (2.79) (2.76) (0.12) (0.14) Shock –0.082 –0.082 –0.112 –0.108 –0.063 –0.064 (–2.63) (–2.50) (–2.63) (–2.57) (–2.19) (–2.29) Parity –0.051 –0.050 –0.108 –0.104 0.030 0.029 (–1.57) (–1.59) (–2.41) (–2.92) (1.06) (1.01) LogEstAge –0.633 –0.633 –0.467 –0.467 –0.503 –0.503 (–10.79) (–10.83) (–8.07) (–8.08) (–12.95) (–12.97) LogSales 0.021 –0.005 0.023 –0.019 –0.001 –0.018 (0.85) (–0.03) (1.15) (–0.14) (–0.06) (–0.15) Leverage 0.077 0.074 –0.035 –0.041 –0.043 –0.043 (0.80) (0.82) (–0.76) (–0.88) (–1.94) (–1.80) LogFirmEmployees 0.062 0.154 0.088 0.355 0.087 0.016 (1.37) (0.81) (2.44) (2.51) (2.15) (0.08) LogSales2 0.001 0.001 0.000 (0.13) (0.31) (0.13) LogFirmEmployees2 –0.005 –0.014 0.004 (–0.43) (–1.77) (0.33) Adjusted R2 0.171 0.171 0.064 0.065 0.076 0.076 Observations 48,896 48,896 48,896 48,896 48,896 48,896 t-test: Shock×Parity+ Shock=0 0.709 0.719 0.406 0.397 0.098 0.091 Table VII. Employment change—white-collar, skilled blue-collar, and unskilled blue-collar employees This table presents OLS estimation results with the log change in the number of white-collar, skilled blue-collar, or unskilled blue-collar employees as the dependent variable. Only establishments with more than 50 employees are included. All regressions control for year- and establishment-fixed effects. T-statistics are reported in parentheses. Standard errors are clustered at the firm level. The table also reports the p-value for the t-test that Shock + Shock × Parity = 0. Dependent variable White-collar employees Skilled blue-collar employees Unskilled blue-collar employees (1) (2) (3) (4) (5) (6) Shock×Parity 0.076 0.076 0.137 0.133 0.005 0.006 (2.23) (2.14) (2.79) (2.76) (0.12) (0.14) Shock –0.082 –0.082 –0.112 –0.108 –0.063 –0.064 (–2.63) (–2.50) (–2.63) (–2.57) (–2.19) (–2.29) Parity –0.051 –0.050 –0.108 –0.104 0.030 0.029 (–1.57) (–1.59) (–2.41) (–2.92) (1.06) (1.01) LogEstAge –0.633 –0.633 –0.467 –0.467 –0.503 –0.503 (–10.79) (–10.83) (–8.07) (–8.08) (–12.95) (–12.97) LogSales 0.021 –0.005 0.023 –0.019 –0.001 –0.018 (0.85) (–0.03) (1.15) (–0.14) (–0.06) (–0.15) Leverage 0.077 0.074 –0.035 –0.041 –0.043 –0.043 (0.80) (0.82) (–0.76) (–0.88) (–1.94) (–1.80) LogFirmEmployees 0.062 0.154 0.088 0.355 0.087 0.016 (1.37) (0.81) (2.44) (2.51) (2.15) (0.08) LogSales2 0.001 0.001 0.000 (0.13) (0.31) (0.13) LogFirmEmployees2 –0.005 –0.014 0.004 (–0.43) (–1.77) (0.33) Adjusted R2 0.171 0.171 0.064 0.065 0.076 0.076 Observations 48,896 48,896 48,896 48,896 48,896 48,896 t-test: Shock×Parity+ Shock=0 0.709 0.719 0.406 0.397 0.098 0.091 Dependent variable White-collar employees Skilled blue-collar employees Unskilled blue-collar employees (1) (2) (3) (4) (5) (6) Shock×Parity 0.076 0.076 0.137 0.133 0.005 0.006 (2.23) (2.14) (2.79) (2.76) (0.12) (0.14) Shock –0.082 –0.082 –0.112 –0.108 –0.063 –0.064 (–2.63) (–2.50) (–2.63) (–2.57) (–2.19) (–2.29) Parity –0.051 –0.050 –0.108 –0.104 0.030 0.029 (–1.57) (–1.59) (–2.41) (–2.92) (1.06) (1.01) LogEstAge –0.633 –0.633 –0.467 –0.467 –0.503 –0.503 (–10.79) (–10.83) (–8.07) (–8.08) (–12.95) (–12.97) LogSales 0.021 –0.005 0.023 –0.019 –0.001 –0.018 (0.85) (–0.03) (1.15) (–0.14) (–0.06) (–0.15) Leverage 0.077 0.074 –0.035 –0.041 –0.043 –0.043 (0.80) (0.82) (–0.76) (–0.88) (–1.94) (–1.80) LogFirmEmployees 0.062 0.154 0.088 0.355 0.087 0.016 (1.37) (0.81) (2.44) (2.51) (2.15) (0.08) LogSales2 0.001 0.001 0.000 (0.13) (0.31) (0.13) LogFirmEmployees2 –0.005 –0.014 0.004 (–0.43) (–1.77) (0.33) Adjusted R2 0.171 0.171 0.064 0.065 0.076 0.076 Observations 48,896 48,896 48,896 48,896 48,896 48,896 t-test: Shock×Parity+ Shock=0 0.709 0.719 0.406 0.397 0.098 0.091 To explain this surprising finding, we hand collect information on the occupational status and the educational and vocational qualification of labor representatives on supervisory boards for 1990, 1999, and 2008, the beginning, the middle, and the end of our sample period, for all firms which provide the relevant information in their annual reports. Table VIII, Panel A, categorizes labor representatives into unskilled blue-collar, skilled blue-collar, white-collar, and union representatives. The occupational status of union representatives is usually not reported, so they form a separate category in the table. In Panel B, we categorize labor representatives as low-qualified, qualified, and highly qualified. Here, we exclude union representatives, because information on union representatives’ educational/vocational qualifications is usually not reported. Panel C shows the occupational status and qualification categories are closely related. Table VIII. Employee representatives by occupational status and qualification Panel A reports the occupational status and Panel B the educational and vocational qualification of labor representatives on supervisory boards. We hand collected this information for all sample firms existing in 1990, 1999, and 2008, the beginning, the middle, and the end of our sample period. The tabulation is based on 97, 231, and 229 labor representatives of 15, 35, and 48 sample firms, respectively, which report the relevant personal information in annual reports. To follow the structure of the IAB data, we categorized labor representatives in Panel A into (1) unskilled blue collar, (2) skilled blue collar, (3) white collar, and (4) union representatives. The occupational status of union representatives is usually not reported, but their occupational status tends to be white collar. Panel B categorizes labor representatives into (1) low qualified, (2) qualified, and (3) highly qualified. In Panel B, union representatives are excluded because their qualification is usually not reported. Panel C presents how the classification based on occupational status corresponds to the breakdown by educational and vocational qualification. It is based on a random sample of 2% of all employees covered by the IAB database between 1975 and 2008 (“Sample of Integrated Labour Market Biographies”). Panel A Occupational status 1990 1999 2008 Unskilled blue collar (%) 0.0 0.0 0.0 Skilled blue collar (%) 42.3 22.5 22.3 White collar (%) 36.1 54.5 56.3 Union