Taken by storm: business financing and survival in the aftermath of Hurricane Katrina

Taken by storm: business financing and survival in the aftermath of Hurricane Katrina Abstract We use Hurricane Katrina’s damage to the Mississippi coast in 2005 as a natural experiment to study business survival in the aftermath of a capital-destruction shock. We find very low survival rates for businesses that incurred physical damage, particularly for small firms and less-productive establishments. Conditional on survival, larger and more-productive businesses that rebuilt their operations hired more workers than their smaller and less-productive counterparts. Auxiliary evidence from the Survey of Business Owners suggests that the differential size effect is tied to the presence of financial constraints, pointing to a socially inefficient level of exits and to distortions of allocative efficiency in response to this negative shock. Over time, the size advantage disappeared and market mechanisms seem to prevail. 1. Introduction Business churn—entry and exit—is an important channel supporting innovation and productivity growth, and as such is an equilibrium feature of efficient markets. This is particularly true in the retail sector (Foster et al., 2006). Nevertheless, adjustments on the extensive margin are costly and frictions can lead to socially inefficient outcomes if relatively productive firms exit while less-productive firms remain in operation in response to a shock (Caballero and Hammour, 1994; Barlevy, 2003; Bertrand et al., 2007).1 An open question is whether financial constraints hamper allocative efficiency by reducing the importance of market fundamentals such as demand and costs, thus perpetuating the presence of relatively low-productivity firms. In this paper, we exploit the quasi-natural experiment created by variation in the degree of damage inflicted by Hurricane Katrina on businesses at different locations. We explore the relative importance of productivity and financial frictions in business responses to idiosyncratic negative capital shocks. We are interested in identifying the extent to which capital shocks merely hasten the exit of less-efficient businesses, as opposed to prompting exit by efficient and viable, but financially constrained, firms. We focus on the extensive margin—business survival—because it is the costliest from a social perspective. We find that smaller and more financially vulnerable firms exhibited lower survival rates in the aftermath of storm damage after controlling for productivity differences. We also find that, even conditional on survival, smaller and more financially vulnerable firms did not rebuild or grow as quickly following storm damage. Our paper is related to a broad literature exploring the impact of financial constraints on real economic activity, particularly relating to small firms. The papers most closely related to ours examine the impact of credit-supply shocks, business-cycle shocks or housing-price shocks on the employment and output of firms of different sizes and ages. For example, Gertler and Gilchrist (1994), Sharpe (1994), and, more recently, Greenstone et al. (2014) show that small firms are more sensitive than large ones to monetary and business-cycle shocks. Adelino et al. (2015) and Fort et al. (2013) find that small firms are more sensitive to housing-price shocks, underscoring the potential importance of less-traditional forms of financing to these firms. A common difficulty with this literature is disentangling cost shocks associated with business cycles, for example as a result of an increase in the cost of financing due to an increase in interest rates or the collapse of collateral values, from demand shocks associated with the same cycles.2 Our identification solves this problem by allowing us to identify highly localized capital-destruction shocks within relatively small areas. Our paper also relates to a literature on the impact of credit constraints on business entry (e.g., Aghion et al. 2007). A fundamental difficulty in studying the determinants of entry is that the pool of potential entrants is rarely observed. Our setting provides a unique opportunity to observe something very close to the pool of potential entrants: the pool of businesses that existed in that area shortly before the shock, and which would likely have continued to operate in the absence of the shock. The destruction of a business’s capital—structure, equipment, intermediate inputs and inventory—results in a cost shock very much like entry cost if the business is to return to operation. Our results therefore support the view that financial constraints affect entry rates, particularly for small businesses. A few other papers use shocks from natural disasters to identify financial effects. Hosono et al. (2012) use detailed firm-level data to estimate the impact of the 1995 Kobe earthquake on the supply of loans. They find that firms whose headquarters were located outside the damaged area but which had borrowing relationships with banks located inside the damaged area fared worse than undamaged firms borrowing from undamaged banks. A similar finding for the 2011 Great Tohoku earthquake is reported in Uchida et al. (2013). Unlike our paper, these papers focus on the ‘bank-lending channel’ and the impact of established lending relationships when the bank suffers a shock. Our focus is on a shock to firms, many of which may not have established lending relationships but rely on more informal and less well-documenting sources of financing: personal loans, loans from friends and family and even credit-card debt. Hurricanes often cause devastation over large geographic areas, which makes it impossible to cleanly identify capital shocks distinctly from demand and infrastructure shocks. To circumvent this problem, our analysis focuses not on Louisiana, where Hurricane Katrina’s impact was most widespread, but on the Mississippi coast, where damage was much more limited and localized, infrastructure was largely unaffected (and where infrastructure was damaged, repair times were fairly short) and population outflow was minimal. Undamaged businesses near the damaged areas serve as our control group. Importantly, our identification does not require the absence of demand or productivity shocks. We allow for the possibility that storm-damaged businesses experience disproportionately large negative shocks to demand (e.g., because consumers are reluctant to travel to damaged areas) or to their supply chains, and only assume that these shocks did not disproportionately impact small or less-productive firms. We use data from the Census Bureau’s Longitudinal Business Database (LBD) on approximately 10,000 business establishments in Mississippi, including over 1500 businesses in four counties that experienced significant storm damage, combined with precise information on the location and extent of the damage from the Federal Emergency Management Administration (FEMA). These data allow us to pinpoint which establishments were damaged or destroyed and which were left intact in the same area. We focus on establishments in the retail, restaurant and hotel sectors, which require a storefront to conduct business. Our identification comes from the randomness of actual damage within a limited geographic area. We find that the storm, which hit in August 2005, generated significant excess exits of physically damaged establishments in the short run and created a 30-point wedge between the survival rate of damaged and undamaged businesses in Mississippi by 2006. This finding is consistent with the notion that, in the short term, distress caused businesses that would otherwise have survived to cease operation. In addition, we find evidence for both efficient and inefficient exit in the short run. Across the board, exiting establishments are less productive than survivors, and this productivity wedge is 50% higher for businesses whose physical structures were destroyed by Katrina. At the same time, even controlling for productivity we find that the brunt of the effect of storm damage on short-run survival fell on smaller firms. Larger firms have higher survival rates in undamaged areas, but having been hit by storm damage triples or quadruples the advantage that these firms have over their smaller, also-damaged, counterparts. In the long run, we find exit rates of large and small firms in damaged areas equalize; we interpret this as possible evidence of an unforeseen endogenous shock induced by the high rate of small-firm exit. In auxiliary analyses, we find direct evidence of the importance of financial constraints to business survival. Businesses that had previously relied on credit-card debt to finance expansion or capital improvements, demonstrating a very high marginal cost of financing, also experienced much lower post-shock survival rates than similar businesses that relied on other forms of financing, including bank loans, for capital projects. On the intensive margin, conditional on survival, physically damaged establishments grew at a lower rate than their undamaged counterparts. Consistent with the survival results, we find that larger and more productive firms were able to rebuild their operations in the short run by hiring more workers than their smaller and less-productive counterparts. However, over time the size advantage disappears; 5 years after the storm, only initial productivity predicts survival and growth. We take this as evidence that, although access to finance may confer an initial advantage at entry and offer protection against shocks, this advantage dissipates over time. The rest of the paper is organized as follows. Section 2 describes our data in detail and provides some preliminary figures highlighting our empirical identification strategy. Our analyses of survival and firm growth are, respectively, in Sections 3 and 4. Section 5 concludes. 2. Background and data Hurricane Katrina made landfall in Louisiana in late August 2005, where it caused massive flooding, and quickly veered into Mississippi. The major source of damage in Mississippi was wind damage, which caused storm surges. Figure 1 shows a map of Mississippi, highlighting the four counties—Hancock, Harrison, Jackson, and Stone Counties—that were most affected by the hurricane.3 Figure 1 View largeDownload slide Mississippi (shaded counties most affected by Katrina). Figure 1 View largeDownload slide Mississippi (shaded counties most affected by Katrina). The primary building block in our analysis is the Census Bureau’s LBD. The LBD is a longitudinal database covering all employer establishments and firms in the U.S. non-farm private economy.4 We use data from the LBD to track the activity and outcomes of all stores, restaurants and hotels operating in Mississippi between 2002 and 2010. The LBD identifies the six-digit North American Industry Classification System (NAICS) code that represents the primary activity of each business establishment.5 We limit our analysis to retail and restaurant businesses and hotels and other accommodation facilities (including casinos) for several reasons.6 First, they represent a very large share of the local economies in the affected counties, approximately 10 times as large as manufacturing.7 This is important since affected areas are small and we need sectors with enough data to conduct the analysis. Second, unlike many other service industries and some non-service industries (e.g., construction), the location of the business is non-fungible. Whereas a lawyer may continue to provide legal services and a janitorial firm may continue to provide cleaning services even if the main office is destroyed, stores, restaurants and hotels provide their services at the business address and cease operations when that location is destroyed. Finally, these sectors serve local (and tourist) demand. Demand for products in other sectors such as manufacturing may extend beyond the local area differentially depending on the size of the business and in ways that we do not observe, making it hard to determine the relative effect of demand and cost shocks for these businesses. Establishments in the LBD are defined to be ‘active’ in a given year if they report positive payroll for any part of the year. In our baseline regression specifications, we identify a surviving business as one that reports payroll either in the current year or in a subsequent year; conversely, an establishment exit is defined by the absence of any reported payroll in the current and all future years. This definition of survival is conservative in that periods of temporary inactivity are consistent with survival. For robustness checks, we have both expanded and narrowed the definition of survival. First, in some robustness checks, we have narrowed the definition of survival to exclude establishments whose payroll falls by more than 90% and establishments that continue to operate but do so under new ownership. This is our ‘restrictive’ survival variable.8 Conversely, we have also created a ‘expansive’ survival variable, which treats establishments that cease to report payroll but continue to report revenue as non-employers as survivors. To find these businesses, we supplement the LBD with data from the integrated LBD (ILBD), which provides data on businesses with revenues but no payroll. In our sample, using employment to identify survivors, approximately 18% of establishments that had payroll in 2004 are no longer in business in 2006, and 22% of establishments still active in 2006 are no longer in business by 2010. These survival rates decrease (increase) by 1–3 percentage points when we restrict (expand) the definition of exit, depending on the year. Our results are robust to these alternative definitions. The firm identifier helps us determine the age and size of the entity that owns the establishment. Firm age is censored from above because we do not know the exact age of firms that existed in 1976, the LBD’s first year, so we also include an indicator for censored ages, I(FirmAgei=T), in all regressions. Our measure of firm size is the number of establishments the firm operates nationwide; it is equal to 1 for single-unit firms and exceeds 1000 for others. Given that this is a study of survival, the accuracy of longitudinal links is key to our analysis. Establishments that maintain the same address and ownership are relatively easy to track over time. If an establishment moves to a new address, but the address is in the same county, the LBD also identifies the establishment as a continuing operation. This implies that establishments that were located in damaged areas but reopened elsewhere within the same county appear as survivors in our data. It is particularly important that we correctly identify the timing of exits. We depend on tax filings for this purpose. For single-unit firms, this is straightforward: if a firm paid no payroll taxes in 2006 or any year thereafter, we consider its last year of operation to be 2005.9 However, multi-unit firms may continue to operate and pay payroll taxes even if one establishment closes. The Census relies on the Company Organization Survey (COS) to identify exits of establishments belonging to multi-unit firms. All multi-unit firms receive the COS in 5-year intervals, and larger multi-unit firms—those with at least 250 employees in total, across all their establishments—receive the COS annually.10 The LBD does not include establishment-level revenue. We use pre-storm revenue information from the 2002 Census of Retail Trade and the 2002 Census of Accommodation and Food Services to construct a measure of labor productivity at the establishment level. In the absence of information on other inputs, such as cost of materials and capital, we calculate labor productivity as the log of the ratio of the establishment’s annual revenue to employment.11 Because this productivity measure is from 2002, we limit our analysis to establishments that had existed in 2002 as well as 2004. We geocode establishments using Geographic Information System tools to assign latitude and longitude based on the business’s address. In a small number of cases the business address may represent the address of an accountant or other hired provider who assists the business with those forms. To minimize this problem, we drop 230 businesses whose addresses were identical to addresses provided by accounting or bookkeeping firms. Not all addresses are of the necessary quality to be able to geocode down to latitude and longitude successfully. Incomplete addresses and non-standard addresses (e.g., rural routes or PO Box addresses) are the main reasons for failures. Rural areas are known to be particularly problematic in this regard. For 2004, in each of the four Mississippi counties that experienced significant direct damage from Katrina, we were able to geocode more than 85% of establishments. Table 1 lists the number of geocoded establishments in each of the four affected counties in comparison with the rest of the state.12 Geocoding rates are typically higher in the damaged counties close to the Gulf than in the more rural inland areas in the rest of the state. Table 2 compares summary statistics of establishment and firm characteristics for geocoded and non-geocoded establishments in 2004. Compared with non-geocoded establishments, geocoded establishments are on average about 1 year younger and more likely to belong to single-unit firms. On all other dimensions geocoded establishments are not statistically distinguishable from non-geocoded establishments. Table 2 Establishment summary statistics: all establishments, 2004 Variable Obs.a All Non-geo-coded Geo-coded t-Testb Single-unit firms (%) 12,300 59.4 55.4 60.3 0.000 Establishments in firm 12,300 460.3 479.1 456.0 0.451 Firm agec 12,300 18.1 18.8 17.9 0.000 Establishment employment 12,300 17.6 16.2 17.9 0.270 Establishment aged 12,300 12.8 12.6 12.9 0.146 Productivitye 12,300 4.5 4.5 4.5 0.576 Variable Obs.a All Non-geo-coded Geo-coded t-Testb Single-unit firms (%) 12,300 59.4 55.4 60.3 0.000 Establishments in firm 12,300 460.3 479.1 456.0 0.451 Firm agec 12,300 18.1 18.8 17.9 0.000 Establishment employment 12,300 17.6 16.2 17.9 0.270 Establishment aged 12,300 12.8 12.6 12.9 0.146 Productivitye 12,300 4.5 4.5 4.5 0.576 aRounded to the nearest hundred. bp-value from t-test for equality of the mean. cRight-censored age of 29 used for 4000 observations. dRight-censored age of 29 used for 1200 observations. eLog ratio of revenue to employment in 2002 for establishments that survived to 2004. Table 2 Establishment summary statistics: all establishments, 2004 Variable Obs.a All Non-geo-coded Geo-coded t-Testb Single-unit firms (%) 12,300 59.4 55.4 60.3 0.000 Establishments in firm 12,300 460.3 479.1 456.0 0.451 Firm agec 12,300 18.1 18.8 17.9 0.000 Establishment employment 12,300 17.6 16.2 17.9 0.270 Establishment aged 12,300 12.8 12.6 12.9 0.146 Productivitye 12,300 4.5 4.5 4.5 0.576 Variable Obs.a All Non-geo-coded Geo-coded t-Testb Single-unit firms (%) 12,300 59.4 55.4 60.3 0.000 Establishments in firm 12,300 460.3 479.1 456.0 0.451 Firm agec 12,300 18.1 18.8 17.9 0.000 Establishment employment 12,300 17.6 16.2 17.9 0.270 Establishment aged 12,300 12.8 12.6 12.9 0.146 Productivitye 12,300 4.5 4.5 4.5 0.576 aRounded to the nearest hundred. bp-value from t-test for equality of the mean. cRight-censored age of 29 used for 4000 observations. dRight-censored age of 29 used for 1200 observations. eLog ratio of revenue to employment in 2002 for establishments that survived to 2004. Damage information comes from FEMA and is described in detail in Jarmin and Miranda (2009). Using remote-sensing technology, FEMA classified damaged areas over the period from August 30 to September 10 using a four-tier damage scale: limited, moderate, extensive and catastrophic.13 We reduce this to a two-tier scale, combining ‘extensive’ and ‘catastrophic’ into one category (‘severe’ damage) and ‘limited’ and ‘moderate’ into a second category (‘mild’ damage). In practice, there was very little extensive damage so almost all of the damage we classify as severe is catastrophic. Critically, damage designations are not based on insurance claims. Because FEMA’s remote-sensing maps focus primarily on developed areas, we may under-estimate the damage in less-developed areas. Following Jarmin and Miranda (2009), we add the FEMA damage information to the LBD to obtain, for each geocoded establishment, the FEMA damage classification of the location containing that establishment. Figure 2 shows an area on the border of Harrison and Hancock counties in Mississippi in which storm damage was widespread and highly variable. Each gray dot on the map represents a single business establishment.14 Establishments in diagonally cross-hatched areas were severely damaged, while those in horizontal and vertical cross-hatched areas were mildly damaged. Establishments in the white areas were physically undamaged. In addition, a handful of business establishments were located in areas that still had standing water 1 week after the storm. These areas are diagonally lined in the figure but are excluded from our analysis due to the very small number of establishments impacted by flooding; none of our results are sensitive to this exclusion. Figure 2 View largeDownload slide Damage area closeup: Harrison and Hancock counties, MS locations are ‘jittered’ to prevent identification of particular establishments. Figure 2 View largeDownload slide Damage area closeup: Harrison and Hancock counties, MS locations are ‘jittered’ to prevent identification of particular establishments. The accuracy of the FEMA damage designations is critical for our estimation. FEMA reports that, of the 150,000 homes it classified using this scale in Katrina’s immediate aftermath, fewer than 10% were mis-classified (Federal Emergency Management Agency, 2011, 42). The smallest area designated in the Mississippi database is about 145 m2 (0.000056 square miles); the median is about 55,600 m (0.02 square miles). The larger areas tend to be closer to the shore, where damage was most severe and several city blocks were effectively destroyed; smaller areas are designated inland where damage intensity to structures is more likely to differ across small distances. To the extent that damage is mismeasured, this would imply measurement error and therefore attenuation bias in our estimated coefficients. Table 1 provides 2004 summary statistics for the four affected counties and an aggregated ‘rest of state’ category. Approximately 350 establishments were in areas later designated by FEMA as having endured severe damage, and 350 more were in areas later designated as having suffered mild damage. We refer to all of these establishments as ‘damaged’. The last two columns in Table 1 provide the approximate percentage of establishments in each of the counties with a damage designation. Very small cells are suppressed to comply with Census Bureau disclosure requirements. Table 3 shows pre-storm summary statistics for the 2004 cross-section of geocoded establishments. The first column, showing the average value of the variable for all geocoded establishments, reproduces column (4) of Table 2. The next two columns show the average value for establishments located in areas that were later damaged and those located in areas that were undamaged. For almost all the variables listed—firm size (number of establishments in firm as well as a single-unit firm indicator), firm age, establishment size (employment) and establishment age—the differences between the damaged and undamaged areas are both small and statistically insignificant. The only statistically significant difference between damaged and undamaged establishments is that damaged establishments have slightly lower measured pre-storm labor productivity. We control for labor productivity in all the reported regressions in the paper. Table 3 Establishment summary statistics: geo-coded establishments, 2004 Variable Obs.a All Undamaged Damaged t-Testb Single-unit firms (%) 10,000 60.3 60.2 60.9 0.727 Establishments in firm 10,000 456.0 451.2 519.1 0.188 Firm agec 10,000 17.9 17.9 17.4 0.153 Establishment employment 10,000 17.9 17.9 18.6 0.788 Establishment aged 10,000 12.9 12.9 12.4 0.115 Productivitye 10,000 4.5 4.5 4.4 0.000 Variable Obs.a All Undamaged Damaged t-Testb Single-unit firms (%) 10,000 60.3 60.2 60.9 0.727 Establishments in firm 10,000 456.0 451.2 519.1 0.188 Firm agec 10,000 17.9 17.9 17.4 0.153 Establishment employment 10,000 17.9 17.9 18.6 0.788 Establishment aged 10,000 12.9 12.9 12.4 0.115 Productivitye 10,000 4.5 4.5 4.4 0.000 aRounded to the nearest hundred. bp-value from t-test for equality of the means. cRight-censored age of 29 used for 3200 observations. dRight-censored age of 29 used for 1000 observations. eLog ratio of revenue to employment in 2002 for establishments that survived to 2004. Table 3 Establishment summary statistics: geo-coded establishments, 2004 Variable Obs.a All Undamaged Damaged t-Testb Single-unit firms (%) 10,000 60.3 60.2 60.9 0.727 Establishments in firm 10,000 456.0 451.2 519.1 0.188 Firm agec 10,000 17.9 17.9 17.4 0.153 Establishment employment 10,000 17.9 17.9 18.6 0.788 Establishment aged 10,000 12.9 12.9 12.4 0.115 Productivitye 10,000 4.5 4.5 4.4 0.000 Variable Obs.a All Undamaged Damaged t-Testb Single-unit firms (%) 10,000 60.3 60.2 60.9 0.727 Establishments in firm 10,000 456.0 451.2 519.1 0.188 Firm agec 10,000 17.9 17.9 17.4 0.153 Establishment employment 10,000 17.9 17.9 18.6 0.788 Establishment aged 10,000 12.9 12.9 12.4 0.115 Productivitye 10,000 4.5 4.5 4.4 0.000 aRounded to the nearest hundred. bp-value from t-test for equality of the means. cRight-censored age of 29 used for 3200 observations. dRight-censored age of 29 used for 1000 observations. eLog ratio of revenue to employment in 2002 for establishments that survived to 2004. We supplement our analysis with data from the 2002 Survey of Business Owners (SBO). The SBO is conducted in Economic Census years and elicits more detailed information about firm operations than what is available in the Economic Census. The questions on the SBO form vary somewhat from year to year. In 2002, a direct measure of capital access comes from the question: ‘During 2002, were any of the following sources used to finance expansion or capital improvements for this business? Mark all that apply.’ The list includes personal or family savings and other assets; credit-card debt; bank loans; government and government-guaranteed loans; and financing from an outside investor. In addition, a check box for ‘no financing needed’ was also provided. Of the approximately 6300 businesses we were able to match to our geocoded Mississippi LBD sample in 2002, about 3500 reported that they needed and obtained some form of financing for capital improvements in the previous year, and nearly 3000 of those survived to 2004 to be included in our exit regressions. We treat the use of credit-card debt to finance expansion or capital improvements as a strong signal of a high marginal cost of financing. Only about 3.5% of establishments in our data belong to firms that reported using a credit card to finance capital improvements or expansions; the majority of these were single-unit firms. Table 4 provides summary statistics for the sample that matches the SBO data. These establishments are quite large, with 29 employees on average, but they are younger than the full sample, 13 years old on average. Establishment and firm characteristics do not differ statistically between establishments that were later damaged and those that were undamaged by Katrina. Table 4 Establishment summary statistics: Survey of Business Owners, 2004 Variable Obs.a All Undamaged Damaged t-Testb Single-unit firms (%) 2400 22.6 22.5 24.4 0.599 Establishments in firm 2400 690.9 675.2 948.1 0.210 Firm agec 2400 22.7 22.7 22.1 0.411 Establishment employment 2400 31.9 31.4 40.1 0.402 Establishment aged 2400 13.6 13.6 13.8 0.803 Productivitye 2400 4.5 4.5 4.4 0.890 Credit-card expansion financing (%)f 2400 4.1 D D 0.585 Variable Obs.a All Undamaged Damaged t-Testb Single-unit firms (%) 2400 22.6 22.5 24.4 0.599 Establishments in firm 2400 690.9 675.2 948.1 0.210 Firm agec 2400 22.7 22.7 22.1 0.411 Establishment employment 2400 31.9 31.4 40.1 0.402 Establishment aged 2400 13.6 13.6 13.8 0.803 Productivitye 2400 4.5 4.5 4.4 0.890 Credit-card expansion financing (%)f 2400 4.1 D D 0.585 aRounded to the nearest hundred. bp-value from t-test for equality of the means. cRight-censored age of 29 used for 1300 observations. dRight-censored age of 29 used for 200 observations. eLog ratio of revenue to employment in 2002 for establishments that survived to 2004. fSome cells are suppressed to comply with disclosure avoidance. Table 4 Establishment summary statistics: Survey of Business Owners, 2004 Variable Obs.a All Undamaged Damaged t-Testb Single-unit firms (%) 2400 22.6 22.5 24.4 0.599 Establishments in firm 2400 690.9 675.2 948.1 0.210 Firm agec 2400 22.7 22.7 22.1 0.411 Establishment employment 2400 31.9 31.4 40.1 0.402 Establishment aged 2400 13.6 13.6 13.8 0.803 Productivitye 2400 4.5 4.5 4.4 0.890 Credit-card expansion financing (%)f 2400 4.1 D D 0.585 Variable Obs.a All Undamaged Damaged t-Testb Single-unit firms (%) 2400 22.6 22.5 24.4 0.599 Establishments in firm 2400 690.9 675.2 948.1 0.210 Firm agec 2400 22.7 22.7 22.1 0.411 Establishment employment 2400 31.9 31.4 40.1 0.402 Establishment aged 2400 13.6 13.6 13.8 0.803 Productivitye 2400 4.5 4.5 4.4 0.890 Credit-card expansion financing (%)f 2400 4.1 D D 0.585 aRounded to the nearest hundred. bp-value from t-test for equality of the means. cRight-censored age of 29 used for 1300 observations. dRight-censored age of 29 used for 200 observations. eLog ratio of revenue to employment in 2002 for establishments that survived to 2004. fSome cells are suppressed to comply with disclosure avoidance. To see how the storm affected establishment counts, we partition the universe of retail, restaurant and hotel establishments in Mississippi with positive payroll and a geocoded address into two subsets: those located in Hancock, Harrison, Jackson and Stone counties (the counties in which FEMA designated most damaged areas) and those located elsewhere in Mississippi. Figure 3(a) shows the log level of the number of restaurants, stores and hotels that had payroll in these two parts of the state from 2005 to 2010, relative to the 2004 level in each region. Unlike the rest of the state, the four counties in which Katrina damage was concentrated experienced a decline in business activity between 2005 and 2006. The 2006 dip was only halfway reversed in 2007 and 2008, after which the economy stagnated. In 2010, while the rest of the state had approximately 5% fewer businesses in the retail, restaurant and hotel sectors relative to pre-storm levels, the affected counties were down approximately 10% from their pre-storm levels.15 Figure 3 View largeDownload slide Log number of Mississippi stores, restaurants and hotels by area damage status, relative to 2004. (a) All Mississippi and (b) damaged counties. Figure 3 View largeDownload slide Log number of Mississippi stores, restaurants and hotels by area damage status, relative to 2004. (a) All Mississippi and (b) damaged counties. Figure 3(b) restricts the analysis to the four damaged counties and partitions those further into areas that were designated by FEMA as: (a) undamaged, (b) mildly damaged or (c) severely damaged. It shows the log level of the number of restaurants, stores and hotels with payroll activity in each of these areas, again relative to the 2004 baseline. This finer partition shows that activity in the undamaged areas of the four counties more than fully recovered by 2007, overtaking the growth rate upstate. By contrast, even areas that had experienced only mild damage had not fully recovered by 2007 and suffered a significant decline in the Great Recession years (2007–2009). Areas classified as severely damaged experienced an even greater decline: the number of active establishments decreased between 2005 and 2006 by approximately 35%, then declined by a further 10% by 2007, and still had not stabilized by the end of our frame in 2010, at which point the decrease exceeded 50%.16 Figure 4(a,b) restricts the analysis to the cohort of establishments that were active in 2004 and also had positive revenue in 2002 (therefore, they were at least 2 years old in 2004).17 The difference between these figures and the previous ones is that we now exclude entrants. The solid line in Figure 4(a) shows establishments in undamaged counties exit at a rate of roughly 9% per year. By 2010, approximately 40% of these establishments had exited. The dashed line shows survival rates for establishments in the four counties with significant damage. The trends are similar prior to the storm, but the count of continuing establishments in the damaged counties drops by 15 percentage points between 2005 and 2006, when the hurricane hit, before settling back to similar trends. Finally, focusing on the damaged counties only, we again find large differences in outcomes by degree of damage. Figure 4(b) shows that undamaged areas experienced survival rates very similar to those upstate, whereas establishments located in storm-hit areas experienced a cumulative exit rate of 80 log points (55%) by 2010. Figure 4 View largeDownload slide Log number of Mississippi stores, restaurants and hotels that existed in 2002 by area damage status, relative to 2004. (a) All Mississippi and (b) damaged counties. Figure 4 View largeDownload slide Log number of Mississippi stores, restaurants and hotels that existed in 2002 by area damage status, relative to 2004. (a) All Mississippi and (b) damaged counties. Finally, restricting the analysis to establishments in areas with severe damage, we partition establishments using two more criteria. First, in Figure 5(a), we partition by firm size, based on the number of establishments the owning firm operated in 2004. We separate single-unit firms from small chains (with up to 100 establishments nationwide) and large chains (with more than 100 establishments). Relative to their 2004 levels, the number of single-unit establishments and establishments in small firms declines by 50–60% by 2006, while the number of establishments in large chains declines by only 20%. By 2010, there are virtually no single-establishment and small-chain stores left in the severely damaged area, but the number of establishments belonging to large chains declines only by a cumulative 40 log points. Second, in Figure 5(b), we partition establishments by their relative position in the 2002 productivity distribution: bottom quartile, interquartile range or top quartile. Here, we see that the exit rate is monotonic in 2002 productivity: by 2006, the number of establishments in the lowest three productivity quartiles declines by 50–60 log points, but in the upper quartile it declines by less than 40 log points. These effects, too, are magnified by 2010. Figure 5 View largeDownload slide Log number of Mississippi stores, restaurants and hotels that existed in 2002 in the severely damaged area by business characteristics, relative to 2004. (a) By firm size and (b) by productivity quartile. Figure 5 View largeDownload slide Log number of Mississippi stores, restaurants and hotels that existed in 2002 in the severely damaged area by business characteristics, relative to 2004. (a) By firm size and (b) by productivity quartile. In the next section, we formalize these findings using regression analysis. 3. Survival and firm characteristics 3.1. Short-run analysis 3.1.1. Firm size and productivity In the absence of frictions in financial markets, we expect to find a socially efficient response to the shock: firms return to operation if and only if the present discounted value of future profits exceeds the lump-sum cost of rebuilding structures, buying new equipment and replenishing inventories. On the other hand, if financial markets are inefficient so that the cost of financing is higher for small firms than large ones, we expect larger firms to return to operation at higher rates than smaller firms. Formally, we estimate a linear probability model of survival, including pre-storm labor productivity as a proxy for future profitability, which we cannot observe directly, as well as a measure of firm size: Survivali=αj(i)N(i)+γn(i)+σln(Firm Size)i+δDamagei+βln(Firm Size)i·Damagei+π·Prodi+φ·Prodi·Damagei+ηln(Firm Age)i+ηTI(Firm Agei=T)+ɛi, (1) where Survival is an indicator that equals 1 if establishment i was in operation in 2006 or returned to operation thereafter, and 0 if it permanently exited the employer universe by 2006. Firm Size is the nationwide count of establishments owned by establishment i’s firm. The sample includes all geocoded Mississippi retail, restaurant and hotel establishments with positive payroll in 2004 and labor productivity estimates from the 2002 Economic Census (the same establishments whose survival is plotted in Figure 4); we use 2004 rather than 2005 data because the shock occurred partway into 2005 and because some observations from 2005 may be missing due to the upheaval caused by the storm. The results are shown in the first column of Table 5. All coefficients are interpreted as marginal survival rates; the baseline is the survival rate of a hypothetical 1-year-old undamaged establishment in a single-establishment firm, whose labor productivity is equal to the average level within its six-digit NAICS sector. Table 5 Difference-in-difference survival regressions: productivity vs. firm size 2004–2006 Variable Baseline Restrictive Expansive 2002–2004 2004–2010 Severe damage −0.3162*** −0.4481*** −0.2924*** 0.1109*** −0.3811*** (0.0413) (0.0407) (0.0397) (0.0394) (0.0817) Mild damage −0.1566 −0.2059* −0.0691 −0.0181 −0.0968 (0.1027) (0.1204) (0.0744) (0.0321) (0.1128) ln(FirmSize) 0.0083*** 0.0067*** 0.0067*** 0.0090*** 0.0202*** (0.0018) (0.0017) (0.0018) (0.0013) (0.0034) ln(FirmAge) 0.0401*** 0.0330*** 0.0362*** 0.0282*** 0.0716*** (0.0057) (0.0070) (0.0052) (0.0065) (0.0092) I(FirmAge=T)b −0.0163 −0.0118 −0.0143 0.0059 −0.0233 (0.0139) (0.0136) (0.0145) (0.0092) (0.0149) Productivity 0.0463*** 0.0398*** 0.0448*** 0.0815*** 0.0619*** (0.0068) (0.0071) (0.0062) (0.0067) (0.0074) Severe damage 0.0169*** 0.0175*** 0.0147*** −0.0030 −0.0015  ×ln(Size) (0.0021) (0.0030) (0.0016) (0.0024) (0.0031) Mild damage 0.0097 0.0058 0.0065 −0.0012 0.0016  ×ln(Size) (0.0086) (0.0088) (0.0090) (0.0031) (0.0189) Severe damage 0.0275*** 0.0520*** 0.0256*** −0.0197*** 0.0456***  ×Prod (0.0061) (0.0052) (0.0046) (0.0071) (0.0113) Mild damage 0.0278 0.0386* 0.0124 0.0011 0.0112  ×Prod (0.0182) (0.0225) (0.0113) (0.0092) (0.0311) County FE×sector FE ✓ ✓ ✓ ✓ ✓ NAICS FE (six-digit) ✓ ✓ ✓ ✓ ✓ Observationsd 10,000 10,000 10,000 10,300 10,000 Percent predicted outside [0,1]c 6% 4% 6% 15% 3% 2004–2006 Variable Baseline Restrictive Expansive 2002–2004 2004–2010 Severe damage −0.3162*** −0.4481*** −0.2924*** 0.1109*** −0.3811*** (0.0413) (0.0407) (0.0397) (0.0394) (0.0817) Mild damage −0.1566 −0.2059* −0.0691 −0.0181 −0.0968 (0.1027) (0.1204) (0.0744) (0.0321) (0.1128) ln(FirmSize) 0.0083*** 0.0067*** 0.0067*** 0.0090*** 0.0202*** (0.0018) (0.0017) (0.0018) (0.0013) (0.0034) ln(FirmAge) 0.0401*** 0.0330*** 0.0362*** 0.0282*** 0.0716*** (0.0057) (0.0070) (0.0052) (0.0065) (0.0092) I(FirmAge=T)b −0.0163 −0.0118 −0.0143 0.0059 −0.0233 (0.0139) (0.0136) (0.0145) (0.0092) (0.0149) Productivity 0.0463*** 0.0398*** 0.0448*** 0.0815*** 0.0619*** (0.0068) (0.0071) (0.0062) (0.0067) (0.0074) Severe damage 0.0169*** 0.0175*** 0.0147*** −0.0030 −0.0015  ×ln(Size) (0.0021) (0.0030) (0.0016) (0.0024) (0.0031) Mild damage 0.0097 0.0058 0.0065 −0.0012 0.0016  ×ln(Size) (0.0086) (0.0088) (0.0090) (0.0031) (0.0189) Severe damage 0.0275*** 0.0520*** 0.0256*** −0.0197*** 0.0456***  ×Prod (0.0061) (0.0052) (0.0046) (0.0071) (0.0113) Mild damage 0.0278 0.0386* 0.0124 0.0011 0.0112  ×Prod (0.0182) (0.0225) (0.0113) (0.0092) (0.0311) County FE×sector FE ✓ ✓ ✓ ✓ ✓ NAICS FE (six-digit) ✓ ✓ ✓ ✓ ✓ Observationsd 10,000 10,000 10,000 10,300 10,000 Percent predicted outside [0,1]c 6% 4% 6% 15% 3% Notes: The sample includes establishments at least 2 years old and with employees in the base year. The LHS variable is an indicator for the establishment surviving from the first to the last year in the range. Robust standard errors in parentheses, clustered by county. aAge of zero occurs only in 2002 sample; later samples are continuers from 2002. bAge of T indicates the firm or one of its original establishments was in operation in 1976. cRounded to the nearest percentage point. dRounded to the nearest hundred. *Significant at 10%; **significant at 5% and ***significant at 1%. Table 5 Difference-in-difference survival regressions: productivity vs. firm size 2004–2006 Variable Baseline Restrictive Expansive 2002–2004 2004–2010 Severe damage −0.3162*** −0.4481*** −0.2924*** 0.1109*** −0.3811*** (0.0413) (0.0407) (0.0397) (0.0394) (0.0817) Mild damage −0.1566 −0.2059* −0.0691 −0.0181 −0.0968 (0.1027) (0.1204) (0.0744) (0.0321) (0.1128) ln(FirmSize) 0.0083*** 0.0067*** 0.0067*** 0.0090*** 0.0202*** (0.0018) (0.0017) (0.0018) (0.0013) (0.0034) ln(FirmAge) 0.0401*** 0.0330*** 0.0362*** 0.0282*** 0.0716*** (0.0057) (0.0070) (0.0052) (0.0065) (0.0092) I(FirmAge=T)b −0.0163 −0.0118 −0.0143 0.0059 −0.0233 (0.0139) (0.0136) (0.0145) (0.0092) (0.0149) Productivity 0.0463*** 0.0398*** 0.0448*** 0.0815*** 0.0619*** (0.0068) (0.0071) (0.0062) (0.0067) (0.0074) Severe damage 0.0169*** 0.0175*** 0.0147*** −0.0030 −0.0015  ×ln(Size) (0.0021) (0.0030) (0.0016) (0.0024) (0.0031) Mild damage 0.0097 0.0058 0.0065 −0.0012 0.0016  ×ln(Size) (0.0086) (0.0088) (0.0090) (0.0031) (0.0189) Severe damage 0.0275*** 0.0520*** 0.0256*** −0.0197*** 0.0456***  ×Prod (0.0061) (0.0052) (0.0046) (0.0071) (0.0113) Mild damage 0.0278 0.0386* 0.0124 0.0011 0.0112  ×Prod (0.0182) (0.0225) (0.0113) (0.0092) (0.0311) County FE×sector FE ✓ ✓ ✓ ✓ ✓ NAICS FE (six-digit) ✓ ✓ ✓ ✓ ✓ Observationsd 10,000 10,000 10,000 10,300 10,000 Percent predicted outside [0,1]c 6% 4% 6% 15% 3% 2004–2006 Variable Baseline Restrictive Expansive 2002–2004 2004–2010 Severe damage −0.3162*** −0.4481*** −0.2924*** 0.1109*** −0.3811*** (0.0413) (0.0407) (0.0397) (0.0394) (0.0817) Mild damage −0.1566 −0.2059* −0.0691 −0.0181 −0.0968 (0.1027) (0.1204) (0.0744) (0.0321) (0.1128) ln(FirmSize) 0.0083*** 0.0067*** 0.0067*** 0.0090*** 0.0202*** (0.0018) (0.0017) (0.0018) (0.0013) (0.0034) ln(FirmAge) 0.0401*** 0.0330*** 0.0362*** 0.0282*** 0.0716*** (0.0057) (0.0070) (0.0052) (0.0065) (0.0092) I(FirmAge=T)b −0.0163 −0.0118 −0.0143 0.0059 −0.0233 (0.0139) (0.0136) (0.0145) (0.0092) (0.0149) Productivity 