representative (%) 21.6 22.9 21.4 Sum (%) 100.0 100.0 100.0 Number of labor representatives 97 231 229 Number of firms 15 35 48 Panel B Qualification 1990 1999 2008 Low qualified (%) 0.0 0.0 0.0 Qualified (%) 86.8 92.1 59.4 Highly qualified (%) 13.2 7.9 40.6 Sum (%) 100.0 100.0 100.0 Number of labor representatives 76 178 180 Number of firms 15 35 48 Panel A Occupational status 1990 1999 2008 Unskilled blue collar (%) 0.0 0.0 0.0 Skilled blue collar (%) 42.3 22.5 22.3 White collar (%) 36.1 54.5 56.3 Union representative (%) 21.6 22.9 21.4 Sum (%) 100.0 100.0 100.0 Number of labor representatives 97 231 229 Number of firms 15 35 48 Panel B Qualification 1990 1999 2008 Low qualified (%) 0.0 0.0 0.0 Qualified (%) 86.8 92.1 59.4 Highly qualified (%) 13.2 7.9 40.6 Sum (%) 100.0 100.0 100.0 Number of labor representatives 76 178 180 Number of firms 15 35 48 Panel C Unskilled blue collar Skilled blue collar White collar Sum Low qualified (%) 15.5 2.2 2.5 20.2 Qualified (%) 9.8 25.6 36.6 72.0 Highly qualified (%) 0.1 0.1 7.7 7.9 Sum (%) 25.4 27.9 46.8 100.0 Panel C Unskilled blue collar Skilled blue collar White collar Sum Low qualified (%) 15.5 2.2 2.5 20.2 Qualified (%) 9.8 25.6 36.6 72.0 Highly qualified (%) 0.1 0.1 7.7 7.9 Sum (%) 25.4 27.9 46.8 100.0 Table VIII. Employee representatives by occupational status and qualification Panel A reports the occupational status and Panel B the educational and vocational qualification of labor representatives on supervisory boards. We hand collected this information for all sample firms existing in 1990, 1999, and 2008, the beginning, the middle, and the end of our sample period. The tabulation is based on 97, 231, and 229 labor representatives of 15, 35, and 48 sample firms, respectively, which report the relevant personal information in annual reports. To follow the structure of the IAB data, we categorized labor representatives in Panel A into (1) unskilled blue collar, (2) skilled blue collar, (3) white collar, and (4) union representatives. The occupational status of union representatives is usually not reported, but their occupational status tends to be white collar. Panel B categorizes labor representatives into (1) low qualified, (2) qualified, and (3) highly qualified. In Panel B, union representatives are excluded because their qualification is usually not reported. Panel C presents how the classification based on occupational status corresponds to the breakdown by educational and vocational qualification. It is based on a random sample of 2% of all employees covered by the IAB database between 1975 and 2008 (“Sample of Integrated Labour Market Biographies”). Panel A Occupational status 1990 1999 2008 Unskilled blue collar (%) 0.0 0.0 0.0 Skilled blue collar (%) 42.3 22.5 22.3 White collar (%) 36.1 54.5 56.3 Union representative (%) 21.6 22.9 21.4 Sum (%) 100.0 100.0 100.0 Number of labor representatives 97 231 229 Number of firms 15 35 48 Panel B Qualification 1990 1999 2008 Low qualified (%) 0.0 0.0 0.0 Qualified (%) 86.8 92.1 59.4 Highly qualified (%) 13.2 7.9 40.6 Sum (%) 100.0 100.0 100.0 Number of labor representatives 76 178 180 Number of firms 15 35 48 Panel A Occupational status 1990 1999 2008 Unskilled blue collar (%) 0.0 0.0 0.0 Skilled blue collar (%) 42.3 22.5 22.3 White collar (%) 36.1 54.5 56.3 Union representative (%) 21.6 22.9 21.4 Sum (%) 100.0 100.0 100.0 Number of labor representatives 97 231 229 Number of firms 15 35 48 Panel B Qualification 1990 1999 2008 Low qualified (%) 0.0 0.0 0.0 Qualified (%) 86.8 92.1 59.4 Highly qualified (%) 13.2 7.9 40.6 Sum (%) 100.0 100.0 100.0 Number of labor representatives 76 178 180 Number of firms 15 35 48 Panel C Unskilled blue collar Skilled blue collar White collar Sum Low qualified (%) 15.5 2.2 2.5 20.2 Qualified (%) 9.8 25.6 36.6 72.0 Highly qualified (%) 0.1 0.1 7.7 7.9 Sum (%) 25.4 27.9 46.8 100.0 Panel C Unskilled blue collar Skilled blue collar White collar Sum Low qualified (%) 15.5 2.2 2.5 20.2 Qualified (%) 9.8 25.6 36.6 72.0 Highly qualified (%) 0.1 0.1 7.7 7.9 Sum (%) 25.4 27.9 46.8 100.0 These tabulations reveal a striking phenomenon: We do not find a single unskilled blue-collar or low-qualified worker among the labor representatives for any firm in any year. Although companies not included in this tabulation may have unskilled blue-collar or low-qualified workers on their boards, the numbers in Table VIII reveal the lack of real representation of unskilled blue-collar or low-qualified workers.30 The implicit insurance contract works only for those who have their own kind represented on the board, an important enforcement mechanism to ensure employers honor the contract. There appears to be an agency problem between unskilled workers and their representatives on supervisory boards. Why unskilled workers cannot remedy this lack of proper representation is puzzling and warrants further investigation, which would require an in-depth analysis of the election process of worker representatives, an issue beyond the scope of this paper. 4.2 Wages The protection against layoffs may not be the result of implementing implicit insurance contracts. The employment protection could be due to employee entrenchment through worker–management collusion (Pagano and Volpin, 2005; Atanassov and Kim, 2009). Whereas the implicit-insurance hypothesis predicts lower wages (Hypothesis 2), the entrenchment hypothesis predicts, if anything, the opposite. Such entrenched employees are unlikely to accept lower wages. Thus, we relate wages to Parity, the coefficient of which measures the wage difference between parity-codetermined and non-parity firms. We use the median wage at each establishment, because the IAB provides only the first quartile, the median, and the third quartile wages. We use two sets of control variables: (1) the same control variables as in Column (2) in Table III and (2) these variables plus the number of employees in the establishment, the median employee age, and the percentage of white-collar employees. Prior research suggests the additional control variables help explain average employee wages (e.g., Oi and Idson, 1999; Brown and Medoff, 1989). We take logs of all level variables when estimating regressions. 