0.0463*** 0.0398*** 0.0448*** 0.0815*** 0.0619*** (0.0068) (0.0071) (0.0062) (0.0067) (0.0074) Severe damage 0.0169*** 0.0175*** 0.0147*** −0.0030 −0.0015  ×ln(Size) (0.0021) (0.0030) (0.0016) (0.0024) (0.0031) Mild damage 0.0097 0.0058 0.0065 −0.0012 0.0016  ×ln(Size) (0.0086) (0.0088) (0.0090) (0.0031) (0.0189) Severe damage 0.0275*** 0.0520*** 0.0256*** −0.0197*** 0.0456***  ×Prod (0.0061) (0.0052) (0.0046) (0.0071) (0.0113) Mild damage 0.0278 0.0386* 0.0124 0.0011 0.0112  ×Prod (0.0182) (0.0225) (0.0113) (0.0092) (0.0311) County FE×sector FE ✓ ✓ ✓ ✓ ✓ NAICS FE (six-digit) ✓ ✓ ✓ ✓ ✓ Observationsd 10,000 10,000 10,000 10,300 10,000 Percent predicted outside [0,1]c 6% 4% 6% 15% 3% Notes: The sample includes establishments at least 2 years old and with employees in the base year. The LHS variable is an indicator for the establishment surviving from the first to the last year in the range. Robust standard errors in parentheses, clustered by county. aAge of zero occurs only in 2002 sample; later samples are continuers from 2002. bAge of T indicates the firm or one of its original establishments was in operation in 1976. cRounded to the nearest percentage point. dRounded to the nearest hundred. *Significant at 10%; **significant at 5% and ***significant at 1%. The identifying assumption in this regression is that, within the counties affected by Katrina, the precise path of the storm and therefore the damage inflicted was random, i.e., uncorrelated with ɛ. Table 3 in Section 2 provides reassurance that observables are distributed similarly in the treated (damaged) and control (undamaged) samples. The only exception is labor productivity, which is slightly lower on average for establishments located in damaged areas, and for which we control directly. Finally, we assume that county-by-sector (retail, restaurants, and hotels) and six-digit industry-fixed effects (α and γ, respectively) fully capture demand shocks following the storm. (The index j(i) indexes the county of establishment i; N(i) is the sector of establishment i—i.e., retail, restaurant, or hotel—and n(i) is for the six-digit NAICS code of establishment i.) The remaining differences between damaged and undamaged establishments can then be attributed to their differential recovery costs. The error term ε is clustered at the county level. Clustering accounts for the fact that business survival is interdependent across the county. The main short-run effect of the storm is a decrease in survival probability in areas that experienced severe (primarily catastrophic) damage. In some cases, although the business was previously viable, the cost of restoring structures, equipment and inventories cannot be justified by expected future profits. Ceteris paribus, establishments in these areas had a 32 percentage-point lower survival probability between 2004 and 2006 than undamaged establishments in the same counties. The effect of mild damage is smaller, about 16 percentage points and less precisely estimated. Size and productivity both act to diminish the negative effect of storm damage. The positive coefficient on the interaction of productivity and severe damage is consistent with the efficient-market hypothesis: the most productive businesses find it worthwhile to rebuild and return to operation, while less-productive ones rationally exit. The standard deviation of productivity within a six-digit NAICS industry is approximately 0.6, so a one-standard-deviation increase in productivity is associated with a 4.4 percentage point increase in the survival probability of a severely damaged establishment. The same productivity difference increases the survival probability of an undamaged establishment by 2.8 percentage points; in other words, high productivity disproportionately protects damaged businesses. At the same time, a doubling of a firm’s size also reduces the exit probability of severely damaged establishments, by 1.7 percentage points, triple the effect of size for undamaged businesses.18 The control variables all have expected signs. Establishments in older and larger firms are more likely to survive, consistent with selection for better management, access to resources and other correlates of survival. The direct effect of productivity is also positive. The results above assume a linear relationship between damage, log size and log productivity, and survival. We have also estimated nonlinear specifications (not shown). We find that the mitigating effect of firm size on the impact of severe damage starts with firms with more than 10 establishments, and is particularly strong for firms with more than 500 establishments. These are the firms that have the most resources, financial as well as managerial, to cope with disasters of this magnitude. In contrast, we do not find a systematic nonlinear pattern for the ameliorative effect of prior productivity on the impact of storm damage. The second and third columns of Table 5 verify that the main conclusion is not sensitive to the definition of the exit variable. In Column (2) of Table 5 we replace the baseline survival variable with a more restrictive variable, which reclassifies some survivors as exiters, and in Column (3) we use an expansive variable, which reclassifies some exiters as survivors. The main effect of both severe and mild damage changes with these redefinitions of survival—increasing in the former case and decreasing in the latter, consistent with excluding or adding marginal firms. The estimated coefficient on the interaction of damage and productivity also increases in the restrictive specification and decreases in the expansive specification; other interaction terms are stable. It is possible that survival rates of small firms are generally lower in the area damaged by Katrina, not because of Katrina, but because the damaged areas somehow favor large chains. If that were the case, survival rates in these areas would also have been lower for small businesses prior to the storm. To test this possibility we re-estimate Equation (1) but, on the left-hand side, we replace exit between 2004 and 2006 with exit between 2002 and 2004. This regression functions as a falsification exercise.19 Results from this specification using the baseline exit variable are shown in Column (4) of Table 5. We find higher survival rates on average in the damaged area prior to Katrina’s landfall, and lower survival rates for less-productive establishments in these areas, but we see no evidence that smaller firms fare worse in these areas than large firms. We have checked the robustness of these results in several unreported regressions. Changing the sample of controls to include only counties immediately adjacent to the damaged counties (Pearl River, Forrest, Perry and George), or to omit those same counties, or to include only the four counties with some severely damaged areas, does not change the results in any meaningful way, although standard errors on some coefficients increase. Similarly, adding establishment age and employment to the regressions has no impact on the qualitative patterns of coefficients. Finally, we have estimated the regression using a probit model; the results are again qualitatively unchanged. To alleviate the concern that differential errors in the exit rates of firms by firm size are driving the results, we also estimate the regression on a sample that includes all single-unit firms as well as establishments belonging to multi-unit firms with at least 250 employees in total. These firms are surveyed annually in the COS and their exit rates are likely to be measured most accurately. Relative to Column (1) of Table 5, these results (not shown, but available upon request) show coefficients that are larger in absolute terms, and equally strong statistically, on all damage variables and interaction terms. 3.1.2. Access to credit Credit constraints are not the only explanation for small firms’ greater sensitivity to the cost shock in the short run. For example, small-business owners may be more risk averse than larger businesses and may have responded more cautiously to uncertainty about the local economy’s rebounding.20,21 The decision to rebuild and reopen may have included considerations other than the success of a particular business establishment, such as public relations or media attention (particularly for large businesses) or attachment to the area (for locally owned businesses). In addition, the larger labor pool available to large firms may provide access to specialized managerial skill and other resources unavailable to their smaller counterparts. Although we cannot directly test for these and other alternative explanations, we can test for the importance of credit constraints. Most Census data sets do not contain any direct information about business balance sheets, banking and credit relationships or access to financial markets. The one exception is the SBO. We link the SBO data to our sample of establishments in Mississippi to explore the role of financing on exit in the aftermath of the storm. We use the 2002 SBO, which covers the cohort of establishments in our data, to explore the extent to which access to credit can explain the differential effects of severe damage by firm size. The SBO does not collect balance-sheet information or any direct measure of assets, but it does include a question about funding sources for capital improvements and expansions undertaken during 2002. We omit from our analysis businesses that report they did not ‘need’ any such funding, not being able to distinguish whether they did not want to make capital improvements or they made no capital improvements because funding was unavailable or too costly. We are agnostic about most sources of financing due to problems with interpretation. For example, a business may have obtained a government or government-guaranteed loan because it is not sufficiently viable to obtain a private loan, or because its owner is savvy and able to exploit any available resources; the former implies a negative relationship between such loans and survival whereas the latter may lead to a positive relationship. Likewise, using personal savings may indicate that a business cannot attract loans or outside investors, or it may be a signal that the owner has the resources to invest in her business and the confidence that it will do well. One source of financing, credit-card debt, stands out as particularly useful for our purposes. Credit cards charge high interest rates. Many small businesses use credit cards for convenience to pay some bills, for short-term cash flow or convenience, but using credit-card debt to finance expansion or capital improvements is a likely signal of a lack of other viable sources of funds.22,23 Whereas a business owner is unlikely to incur such an unsecured expense without reasonable expectation that the investment will justify itself, a surprise on the order of the demolition of the business by storm surge could prove particularly fatal to a credit-card-financed business. We estimate a model with interactions of damage with past use of credit cards as follows: Survivali=αj(i)+γn(i)+σln(Firm Size)i+δDamagei+βln(Firm Size)i·Damagei+λCreditCardi+μCreditCardi·Damagei+π·Prodi+φ·Prodi·Damagei+ηln(Firm Age)i+ηTI(FirmAgei=T)+ɛi. (2) The results are reported in Table 6. The first column reports the results of a regression with the SBO sample but without the credit-card variables. As with the full sample in the previous section, we find size and productivity predict survival in the undamaged areas. The estimated effects are of similar magnitude. The coefficient on the interaction of productivity and severe damage is positive and large as before, although not significant. The interaction between firm size and severe damage is statistically significant, though its magnitude is only about a quarter of the size in Table 5 (0.005 vs. 0.021). This may be a result of the fact that the SBO sample has fewer single-unit firms relative to the sample in the earlier regressions.24 In the second column, we add the credit-card variable and its interaction with the damage vector. The third and fourth columns repeat this analysis using the restrictive and expansive survival variables, respectively. Across specifications, the coefficient on the credit-card variable is small and statistically insignificant. On its face, this may seem puzzling, given our contention that credit-card debt is a strong signal of a financially weak business. However, the credit-card financing question refers to 2002, and our sample conditions on the business having survived to 2004. Credit-card-reliant businesses that survived 2 years may be a selected sample, no weaker than its counterparts which relied on other sources of financing, at least absent any additional shocks. The SBO questionnaire does not allow us to distinguish between businesses that went into significant credit-card debt and those that used their cards more on a more limited basis. In addition, the SBO does not ask businesses to rank the relative or absolutely importance of their various financing sources. Most businesses that used credit cards relied on other sources of financing as well: approximately half also reported using personal savings, and others also received bank loans. Although credit-card usage in 2002 does not predict survival for undamaged businesses, by and large, businesses that relied on this expensive form of financing and were hit by severe storm damage were unable to recover from this shock. The coefficient on the interaction of credit-card usage and damage is very large in absolute value and statistically significant: conditional on severe storm damage, an establishment whose owner reports having used a credit card for expansion or capital improvements in 2002 was 73 percentage points less likely to survive from 2004 to 2006 than one whose owner did not use a credit card for these purposes. In other words, although businesses that had previously signaled a high marginal cost of financial capital were able to continue operating as long as no major cost shocks arose, they could not adjust to a major shock. Interestingly, the direct effect of size is unchanged in this specification, but the interaction effect of size and damage disappears entirely, becoming negative. The addition of the credit-card variable appears to lend precision to some of the estimated coefficients. 3.2. Long-run analysis We return to the full sample of stores, restaurants and hotels to estimate the probability that a business that did not exit between 2004 and 2006, nevertheless exited between 2004 and 2010. As before, we contrast damaged and undamaged establishments. We estimate the same model as in Equation (1), but replace the LHS variable with an indicator for survival between 2004 and 2010. The results are reported in the last column of Table 5. By late 2007, demand was suppressed by reduced tourism, a consequence of the Great Recession. Businesses in severely damaged areas, however, fared even worse than those in undamaged areas: the coefficient on severe damage is six and half points larger, in absolute terms, than in Column (1).25 This long-run effect of the storm is striking in part because of the massive federal, state, local and private funds that poured into the area for rebuilding efforts. Moreover, the interaction effects now tell a different story. On the one hand, the interaction of productivity and severe damage continues to be positive and even increases in magnitude, implying that the most productive stores, restaurants and hotels continue to be partially shielded from the effects of the storm. But the coefficient on the interaction of size and damage attenuates to zero: controlling for productivity, larger firms located in severely damaged areas were no more likely to survive to 2010 than smaller firms. Put differently, larger firms, which survived the initial shock, were more likely than small firms to exit between 2006 and 2010. These establishments experienced severe damage in 2005, invested considerable resources rebuilding and were active again by 2006. Why did these larger businesses exit later? Common shocks, such as the Great Recession, which started in 2007, alone cannot explain why businesses in the damaged area experienced different survival rates than businesses in other areas, nor why these rates differ for small and large firms. Delayed exit by large businesses is consistent with differential access to credit in the aftermath of the storm, which induced a stronger selection of the smaller businesses. Under this scenario only ‘superstar’ small businesses were able to return to operation. Large firms, having had relatively easier access to internal resources, collateral or established banking relations, experienced less selection based on expected future performance. If this is the case, surviving establishments belonging to small firms must be more profitable, on average, than surviving unconstrained establishments.26 Consequently, they are also less vulnerable to a continued shock. However, this explanation implies that overall survival rates from 2006 to 2010 should have been higher in the damaged area, where businesses were put through a sort of ‘stress test’ in 2005, than in other areas in the same counties barring any additional shock. This is not the case. What secondary shock could have affected damaged areas relative to undamaged areas leading establishments of large firms to disproportionally exit? The most likely explanation is that the impact of the storm compounded over time, because of endogenous local demand shock induced by the destruction and consequent closure of many neighboring businesses. The failure of so many businesses may have had an aggregate effect by negatively impacting the local economy. It may also have had a localized indirect effect by reducing customer traffic to their surviving neighbors, in turn increasing their failure rate. These effects might have been difficult to foresee by large businesses that poured in resources to rebuilt and return to operations. This would have been particularly problematic for the less-productive large firms. This explanation is consistent with evidence on the importance of externalities in shopping malls (see, e.g., Pashigian and Gould, 1998; Gould et al., 2005) and with evidence of agglomeration effects in an urban setting (Zhu et al., 2011).27 Note in this regard that the Great Recession may have had a disproportionate effect in the damaged areas by also suppressing entry rates, which otherwise would be expected to offset exits. (See Figure 3 for the trends in the overall count of establishments by area damage status.) 