4.2.a. Wages for all employees The first two columns in Table IX report estimation results for all employees. The coefficients on Parity are negative and highly significant. The point estimate in Column (2) indicates employees of parity-codetermined firms receive on average about 3.3% lower wages, consistent with Hypothesis 2 and implicit contract theory. Table IX. Wages—all, highly qualified, qualified, and low-qualified employees This table presents OLS estimation results with median wages of all, highly qualified, qualified, or low-qualified employees as the dependent variable. The wage variables are defined as the log of median gross average daily wage for (1) all full-time employees, (2) with higher educational qualifications, (3) with educational/vocational qualifications, and (4) without educational/vocational qualifications. Only establishments with more than 50 employees are included. All regressions control for year- and establishment-fixed effects. The t-statistics are reported in parentheses. Standard errors are clustered at the firm level. Dependent variable All employees Highly qualified Qualified Low-qualified (1) (2) (3) (4) (5) (6) (7) (8) Parity –0.031 –0.033 –0.031 –0.031 –0.029 –0.032 –0.033 –0.034 (–3.10) (–3.78) (–2.49) (–2.56) (–3.01) (–3.70) (–1.59) (–1.66) LogEstAge 0.049 0.049 0.059 0.060 0.050 0.051 0.031 0.030 (3.55) (3.42) (6.18) (6.06) (3.61) (3.49) (1.90) (1.79) LogSales –0.212 –0.194 –0.022 –0.018 –0.244 –0.229 –0.058 –0.046 (–2.27) (–2.37) (–0.37) (–0.32) (–2.70) (–2.85) (–0.60) (–0.52) Leverage –0.024 –0.023 0.005 0.006 –0.016 –0.015 –0.075 –0.073 (–0.93) (–0.90) (0.23) (0.29) (–0.65) (–0.64) (–2.88) (–2.85) LogFirmEmployees 0.032 0.067 –0.062 –0.053 0.039 0.065 0.016 0.023 (0.34) (0.77) (–1.04) (–0.89) (0.44) (0.78) (0.14) (0.22) LogSales2 0.006 0.005 0.001 0.001 0.006 0.006 0.002 0.002 (2.49) (2.61) (0.73) (0.68) (2.83) (3.00) (0.90) (0.85) LogFirmEmployees2 –0.003 –0.004 0.004 0.003 –0.003 –0.003 –0.002 –0.002 (–0.48) (–0.73) (0.94) (0.88) (–0.49) (–0.64) (–0.27) (–0.28) Log#Employees –0.033 –0.009 –0.032 –0.015 (–4.04) (–1.58) (–3.80) (–1.84) LogMedianEmplAge 0.185 0.071 0.194 0.206 (3.98) (2.45) (5.65) (4.86) RatioWhiteCollar 0.153 0.022 0.072 0.047 (2.89) (1.03) (1.59) (0.75) Adjusted R2 0.942 0.945 0.825 0.826 0.926 0.929 0.800 0.801 Observations 51,205 51,205 38,670 38,670 51,060 51,060 42,336 42,336 Dependent variable All employees Highly qualified Qualified Low-qualified (1) (2) (3) (4) (5) (6) (7) (8) Parity –0.031 –0.033 –0.031 –0.031 –0.029 –0.032 –0.033 –0.034 (–3.10) (–3.78) (–2.49) (–2.56) (–3.01) (–3.70) (–1.59) (–1.66) LogEstAge 0.049 0.049 0.059 0.060 0.050 0.051 0.031 0.030 (3.55) (3.42) (6.18) (6.06) (3.61) (3.49) (1.90) (1.79) LogSales –0.212 –0.194 –0.022 –0.018 –0.244 –0.229 –0.058 –0.046 (–2.27) (–2.37) (–0.37) (–0.32) (–2.70) (–2.85) (–0.60) (–0.52) Leverage –0.024 –0.023 0.005 0.006 –0.016 –0.015 –0.075 –0.073 (–0.93) (–0.90) (0.23) (0.29) (–0.65) (–0.64) (–2.88) (–2.85) LogFirmEmployees 0.032 0.067 –0.062 –0.053 0.039 0.065 0.016 0.023 (0.34) (0.77) (–1.04) (–0.89) (0.44) (0.78) (0.14) (0.22) LogSales2 0.006 0.005 0.001 0.001 0.006 0.006 0.002 0.002 (2.49) (2.61) (0.73) (0.68) (2.83) (3.00) (0.90) (0.85) LogFirmEmployees2 –0.003 –0.004 0.004 0.003 –0.003 –0.003 –0.002 –0.002 (–0.48) (–0.73) (0.94) (0.88) (–0.49) (–0.64) (–0.27) (–0.28) Log#Employees –0.033 –0.009 –0.032 –0.015 (–4.04) (–1.58) (–3.80) (–1.84) LogMedianEmplAge 0.185 0.071 0.194 0.206 (3.98) (2.45) (5.65) (4.86) RatioWhiteCollar 0.153 0.022 0.072 0.047 (2.89) (1.03) (1.59) (0.75) Adjusted R2 0.942 0.945 0.825 0.826 0.926 0.929 0.800 0.801 Observations 51,205 51,205 38,670 38,670 51,060 51,060 42,336 42,336 Table IX. Wages—all, highly qualified, qualified, and low-qualified employees This table presents OLS estimation results with median wages of all, highly qualified, qualified, or low-qualified employees as the dependent variable. The wage variables are defined as the log of median gross average daily wage for (1) all full-time employees, (2) with higher educational qualifications, (3) with educational/vocational qualifications, and (4) without educational/vocational qualifications. Only establishments with more than 50 employees are included. All regressions control for year- and establishment-fixed effects. The t-statistics are reported in parentheses. Standard errors are clustered at the firm level. Dependent variable All employees Highly qualified Qualified Low-qualified (1) (2) (3) (4) (5) (6) (7) (8) Parity –0.031 –0.033 –0.031 –0.031 –0.029 –0.032 –0.033 –0.034 (–3.10) (–3.78) (–2.49) (–2.56) (–3.01) (–3.70) (–1.59) (–1.66) LogEstAge 0.049 0.049 0.059 0.060 0.050 0.051 0.031 0.030 (3.55) (3.42) (6.18) (6.06) (3.61) (3.49) (1.90) (1.79) LogSales –0.212 –0.194 –0.022 –0.018 –0.244 –0.229 –0.058 –0.046 (–2.27) (–2.37) (–0.37) (–0.32) (–2.70) (–2.85) (–0.60) (–0.52) Leverage –0.024 –0.023 0.005 0.006 –0.016 –0.015 –0.075 –0.073 (–0.93) (–0.90) (0.23) (0.29) (–0.65) (–0.64) (–2.88) (–2.85) LogFirmEmployees 0.032 0.067 –0.062 –0.053 0.039 0.065 0.016 0.023 (0.34) (0.77) (–1.04) (–0.89) (0.44) (0.78) (0.14) (0.22) LogSales2 0.006 0.005 0.001 0.001 0.006 0.006 0.002 0.002 (2.49) (2.61) (0.73) (0.68) (2.83) (3.00) (0.90) (0.85) LogFirmEmployees2 –0.003 –0.004 0.004 0.003 –0.003 –0.003 –0.002 –0.002 (–0.48) (–0.73) (0.94) (0.88) (–0.49) (–0.64) (–0.27) (–0.28) Log#Employees –0.033 –0.009 –0.032 –0.015 (–4.04) (–1.58) (–3.80) (–1.84) LogMedianEmplAge 0.185 0.071 0.194 0.206 (3.98) (2.45) (5.65) (4.86) RatioWhiteCollar 0.153 0.022 0.072 0.047 (2.89) (1.03) (1.59) (0.75) Adjusted R2 0.942 0.945 0.825 0.826 0.926 0.929 0.800 0.801 Observations 51,205 51,205 38,670 38,670 51,060 51,060 42,336 42,336 Dependent variable All employees Highly qualified Qualified Low-qualified (1) (2) (3) (4) (5) (6) (7) (8) Parity –0.031 –0.033 –0.031 –0.031 –0.029 –0.032 –0.033 –0.034 (–3.10) (–3.78) (–2.49) (–2.56) (–3.01) (–3.70) (–1.59) (–1.66) LogEstAge 0.049 0.049 0.059 0.060 0.050 0.051 