4. Growth Examination of survival patterns post Katrina suggests establishment exits were concentrated among low-productivity and financially constrained firms. Our results so far suggest small firms were less likely to survive unless they were highly productive. A natural question to ask then is whether, conditional on survival, initial size and productivity predict the impact of the storm on business’s growth rate. To answer this question, we use all observations that survived from 2004 to 2006 based on the baseline survival definition, and estimate Growthi=αj(i)N(i)+γn(i)+σln(Firm Size)i+δDamagei+βln(Firm Size)i·Damagei+π·Prodi+φ·Prodi·Damagei+ηln(Firm Age)i+ηTI(FirmAgei=T)+ɛi, (3) where Growth is defined as the log change in establishment employment from 2004 to 2006, and all the other variables are as defined above. Results are shown in the first column of Table 7. Table 7 Difference-in-difference growth regressions: productivity vs. firm size 2004–2006 2002–2004 2004–2010 Severe damage −0.8802*** 0.0838 −0.3554*** (0.1030) (0.1378) (0.1244) Mild damage −0.0028 0.0109 0.0378 (0.2456) (0.0495) (0.2144) ln(FirmSize) 0.0127*** −0.0039 0.0189*** (0.0031) (0.0027) (0.0037) ln(FirmAge) −0.0301** −0.0200** −0.0230 (0.0116) (0.0090) (0.0161) I(FirmAge=T)b 0.0404* 0.0098 −0.0111 (0.0224) (0.0153) (0.0197) Productivity 0.0378*** 0.1784*** 0.1141*** (0.0095) (0.0112) (0.0153) Severe damage 0.0132*** −0.0196*** −0.0029  ×ln(Size) (0.0041) (0.0029) (0.0041) Mild damage −0.0036 −0.0239*** 0.0189***  ×ln(Size) (0.0210) (0.0025) (0.0065) Severe damage 0.1184*** 0.0110 0.0633**  ×Prod (0.0122) (0.0318) (0.0312) Mild damage −0.0128 0.0195 −0.0243  ×Prod (0.0454) (0.0158) (0.0449) County FE×sector FE ✓ ✓ ✓ NAICS FE (six-digit) ✓ ✓ ✓ Observationsc 8300 9100 6600 2004–2006 2002–2004 2004–2010 Severe damage −0.8802*** 0.0838 −0.3554*** (0.1030) (0.1378) (0.1244) Mild damage −0.0028 0.0109 0.0378 (0.2456) (0.0495) (0.2144) ln(FirmSize) 0.0127*** −0.0039 0.0189*** (0.0031) (0.0027) (0.0037) ln(FirmAge) −0.0301** −0.0200** −0.0230 (0.0116) (0.0090) (0.0161) I(FirmAge=T)b 0.0404* 0.0098 −0.0111 (0.0224) (0.0153) (0.0197) Productivity 0.0378*** 0.1784*** 0.1141*** (0.0095) (0.0112) (0.0153) Severe damage 0.0132*** −0.0196*** −0.0029  ×ln(Size) (0.0041) (0.0029) (0.0041) Mild damage −0.0036 −0.0239*** 0.0189***  ×ln(Size) (0.0210) (0.0025) (0.0065) Severe damage 0.1184*** 0.0110 0.0633**  ×Prod (0.0122) (0.0318) (0.0312) Mild damage −0.0128 0.0195 −0.0243  ×Prod (0.0454) (0.0158) (0.0449) County FE×sector FE ✓ ✓ ✓ NAICS FE (six-digit) ✓ ✓ ✓ Observationsc 8300 9100 6600 Notes: The sample includes establishments at least 2 years old in the base year, with employees in both the base and end years. The LHS variable is the establishment’s employment growth rate between the 2 years. Robust standard errors in parentheses, clustered by county. aAge of zero occurs only in 2002 sample; later samples are continuers from 2002. bAge of T indicates the firm or one of its original establishments was in operation in 1976. cRounded to the nearest hundred. *Significant at 10%; **significant at 5% and ***significant at 1%. Table 7 Difference-in-difference growth regressions: productivity vs. firm size 2004–2006 2002–2004 2004–2010 Severe damage −0.8802*** 0.0838 −0.3554*** (0.1030) (0.1378) (0.1244) Mild damage −0.0028 0.0109 0.0378 (0.2456) (0.0495) (0.2144) ln(FirmSize) 0.0127*** −0.0039 0.0189*** (0.0031) (0.0027) (0.0037) ln(FirmAge) −0.0301** −0.0200** −0.0230 (0.0116) (0.0090) (0.0161) I(FirmAge=T)b 0.0404* 0.0098 −0.0111 (0.0224) (0.0153) (0.0197) Productivity 0.0378*** 0.1784*** 0.1141*** (0.0095) (0.0112) (0.0153) Severe damage 0.0132*** −0.0196*** −0.0029  ×ln(Size) (0.0041) (0.0029) (0.0041) Mild damage −0.0036 −0.0239*** 0.0189***  ×ln(Size) (0.0210) (0.0025) (0.0065) Severe damage 0.1184*** 0.0110 0.0633**  ×Prod (0.0122) (0.0318) (0.0312) Mild damage −0.0128 0.0195 −0.0243  ×Prod (0.0454) (0.0158) (0.0449) County FE×sector FE ✓ ✓ ✓ NAICS FE (six-digit) ✓ ✓ ✓ Observationsc 8300 9100 6600 2004–2006 2002–2004 2004–2010 Severe damage −0.8802*** 0.0838 −0.3554*** (0.1030) (0.1378) (0.1244) Mild damage −0.0028 0.0109 0.0378 (0.2456) (0.0495) (0.2144) ln(FirmSize) 0.0127*** −0.0039 0.0189*** (0.0031) (0.0027) (0.0037) ln(FirmAge) −0.0301** −0.0200** −0.0230 (0.0116) (0.0090) (0.0161) I(FirmAge=T)b 0.0404* 0.0098 −0.0111 (0.0224) (0.0153) (0.0197) Productivity 0.0378*** 0.1784*** 0.1141*** (0.0095) (0.0112) (0.0153) Severe damage 0.0132*** −0.0196*** −0.0029  ×ln(Size) (0.0041) (0.0029) (0.0041) Mild damage −0.0036 −0.0239*** 0.0189***  ×ln(Size) (0.0210) (0.0025) (0.0065) Severe damage 0.1184*** 0.0110 0.0633**  ×Prod (0.0122) (0.0318) (0.0312) Mild damage −0.0128 0.0195 −0.0243  ×Prod (0.0454) (0.0158) (0.0449) County FE×sector FE ✓ ✓ ✓ NAICS FE (six-digit) ✓ ✓ ✓ Observationsc 8300 9100 6600 Notes: The sample includes establishments at least 2 years old in the base year, with employees in both the base and end years. The LHS variable is the establishment’s employment growth rate between the 2 years. Robust standard errors in parentheses, clustered by county. aAge of zero occurs only in 2002 sample; later samples are continuers from 2002. bAge of T indicates the firm or one of its original establishments was in operation in 1976. cRounded to the nearest hundred. *Significant at 10%; **significant at 5% and ***significant at 1%. Our first finding is that establishments that were severely damaged have much lower growth rates from 2004 to 2006, even conditional on survival. Although the sign is not surprising, the magnitude is notable: employment growth at severely damaged establishments is 88% log points lower than in undamaged establishments. The difference is likely accounted for by job loss during the rebuilding phase as well as diverted demand toward undamaged businesses, both of which would have suppressed employment.28 Consistent with our survival results, establishments located in areas subject to mild damage grew at rates similar to undamaged establishments. In addition, ceteris paribus, establishments the belonged to larger firms and establishments with higher 2002 productivity grew at higher rates from 2004 to 2006. More interesting from our perspective are the interaction effects. Our prior is that more-productive firms are able to disproportionally take advantage of the new conditions and grow faster. Our findings provide support for this interpretation. We find that establishments subject to severe damage but belonging to larger chains, and those with higher prior productivity, experienced a lesser negative shock to growth than similarly damaged establishments that belonged to smaller firms and had lower prior productivity. Our estimates imply that a severely damaged establishment belonging to a 1000-establishment firm grew on average by 7.8 points more than a single-establishment firm subject to the same damage level between 2004 and 2006; in the undamaged areas, the difference was only 3.8. In other words, the effect of size on growth is doubled for establishments that experienced severe damage. This is true even after controlling for productivity. We interpret this as evidence that establishments of large firms had an advantage rebuilding their capital and operations. Even more strikingly, a one-standard-deviation increase in productivity increased a severely damaged business’s 2-year employment growth by 9.4 points; in the undamaged areas, the difference was only 2.3 points (a 4-fold effect). The next two columns show results, from, first, 2002–2004 regressions (a falsification test) and, second, 2004–2010 regressions (long-run analysis). In the 2002–2004 period, we find that productivity was a strong predictor of growth. The interaction of chain size and future damage is negative, suggesting that large chains may have been relatively ill suited to local demand conditions in the area that later experienced the most severe storm damage. If we are willing to assume that the 2002–2004 results would have been replicated in 2004–2006 but for the storm, the falsification exercise can also be used to provide a triple-difference interpretation of our main results. In this interpretation, the protective impact of chain size on damaged businesses is even greater. The last column shows long-run results up to 5 years after the storm. Results are conditional on surviving to 2010, and as such should be interpreted as a selected sample. We find that initial productivity and size are strong predictors of survival for all businesses. Looking at the damaged areas specifically, however, we find that size no longer provides an advantage. By contrast, productivity does confer an advantage. These results mirror our findings on exits from Table 5. Although size confers an initial advantage, even conditional on productivity, in the long run this advantage disappears and the most productive establishments, regardless of size, are the ones that are able to take advantage of opportunities and grow. 5. Concluding remarks Our analysis uses Hurricane Katrina as a natural experiment to examine the impact of an external cost shock due to capital destruction on business activity.29 Consistent with a ‘cleansing’ hypothesis, we find that less-productive establishments exited disproportionately following the initial shock. But even after controlling for productivity, we find that establishments belonging to small firms were disproportionately affected. Business owners who reported relying, at least in part, on credit-card debt to finance capital projects were particularly vulnerable to exit following severe damage. Focusing on survivors, we find that large firms had an advantage rebuilding their operations quickly after the storm, as did the more productive firms of all sizes. Five years after the storm, size no longer confers an advantage on the initial cohort for either survival or growth. Although large firms may have been able to disproportionally survive the storm and rebuild initially, they did not perform better over time. The short-run results suggest that binding constraints other than those captured by an establishment’s productivity serve as a selection mechanism for small businesses following a cost shock. Small businesses survive the initial shock at much lower rates than large ones. They are also not as quick to rebuild their operations. As businesses age and grow, this selection mechanism diminishes in importance and eventually disappears, consistent with the idea that access to finance gives an initial advantage at entry but this advantage dissipates over time to the benefit of productivity.30 But small businesses that face a major cost shock early in their development cannot reach this later phase. Our data do not allow us to identify the precise channel by which financial constraints impact the survival of firms. For example, large firms may have an advantage over small ones due to access to internal capital markets (e.g., Fazzari et al., 1988; Lamont, 1997; Campello et al., 2010); this explanation implies some sort of failure in external capital markets, for example due to an information asymmetry. Alternatively, small and credit-constrained firms may be under-insured compared with larger firms, for example because they are less risk averse, uninformed about their risks, face higher insurance premia or are too illiquid to pay regular insurance premia.31 Although we cannot explicitly test for risk aversion, because small firms are less diversified than large ones they are unlikely to be less risk averse; all the other explanations imply some sort of financial friction. Since the channel determines the policy implications of our findings, more work is needed to determine the relative importance of these and other possible mechanisms. Our findings have important implications for understanding the impact of large cost shocks. First, cost shocks can take a long time to dissipate. In the case of Hurricane Katrina, the recovery took years despite a major effort by the federal government and others to aid in the recovery. Second, there is wide heterogeneity in business response to cost shocks. Small, credit-constrained firms are disproportionately affected and exit immediately. Some of these small firms are highly productive and may have otherwise survived. Third, in the case of Katrina, exiting businesses made up a large share of the local economy, and represent a loss not only of structures but also of entrepreneurial and social capital, which have been difficult to replace and rebuild. Fourth, the impact on the local economy from the high level of small-firm exit may itself have induced a second wave of business exits, further impeding and delaying recovery. Our long-run analysis shows that the first wave of small-business exit was followed by a second wave of large-firm establishment exit, particularly of less-productive large firms. We interpret this as evidence that market mechanisms eventually prevail. These results are particularly relevant when developing strategies for a prompt recovery in the aftermath of an economic shock whether caused by a natural disaster and likely otherwise. The Great Recession hit in 2007, less than 2 years after Katrina; by 2010 the region had yet to return to pre-2005 levels of economic activity. Measures that can directly address these additional risk factors may be able to speed up recovery when the next natural disaster hits. Figure B1 View largeDownload slide Log number of Mississippi stores, restaurants, and hotels by area damage status, relative to 2002. (a) All Mississippi and (b) damaged counties. Figure B1 View largeDownload slide Log number of Mississippi stores, restaurants, and hotels by area damage status, relative to 2002. (a) All Mississippi and (b) damaged counties. Figure B2 View largeDownload slide Log number of Mississippi stores, restaurants, and hotels that existed in 2002 by area damage status, relative to 2002. (a) All Mississippi and (b) damaged counties. Figure B2 View largeDownload slide Log number of Mississippi stores, restaurants, and hotels that existed in 2002 by area damage status, relative to 2002. (a) All Mississippi and (b) damaged counties. Funding Javier Miranda received no direct financial support from any organization for this paper. Emek Basker was partially funded from an ASA/NSF/Census Bureau Fellowship. The funding agencies are the American Statistical Association, the National Science Foundation, and the U.S. Census Bureau. There is no grant number. The fellowship is describe here: https://www.census.gov/srd/www/fellweb.html. Footnotes 1 The debate on whether shocks lead to productive cleansing or counterproductive destruction can be traced back to Schumpeter’s creative destruction hypothesis (Schumpeter, 1939, 1942). 2 For example, Mian and Sufi (2014a, 2014b) argue that a decline in consumer demand, not a decline in credit supply, was responsible for unemployment during the Great Recession. 3 More detail on Katrina’s impact on the Mississippi coast is available in Appendix A. 4 For more information on the LBD, see Jarmin and Miranda (2002). 5 This is an extremely fine classification. For example, among car dealerships this classification distinguishes between new- and used-car dealerships and between both of those and motorcycle dealerships; in the home-furnishings sector, it distinguishes between stores specializing in floor coverings, window treatments, and other home furnishings; and in the apparel sector, it distinguishes between men’s-, women’s-, children’s-, and family-clothing stores. 6 These business establishments correspond to NAICS 44-45 and 721-722. We exclude from the analysis non-store retailers such as catalog companies and vending-machine operators, NAICS 454, as well as caterers and mobile food-service providers, NAICS 72232 and 72233. 7 In 2004, according to published numbers from County Business Patterns, the four damaged counties, combined, had 2362 retail and accommodation establishments but only 247 manufacturing establishments. 8 The 90% threshold on payroll reduction is arbitrary. We have checked the robustness of our results using various alternative thresholds, as low as 50%, and continue to find qualitatively similar results. 9 The ILBD data allow us to relax this definition in the event that a firm continues to earn revenue in 2006. 10 This threshold was increased to 500 in more recent years. 11 Establishment revenue is not available in annual data sets, but only in the quinquennial (5-year) Economic Censuses. Employment is measured as of the week of 12 March 2002. We drop the top and bottom 1% of our productivity measures to remove the influence of outliers. Our ratio measure is also used in Foster et al. (2002) and Doms et al. (2004). Basker (2012) uses the ratio of revenue to payroll as an alternative measure of productivity. See Foster et al. (2002), Haskel and Sadun (2009), and Betancourt (2005) for further discussion. 12 All establishment counts throughout the paper are rounded to the nearest hundred. This is done to ensure that no confidential information is disclosed in the event that revisions require us to change the sample in small ways. Table 1 County summary statistics, 2004 State County Estabsa Geo-codeda Severe damage Mild damage MS Hancock 200 200 10.7% 68.0% MS Harrison 1000 900 35.2% 16.9% MS Jackson 500 400 6.9% 18.1% MS Stone 100 <100 0.0% MS Rest of state 10,600 8500 0.3% Totalb 12,300 10,000 3.5% 3.5% State County Estabsa Geo-codeda Severe damage Mild damage MS Hancock 200 200 10.7% 68.0% MS Harrison 1000 900 35.2% 16.9% MS Jackson 500 400 6.9% 18.1% MS Stone 100 <100 0.0% MS Rest of state 10,600 8500 0.3% Totalb 12,300 10,000 3.5% 3.5% Damage percentages are of geo-coded establishments. Blank cells indicate fewer than 10 establishments in damage zone. aEstablishment counts represent the retail, restaurant and hotel sectors, rounded to the nearest hundred. bMay not match sum due to rounding. Table 1 County summary statistics, 2004 State County Estabsa Geo-codeda Severe damage Mild damage MS Hancock 200 200 10.7% 68.0% MS Harrison 1000 900 35.2% 16.9% MS Jackson 500 400 6.9% 18.1% MS Stone 100 <100 0.0% MS Rest of state 10,600 8500 0.3% Totalb 12,300 10,000 3.5% 3.5% State County Estabsa Geo-codeda Severe damage Mild damage MS Hancock 200 200 10.7% 68.0% MS Harrison 1000 900 35.2% 16.9% MS Jackson 500 400 6.9% 18.1% MS Stone 100 <100 0.0% MS Rest of state 10,600 8500 0.3% Totalb 12,300 10,000 3.5% 3.5% Damage percentages are of geo-coded establishments. Blank cells indicate fewer than 10 establishments in damage zone. aEstablishment counts represent the retail, restaurant and hotel sectors, rounded to the nearest hundred. bMay not match sum due to rounding. 13 FEMA’s damage classification defines damage categories as follows. ‘Limited Damage: Generally superficial damage to solid structures (e.g., loss of tiles or roof shingles); some mobile homes and light structures are damaged or displaced. Moderate Damage: Solid structures sustain exterior damage (e.g., missing roofs or roof segments); some mobile homes and light structures are destroyed, many are damaged or displaced. Extensive Damage: Some solid structures are destroyed; most sustain exterior and interior damage (roofs missing, interior walls exposed); most mobile homes and light structures are destroyed. Catastrophic Damage: Most solid and all light or mobile home structures destroyed.’ 14 These dots were ‘jittered’ in compliance with Census Bureau disclosure procedures to prevent identification of particular establishments. 15 These figures are reproduced in the Appendix using 2002 as the base year. 16 These figures are reproduced in the Appendix using 2002 as the base year. 17 We need them to have revenue in 2002 so we can control for productivity later on. 18 An alternative to interacting firm size and damage is to interact firm age and damage, or include both interactions. Despite recent evidence that firm age may be a better indicator of a firm’s ability to withstand a serious shock (Haltiwanger et al., 2013), we prefer using firm size in this setting with a limited sample size for two reasons. First, firm size has a technical advantage over firm age, in that it is never censored and can take on any integer value. Firm age is right-censored for about a third of the establishments in our sample, dramatically limiting the explanatory variable of the continuous variable ln(Firm Age). Second, on a conceptual level, since the shock we consider here is destruction of capital stock, the total number of establishments the firm operates provides a measure of the fraction of the capital stock actually destroyed; all single-establishment firms, whether young or old, experienced a 100% capital-stock destruction if Katrina’s winds and storm surge destroyed their one establishment. 19 To maintain the same age distribution of firms in the 2002 sample, we drop firms with ages 0 or 1 in 2002. 20 In this context, it is interesting to note that Dessaint and Matray (2013), using data from large publicly traded firms, find evidence that managers tend to over-react to hurricane risks. 21 As noted in Appendix A, out-migration was very limited. 22 We are not aware of recent survey data on credit-card usage by small businesses. In the 1993 and 1998 waves of the National Survey of Small Business Finances, respectively, 29% and 34% of business owners reported using a business credit card; over 40% reported using a personal credit card for business expenses (Blanchflower et al., 2003). However, Berger and Udell (1998) report that, after accounting for paid bills, less than 1% of small-business debt in the 1993 wave came from credit-card debt. This finding is consistent with small businesses using credit cards as a payment method but, for the most part, not relying on them for long-term financing. 23 One might be concerned that use of credit card for capital improvements and expansion in the SBO is not a signal of credit constraints. For example, a restaurant that buys a new oven using a credit card could check the box indicating it used a credit card to finance a capital improvement. If this is the case it will cause attenuation bias in our estimated coefficients. 24 Only 30% of SBO establishments, compared with 60% of establishments in the full LBD sample, are single-units. As our previous results show, larger firms were less affected by severe damage in the short run. Table 6 Difference-in-difference survival regressions: auxiliary analysis, 2004–2006 Baseline Baseline Restrictive Expansive Severe damage −0.2923*** −0.2753*** −0.3838*** −0.2967*** (0.0595) (0.0542) (0.1132) (0.0466) Mild damage −0.1309 −0.1006 −0.2307*** −0.1038 (0.1776) (0.1039) (0.0557) (0.1031) ln(FirmSize) 0.0079* 0.0078* 0.0087** 0.0078* (0.0044) (0.0041) (0.0040) (0.0042) ln(FirmAge) 0.0296* 0.0288 −0.0032 0.0279 (0.0176) (0.0174) (0.0193) (0.0169) I(FirmAge=T)a −0.0087 −0.0075 0.0199 −0.0069 (0.0277) (0.0281) (0.0306) (0.0278) Productivity 0.0218 0.0232* 0.0309** 0.0184 (0.0142) (0.0135) (0.0144) (0.0127) Credit cardb −0.0116 0.0015 −0.0194 (0.0366) (0.0365) (0.0378) Severe damage 0.0046 −0.0099** −0.0125 −0.0084*  ×ln(Size) (0.0032) (0.0041) (0.0121) (0.0045) Mild damage −0.0103 −0.0123 −0.0151 −0.0123  ×ln(Size) (0.0074) (0.0077) (0.0137) (0.0078) Severe damage 0.0248 0.0426** 0.0640*** 0.0449***  ×Prod (0.0196) (0.0173) (0.0231) (0.0146) Mild damage 0.0580 0.0544** 0.0702*** 0.0552**  ×Prod (0.0356) (0.0263) (0.0130) (0.0263) Severe damage −0.7316*** −0.7049*** −0.7124***  ×CreditCard (0.0536) (0.0515) (0.0515) Mild damage D D D  ×CreditCarde County FE ✓ ✓ ✓ ✓ NAICS FE (six-digit) ✓ ✓ ✓ ✓ Observationsd 2400 2400 2400 2400 Percent predicted outside [0,1]c 10% 10% 6% 10% Baseline Baseline Restrictive Expansive Severe damage −0.2923*** −0.2753*** −0.3838*** −0.2967*** (0.0595) (0.0542) (0.1132) (0.0466) Mild damage −0.1309 −0.1006 −0.2307*** −0.1038 (0.1776) (0.1039) (0.0557) (0.1031) ln(FirmSize) 0.0079* 0.0078* 0.0087** 0.0078* (0.0044) (0.0041) (0.0040) (0.0042) ln(FirmAge) 0.0296* 0.0288 −0.0032 0.0279 (0.0176) (0.0174) (0.0193) (0.0169) I(FirmAge=T)a −0.0087 −0.0075 0.0199 −0.0069 (0.0277) (0.0281) (0.0306) (0.0278) Productivity 0.0218 0.0232* 0.0309** 0.0184 (0.0142) (0.0135) (0.0144) (0.0127) Credit cardb −0.0116 0.0015 −0.0194 (0.0366) (0.0365) (0.0378) Severe damage 0.0046 −0.0099** −0.0125 −0.0084*  ×ln(Size) (0.0032) (0.0041) (0.0121) (0.0045) Mild damage −0.0103 −0.0123 −0.0151 −0.0123  ×ln(Size) (0.0074) (0.0077) (0.0137) (0.0078) Severe damage 0.0248 0.0426** 0.0640*** 0.0449***  ×Prod (0.0196) (0.0173) (0.0231) (0.0146) Mild damage 0.0580 0.0544** 0.0702*** 0.0552**  ×Prod (0.0356) (0.0263) (0.0130) (0.0263) Severe damage −0.7316*** −0.7049*** −0.7124***  ×CreditCard (0.0536) (0.0515) (0.0515) Mild damage D D D  ×CreditCarde County FE ✓ ✓ ✓ ✓ NAICS FE (six-digit) ✓ ✓ ✓ ✓ Observationsd 2400 2400 2400 2400 Percent predicted outside [0,1]c 10% 10% 6% 10% Notes: The sample includes establishments at least 2 years old in the base year, in the SBO, with employees in 2004. The LHS variable is an indicator for the establishment surviving from 2004 to 2006. Robust standard errors in parentheses, clustered by county. aAge of T indicates the firm or one of its original establishments was in operation in 1976. bUsed a credit card to finance capital improvements or expansion in 2002. cRounded to the nearest percentage point. dRounded to the nearest hundred. eCoefficients on this variable are statistically insignificant and suppressed to comply with disclosure avoidance. *Significant at 10%; **significant at 5% and ***significant at 1%. Table 6 Difference-in-difference survival regressions: auxiliary analysis, 2004–2006 Baseline Baseline Restrictive Expansive Severe damage −0.2923*** −0.2753*** −0.3838*** −0.2967*** (0.0595) (0.0542) (0.1132) (0.0466) Mild damage −0.1309 −0.1006 −0.2307*** −0.1038 (0.1776) (0.1039) (0.0557) (0.1031) ln(FirmSize) 0.0079* 0.0078* 0.0087** 0.0078* (0.0044) (0.0041) (0.0040) (0.0042) ln(FirmAge) 0.0296* 0.0288 −0.0032 0.0279 (0.0176) (0.0174) (0.0193) (0.0169) I(FirmAge=T)a −0.0087 −0.0075 0.0199 −0.0069 (0.0277) (0.0281) (0.0306) (0.0278) Productivity 0.0218 0.0232* 0.0309** 0.0184 (0.0142) (0.0135) (0.0144) (0.0127) Credit cardb −0.0116 0.0015 −0.0194 (0.0366) (0.0365) (0.0378) Severe damage 0.0046 −0.0099** −0.0125 −0.0084*  ×ln(Size) (0.0032) (0.0041) (0.0121) (0.0045) Mild damage −0.0103 −0.0123 −0.0151 −0.0123  ×ln(Size) (0.0074) (0.0077) (0.0137) (0.0078) Severe damage 0.0248 0.0426** 0.0640*** 0.0449***  ×Prod (0.0196) (0.0173) (0.0231) (0.0146) Mild damage 0.0580 0.0544** 0.0702*** 0.0552**  ×Prod (0.0356) (0.0263) (0.0130) (0.0263) Severe damage −0.7316*** −0.7049*** −0.7124***  ×CreditCard (0.0536) (0.0515) (0.0515) Mild damage D D D  ×CreditCarde County FE ✓ ✓ ✓ ✓ NAICS FE (six-digit) ✓ ✓ ✓ ✓ Observationsd 2400 2400 2400 2400 Percent predicted outside [0,1]c 10% 10% 6% 10% Baseline Baseline Restrictive Expansive Severe damage −0.2923*** −0.2753*** −0.3838*** −0.2967*** (0.0595) (0.0542) (0.1132) (0.0466) Mild damage −0.1309 −0.1006 −0.2307*** −0.1038 (0.1776) (0.1039) (0.0557) (0.1031) ln(FirmSize) 0.0079* 0.0078* 0.0087** 0.0078* (0.0044) (0.0041) (0.0040) (0.0042) ln(FirmAge) 0.0296* 0.0288 −0.0032 0.0279 (0.0176) (0.0174) (0.0193) (0.0169) I(FirmAge=T)a −0.0087 −0.0075 0.0199 −0.0069 (0.0277) (0.0281) (0.0306) (0.0278) Productivity 0.0218 0.0232* 0.0309** 0.0184 (0.0142) (0.0135) (0.0144) (0.0127) Credit cardb −0.0116 0.0015 −0.0194 (0.0366) (0.0365) (0.0378) Severe damage 0.0046 −0.0099** −0.0125 −0.0084*  ×ln(Size) (0.0032) (0.0041) (0.0121) (0.0045) Mild damage −0.0103 −0.0123 −0.0151 −0.0123  ×ln(Size) (0.0074) (0.0077) (0.0137) (0.0078) Severe damage 0.0248 0.0426** 0.0640*** 0.0449***  ×Prod (0.0196) (0.0173) (0.0231) (0.0146) Mild damage 0.0580 0.0544** 0.0702*** 0.0552**  ×Prod (0.0356) (0.0263) (0.0130) (0.0263) Severe damage −0.7316*** −0.7049*** −0.7124***  ×CreditCard (0.0536) (0.0515) (0.0515) Mild damage D D D  ×CreditCarde County FE ✓ ✓ ✓ ✓ NAICS FE (six-digit) ✓ ✓ ✓ ✓ Observationsd 2400 2400 2400 2400 Percent predicted outside [0,1]c 10% 10% 6% 10% Notes: The sample includes establishments at least 2 years old in the base year, in the SBO, with employees in 2004. The LHS variable is an indicator for the establishment surviving from 2004 to 2006. Robust standard errors in parentheses, clustered by county. aAge of T indicates the firm or one of its original establishments was in operation in 1976. bUsed a credit card to finance capital improvements or expansion in 2002. cRounded to the nearest percentage point. dRounded to the nearest hundred. eCoefficients on this variable are statistically insignificant and suppressed to comply with disclosure avoidance. *Significant at 10%; **significant at 5% and ***significant at 1%. 25 We have also estimated this regression on the sample selected to have survived from 2004 to 2006, and find statistically and economically lower survival rates by damaged businesses between 2006 and 2010: survival rates of continuing business in these areas were 22 percentage points lower than those in undamaged areas. 26 Unfortunately, we cannot test this with the data we currently have available to us. 27 The idea is similar to the contagion effect of home foreclosures (Harding et al., 2009; Towe and Lawley, 2013), which may create additional externalities, like higher crime rates (Cui and Walsh, 2015). For a general discussion of agglomeration externalities, see Rosenthal and Strange (2003). 28 Katrina hit in August of 2005 and our employment measure in 2006 is for the week of March 12. This gives establishments seven months to rebuild their structures and restore staffing levels. 29 There are several recent studies of the effects of Katrina on population and labor-market outcomes. Among them, Deryugina et al. (forthcoming) study the effects of Katrina on Louisiana residents, and Groen et al. (2013) study Katrina survivors from a broader geographic area. 30 Foster et al. (2016) show that, as a rule, single-unit retailers that contract almost always exit entirely, whereas establishments in large firms are much more likely to contract without exiting, implying they have additional resources or other margins on which to adjust. 31 Kunreuther (1996, 2006) explains households’ low rates of disaster-insurance coverage with a combination of underestimation of the probability of disaster, above-market discount rates and binding budget constraints. These explanations may also apply to small businesses. 32 The SBA approved over 13,400 disaster loans for businesses of all sizes affected by the hurricanes from fiscal years 2005 to 2009, and more than 10,700 of these loans were identified as having assisted small businesses. A total of 2362 of these small-business loans went to Mississippi. Most of these loans were specifically directed to small businesses that were not able to obtain credit elsewhere (Small Business Administration, 2008). Unfortunately, we are unable to link information about loans to the LBD because the loans were issued in the name of the owner, not the business; moreover, many owners provided out-of-state addresses to the SBA. Acknowledgements Any opinions and conclusions expressed herein are those of the authors and do not necessarily represent the views of the U.S. Census Bureau. All results have been reviewed to ensure that no confidential information is disclosed. We thank Manuel Adelino, Saku Aura, Susanto Basu, Randy Becker, David Brown, Jeff Brown, Jeff Czajkowski, Tatyana Deryugina, Tim Dunne, Steve Fazzari, Teresa Fort, Lucia Foster, Etienne Gagnon, Jeff Groen, Hanna Halaburda, John Haltiwanger, Ron Jarmin, Bill Kerr, Shawn Klimek, Mark Kutzbach, Traci Mach, David Matsa, Erika McEntarfer, Guy Michaels, Peter Mueser, Justin Pierce, Allison Plyer, Anne Polivka, Andrea Pozzi, Melissa Schigoda, Antoinette Schoar, Chad Syverson and seminar and conference participants for helpful comments and conversations. This research was started while Basker was an ASA/NSF/Census Bureau Fellow visiting the Center for Economic Studies (CES) at the U.S. Census Bureau. E.B. thanks the funding agencies for their generous support and the economists at CES for their hospitality. References Adelino M. , Schoar A. , Severino F. ( 2015 ) House prices, collateral and self-employment . Journal of Financial Economics , 117 : 288 – 306 . Google Scholar CrossRef Search ADS Aghion P. , Fally T. , Scarpetta S. ( 2007 ) Credit constraints as a barrier to the entry and post-entry growth of firms . Economic Policy , 22 : 731 – 779 . Google Scholar CrossRef Search ADS Barlevy G. ( 2003 ) Credit market frictions and the allocation of resources over the business cycle . Journal of Economics and Management Strategy , 50 : 1795 – 1818 . Basker E. ( 2012 ) Raising the barcode scanner: technology and productivity in the retail sector . American Economic Journal: Applied Economics , 4 : 1 – 29 . Google Scholar CrossRef Search ADS Berger A. N. , Udell G. F. ( 1998 ) The economics of small business finance: the roles of private equity and debt markets in the financial growth cycle . Journal of Banking & Finance , 22 : 613 – 673 . Google Scholar CrossRef Search ADS Bertrand M. , Schoar A. , Thesmar D. ( 2007 ) Banking deregulation and industry structure: evidence from the French banking reforms of 1985 . Journal of Finance , 62 : 597 – 628 . Google Scholar CrossRef Search ADS Betancourt R. R. ( 2005 ) The Economics of Retailing and Distribution . Cheltenham, UK : Edward Elgar . Blanchflower D. G. , Levine P. B. , Zimmerman D. J. ( 2003 ) Discrimination in the small business credit market . Review of Economics and Statistics , 85 : 930 – 943 . Google Scholar CrossRef Search ADS Caballero R. J. , Hammour M. L. ( 1994 ) The cleansing effect of recessions . American Economic Review , 84 : 1350 – 1368 . Campello M. , Graham J. R. , Harvey C. R. ( 2010 ) The real effects of financial constraints: evidence from a financial crisis . Journal of Financial Economics , 97 : 470 – 487 . Google Scholar CrossRef Search ADS Cui L. , Walsh R. ( 2015 ) Foreclosure, vacancy and crime . Journal of Urban Economics , 87 : 72 – 84 . Google Scholar CrossRef Search ADS Deryugina T. , Kawano L. , Levitt S. ( forthcoming ) The economic impact of Hurricane Katrina on its victims: evidence from individual tax returns. American Economic Journal: Applied Economics . Dessaint O. , Matray A. ( 2013 ) Do managers overreact to salient risks? Evidence from Hurricane Strikes. HEC Paris Research Paper FIN-2013-1026. Doms M. E. , Jarmin R. S. , Klimek S. D. ( 2004 ) Information technology investment and firm performance in U.S. retail trade . Economics of Innovation and New Technology , 13 : 595 – 613 . Google Scholar CrossRef Search ADS Fazzari S. M. , Hubbard R. G. , Petersen B. C. ( 1988 ) Financing constraints and corporate investment . Brookings Papers on Economic Activity , 1988 : 141 – 206 . Google Scholar CrossRef Search ADS Federal Emergency Management Agency . ( 2011 ) Federal interagency geospatial concept of operations (GeoCONOPS), version 3.0. Fort T. , Haltiwanger J. , Jarmin R. , Miranda J. ( 2013 ) How firms respond to business cycles: the role of firm age and firm size . IMF Economic Review , 61 : 520 – 559 . Google Scholar CrossRef Search ADS Foster L. , Haltiwanger J. , Klimek S. , Krizan C. J. , Ohlmacher S. ( 2016 ) The evolution of national retail chains: how we got here. In Basker E. (ed.) Handbook on the Economics of Retailing and Distribution , pp. 7 – 37 . Cheltenham, UK : Edward Elgar . Foster L. , Haltiwanger J. , Krizan C. J. ( 2002 ) The link between aggregate and micro productivity growth: evidence from retail trade. National Bureau of Economic Research Working Paper 9120. Foster L. , Haltiwanger J. , Krizan C. J. ( 2006 ) Market selection, reallocation and restructuring in the U.S. retail trade sector in the 1990s . Review of Economics and Statistics , 88 : 748 – 758 . Google Scholar CrossRef Search ADS Gertler M. , Gilchrist S. ( 1994 ) Monetary policy, business cycles, and the behavior of small manufacturing firms . Quarterly Journal of Economics , 109 : 309 – 340 . Google Scholar CrossRef Search ADS Gould E. D. , Pashigian B. P. , Prendergast C. J. ( 2005 ) Contracts, externalities, and incentives in shopping malls . Review of Economics and Statistics , 87 : 411 – 422 . Google Scholar CrossRef Search ADS Greenstone M. , Mas A. , Nguyen H.-L. ( 2014 ) Do credit market shocks affect the real economy? Quasi-experimental evidence from the great recession and ‘normal’ economic times. National Bureau of Economic Research Working Paper 20704. Groen J. , Kutzbach M. , Polivka A. ( 2013 ) Storms and Jobs: The Effect of Hurricanes on Individuals’ Employment and Earnings over the Long Term. Unpublished paper, U.S. Census Bureau. Haltiwanger J. , Jarmin R. , Miranda J. ( 2013 ) Who creates jobs? Small vs. large vs. young . Review of Economics and Statistics , 95 : 347 – 361 . Google Scholar CrossRef Search ADS Harding J. P. , Rosenblatt E. , Yao V. W. ( 2009 ) The contagion effect of foreclosed properties . Journal of Urban Economics , 66 : 164 – 178 . Google Scholar CrossRef Search ADS Haskel J. , Sadun R. ( 2009 ) Entry, exit and labour productivity in UK retailing: evidence from micro data. In Jensen J. B. , Dunne T. , Roberts M. J. (eds) Producer Dynamics: New Evidence from Micro Data . University of Chicago Press , Chicago. Hosono K. , Miyakawa D. , Uchino T. , Hazama M. , Ono A. , Uchida H. , Uesugi I. ( 2012 ) Natural Disasters, Damage to Banks, and Firm Investment, Gakushuin University, Tokyo. Unpublished Paper. Jarmin R. S. , Miranda J. ( 2002 ) The Longitudinal Business Database. Unpublished Paper, U.S. Census Bureau. Jarmin R. S. , Miranda J. ( 2009 ) The Impact of Hurricanes Katrina, Rita and Wilma on Business Establishments . Journal of Business Valuation and Economic Loss Analysis , 4 : article 7. Kast S. ( 2005 ) Disaster bridge loan deadline extended to Jan. 31 for southernmost counties. US Fed News. Kunreuther H. ( 1996 ) Mitigating disaster losses through insurance . Journal of Risk and Uncertainty , 12 : 171 – 187 . Google Scholar CrossRef Search ADS Kunreuther H. ( 2006 ) Disaster mitigation and insurance: learning from Katrina . Annals of the American Academy of Political and Social Science , 604 : 208 – 227 . Google Scholar CrossRef Search ADS Lamont O. ( 1997 ) Cash flow and investment: evidence from internal capital markets . Journal of Finance , 52 : 83 – 109 . Google Scholar CrossRef Search ADS Mian A. , Sufi A. ( 2014a ) House of Debt . Chicago, IL : University of Chicago Press . Mian A. , Sufi A. ( 2014b ) What explains the 2007–2009 drop in employment? Econometrica , 82 : 2197 – 2223 . Google Scholar CrossRef Search ADS Pashigian B. P. , Gould E. D. ( 1998 ) Internalizing externalities: the pricing of space in shopping malls . Journal of Law and Economics , 41 : 115 – 142 . Google Scholar CrossRef Search ADS Rosenthal S. S. , Strange W. C. ( 2003 ) Geography, industrial organization, and agglomeration . Review of Economics and Statistics , 85 : 377 – 393 . Google Scholar CrossRef Search ADS Sayre E. A. , Butler D. ( 2011 ) The Geography of Recovery: An Analysis of the Mississippi Gulf Coast after Hurricane Katrina. Unpublished paper, University of Southern Mississippi. Schumpeter J. A. ( 1939 ) Business Cycles: A Theoretical, Historical and Statistical Analysis of the Capitalist Process . New York, NY : McGraw-Hill . Schumpeter J. A. ( 1942 ) Capitalism, Socialism and Democracy . New York, NY : Harper . Sharpe S. ( 1994 ) Financial market imperfections, firm leverage, and the cyclicality of employment . American Economic Review , 84 : 1060 – 1074 . Small Business Administration . ( 2008 ) Annual performance report. Discussion Paper GAO-041-076, SBA. Smith J. P. ( 2012 ) Hurricane Katrina: The Mississippi Story . Jackson, MS : University Press of Mississippi . Google Scholar CrossRef Search ADS Towe C. , Lawley C. ( 2013 ) The contagion effect of neighboring foreclosures . American Economic Journal: Economic Policy , 5 : 313 – 335 . Google Scholar CrossRef Search ADS Uchida H. , Miyakawa D. , Hosono K. , Ono A. , Uchino T. , Uesugi I. ( 2013 ) Natural disaster and natural selection. Research Institute of Economy, Trade, and Industry (RIETI) Discussion Paper 12-E-062. Zhu T. , Singh V. , Dukes A. ( 2011 ) Local competition, entry, and agglomeration . Quantitative Marketing and Economics , 9 : 129 – 154 . Google Scholar CrossRef Search ADS Appendix A. Katrina’s effect on the Mississippi Coast Hurricane Katrina was the most damaging hurricane of a particularly active 2005 hurricane season. Katrina struck several locations in Florida before veering into the Gulf of Mexico and making landfall again in New Orleans on 29 August 2005 as a Category 3 hurricane. Katrina caused damage in several states, including Alabama and Florida, but the most severe damage to businesses was in Louisiana (primarily due to flooding) and along the Mississippi coast (primarily due to high winds and storm surge). In Louisiana, flood waters did not completely recede for several weeks. The quick recovery of the Mississippi coast depended heavily on two sectors, military and casinos. Uncertainty about whether Keesler Air Force Base, which was heavily damaged, would be rebuilt was resolved within three weeks of the hurricane, when Air Force Secretary Pete Geren visited the base and promised to spend a billion dollars to fully restore it. The casinos, which had been barred from land and therefore operated on floating barges, presented a bigger challenge when they threatened not to rebuild unless they were allowed on land. Their threat was heeded: a month after the storm, on 30 September 2005, a land-based casino bill made it through the Mississippi legislature (Smith, 2012, 218–231). The combination of the Federal government’s explicit commitment to rebuilding Keesler and the casino bill seemed to seal the return of the Mississippi Gulf Coast. Figure A1 shows air travel to/from Gulfport-Biloxi International Airport relative to its January 2005 level, using monthly data from the Bureau of Transportation Statistics. There was a large negative shock of approximately 120 log points in September 2005, after which the air travel recovered relatively quickly; by early 2007 it was back to national trend. Figure A1 View largeDownload slide Air travel to/from Gulfport–Biloxi vs. all other domestic travel. Figure A1 View largeDownload slide Air travel to/from Gulfport–Biloxi vs. all other domestic travel. Business recovery was also aided by a web of government programs that provided post-storm support to residents and business owners. The most substantial program directed at business owners was a loan program administered by the Small Business Administration (SBA). Access to this program was not restricted to small business and it offered lower interest rates and longer terms than conventional loans.32 In addition, Mississippi offered small businesses in the worst-hit areas a 180-day, no-interest loan program; by the end of 2005, 392 small businesses had taken loans totaling over $9 million under this program (Kast, 2005). Smith (2012) reports that ‘by January 2006, three of the 13 destroyed casinos were back in business [….] Seven more casinos were scheduled to reopen by the end of 2006’ (231). The large investments these casinos made clearly signals that they, like the rest of Mississippi, expected a swift return to pre-Katrina conditions. Mississippi’s population remained largely in place in the aftermath of Katrina. Table A1 lists the 2000 and 2010 population in the affected counties and the rest of the state. Population changes between 2000 and 2010 were generally modest in Mississippi. The only exception is for one of the damaged counties, Stone County, which saw a population gain of nearly 27%. Stone County is very small, however, and accounts for little economic activity; fewer than 0.5% of our observations are from Stone County. In addition, the local unemployment rate, which rose in Hancock, Harrison and Jackson counties in 2005 and 2006, had returned to its pre-storm levels by 2007 (Sayre and Butler, 2011). Table A1 Population of selected Mississippi counties 2000–2010 2000 2010 Log County Population Population Change Hancock 42,967 43,929 +2.2% Harrison 189,601 187,105 –1.3% Jackson 131,420 139,668 +6.1% Stone 13,622 17,786 +26.7% Rest of state 2,467,048 2,578,809 +4.4% 2000 2010 Log County Population Population Change Hancock 42,967 43,929 +2.2% Harrison 189,601 187,105 –1.3% Jackson 131,420 139,668 +6.1% Stone 13,622 17,786 +26.7% Rest of state 2,467,048 2,578,809 +4.4% Source: Authors’ calculations from population census, 2000 and 2010. Table A1 Population of selected Mississippi counties 2000–2010 2000 2010 Log County Population Population Change Hancock 42,967 43,929 +2.2% Harrison 189,601 187,105 –1.3% Jackson 131,420 139,668 +6.1% Stone 13,622 17,786 +26.7% Rest of state 2,467,048 2,578,809 +4.4% 2000 2010 Log County Population Population Change Hancock 42,967 43,929 +2.2% Harrison 189,601 187,105 –1.3% Jackson 131,420 139,668 +6.1% Stone 13,622 17,786 +26.7% Rest of state 2,467,048 2,578,809 +4.4% Source: Authors’ calculations from population census, 2000 and 2010. Appendix B. Additional figures Published by Oxford University Press 2017. This work is written by US Government employees and is in the public domain in the US. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Journal of Economic Geography Oxford University Press

Taken by storm: business financing and survival in the aftermath of Hurricane Katrina

Loading next page...