0.031 0.030 (3.55) (3.42) (6.18) (6.06) (3.61) (3.49) (1.90) (1.79) LogSales –0.212 –0.194 –0.022 –0.018 –0.244 –0.229 –0.058 –0.046 (–2.27) (–2.37) (–0.37) (–0.32) (–2.70) (–2.85) (–0.60) (–0.52) Leverage –0.024 –0.023 0.005 0.006 –0.016 –0.015 –0.075 –0.073 (–0.93) (–0.90) (0.23) (0.29) (–0.65) (–0.64) (–2.88) (–2.85) LogFirmEmployees 0.032 0.067 –0.062 –0.053 0.039 0.065 0.016 0.023 (0.34) (0.77) (–1.04) (–0.89) (0.44) (0.78) (0.14) (0.22) LogSales2 0.006 0.005 0.001 0.001 0.006 0.006 0.002 0.002 (2.49) (2.61) (0.73) (0.68) (2.83) (3.00) (0.90) (0.85) LogFirmEmployees2 –0.003 –0.004 0.004 0.003 –0.003 –0.003 –0.002 –0.002 (–0.48) (–0.73) (0.94) (0.88) (–0.49) (–0.64) (–0.27) (–0.28) Log#Employees –0.033 –0.009 –0.032 –0.015 (–4.04) (–1.58) (–3.80) (–1.84) LogMedianEmplAge 0.185 0.071 0.194 0.206 (3.98) (2.45) (5.65) (4.86) RatioWhiteCollar 0.153 0.022 0.072 0.047 (2.89) (1.03) (1.59) (0.75) Adjusted R2 0.942 0.945 0.825 0.826 0.926 0.929 0.800 0.801 Observations 51,205 51,205 38,670 38,670 51,060 51,060 42,336 42,336 We cannot completely rule out that employees offer wage cuts in exchange for other benefits we do not observe; for example, for better working conditions. While possible, we believe our explanation is more parsimonious. There is no reason to believe that other aspects of the labor contract, such as working conditions, safety, etc. change abruptly at the 2,000 threshold of employees in Germany. Working conditions are negotiated with work councils at the establishment level and usually not part of board-level negotiations. The discontinuous change at the 2,000 threshold documented through the placebo tests and RD analysis above speaks more in favor of a quid pro quo in exchange for employment insurance. Estimated coefficients on controls are mostly consistent with intuition. Older employees and employees working in older establishments and establishments with a greater proportion of white collar workers are paid more. However, the number of employees in establishments is associated with lower wages. This is somewhat surprising given the finding in Brown and Medoff (1989) that an increase in the number of employees is associated with higher wages. The difference could be due to differences in sample and specification. Our sample is at the establishment level, heavily skewed toward large firms, and our regression contains a number of other firm size variables, whereas Brown and Medoff (1989) rely on firm-level data, include small businesses, and their specification contains fewer size controls. We do not include Shock and Shock × Parity in the regressions, because we are primarily interested in the difference in average wages between parity and non-parity firms. For consistency with the employment regressions, we include Shock and Shock × Parity in the wage regressions and report results in Table A-6, Panel B, Columns (1) and (2). The results are robust; employees of parity firms receive 3.0–3.1% lower wages during non-shock periods and the difference is significant. The coefficients on Shock are insignificant. The Shock × Parity term shows positive but insignificant coefficients. These results confirm our prior that German wages are sticky and do not respond to industry shocks in significant ways. 4.2.b. Wages by qualification The IAB does not provide the wage data broken down by occupational status. It provides only the breakdown by qualifications. To see how the classification based on qualifications corresponds to occupational status, at our request, IAB cross-tabulated the percentage of employees belonging to each type of occupational status and qualification based on a random sample of 2% of all employees covered by its database between 1975 and 2008 (Sample of Integrated Labour Market Biographies). The tabulation is shown in Table VIII, Panel C. Most highly qualified workers tend to be white-collar workers; most qualified workers are either white collar or skilled blue-collar workers; and most low-qualified workers are unskilled blue-collar workers. Columns (3)–(8) in Table IX report separate estimates for employees grouped by qualification.31 Coefficients on Parity range from –2.9% to –3.4% for all three qualification levels. The coefficients are highly significant for the qualified and highly qualified groups, suggesting that skilled blue-collar and white-collar employees of parity firms receive significantly lower wages. For low-qualified employees, the coefficient on Parity is not significant, even though the size of the coefficient is similar. This group of employees has larger standard errors, probably because roughly one-fourth belongs to skilled blue-collar or white-collar workers (see Table VIII, Panel C). 4.2.c. Identification issues Placebo tests. As in the employment analysis, we re-estimate the wage results while replacing the parity dummy with placebo dummies. The results are reported in Table A-8 of the Online Appendix, separately for parity and non-parity firms. Coefficients on placebo dummies are statistically zero for all four regressions. We also repeat placebo regressions for all employees with Placebo equal to one if employment exceeds a threshold that varies from 1,500 to 2,500 in increments of 50. Figure 7 shows the regression coefficients on Placebo (right axis, gray line) and the t-statistic (left axis, black line). The lowest values are reached around the 2,000 threshold, supporting the conclusion that the relevance of this threshold is genuine.32 Figure 7. View largeDownload slide Wages—placebo tests. Notes: This figure shows the coefficients on Placebo (right-hand axis, gray line) and the corresponding t-statistics (left-hand axis, black line) from estimating Column (2) in Table IX with Placebo instead of Parity. We vary the placebo threshold in steps of 50 from 1,500 to 2,500 and let Placebo be equal to one for any establishment of a firm with employment in Germany above the corresponding placebo threshold. Figure 7. View largeDownload slide Wages—placebo tests. Notes: This figure shows the coefficients on Placebo (right-hand axis, gray line) and the