 
/lp/ou_press/taken-by-storm-business-financing-and-survival-in-the-aftermath-of-IcPTTrbEMD
Publisher
Oxford University Press
Copyright
Published by Oxford University Press 2017. This work is written by US Government employees and is in the public domain in the US.
ISSN
1468-2702
eISSN
1468-2710
D.O.I.
10.1093/jeg/lbx023
Publisher site
See Article on Publisher Site

Abstract

Abstract We use Hurricane Katrina’s damage to the Mississippi coast in 2005 as a natural experiment to study business survival in the aftermath of a capital-destruction shock. We find very low survival rates for businesses that incurred physical damage, particularly for small firms and less-productive establishments. Conditional on survival, larger and more-productive businesses that rebuilt their operations hired more workers than their smaller and less-productive counterparts. Auxiliary evidence from the Survey of Business Owners suggests that the differential size effect is tied to the presence of financial constraints, pointing to a socially inefficient level of exits and to distortions of allocative efficiency in response to this negative shock. Over time, the size advantage disappeared and market mechanisms seem to prevail. 1. Introduction Business churn—entry and exit—is an important channel supporting innovation and productivity growth, and as such is an equilibrium feature of efficient markets. This is particularly true in the retail sector (Foster et al., 2006). Nevertheless, adjustments on the extensive margin are costly and frictions can lead to socially inefficient outcomes if relatively productive firms exit while less-productive firms remain in operation in response to a shock (Caballero and Hammour, 1994; Barlevy, 2003; Bertrand et al., 2007).1 An open question is whether financial constraints hamper allocative efficiency by reducing the importance of market fundamentals such as demand and costs, thus perpetuating the presence of relatively low-productivity firms. In this paper, we exploit the quasi-natural experiment created by variation in the degree of damage inflicted by Hurricane Katrina on businesses at different locations. We explore the relative importance of productivity and financial frictions in business responses to idiosyncratic negative capital shocks. We are interested in identifying the extent to which capital shocks merely hasten the exit of less-efficient businesses, as opposed to prompting exit by efficient and viable, but financially constrained, firms. We focus on the extensive margin—business survival—because it is the costliest from a social perspective. We find that smaller and more financially vulnerable firms exhibited lower survival rates in the aftermath of storm damage after controlling for productivity differences. We also find that, even conditional on survival, smaller and more financially vulnerable firms did not rebuild or grow as quickly following storm damage. Our paper is related to a broad literature exploring the impact of financial constraints on real economic activity, particularly relating to small firms. The papers most closely related to ours examine the impact of credit-supply shocks, business-cycle shocks or housing-price shocks on the employment and output of firms of different sizes and ages. For example, Gertler and Gilchrist (1994), Sharpe (1994), and, more recently, Greenstone et al. (2014) show that small firms are more sensitive than large ones to monetary and business-cycle shocks. Adelino et al. (2015) and Fort et al. (2013) find that small firms are more sensitive to housing-price shocks, underscoring the potential importance of less-traditional forms of financing to these firms. A common difficulty with this literature is disentangling cost shocks associated with business cycles, for example as a result of an increase in the cost of financing due to an increase in interest rates or the collapse of collateral values, from demand shocks associated with the same cycles.2 Our identification solves this problem by allowing us to identify highly localized capital-destruction shocks within relatively small areas. Our paper also relates to a literature on the impact of credit constraints on business entry (e.g., Aghion et al. 2007). A fundamental difficulty in studying the determinants of entry is that the pool of potential entrants is rarely observed. Our setting provides a unique opportunity to observe something very close to the pool of potential entrants: the pool of businesses that existed in that area shortly before the shock, and which would likely have continued to operate in the absence of the shock. The destruction of a business’s capital—structure, equipment, intermediate inputs and inventory—results in a cost shock very much like entry cost if the business is to return to operation. Our results therefore support the view that financial constraints affect entry rates, particularly for small businesses. A few other papers use shocks from natural disasters to identify financial effects. Hosono et al. (2012) use detailed firm-level data to estimate the impact of the 1995 Kobe earthquake on the supply of loans. They find that firms whose headquarters were located outside the damaged area but which had borrowing relationships with banks located inside the damaged area fared worse than undamaged firms borrowing from undamaged banks. A similar finding for the 2011 Great Tohoku earthquake is reported in Uchida et al. (2013). Unlike our paper, these papers focus on the ‘bank-lending channel’ and the impact of established lending relationships when the bank suffers a shock. Our focus is on a shock to firms, many of which may not have established lending relationships but rely on more informal and less well-documenting sources of financing: personal loans, loans from friends and family and even credit-card debt. Hurricanes often cause devastation over large geographic areas, which makes it impossible to cleanly identify capital shocks distinctly from demand and infrastructure shocks. To circumvent this problem, our analysis focuses not on Louisiana, where Hurricane Katrina’s impact was most widespread, but on the Mississippi coast, where damage was much more limited and localized, infrastructure was largely unaffected (and where infrastructure was damaged, repair times were fairly short) and population outflow was minimal. Undamaged businesses near the damaged areas serve as our control group. Importantly, our identification does not require the absence of demand or productivity shocks. We allow for the possibility that storm-damaged businesses experience disproportionately large negative shocks to demand (e.g., because consumers are reluctant to travel to damaged areas) or to their supply chains, and only assume that these shocks did not disproportionately impact small or less-productive firms. We use data from the Census Bureau’s Longitudinal Business Database (LBD) on approximately 10,000 business establishments in Mississippi, including over 1500 businesses in four counties that experienced significant storm damage, combined with precise information on the location and extent of the damage from the Federal Emergency Management Administration (FEMA). These data allow us to pinpoint which establishments were damaged or destroyed and which were left intact in the same area. We focus on establishments in the retail, restaurant and hotel sectors, which require a storefront to conduct business. Our identification comes from the randomness of actual damage within a limited geographic area. We find that the storm, which hit in August 2005, generated significant excess exits of physically damaged establishments in the short run and created a 30-point wedge between the survival rate of damaged and undamaged businesses in Mississippi by 2006. This finding is consistent with the notion that, in the short term, distress caused businesses that would otherwise have survived to cease operation. In addition, we find evidence for both efficient and inefficient exit in the short run. Across the board, exiting establishments are less productive than survivors, and this productivity wedge is 50% higher for businesses whose physical structures were destroyed by Katrina. At the same time, even controlling for productivity we find that the brunt of the effect of storm damage on short-run survival fell on smaller firms. Larger firms have higher survival rates in undamaged areas, but having been hit by storm damage triples or quadruples the advantage that these firms have over their smaller, also-damaged, counterparts. In the long run, we find exit rates of large and small firms in damaged areas equalize; we interpret this as possible evidence of an unforeseen endogenous shock induced by the high rate of small-firm exit. In auxiliary analyses, we find direct evidence of the importance of financial constraints to business survival. Businesses that had previously relied on credit-card debt to finance expansion or capital improvements, demonstrating a very high marginal cost of financing, also experienced much lower post-shock survival rates than similar businesses that relied on other forms of financing, including bank loans, for capital projects. On the intensive margin, conditional on survival, physically damaged establishments grew at a lower rate than their undamaged counterparts. Consistent with the survival results, we find that larger and more productive firms were able to rebuild their operations in the short run by hiring more workers than their smaller and less-productive counterparts. However, over time the size advantage disappears; 5 years after the storm, only initial productivity predicts survival and growth. We take this as evidence that, although access to finance may confer an initial advantage at entry and offer protection against shocks, this advantage dissipates over time. The rest of the paper is organized as follows. Section 2 describes our data in detail and provides some preliminary figures highlighting our empirical identification strategy. Our analyses of survival and firm growth are, respectively, in Sections 3 and 4. Section 5 concludes. 2. Background and data Hurricane Katrina made landfall in Louisiana in late August 2005, where it caused massive flooding, and quickly veered into Mississippi. The major source of damage in Mississippi was wind damage, which caused storm surges. Figure 1 shows a map of Mississippi, highlighting the four counties—Hancock, Harrison, Jackson, and Stone Counties—that were most affected by the hurricane.3 Figure 1 View largeDownload slide Mississippi (shaded counties most affected by Katrina). Figure 1 View largeDownload slide Mississippi (shaded counties most affected by Katrina). The primary building block in our analysis is the Census Bureau’s LBD. The LBD is a longitudinal database covering all employer establishments and firms in the U.S. non-farm private economy.4 We use data from the LBD to track the activity and outcomes of all stores, restaurants and hotels operating in Mississippi between 2002 and 2010. The LBD identifies the six-digit North American Industry Classification System (NAICS) code that represents the primary activity of each business establishment.5 We limit our analysis to retail and restaurant businesses and hotels and other accommodation facilities (including casinos) for several reasons.6 First, they represent a very large share of the local economies in the affected counties, approximately 10 times as large as manufacturing.7 This is important since affected areas are small and we need sectors with enough data to conduct the analysis. Second, unlike many other service industries and some non-service industries (e.g., construction), the location of the business is non-fungible. Whereas a lawyer may continue to provide legal services and a janitorial firm may continue to provide cleaning services even if the main office is destroyed, stores, restaurants and hotels provide their services at the business address and cease operations when that location is destroyed. Finally, these sectors serve local (and tourist) demand. Demand for products in other sectors such as manufacturing may extend beyond the local area differentially depending on the size of the business and in ways that we do not observe, making it hard to determine the relative effect of demand and cost shocks for these businesses. Establishments in the LBD are defined to be ‘active’ in a given year if they report positive payroll for any part of the year. In our baseline regression specifications, we identify a surviving business as one that reports payroll either in the current year or in a subsequent year; conversely, an establishment exit is defined by the absence of any reported payroll in the current and all future years. This definition of survival is conservative in that periods of temporary inactivity are consistent with survival. For robustness checks, we have both expanded and narrowed the definition of survival. First, in some robustness checks, we have narrowed the definition of survival to exclude establishments whose payroll falls by more than 90% and establishments that continue to operate but do so under new ownership. This is our ‘restrictive’ survival variable.8 Conversely, we have also created a ‘expansive’ survival variable, which treats establishments that cease to report payroll but continue to report revenue as non-employers as survivors. To find these businesses, we supplement the LBD with data from the integrated LBD (ILBD), which provides data on businesses with revenues but no payroll. In our sample, using employment to identify survivors, approximately 18% of establishments that had payroll in 2004 are no longer in business in 2006, and 22% of establishments still active in 2006 are no longer in business by 2010. These survival rates decrease (increase) by 1–3 percentage points when we restrict (expand) the definition of exit, depending on the year. Our results are robust to these alternative definitions. The firm identifier helps us determine the age and size of the entity that owns the establishment. Firm age is censored from above because we do not know the exact age of firms that existed in 1976, the LBD’s first year, so we also include an indicator for censored ages, I(FirmAgei=T), in all regressions. Our measure of firm size is the number of establishments the firm operates nationwide; it is equal to 1 for single-unit firms and exceeds 1000 for others. Given that this is a study of survival, the accuracy of longitudinal links is key to our analysis. Establishments that maintain the same address and ownership are relatively easy to track over time. If an establishment moves to a new address, but the address is in the same county, the LBD also identifies the establishment as a continuing operation. This implies that establishments that were located in damaged areas but reopened elsewhere within the same county appear as survivors in our data. It is particularly important that we correctly identify the timing of exits. We depend on tax filings for this purpose. For single-unit firms, this is straightforward: if a firm paid no payroll taxes in 2006 or any year thereafter, we consider its last year of operation to be 2005.9 However, multi-unit firms may continue to operate and pay payroll taxes even if one establishment closes. The Census relies on the Company Organization Survey (COS) to identify exits of establishments belonging to multi-unit firms. All multi-unit firms receive the COS in 5-year intervals, and larger multi-unit firms—those with at least 250 employees in total, across all their establishments—receive the COS annually.10 The LBD does not include establishment-level revenue. We use pre-storm revenue information from the 2002 Census of Retail Trade and the 2002 Census of Accommodation and Food Services to construct a measure of labor productivity at the establishment level. In the absence of information on other inputs, such as cost of materials and capital, we calculate labor productivity as the log of the ratio of the establishment’s annual revenue to employment.11 Because this productivity measure is from 2002, we limit our analysis to establishments that had existed in 2002 as well as 2004. We geocode establishments using Geographic Information System tools to assign latitude and longitude based on the business’s address. In a small number of cases the business address may represent the address of an accountant or other hired provider who assists the business with those forms. To minimize this problem, we drop 230 businesses whose addresses were identical to addresses provided by accounting or bookkeeping firms. Not all addresses are of the necessary quality to be able to geocode down to latitude and longitude successfully. Incomplete addresses and non-standard addresses (e.g., rural routes or PO Box addresses) are the main reasons for failures. Rural areas are known to be particularly problematic in this regard. For 2004, in each of the four Mississippi counties that experienced significant direct damage from Katrina, we were able to geocode more than 85% of establishments. Table 1 lists the number of geocoded establishments in each of the four affected counties in comparison with the rest of the state.12 Geocoding rates are typically higher in the damaged counties close to the Gulf than in the more rural inland areas in the rest of the state. Table 2 compares summary statistics of establishment and firm characteristics for geocoded and non-geocoded establishments in 2004. Compared with non-geocoded establishments, geocoded establishments are on average about 1 year younger and more likely to belong to single-unit firms. On all other dimensions geocoded establishments are not statistically distinguishable from non-geocoded establishments. Table 2 Establishment summary statistics: all establishments, 2004 Variable Obs.a All Non-geo-coded Geo-coded t-Testb Single-unit firms (%) 12,300 59.4 55.4 60.3 0.000 Establishments in firm 12,300 460.3 479.1 456.0 0.451 Firm agec 12,300 18.1 18.8 17.9 0.000 Establishment employment 12,300 17.6 16.2 17.9 0.270 Establishment aged 12,300 12.8 12.6 12.9 0.146 Productivitye 12,300 4.5 4.5 4.5 0.576 Variable Obs.a All Non-geo-coded Geo-coded t-Testb Single-unit firms (%) 12,300 59.4 55.4 60.3 0.000 Establishments in firm 12,300 460.3 479.1 456.0 0.451 Firm agec 12,300 18.1 18.8 17.9 0.000 Establishment employment 12,300 17.6 16.2 17.9 0.270 Establishment aged 12,300 12.8 12.6 12.9 0.146 Productivitye 12,300 4.5 4.5 4.5 0.576 aRounded to the nearest hundred. bp-value from t-test for equality of the mean. cRight-censored age of 29 used for 4000 observations. dRight-censored age of 29 used for 1200 observations. eLog ratio of revenue to employment in 2002 for establishments that survived to 2004. Table 2 Establishment summary statistics: all establishments, 2004 Variable Obs.a All Non-geo-coded Geo-coded t-Testb Single-unit firms (%) 12,300 59.4 55.4 60.3 0.000 Establishments in firm 12,300 460.3 479.1 456.0 0.451 Firm agec 12,300 18.1 18.8 17.9 0.000 Establishment employment 12,300 17.6 16.2 17.9 0.270 Establishment aged 12,300 12.8 12.6 12.9 0.146 Productivitye 12,300 4.5 4.5 4.5 0.576 Variable Obs.a All Non-geo-coded Geo-coded t-Testb Single-unit firms (%) 12,300 59.4 55.4 60.3 0.000 Establishments in firm 12,300 460.3 479.1 456.0 0.451 Firm agec 12,300 18.1 18.8 17.9 0.000 Establishment employment 12,300 17.6 16.2 17.9 0.270 Establishment aged 12,300 12.8 12.6 12.9 0.146 Productivitye 12,300 4.5 4.5 4.5 0.576 aRounded to the nearest hundred. bp-value from t-test for equality of the mean. cRight-censored age of 29 used for 4000 observations. dRight-censored age of 29 used for 1200 observations. eLog ratio of revenue to employment in 2002 for establishments that survived to 2004. Damage information comes from FEMA and is described in detail in Jarmin and Miranda (2009). Using remote-sensing technology, FEMA classified damaged areas over the period from August 30 to September 10 using a four-tier damage scale: limited, moderate, extensive and catastrophic.13 We reduce this to a two-tier scale, combining ‘extensive’ and ‘catastrophic’ into one category (‘severe’ damage) and ‘limited’ and ‘moderate’ into a second category (‘mild’ damage). In practice, there was very little extensive damage so almost all of the damage we classify as severe is catastrophic. Critically, damage designations are not based on insurance claims. Because FEMA’s remote-sensing maps focus primarily on developed areas, we may under-estimate the damage in less-developed areas. Following Jarmin and Miranda (2009), we add the FEMA damage information to the LBD to obtain, for each geocoded establishment, the FEMA damage classification of the location containing that establishment. Figure 2 shows an area on the border of Harrison and Hancock counties in Mississippi in which storm damage was widespread and highly variable. Each gray dot on the map represents a single business establishment.14 Establishments in diagonally cross-hatched areas were severely damaged, while those in horizontal and vertical cross-hatched areas were mildly damaged. Establishments in the white areas were physically undamaged. In addition, a handful of business establishments were located in areas that still had standing water 1 week after the storm. These areas are diagonally lined in the figure but are excluded from our analysis due to the very small number of establishments impacted by flooding; none of our results are sensitive to this exclusion. Figure 2 View largeDownload slide Damage area closeup: Harrison and Hancock counties, MS locations are ‘jittered’ to prevent identification of particular establishments. Figure 2 View largeDownload slide Damage area closeup: Harrison and Hancock counties, MS locations are ‘jittered’ to prevent identification of particular establishments. The accuracy of the FEMA damage designations is critical for our estimation. FEMA reports that, of the 150,000 homes it classified using this scale in Katrina’s immediate aftermath, fewer than 10% were mis-classified (Federal Emergency