corresponding t-statistics (left-hand axis, black line) from estimating Column (2) in Table IX with Placebo instead of Parity. We vary the placebo threshold in steps of 50 from 1,500 to 2,500 and let Placebo be equal to one for any establishment of a firm with employment in Germany above the corresponding placebo threshold. RD analysis. We also test whether our conclusions regarding wages hold locally around the 2,000-employee cut-off with RD analyses. We follow the same steps as in the employment analysis. Unlike the employment analysis, however, we do not see a drop in sample size because the estimation is done for the entire cycle of shock and non-shock periods. Results are reported in Table X, with Figure 8 plotting local polynomial regressions around the 2,000 threshold with 95% confidence intervals in dashed lines. The drop in wages above the 2,000 threshold is significant for the optimal bandwidth, implying that parity firms’ employees are paid less than those at non-parity firms. The estimate for the insurance premium at the optimal bandwidth is about 8%, much larger than those estimated via the difference-in-differences approach, because the RD analyses do not contain the control variables used in the difference-in-differences approach. Table X. Wages—RD analysis This table presents results for a kernel regression using a triangular kernel and the optimized (Imbens and Kalyanaraman, 2012) bandwidth (bw = 1). The dependent variable is log median daily wage. We modify the optimized bandwidth by factors of 0.5 and 2 to check robustness. A sharp RD design is assumed, where the treatment variable (Parity) changes from zero to one at 2,000 employees (firm level). The z-statistics for the coefficient estimates are reported in parentheses below the estimates. Dependent variable Log median daily wage Parity (bw = 1) –0.079 (–1.95) Parity (bw = 0.5) –0.148 (–2.58) Parity (bw = 2) –0.039 (–1.26) Observations 4,962 Dependent variable Log median daily wage Parity (bw = 1) –0.079 (–1.95) Parity (bw = 0.5) –0.148 (–2.58) Parity (bw = 2) –0.039 (–1.26) Observations 4,962 Table X. Wages—RD analysis This table presents results for a kernel regression using a triangular kernel and the optimized (Imbens and Kalyanaraman, 2012) bandwidth (bw = 1). The dependent variable is log median daily wage. We modify the optimized bandwidth by factors of 0.5 and 2 to check robustness. A sharp RD design is assumed, where the treatment variable (Parity) changes from zero to one at 2,000 employees (firm level). The z-statistics for the coefficient estimates are reported in parentheses below the estimates. Dependent variable Log median daily wage Parity (bw = 1) –0.079 (–1.95) Parity (bw = 0.5) –0.148 (–2.58) Parity (bw = 2) –0.039 (–1.26) Observations 4,962 Dependent variable Log median daily wage Parity (bw = 1) –0.079 (–1.95) Parity (bw = 0.5) –0.148 (–2.58) Parity (bw = 2) –0.039 (–1.26) Observations 4,962 Figure 8. View largeDownload slide Discontinuity at the 2,000-employee cutoff: wages. Notes: This figure shows the results of two local polynomial regressions for log median daily wage around the parity cutoff of 2,000 employees. Figure 8. View largeDownload slide Discontinuity at the 2,000-employee cutoff: wages. Notes: This figure shows the results of two local polynomial regressions for log median daily wage around the parity cutoff of 2,000 employees. Alternative sample constructions. As with the employment regressions, we re-estimate our baseline regressions while (1) lowering the threshold of the minimum number of employees in an establishment for inclusion in our sample to ten employees and (2) truncating the sample at the 1st and 99th percentile of employment growth. The re-estimation results reported in Panels C and D of Table A-6 of the Online Appendix are robust. 5. Operating Risks and Firm Performance In this section, we examine how the employment insurance affects firm risk and performance over the cycle of non-shock and shock periods. These tests require regression estimation at the firm level, so we redefine our shock measure as FirmShock, the proportion of a firm’s employees working in establishments in industries for which Shock = 1. FirmShock is a weighted average of Shock in a given firm-year, ranging between 0 and 1. For example, if 60% of a firm’s employees work in industries in which Shock equals 1, and the remaining 40% work in industries not subject to a shock in that year, then FirmShock equals 0.6. 5.1 Operating Leverage Providing job guarantees limits firms’ ability to reduce payroll in response to changes in technology, consumer taste, or general business conditions. Then wages become more fixed costs rather than variable costs, increasing operating leverage. To test this prediction, we relate Parity to two different measures of operating leverage. First, we follow Mandelker and Rhee (1984) and Chen, Kacperczyk, and Ortiz-Molina (2011) and use the elasticity of a firm’s operating income (after depreciation) with respect to sales, the Mandelker and Rhee degree of operating leverage (MRDOL). We estimate it for firms with positive EBIT using a time-series regression of log (EBIT) on log (Sales) using the five most recent annual observations. Second, we use the Novy-Marx (2011) operating leverage measure, which is cost of goods sold plus selling, general and administrative expenses, scaled by the book value of assets (NMOL). For each measure of operating leverage, we use two different specifications, one with year- and industry-fixed effects; the other with year- and firm-fixed effects. The control variables are similar to Chen, Kacperczyk, and Ortiz-Molina (2011). Results are reported in Table XI, Panel A without control variables. The coefficient on Parity is positive and statistically significant, regardless of the specification. It appears employment insurance through parity codetermination increases operating leverage. Table XI. Firm-level regressions: operating leverage, effects on ROA and Q Panel A presents OLS estimation results with (1) the Mandelker and Rhee (1984) measure of operating leverage (MRDOL) and (2) the Novy-Marx (2011) measure of operating leverage (NMOL) as the dependent variable. Following Mandelker and Rhee (1984), we estimate for firms with positive EBIT the degree of operating leverage using a time-series regression of log(EBIT) on log(Sales) using the five most recent annual observations. Novy-Marx (2011) defines operating leverage as cost of goods sold plus selling, general and administrative expenses, scaled by the book value of assets. Control variables (LogAssets, LogTobinsQ, FixedAssets/Assets, Leverage, LogFirmAge, and FirmEmployees/Assets) are not reported to conserve space. The t-statistics are reported in parentheses. Standard errors are clustered at the firm level. Panel B presents OLS estimation results with (1) ROA and (2) log Tobin’s Q as the dependent variable. The FirmShock variable is defined as the weighted average of Shock across all establishments in a firm-year. All regressions control for year- and firm-fixed effects. The t-statistics are reported in parentheses. Standard errors are clustered at the firm level. Panel A Dependent variable MRDOL NMOL (1) (2) (3) (4) Parity 0.177 0.135 0.297 0.127 (2.16) (1.79) (4.67) (2.60) Adjusted R2 0.146 0.241 0.736 0.913 Observations 1,442 1,442 953 953 Year F.E. Yes Yes Yes Yes Industry F.E. Yes No Yes No Firm F.E. No Yes No Yes Panel B Dependent variable ROA LogTobinsQ (1) (2) (3) (4) FirmShock×Parity –0.036 –0.092 (–2.59) (–1.80) FirmShock –0.030 –0.075 (–2.40) (–1.70) Parity 0.003 –0.003 0.031 0.025 (0.29) (–0.38) (1.58) (1.27) LogFirmAge –0.005 –0.006 –0.037 –0.038 (–0.75) (–0.83) (–2.26) (–2.32) LogSales –0.029 –0.029 –0.747 –0.757 (–0.65) (–0.67) (–8.39) (–8.53) Leverage –0.056 –0.055 –0.249 –0.249 (–5.15) (–5.06) (–9.72) (–9.72) LogFirmEmployees –0.031 –0.030 0.277 0.280 (–1.85) (–1.76) (6.32) (6.39) LogSales2 0.001 0.001 0.018 0.018 (1.36) (1.38) (8.38) (8.53) LogFirmEmployees2 0.002 0.002 –0.019 –0.019 (1.32) (1.25) (–5.95) (–6.03) Adjusted R2 0.545 0.543 0.682 0.681 Observations 1,734 1,734 1,885 1,885 Panel A Dependent variable MRDOL NMOL (1) (2) (3) (4) Parity 0.177 0.135 0.297 0.127 (2.16) (1.79) (4.67) (2.60) Adjusted R2 0.146 0.241 0.736 0.913 Observations 1,442 1,442 953 953 Year F.E. Yes Yes Yes Yes Industry F.E. Yes No Yes No Firm F.E. No Yes No Yes Panel B Dependent variable ROA LogTobinsQ (1) (2) (3) (4) FirmShock×Parity –0.036 –0.092 (–2.59) (–1.80) FirmShock –0.030 –0.075 (–2.40) (–1.70) Parity 0.003 –0.003 0.031 0.025 (0.29) (–0.38) (1.58) (1.27) LogFirmAge –0.005 –0.006 –0.037 –0.038 (–0.75) (–0.83) (–2.26) (–2.32) LogSales –0.029 –0.029 –0.747 –0.757 (–0.65) (–0.67) (–8.39) (–8.53) Leverage –0.056 –0.055 –0.249 –0.249 (–5.15) (–5.06) (–9.72) (–9.72) LogFirmEmployees –0.031 –0.030 0.277 0.280 (–1.85) (–1.76) (6.32) (6.39) LogSales2 0.001 0.001 0.018 0.018 (1.36) (1.38) (8.38) (8.53) LogFirmEmployees2 0.002 0.002 –0.019 –0.019 (1.32) (1.25) (–5.95) (–6.03) Adjusted R2 0.545 0.543 0.682 0.681 Observations 1,734 1,734 1,885 1,885 Table XI. Firm-level regressions: operating leverage, effects on ROA and Q Panel A presents OLS estimation results with (1) the Mandelker and Rhee (1984) measure of operating leverage (MRDOL) and (2) the Novy-Marx (2011) measure of operating leverage (NMOL) as the dependent variable. Following Mandelker and Rhee (1984), we estimate for firms with positive EBIT the degree of operating leverage using a time-series regression of log(EBIT) on log(Sales) using the five most recent annual observations. Novy-Marx (2011) defines operating leverage as cost of goods sold plus selling, general and administrative expenses, scaled by the book value of assets. Control variables (LogAssets, LogTobinsQ, FixedAssets/Assets, Leverage, LogFirmAge, and FirmEmployees/Assets) are not reported to conserve space. The t-statistics are reported in parentheses. Standard errors are clustered at the firm level. Panel B presents OLS estimation results with (1) ROA and (2) log Tobin’s Q as the dependent variable. The FirmShock variable is defined as the weighted average of Shock across all establishments in a firm-year. All regressions control for year- and firm-fixed effects. The t-statistics are reported in parentheses. Standard errors are clustered at the firm level. Panel A Dependent variable MRDOL NMOL (1) (2) (3) (4) Parity 0.177 0.135 0.297 0.127 (2.16) (1.79) (4.67) (2.60) Adjusted R2 0.146 0.241 0.736 0.913 Observations 1,442 1,442 953 953 Year F.E. Yes Yes Yes Yes Industry F.E. Yes No Yes No Firm F.E. No Yes No Yes Panel B Dependent variable ROA LogTobinsQ (1) (2) (3) (4) FirmShock×Parity –0.036 –0.092 (–2.59) (–1.80) FirmShock –0.030 –0.075 (–2.40) (–1.70) Parity 0.003 –0.003 0.031 0.025 (0.29) (–0.38) (1.58) (1.27) LogFirmAge –0.005 –0.006 –0.037 –0.038 (–0.75) (–0.83) (–2.26) (–2.32) LogSales –0.029 –0.029 –0.747 –0.757 (–0.65) (–0.67) (–8.39) (–8.53) Leverage –0.056 –0.055 –0.249 –0.249 (–5.15) (–5.06) (–9.72) (–9.72) LogFirmEmployees –0.031 –0.030 0.277 0.280 (–1.85) (–1.76) (6.32) (6.39) LogSales2 0.001 0.001 0.018 0.018 (1.36) (1.38) (8.38) (8.53) LogFirmEmployees2 0.002 0.002 –0.019 –0.019 (1.32) (1.25) (–5.95) (–6.03) Adjusted R2 0.545 0.543 0.682 0.681 Observations 1,734 1,734 1,885 1,885 Panel A Dependent variable MRDOL NMOL (1) (2) (3) (4) Parity 0.177 0.135 0.297 0.127 (2.16) (1.79) (4.67) (2.60) Adjusted R2 0.146 0.241 0.736 0.913 Observations 1,442 1,442 953 953 Year F.E. Yes Yes Yes Yes Industry F.E. Yes No Yes No Firm F.E. No Yes No Yes Panel B Dependent variable ROA LogTobinsQ (1) (2) (3) (4) FirmShock×Parity –0.036 –0.092 (–2.59) (–1.80) FirmShock –0.030 –0.075 (–2.40) (–1.70) Parity 0.003 –0.003 0.031 0.025 (0.29) (–0.38) (1.58) (1.27) LogFirmAge –0.005 –0.006 –0.037 –0.038 (–0.75) (–0.83) (–2.26) (–2.32) LogSales –0.029 –0.029 –0.747 –0.757 (–0.65) (–0.67) (–8.39) (–8.53) Leverage –0.056 –0.055 –0.249 –0.249 (–5.15) (–5.06) (–9.72) (–9.72) LogFirmEmployees –0.031 –0.030 0.277 0.280 (–1.85) (–1.76) (6.32) (6.39) LogSales2 0.001 0.001 0.018 0.018 (1.36) (1.38) (8.38) (8.53) LogFirmEmployees2 0.002 0.002 –0.019 –0.019 (1.32) (1.25) (–5.95) (–6.03) Adjusted R2 0.545 0.543 0.682 0.681 Observations 1,734 1,734 1,885 1,885 The higher operating leverage implies that parity firms will suffer greater relative reductions in profitability and valuation from industry shocks than non-parity firms. To test these implications, we estimate difference-in-differences in ROA and the logarithm of Tobin’s Q, our measures of profitability and valuation, respectively. ROA is defined as EBITDA/Total Assets to avoid changes in depreciation and amortization methods during our sample period affecting estimation results. Tobin’s Q is defined as the market value of common equity plus total assets minus the book value of common equity divided by total assets. Our main interest is again the coefficient of FirmShock × Parity, which we expect to be negative for ROA and Q. The estimation results are reported in Table XI, Panel B, Columns (1) and (3). The coefficient on FirmShock × Parity is significant and negative in both specifications. The economic magnitude is also large. The estimates for ROA show that profitability of parity-codetermined firms falls by 3.6% more if all employees of a firm are affected by a shock. This number is substantial; Table I shows that the mean (median) ROA of all firm-year observations in the sample is 7.5% (6.9%) and the coefficient on FirmShock for non-parity firms is only –3.0%. The incremental decline in Tobin’s Q for parity firms is 9.2% if all employees are affected by a shock, which is larger than the decline of 7.5% for non-parity firms. These estimates for ROA and Tobin’s Q suggest that parity codetermination more than doubles the negative impact of shocks on profitability and valuation relative to non-parity firms. 5.2 Are Shareholders Properly Compensated for Bearing the Risk? These results imply that the employment insurance increases shareholder risk—the risk of getting hurt more during industry downturns. Is the insurance premium in the form of lower wages sufficient to compensate shareholders for the risk? More generally, are shareholders worse or better off as the result of parity codetermination? We investigate this issue by reestimating the ROA and Q regressions while omitting FirmShock and FirmShock ×Parity and report the re-estimated coefficients of Parity in Columns (2) and (4) of Table XI, Panel B. The coefficient on Parity is insignificant for both ROA and Q, implying operating profits and shareholder value through the cycle are unrelated to parity codetermination. This conclusion is in line with previous findings in the literature on German codetermination.33 It appears workers capture all the efficiency gains from the improved risk-sharing arrangement through the implicit insurance contract with firms; however, they do pay an insurance premium that is on average sufficient to cover the risks of providing the insurance. Put differently, the wage concessions workers make to firms seem to be just sufficient to compensate shareholders for higher operating leverage. 6. Conclusion We find parity-codetermined firms provide skilled employees greater protection against layoffs during adverse industry shocks. There are two alternative explanations. According to the insurance hypothesis, parity-codetermination serves as an enforcement mechanism to ensure firms honor implicit insurance contracts, whereby workers receive protection against adverse shocks in return for accepting lower wages. The alternative explanation, employee entrenchment, suggests the worker control rights bestowed by parity-codetermination allow employees to prevent layoffs when the industry suffers adverse shocks, but without offering adequate concessions in the form of lower wages. We distinguish the two competing explanations by examining the wage differential between parity and non-parity firms and by analyzing firm valuation and performance. We find employees of parity firms receive significantly lower wages relative to those working for non-parity firms. Our firm performance analyses indicate that shareholders do not gain or lose from improved risk sharing and workers reap all the benefits from this arrangement. Both findings favor the insurance hypothesis. However, only skilled blue-collar and white-collar workers benefit from employment insurance, whereas unskilled blue-collar workers do not. We attribute this finding to the lack of real representation of unskilled blue-collar workers on supervisory boards. Supplementary Material Supplementary data are available at Review of Finance online. Footnotes * We are grateful for helpful comments and suggestions made by Daniel Ferreira, Fangjian Fu, Yao Lu, David Matsa, Marco Pagano, Page Quimet, David Sraer, Alex Stomper, Gary Twite, and participants at the Ackerman Conference on Corporate Governance at Bar Ilan University, the Duke-UNC Conference on Corporate Finance, the Singapore International Conference on Finance, the Rothschild Caesarea Center 11th Annual Conference, the CEPR/Study Center Gerzensee European Summer Symposium, the China International Conference in Finance, CSEF Conference on Finance and Labor, 2014 American Finance Association Annual Meetings, and seminar participants at the University of Amsterdam, University of British Columbia, University of Hong Kong, University of Lausanne, University of Mannheim, University of Maryland, University of Michigan, UNSW, and Nanyang Technological University. We are also grateful to Stefan Bender and his team at the Institut für Arbeitsmarkt- und Berufsforschung for providing access to their data. 1 Exceptions are, for example, Fama and Schwert (1977), Campbell (1996), and Jagannathan and Wang (1996). 2 On risk-return trade-offs see Palacios-Huerta (2003) and Lustig and Van Nieuwerburgh (2008). On financial distress, optimal capital structure, and restructuring, see Berk, Stanton, and Zechner (2010); Agrawal and Matsa (2013); Graham et al. (2016); and Agrawal and Tambe (2016). These studies are selected from a much larger literature on finance and labor. We do not provide a systematic and comprehensive survey of this literature here. 3 Azariadis (1975) assumes that firms that do not honor implicit contracts would “suffer a catastrophic loss in reputation” (p. 1187) and Rudanko (2011) invokes a similar assumption with the claim “equilibrium contracts are likely to be self-enforcing for a range of reasonable parameterizations.” (pp. 2823–2824). Thomas and Worrall (1988) assume that agents can always renege on a contract, but at an exogenously imposed cost: they will be excluded from all trading in the future if they renege on a contract. 