Management Agency, 2011, 42). The smallest area designated in the Mississippi database is about 145 m2 (0.000056 square miles); the median is about 55,600 m (0.02 square miles). The larger areas tend to be closer to the shore, where damage was most severe and several city blocks were effectively destroyed; smaller areas are designated inland where damage intensity to structures is more likely to differ across small distances. To the extent that damage is mismeasured, this would imply measurement error and therefore attenuation bias in our estimated coefficients. Table 1 provides 2004 summary statistics for the four affected counties and an aggregated ‘rest of state’ category. Approximately 350 establishments were in areas later designated by FEMA as having endured severe damage, and 350 more were in areas later designated as having suffered mild damage. We refer to all of these establishments as ‘damaged’. The last two columns in Table 1 provide the approximate percentage of establishments in each of the counties with a damage designation. Very small cells are suppressed to comply with Census Bureau disclosure requirements. Table 3 shows pre-storm summary statistics for the 2004 cross-section of geocoded establishments. The first column, showing the average value of the variable for all geocoded establishments, reproduces column (4) of Table 2. The next two columns show the average value for establishments located in areas that were later damaged and those located in areas that were undamaged. For almost all the variables listed—firm size (number of establishments in firm as well as a single-unit firm indicator), firm age, establishment size (employment) and establishment age—the differences between the damaged and undamaged areas are both small and statistically insignificant. The only statistically significant difference between damaged and undamaged establishments is that damaged establishments have slightly lower measured pre-storm labor productivity. We control for labor productivity in all the reported regressions in the paper. Table 3 Establishment summary statistics: geo-coded establishments, 2004 Variable Obs.a All Undamaged Damaged t-Testb Single-unit firms (%) 10,000 60.3 60.2 60.9 0.727 Establishments in firm 10,000 456.0 451.2 519.1 0.188 Firm agec 10,000 17.9 17.9 17.4 0.153 Establishment employment 10,000 17.9 17.9 18.6 0.788 Establishment aged 10,000 12.9 12.9 12.4 0.115 Productivitye 10,000 4.5 4.5 4.4 0.000 Variable Obs.a All Undamaged Damaged t-Testb Single-unit firms (%) 10,000 60.3 60.2 60.9 0.727 Establishments in firm 10,000 456.0 451.2 519.1 0.188 Firm agec 10,000 17.9 17.9 17.4 0.153 Establishment employment 10,000 17.9 17.9 18.6 0.788 Establishment aged 10,000 12.9 12.9 12.4 0.115 Productivitye 10,000 4.5 4.5 4.4 0.000 aRounded to the nearest hundred. bp-value from t-test for equality of the means. cRight-censored age of 29 used for 3200 observations. dRight-censored age of 29 used for 1000 observations. eLog ratio of revenue to employment in 2002 for establishments that survived to 2004. Table 3 Establishment summary statistics: geo-coded establishments, 2004 Variable Obs.a All Undamaged Damaged t-Testb Single-unit firms (%) 10,000 60.3 60.2 60.9 0.727 Establishments in firm 10,000 456.0 451.2 519.1 0.188 Firm agec 10,000 17.9 17.9 17.4 0.153 Establishment employment 10,000 17.9 17.9 18.6 0.788 Establishment aged 10,000 12.9 12.9 12.4 0.115 Productivitye 10,000 4.5 4.5 4.4 0.000 Variable Obs.a All Undamaged Damaged t-Testb Single-unit firms (%) 10,000 60.3 60.2 60.9 0.727 Establishments in firm 10,000 456.0 451.2 519.1 0.188 Firm agec 10,000 17.9 17.9 17.4 0.153 Establishment employment 10,000 17.9 17.9 18.6 0.788 Establishment aged 10,000 12.9 12.9 12.4 0.115 Productivitye 10,000 4.5 4.5 4.4 0.000 aRounded to the nearest hundred. bp-value from t-test for equality of the means. cRight-censored age of 29 used for 3200 observations. dRight-censored age of 29 used for 1000 observations. eLog ratio of revenue to employment in 2002 for establishments that survived to 2004. We supplement our analysis with data from the 2002 Survey of Business Owners (SBO). The SBO is conducted in Economic Census years and elicits more detailed information about firm operations than what is available in the Economic Census. The questions on the SBO form vary somewhat from year to year. In 2002, a direct measure of capital access comes from the question: ‘During 2002, were any of the following sources used to finance expansion or capital improvements for this business? Mark all that apply.’ The list includes personal or family savings and other assets; credit-card debt; bank loans; government and government-guaranteed loans; and financing from an outside investor. In addition, a check box for ‘no financing needed’ was also provided. Of the approximately 6300 businesses we were able to match to our geocoded Mississippi LBD sample in 2002, about 3500 reported that they needed and obtained some form of financing for capital improvements in the previous year, and nearly 3000 of those survived to 2004 to be included in our exit regressions. We treat the use of credit-card debt to finance expansion or capital improvements as a strong signal of a high marginal cost of financing. Only about 3.5% of establishments in our data belong to firms that reported using a credit card to finance capital improvements or expansions; the majority of these were single-unit firms. Table 4 provides summary statistics for the sample that matches the SBO data. These establishments are quite large, with 29 employees on average, but they are younger than the full sample, 13 years old on average. Establishment and firm characteristics do not differ statistically between establishments that were later damaged and those that were undamaged by Katrina. Table 4 Establishment summary statistics: Survey of Business Owners, 2004 Variable Obs.a All Undamaged Damaged t-Testb Single-unit firms (%) 2400 22.6 22.5 24.4 0.599 Establishments in firm 2400 690.9 675.2 948.1 0.210 Firm agec 2400 22.7 22.7 22.1 0.411 Establishment employment 2400 31.9 31.4 40.1 0.402 Establishment aged 2400 13.6 13.6 13.8 0.803 Productivitye 2400 4.5 4.5 4.4 0.890 Credit-card expansion financing (%)f 2400 4.1 D D 0.585 Variable Obs.a All Undamaged Damaged t-Testb Single-unit firms (%) 2400 22.6 22.5 24.4 0.599 Establishments in firm 2400 690.9 675.2 948.1 0.210 Firm agec 2400 22.7 22.7 22.1 0.411 Establishment employment 2400 31.9 31.4 40.1 0.402 Establishment aged 2400 13.6 13.6 13.8 0.803 Productivitye 2400 4.5 4.5 4.4 0.890 Credit-card expansion financing (%)f 2400 4.1 D D 0.585 aRounded to the nearest hundred. bp-value from t-test for equality of the means. cRight-censored age of 29 used for 1300 observations. dRight-censored age of 29 used for 200 observations. eLog ratio of revenue to employment in 2002 for establishments that survived to 2004. fSome cells are suppressed to comply with disclosure avoidance. Table 4 Establishment summary statistics: Survey of Business Owners, 2004 Variable Obs.a All Undamaged Damaged t-Testb Single-unit firms (%) 2400 22.6 22.5 24.4 0.599 Establishments in firm 2400 690.9 675.2 948.1 0.210 Firm agec 2400 22.7 22.7 22.1 0.411 Establishment employment 2400 31.9 31.4 40.1 0.402 Establishment aged 2400 13.6 13.6 13.8 0.803 Productivitye 2400 4.5 4.5 4.4 0.890 Credit-card expansion financing (%)f 2400 4.1 D D 0.585 Variable Obs.a All Undamaged Damaged t-Testb Single-unit firms (%) 2400 22.6 22.5 24.4 0.599 Establishments in firm 2400 690.9 675.2 948.1 0.210 Firm agec 2400 22.7 22.7 22.1 0.411 Establishment employment 2400 31.9 31.4 40.1 0.402 Establishment aged 2400 13.6 13.6 13.8 0.803 Productivitye 2400 4.5 4.5 4.4 0.890 Credit-card expansion financing (%)f 2400 4.1 D D 0.585 aRounded to the nearest hundred. bp-value from t-test for equality of the means. cRight-censored age of 29 used for 1300 observations. dRight-censored age of 29 used for 200 observations. eLog ratio of revenue to employment in 2002 for establishments that survived to 2004. fSome cells are suppressed to comply with disclosure avoidance. To see how the storm affected establishment counts, we partition the universe of retail, restaurant and hotel establishments in Mississippi with positive payroll and a geocoded address into two subsets: those located in Hancock, Harrison, Jackson and Stone counties (the counties in which FEMA designated most damaged areas) and those located elsewhere in Mississippi. Figure 3(a) shows the log level of the number of restaurants, stores and hotels that had payroll in these two parts of the state from 2005 to 2010, relative to the 2004 level in each region. Unlike the rest of the state, the four counties in which Katrina damage was concentrated experienced a decline in business activity between 2005 and 2006. The 2006 dip was only halfway reversed in 2007 and 2008, after which the economy stagnated. In 2010, while the rest of the state had approximately 5% fewer businesses in the retail, restaurant and hotel sectors relative to pre-storm levels, the affected counties were down approximately 10% from their pre-storm levels.15 Figure 3 View largeDownload slide Log number of Mississippi stores, restaurants and hotels by area damage status, relative to 2004. (a) All Mississippi and (b) damaged counties. Figure 3 View largeDownload slide Log number of Mississippi stores, restaurants and hotels by area damage status, relative to 2004. (a) All Mississippi and (b) damaged counties. Figure 3(b) restricts the analysis to the four damaged counties and partitions those further into areas that were designated by FEMA as: (a) undamaged, (b) mildly damaged or (c) severely damaged. It shows the log level of the number of restaurants, stores and hotels with payroll activity in each of these areas, again relative to the 2004 baseline. This finer partition shows that activity in the undamaged areas of the four counties more than fully recovered by 2007, overtaking the growth rate upstate. By contrast, even areas that had experienced only mild damage had not fully recovered by 2007 and suffered a significant decline in the Great Recession years (2007–2009). Areas classified as severely damaged experienced an even greater decline: the number of active establishments decreased between 2005 and 2006 by approximately 35%, then declined by a further 10% by 2007, and still had not stabilized by the end of our frame in 2010, at which point the decrease exceeded 50%.16 Figure 4(a,b) restricts the analysis to the cohort of establishments that were active in 2004 and also had positive revenue in 2002 (therefore, they were at least 2 years old in 2004).17 The difference between these figures and the previous ones is that we now exclude entrants. The solid line in Figure 4(a) shows establishments in undamaged counties exit at a rate of roughly 9% per year. By 2010, approximately 40% of these establishments had exited. The dashed line shows survival rates for establishments in the four counties with significant damage. The trends are similar prior to the storm, but the count of continuing establishments in the damaged counties drops by 15 percentage points between 2005 and 2006, when the hurricane hit, before settling back to similar trends. Finally, focusing on the damaged counties only, we again find large differences in outcomes by degree of damage. Figure 4(b) shows that undamaged areas experienced survival rates very similar to those upstate, whereas establishments located in storm-hit areas experienced a cumulative exit rate of 80 log points (55%) by 2010. Figure 4 View largeDownload slide Log number of Mississippi stores, restaurants and hotels that existed in 2002 by area damage status, relative to 2004. (a) All Mississippi and (b) damaged counties. Figure 4 View largeDownload slide Log number of Mississippi stores, restaurants and hotels that existed in 2002 by area damage status, relative to 2004. (a) All Mississippi and (b) damaged counties. Finally, restricting the analysis to establishments in areas with severe damage, we partition establishments using two more criteria. First, in Figure 5(a), we partition by firm size, based on the number of establishments the owning firm operated in 2004. We separate single-unit firms from small chains (with up to 100 establishments nationwide) and large chains (with more than 100 establishments). Relative to their 2004 levels, the number of single-unit establishments and establishments in small firms declines by 50–60% by 2006, while the number of establishments in large chains declines by only 20%. By 2010, there are virtually no single-establishment and small-chain stores left in the severely damaged area, but the number of establishments belonging to large chains declines only by a cumulative 40 log points. Second, in Figure 5(b), we partition establishments by their relative position in the 2002 productivity distribution: bottom quartile, interquartile range or top quartile. Here, we see that the exit rate is monotonic in 2002 productivity: by 2006, the number of establishments in the lowest three productivity quartiles declines by 50–60 log points, but in the upper quartile it declines by less than 40 log points. These effects, too, are magnified by 2010. Figure 5 View largeDownload slide Log number of Mississippi stores, restaurants and hotels that existed in 2002 in the severely damaged area by business characteristics, relative to 2004. (a) By firm size and (b) by productivity quartile. Figure 5 View largeDownload slide Log number of Mississippi stores, restaurants and hotels that existed in 2002 in the severely damaged area by business characteristics, relative to 2004. (a) By firm size and (b) by productivity quartile. In the next section, we formalize these findings using regression analysis. 3. Survival and firm characteristics 3.1. Short-run analysis 3.1.1. Firm size and productivity In the absence of frictions in financial markets, we expect to find a socially efficient response to the shock: firms return to operation if and only if the present discounted value of future profits exceeds the lump-sum cost of rebuilding structures, buying new equipment and replenishing inventories. On the other hand, if financial markets are inefficient so that the cost of financing is higher for small firms than large ones, we expect larger firms to return to operation at higher rates than smaller firms. Formally, we estimate a linear probability model of survival, including pre-storm labor productivity as a proxy for future profitability, which we cannot observe directly, as well as a measure of firm size: Survivali=αj(i)N(i)+γn(i)+σln(Firm Size)i+δDamagei+βln(Firm Size)i·Damagei+π·Prodi+φ·Prodi·Damagei+ηln(Firm Age)i+ηTI(Firm Agei=T)+ɛi, (1) where Survival is an indicator that equals 1 if establishment i was in operation in 2006 or returned to operation thereafter, and 0 if it permanently exited the employer universe by 2006. Firm Size is the nationwide count of establishments owned by establishment i’s firm. The sample includes all geocoded Mississippi retail, restaurant and hotel establishments with positive payroll in 2004 and labor productivity estimates from the 2002 Economic Census (the same establishments whose survival is plotted in Figure 4); we use 2004 rather than 2005 data because the shock occurred partway into 2005 and because some observations from 2005 may be missing due to the upheaval caused by the storm. The results are shown in the first column of Table 5. All coefficients are interpreted as marginal survival rates; the baseline is the survival rate of a hypothetical 1-year-old undamaged establishment in a single-establishment firm, whose labor productivity is equal to the average level within its six-digit NAICS sector. Table 5 Difference-in-difference survival regressions: productivity vs. firm size 2004–2006 Variable Baseline Restrictive Expansive 2002–2004 2004–2010 Severe damage −0.3162*** −0.4481*** −0.2924*** 0.1109*** −0.3811*** (0.0413) (0.0407) (0.0397) (0.0394) (0.0817) Mild damage −0.1566 −0.2059* −0.0691 −0.0181 −0.0968 (0.1027) (0.1204) (0.0744) (0.0321) (0.1128) ln(FirmSize) 0.0083*** 0.0067*** 0.0067*** 0.0090*** 0.0202*** (0.0018) (0.0017) (0.0018) (0.0013) (0.0034) ln(FirmAge) 0.0401*** 0.0330*** 0.0362*** 0.0282*** 0.0716*** (0.0057) (0.0070) (0.0052) (0.0065) (0.0092) I(FirmAge=T)b −0.0163 −0.0118 −0.0143 0.0059 −0.0233 (0.0139) (0.0136) (0.0145) (0.0092) (0.0149) Productivity 0.0463*** 0.0398*** 0.0448*** 0.0815*** 0.0619*** (0.0068) (0.0071) (0.0062) (0.0067) (0.0074) Severe damage 0.0169*** 0.0175*** 0.0147*** −0.0030 −0.0015  ×ln(Size) (0.0021) (0.0030) (0.0016) (0.0024) (0.0031) Mild damage 0.0097 0.0058 0.0065 −0.0012 0.0016  ×ln(Size) (0.0086) (0.0088) (0.0090) (0.0031) (0.0189) Severe damage 0.0275*** 0.0520*** 0.0256*** −0.0197*** 0.0456***  ×Prod (0.0061) (0.0052) (0.0046) (0.0071) (0.0113) Mild damage 0.0278 0.0386* 0.0124 0.0011 0.0112  ×Prod (0.0182) (0.0225) (0.0113) (0.0092) (0.0311) County FE×sector FE ✓ ✓ ✓ ✓ ✓ NAICS FE (six-digit) ✓ ✓ ✓ ✓ ✓ Observationsd 10,000 10,000 10,000 10,300 10,000 Percent predicted outside [0,1]c 6% 4% 6% 15% 3% 2004–2006 Variable Baseline Restrictive Expansive 2002–2004 2004–2010 Severe damage −0.3162*** −0.4481*** −0.2924*** 0.1109*** −0.3811*** (0.0413) (0.0407) (0.0397) (0.0394) (0.0817) Mild damage −0.1566 −0.2059* −0.0691 −0.0181 −0.0968 (0.1027) (0.1204) (0.0744) (0.0321) (0.1128) ln(FirmSize) 0.0083*** 0.0067*** 0.0067*** 0.0090*** 0.0202*** (0.0018) (0.0017) (0.0018) (0.0013) (0.0034) ln(FirmAge) 0.0401*** 0.0330*** 0.0362*** 0.0282*** 0.0716*** (0.0057) (0.0070) (0.0052) (0.0065) (0.0092) I(FirmAge=T)b −0.0163 −0.0118 −0.0143 0.0059 −0.0233 (0.0139) (0.0136) (0.0145) (0.0092) (0.0149) Productivity 0.0463*** 0.0398*** 0.0448*** 0.0815*** 0.0619*** (0.0068) (0.0071) (0.0062) (0.0067) (0.0074) Severe damage 0.0169*** 0.0175*** 0.0147*** −0.0030 −0.0015  ×ln(Size) (0.0021) (0.0030) (0.0016) (0.0024) (0.0031) Mild damage 0.0097 0.0058 0.0065 −0.0012 0.0016  ×ln(Size) (0.0086) (0.0088) (0.0090) (0.0031) (0.0189) Severe damage 0.0275*** 0.0520*** 0.0256*** −0.0197*** 0.0456***  ×Prod (0.0061) (0.0052) (0.0046) (0.0071) (0.0113) Mild damage 0.0278 0.0386* 0.0124 0.0011 0.0112  ×Prod (0.0182) (0.0225) (0.0113) (0.0092) (0.0311) County FE×sector FE ✓ ✓ ✓ ✓ ✓ NAICS FE (six-digit) ✓ ✓ ✓ ✓ ✓ Observationsd 10,000 10,000 10,000 10,300 10,000 Percent predicted outside [0,1]c 6% 4% 6% 15% 3% Notes: The sample includes establishments at least 2 years old and with employees in the base year. The LHS variable is an indicator for the establishment surviving from the first to the last year in the range. Robust standard errors in parentheses, clustered by county. aAge of zero occurs only in 2002 sample; later samples are continuers from 2002. bAge of T indicates the firm or one of its original establishments was in operation in 1976. cRounded to the nearest percentage point. dRounded to the nearest hundred. *Significant at 10%; **significant at 5% and ***significant at 1%. Table 5 Difference-in-difference survival regressions: productivity vs. firm size 2004–2006 Variable Baseline Restrictive Expansive 2002–2004 2004–2010 Severe damage −0.3162*** −0.4481*** −0.2924*** 0.1109*** −0.3811*** (0.0413) (0.0407) (0.0397) (0.0394) (0.0817) Mild damage −0.1566 −0.2059* −0.0691 −0.0181 −0.0968 (0.1027) (0.1204) (0.0744) (0.0321) (0.1128) ln(FirmSize) 0.0083*** 0.0067*** 0.0067*** 0.0090*** 0.0202*** (0.0018) (0.0017) (0.0018) (0.0013) (0.0034) ln(FirmAge) 0.0401*** 0.0330*** 0.0362*** 0.0282*** 0.0716*** (0.0057) (0.0070) (0.0052) (0.0065) (0.0092) I(FirmAge=T)b −0.0163 −0.0118 −0.0143 0.0059 −0.0233 (0.0139) (0.0136) (0.0145) (0.0092) (0.0149) Productivity 0.0463*** 0.0398*** 0.0448*** 0.0815*** 0.0619*** (0.0068) (0.0071) (0.0062) (0.0067) (0.0074) Severe damage 0.0169*** 0.0175*** 0.0147*** −0.0030 −0.0015  ×ln(Size) (0.0021) (0.0030) (0.0016) (0.0024) (0.0031) Mild damage 0.0097 0.0058 0.0065 −0.0012 0.0016  ×ln(Size) (0.0086) (0.0088) (0.0090) (0.0031) (0.0189) Severe damage 0.0275*** 0.0520*** 0.0256*** −0.0197*** 0.0456***  ×Prod (0.0061) (0.0052) (0.0046) (0.0071) (0.0113) Mild damage 0.0278 0.0386* 0.0124 0.0011 0.0112  ×Prod (0.0182) (0.0225) (0.0113) (0.0092) (0.0311) County FE×sector FE ✓ ✓ ✓ ✓ ✓ NAICS FE (six-digit) ✓ ✓ ✓ ✓ ✓ Observationsd 10,000 10,000 10,000 10,300 10,000 Percent predicted outside [0,1]c 6% 4% 6% 15% 3% 2004–2006 Variable Baseline Restrictive Expansive 2002–2004 2004–2010 Severe damage −0.3162*** −0.4481*** −0.2924*** 0.1109*** −0.3811*** (0.0413) (0.0407) (0.0397) (0.0394) (0.0817) Mild damage −0.1566 −0.2059* −0.0691 −0.0181 −0.0968 (0.1027) (0.1204) (0.0744) (0.0321) (0.1128) ln(FirmSize) 0.0083*** 0.0067*** 0.0067*** 0.0090*** 0.0202*** (0.0018) (0.0017) (0.0018) (0.0013) (0.0034) ln(FirmAge) 0.0401*** 0.0330*** 0.0362*** 0.0282*** 0.0716*** (0.0057) (0.0070) (0.0052) (0.0065) (0.0092) I(FirmAge=T)b −0.0163 −0.0118 −0.0143 0.0059 −0.0233 (0.0139) (0.0136) (0.0145) (0.0092) (0.0149) Productivity 0.0463*** 0.0398*** 0.0448*** 0.0815*** 0.0619*** (0.0068) (0.0071) (0.0062) (0.0067) (0.0074) Severe damage 0.0169*** 0.0175*** 0.0147*** −0.0030 −0.0015  ×ln(Size) (0.0021) (0.0030) (0.0016) (0.0024) (0.0031) Mild damage 0.0097 0.0058 0.0065 −0.0012 0.0016  ×ln(Size) (0.0086) (0.0088) (0.0090) (0.0031) (0.0189) Severe damage 0.0275*** 0.0520*** 0.0256*** −0.0197*** 0.0456***  ×Prod (0.0061) (0.0052) (0.0046) (0.0071) (0.0113) Mild damage 0.0278 0.0386* 0.0124 0.0011 0.0112  ×Prod (0.0182) (0.0225) (0.0113) (0.0092) (0.0311) County FE×sector FE ✓ ✓ ✓ ✓ ✓ NAICS FE (six-digit) ✓ ✓ ✓ ✓ ✓ Observationsd 10,000 10,000 10,000 10,300 10,000 Percent predicted outside [0,1]c 6% 4% 6% 15% 3% Notes: The sample includes establishments at least 2 years old and with employees in the base year. The LHS variable is an indicator for the establishment surviving from the first to the last year in the range. Robust standard errors in parentheses, clustered by county. aAge of zero occurs only in 2002 sample; later samples are continuers from 2002. bAge of T indicates the firm or one of its original establishments was in operation in 