4 Maybe the closest to our approach is Malcomson (1983), who argues unions enforce implicit contracts. 5 Data are from the CBR-LRI dataset (Adams et al., 2017) and refer to earlier years if data for 2016 were not available. Of the 117 countries covered, 27 grant workers board-level representation, albeit to varying degrees. Between 1970 and 2016, there were 28 changes of these laws in 19 countries, of which 22 led to more, and only six to lower, worker representation. Moreover, 92 countries provide workers with influence at the plant level or firm level through work councils. These laws were amended 78 times in 58 countries since 1970, leading to more influence for workers in 60 cases and less influence in 18 cases. 6 See Section 3.3.2 for a detailed description of how industry shocks are defined and constructed. 7 Guiso, Pistaferri, and Schivardi (2005) provide a calibration based on a parameterization of Lucas’s tree model following Clark, Leslie, and Symons (1994). They estimate that workers would be indifferent between the partial insurance the authors find for Italian firms, and uninsured consumption that is on average 9% higher than insured consumption. 8 Renaud (2007) surveys 13 studies investigating the impact of codetermination on company performance using different methodological approaches, sample constructions, and performance variables. The overall evidence seems inconclusive, suggesting no systematic relation between parity codetermination and firm performance. 9 See Azariadis and Stiglitz (1983) and Rosen (1985) for early surveys of this literature. 10 Bach and Serrano-Velarde (2015) argue that it is family links between CEOs and their successors that enhances firms’ ability to commit to implicit contracts. 11 Ellul, Pagano, and Schivardi (2017) estimate this premium to be 6% based on firm-level data. 12 Without frictions in the labor market, only partial insurance is feasible, because workers always receive pay increases if their marginal product rises above their wage. Harris and Holmstrom (1982) and Thomas and Worrall (1988) discuss contracting problems in such a setting. 13 Berk and Walden (2013) argue that firms insure workers’ human capital risk and investors spread this risk by investing in diversified portfolios. Workers could insure themselves directly through participation in capital markets. However, Berk and Walden (2013) show indirect insurance through firms is sufficiently close to being optimal so that workers prefer it to direct participation in capital markets even if the costs of direct participation are small. 14 For firms with more than 2,000 and up to 10,000 employees, the supervisory board has twelve seats, of which six are shareholder representatives, three are from the firm’s workers, one is from middle management, and two are from unions. For firms with up to 20,000 (over 20,000 employees), the size of the supervisory board increases to 16 (20) and the number of representatives from each constituency becomes: shareholders 8 (10), workers 5 (6), middle management 1 (1), and union representatives 2 (3). 15 The industry classification is based on the Statistical Classification of Economic Activities in the European Community (NACE), a six-digit industry classification. NACE is similar to North American Industry Classification System (NAICS). The first four levels are the same for all European countries. The IAB database contains different versions of the NACE classification. We use NACE Revision 1.1, which is based on the International Standard Industrial Classification (ISIC Rev. 3) of the United Nations. 16 Other groups are employees in vocational training, home workers, master craftsmen, and part-time employees. We do not analyze these groups of employees because they usually form only a small fraction of employees and are present in relatively few establishments. 17 In Germany, a relatively small fraction of students obtains an upper secondary school degree (high school, Abitur), although the fraction rose from 31% in 1992 to 45% in 2008. The typical career path in Germany is to leave school after tenth grade to enter vocational training. In 2009, 57.8% of the German population had such a vocational qualification, while 27.8% had none of these qualifications. IAB classifies the former as qualified employees and the latter as low-qualified employees (see Hethey-Maier and Seth, 2010). 18 Gorton and Schmid (2004) and several of the studies surveyed by Renaud (2007) also use parity codetermination as their main variable for labor representation. 19 We experimented with two other definitions of shocks. The first alternative makes shocks comparable across industries with different cash-flow volatilities by scaling shocks with the standard deviation of the industry-growth rate of employment, so that a lower cut-off applies to more volatile industries. The results are qualitatively similar, but statistically weaker. The second alternative uses sales growth or growth in operating income of firms from other European countries to define industry-level shocks. These analyses mostly yield insignificant estimates on the shock variable [insignificant θ in Equation (1)]. 20 Examples include Bosch, Aldi, Boehringer Ingelheim, Edeka, Rewe Group, Haniel, Shell Germany, BP Germany, Ford, Coca Cola, Procter and Gamble, Dow Chemical, Pfizer, IBM, Hewlett-Packard, ExxonMobil, Vodafone, Gazprom Germania, Sanofi-Aventis Germany, Telefónica Germany, and Fujitsu. 21 If parity firms hold on to more employees during negative shock periods, they may hire fewer new employees in expansion periods. Thus, the trends may not to be strictly parallel