1976. cRounded to the nearest percentage point. dRounded to the nearest hundred. *Significant at 10%; **significant at 5% and ***significant at 1%. The identifying assumption in this regression is that, within the counties affected by Katrina, the precise path of the storm and therefore the damage inflicted was random, i.e., uncorrelated with ɛ. Table 3 in Section 2 provides reassurance that observables are distributed similarly in the treated (damaged) and control (undamaged) samples. The only exception is labor productivity, which is slightly lower on average for establishments located in damaged areas, and for which we control directly. Finally, we assume that county-by-sector (retail, restaurants, and hotels) and six-digit industry-fixed effects (α and γ, respectively) fully capture demand shocks following the storm. (The index j(i) indexes the county of establishment i; N(i) is the sector of establishment i—i.e., retail, restaurant, or hotel—and n(i) is for the six-digit NAICS code of establishment i.) The remaining differences between damaged and undamaged establishments can then be attributed to their differential recovery costs. The error term ε is clustered at the county level. Clustering accounts for the fact that business survival is interdependent across the county. The main short-run effect of the storm is a decrease in survival probability in areas that experienced severe (primarily catastrophic) damage. In some cases, although the business was previously viable, the cost of restoring structures, equipment and inventories cannot be justified by expected future profits. Ceteris paribus, establishments in these areas had a 32 percentage-point lower survival probability between 2004 and 2006 than undamaged establishments in the same counties. The effect of mild damage is smaller, about 16 percentage points and less precisely estimated. Size and productivity both act to diminish the negative effect of storm damage. The positive coefficient on the interaction of productivity and severe damage is consistent with the efficient-market hypothesis: the most productive businesses find it worthwhile to rebuild and return to operation, while less-productive ones rationally exit. The standard deviation of productivity within a six-digit NAICS industry is approximately 0.6, so a one-standard-deviation increase in productivity is associated with a 4.4 percentage point increase in the survival probability of a severely damaged establishment. The same productivity difference increases the survival probability of an undamaged establishment by 2.8 percentage points; in other words, high productivity disproportionately protects damaged businesses. At the same time, a doubling of a firm’s size also reduces the exit probability of severely damaged establishments, by 1.7 percentage points, triple the effect of size for undamaged businesses.18 The control variables all have expected signs. Establishments in older and larger firms are more likely to survive, consistent with selection for better management, access to resources and other correlates of survival. The direct effect of productivity is also positive. The results above assume a linear relationship between damage, log size and log productivity, and survival. We have also estimated nonlinear specifications (not shown). We find that the mitigating effect of firm size on the impact of severe damage starts with firms with more than 10 establishments, and is particularly strong for firms with more than 500 establishments. These are the firms that have the most resources, financial as well as managerial, to cope with disasters of this magnitude. In contrast, we do not find a systematic nonlinear pattern for the ameliorative effect of prior productivity on the impact of storm damage. The second and third columns of Table 5 verify that the main conclusion is not sensitive to the definition of the exit variable. In Column (2) of Table 5 we replace the baseline survival variable with a more restrictive variable, which reclassifies some survivors as exiters, and in Column (3) we use an expansive variable, which reclassifies some exiters as survivors. The main effect of both severe and mild damage changes with these redefinitions of survival—increasing in the former case and decreasing in the latter, consistent with excluding or adding marginal firms. The estimated coefficient on the interaction of damage and productivity also increases in the restrictive specification and decreases in the expansive specification; other interaction terms are stable. It is possible that survival rates of small firms are generally lower in the area damaged by Katrina, not because of Katrina, but because the damaged areas somehow favor large chains. If that were the case, survival rates in these areas would also have been lower for small businesses prior to the storm. To test this possibility we re-estimate Equation (1) but, on the left-hand side, we replace exit between 2004 and 2006 with exit between 2002 and 2004. This regression functions as a falsification exercise.19 Results from this specification using the baseline exit variable are shown in Column (4) of Table 5. We find higher survival rates on average in the damaged area prior to Katrina’s landfall, and lower survival rates for less-productive establishments in these areas, but we see no evidence that smaller firms fare worse in these areas than large firms. We have checked the robustness of these results in several unreported regressions. Changing the sample of controls to include only counties immediately adjacent to the damaged counties (Pearl River, Forrest, Perry and George), or to omit those same counties, or to include only the four counties with some severely damaged areas, does not change the results in any meaningful way, although standard errors on some coefficients increase. Similarly, adding establishment age and employment to the regressions has no impact on the qualitative patterns of coefficients. Finally, we have estimated the regression using a probit model; the results are again qualitatively unchanged. To alleviate the concern that differential errors in the exit rates of firms by firm size are driving the results, we also estimate the regression on a sample that includes all single-unit firms as well as establishments belonging to multi-unit firms with at least 250 employees in total. These firms are surveyed annually in the COS and their exit rates are likely to be measured most accurately. Relative to Column (1) of Table 5, these results (not shown, but available upon request) show coefficients that are larger in absolute terms, and equally strong statistically, on all damage variables and interaction terms. 3.1.2. Access to credit Credit constraints are not the only explanation for small firms’ greater sensitivity to the cost shock in the short run. For example, small-business owners may be more risk averse than larger businesses and may have responded more cautiously to uncertainty about the local economy’s rebounding.20,21 The decision to rebuild and reopen may have included considerations other than the success of a particular business establishment, such as public relations or media attention (particularly for large businesses) or attachment to the area (for locally owned businesses). In addition, the larger labor pool available to large firms may provide access to specialized managerial skill and other resources unavailable to their smaller counterparts. Although we cannot directly test for these and other alternative explanations, we can test for the importance of credit constraints. Most Census data sets do not contain any direct information about business balance sheets, banking and credit relationships or access to financial markets. The one exception is the SBO. We link the SBO data to our sample of establishments in Mississippi to explore the role of financing on exit in the aftermath of the storm. We use the 2002 SBO, which covers the cohort of establishments in our data, to explore the extent to which access to credit can explain the differential effects of severe damage by firm size. The SBO does not collect balance-sheet information or any direct measure of assets, but it does include a question about funding sources for capital improvements and expansions undertaken during 2002. We omit from our analysis businesses that report they did not ‘need’ any such funding, not being able to distinguish whether they did not want to make capital improvements or they made no capital improvements because funding was unavailable or too costly. We are agnostic about most sources of financing due to problems with interpretation. For example, a business may have obtained a government or government-guaranteed loan because it is not sufficiently viable to obtain a private loan, or because its owner is savvy and able to exploit any available resources; the former implies a negative relationship between such loans and survival whereas the latter may lead to a positive relationship. Likewise, using personal savings may indicate that a business cannot attract loans or outside investors, or it may be a signal that the owner has the resources to invest in her business and the confidence that it will do well. One source of financing, credit-card debt, stands out as particularly useful for our purposes. Credit cards charge high interest rates. Many small businesses use credit cards for convenience to pay some bills, for short-term cash flow or convenience, but using credit-card debt to finance expansion or capital improvements is a likely signal of a lack of other viable sources of funds.22,23 Whereas a business owner is unlikely to incur such an unsecured expense without reasonable expectation that the investment will justify itself, a surprise on the order of the demolition of the business by storm surge could prove particularly fatal to a credit-card-financed business. We estimate a model with interactions of damage with past use of credit cards as follows: Survivali=αj(i)+γn(i)+σln(Firm Size)i+δDamagei+βln(Firm Size)i·Damagei+λCreditCardi+μCreditCardi·Damagei+π·Prodi+φ·Prodi·Damagei+ηln(Firm Age)i+ηTI(FirmAgei=T)+ɛi. (2) The results are reported in Table 6. The first column reports the results of a regression with the SBO sample but without the credit-card variables. As with the full sample in the previous section, we find size and productivity predict survival in the undamaged areas. The estimated effects are of similar magnitude. The coefficient on the interaction of productivity and severe damage is positive and large as before, although not significant. The interaction between firm size and severe damage is statistically significant, though its magnitude is only about a quarter of the size in Table 5 (0.005 vs. 0.021). This may be a result of the fact that the SBO sample has fewer single-unit firms relative to the sample in the earlier regressions.24 In the second column, we add the credit-card variable and its interaction with the damage vector. The third and fourth columns repeat this analysis using the restrictive and expansive survival variables, respectively. Across specifications, the coefficient on the credit-card variable is small and statistically insignificant. On its face, this may seem puzzling, given our contention that credit-card debt is a strong signal of a financially weak business. However, the credit-card financing question refers to 2002, and our sample conditions on the business having survived to 2004. Credit-card-reliant businesses that survived 2 years may be a selected sample, no weaker than its counterparts which relied on other sources of financing, at least absent any additional shocks. The SBO questionnaire does not allow us to distinguish between businesses that went into significant credit-card debt and those that used their cards more on a more limited basis. In addition, the SBO does not ask businesses to rank the relative or absolutely importance of their various financing sources. Most businesses that used credit cards relied on other sources of financing as well: approximately half also reported using personal savings, and others also received bank loans. Although credit-card usage in 2002 does not predict survival for undamaged businesses, by and large, businesses that relied on this expensive form of financing and were hit by severe storm damage were unable to recover from this shock. The coefficient on the interaction of credit-card usage and damage is very large in absolute value and statistically significant: conditional on severe storm damage, an establishment whose owner reports having used a credit card for expansion or capital improvements in 2002 was 73 percentage points less likely to survive from 2004 to 2006 than one whose owner did not use a credit card for these purposes. In other words, although businesses that had previously signaled a high marginal cost of financial capital were able to continue operating as long as no major cost shocks arose, they could not adjust to a major shock. Interestingly, the direct effect of size is unchanged in this specification, but the interaction effect of size and damage disappears entirely, becoming negative. The addition of the credit-card variable appears to lend precision to some of the estimated coefficients. 3.2. Long-run analysis We return to the full sample of stores, restaurants and hotels to estimate the probability that a business that did not exit between 2004 and 2006, nevertheless exited between 2004 and 2010. As before, we contrast damaged and undamaged establishments. We estimate the same model as in Equation (1), but replace the LHS variable with an indicator for survival between 2004 and 2010. The results are reported in the last column of Table 5. By late 2007, demand was suppressed by reduced tourism, a consequence of the Great Recession. Businesses in severely damaged areas, however, fared even worse than those in undamaged areas: the coefficient on severe damage is six and half points larger, in absolute terms, than in Column (1).25 This long-run effect of the storm is striking in part because of the massive federal, state, local and private funds that poured into the area for rebuilding efforts. Moreover, the interaction effects now tell a different story. On the one hand, the interaction of productivity and severe damage continues to be positive and even increases in magnitude, implying that the most productive stores, restaurants and hotels continue to be partially shielded from the effects of the storm. But the coefficient on the interaction of size and damage attenuates to zero: controlling for productivity, larger firms located in severely damaged areas were no more likely to survive to 2010 than smaller firms. Put differently, larger firms, which survived the initial shock, were more likely than small firms to exit between 2006 and 2010. These establishments experienced severe damage in 2005, invested considerable resources rebuilding and were active again by 2006. Why did these larger businesses exit later? Common shocks, such as the Great Recession, which started in 2007, alone cannot explain why businesses in the damaged area experienced different survival rates than businesses in other areas, nor why these rates differ for small and large firms. Delayed exit by large businesses is consistent with differential access to credit in the aftermath of the storm, which induced a stronger selection of the smaller businesses. Under this scenario only ‘superstar’ small businesses were able to return to operation. Large firms, having had relatively easier access to internal resources, collateral or established banking relations, experienced less selection based on expected future performance. If this is the case, surviving establishments belonging to small firms must be more profitable, on average, than surviving unconstrained establishments.26 Consequently, they are also less vulnerable to a continued shock. However, this explanation implies that overall survival rates from 2006 to 2010 should have been higher in the damaged area, where businesses were put through a sort of ‘stress test’ in 2005, than in other areas in the same counties barring any additional shock. This is not the case. What secondary shock could have affected damaged areas relative to undamaged areas leading establishments of large firms to disproportionally exit? The most likely explanation is that the impact of the storm compounded over time, because of endogenous local demand shock induced by the destruction and consequent closure of many neighboring businesses. The failure of so many businesses may have had an aggregate effect by negatively impacting the local economy. It may also have had a localized indirect effect by reducing customer traffic to their surviving neighbors, in turn increasing their failure rate. These effects might have been difficult to foresee by large businesses that poured in resources to rebuilt and return to operations. This would have been particularly problematic for the less-productive large firms. This explanation is consistent with evidence on the importance of externalities in shopping malls (see, e.g., Pashigian and Gould, 1998; Gould et al., 2005) and with evidence of agglomeration effects in an urban setting (Zhu et al., 2011).27 Note in this regard that the Great Recession may have had a disproportionate effect in the damaged areas by also suppressing entry rates, which otherwise would be expected to offset exits. (See Figure 3 for the trends in the overall count of establishments by area damage status.) 4. Growth Examination of survival patterns post Katrina suggests establishment exits were concentrated among low-productivity and financially constrained firms. Our results so far suggest small firms were less likely to survive unless they were highly productive. A natural question to ask then is whether, conditional on survival, initial size and productivity predict the impact of the storm on business’s growth rate. To answer this question, we use all observations that survived from 2004 to 2006 based on the baseline survival definition, and estimate Growthi=αj(i)N(i)+γn(i)+σln(Firm Size)i+δDamagei+βln(Firm Size)i·Damagei+π·Prodi+φ·Prodi·Damagei+ηln(Firm Age)i+ηTI(FirmAgei=T)+ɛi, (3) where Growth is defined as the log change in establishment employment from 2004 to 2006, and all the other variables are as defined above. Results are shown in the first column of Table 7. Table 7 Difference-in-difference growth regressions: productivity vs. firm size 2004–2006 2002–2004 2004–2010 Severe damage −0.8802*** 0.0838 −0.3554*** (0.1030) (0.1378) (0.1244) Mild damage −0.0028 0.0109 0.0378 (0.2456) (0.0495) (0.2144) ln(FirmSize) 0.0127*** −0.0039 0.0189*** (0.0031) (0.0027) (0.0037) ln(FirmAge) −0.0301** −0.0200** −0.0230 (0.0116) (0.0090) (0.0161) I(FirmAge=T)b 0.0404* 0.0098 −0.0111 (0.0224) (0.0153) (0.0197) Productivity 0.0378*** 0.1784*** 0.1141*** (0.0095) (0.0112) (0.0153) Severe damage 0.0132*** −0.0196*** −0.0029  ×ln(Size) (0.0041) (0.0029) (0.0041) Mild damage −0.0036 −0.0239*** 0.0189***  ×ln(Size) (0.0210) (0.0025) (0.0065) Severe damage 0.1184*** 0.0110 0.0633**  ×Prod (0.0122) (0.0318) (0.0312) Mild damage −0.0128 0.0195 −0.0243  ×Prod (0.0454) (0.0158) (0.0449) County FE×sector FE ✓ ✓ ✓ NAICS FE (six-digit) ✓ ✓ ✓ Observationsc 8300 9100 6600 2004–2006 2002–2004 2004–2010 Severe damage −0.8802*** 0.0838 −0.3554*** (0.1030) (0.1378) (0.1244) Mild damage −0.0028 0.0109 0.0378 (0.2456) (0.0495) (0.2144) ln(FirmSize) 0.0127*** −0.0039 0.0189*** (0.0031) (0.0027) (0.0037) ln(FirmAge) −0.0301** −0.0200** −0.0230 (0.0116) (0.0090) (0.0161) I(FirmAge=T)b 0.0404* 0.0098 −0.0111 (0.0224) (0.0153) (0.0197) Productivity 0.0378*** 0.1784*** 0.1141*** (0.0095) (0.0112) (0.0153) Severe damage 0.0132*** −0.0196*** −0.0029  ×ln(Size) (0.0041) (0.0029) (0.0041) Mild damage −0.0036 −0.0239*** 0.0189***  ×ln(Size) (0.0210) (0.0025) (0.0065) Severe damage 0.1184*** 0.0110 0.0633**  ×Prod (0.0122) (0.0318) (0.0312) Mild damage −0.0128 0.0195 −0.0243  ×Prod (0.0454) (0.0158) (0.0449) County FE×sector FE ✓ ✓ ✓ NAICS FE (six-digit) ✓ ✓ ✓ Observationsc 8300 9100 6600 Notes: The sample includes establishments at least 2 years old in the base year, with employees in both the base and end years. The LHS variable is the establishment’s employment growth rate between the 2 years. Robust standard errors in parentheses, clustered by county. aAge of zero occurs only in 2002 sample; later samples are continuers from 2002. bAge of T indicates the firm or one of its original establishments was in operation in 1976. cRounded to the nearest hundred. *Significant at 10%; **significant at 5% and ***significant at 1%. Table 7 Difference-in-difference growth regressions: productivity vs. firm size 2004–2006 2002–2004 2004–2010 Severe damage −0.8802*** 0.0838 −0.3554*** (0.1030) (0.1378) (0.1244) Mild damage −0.0028 0.0109 0.0378 (0.2456) (0.0495) (0.2144) ln(FirmSize) 0.0127*** −0.0039 0.0189*** (0.0031) (0.0027) (0.0037) ln(FirmAge) −0.0301** −0.0200** −0.0230 (0.0116) (0.0090) (0.0161) I(FirmAge=T)b 0.0404* 0.0098 −0.0111 (0.0224) (0.0153) (0.0197) Productivity 0.0378*** 0.1784*** 0.1141*** (0.0095) (0.0112) (0.0153) Severe damage 0.0132*** −0.0196*** −0.0029  ×ln(Size) (0.0041) (0.0029) (0.0041) Mild damage −0.0036 −0.0239*** 0.0189***  ×ln(Size) (0.0210) (0.0025) (0.0065) Severe damage 0.1184*** 0.0110 0.0633**  ×Prod (0.0122) (0.0318) (0.0312) Mild damage −0.0128 0.0195 −0.0243  ×Prod (0.0454) (0.0158) (0.0449) County FE×sector FE ✓ ✓ ✓ NAICS FE (six-digit) ✓ ✓ ✓ Observationsc 8300 9100 6600 2004–2006 2002–2004 2004–2010 Severe damage −0.8802*** 0.0838 −0.3554*** (0.1030) (0.1378) (0.1244) Mild damage −0.0028 0.0109 0.0378 (0.2456) (0.0495) (0.2144) ln(FirmSize) 0.0127*** −0.0039 0.0189*** (0.0031) (0.0027) (0.0037) ln(FirmAge) −0.0301** −0.0200** −0.0230 (0.0116) (0.0090) (0.0161) I(FirmAge=T)b 0.0404* 0.0098 −0.0111 (0.0224) (0.0153) (0.0197) Productivity 0.0378*** 0.1784*** 0.1141*** (0.0095) (0.0112) (0.0153) Severe damage 0.0132*** −0.0196*** −0.0029  ×ln(Size) (0.0041) (0.0029) (0.0041) Mild damage −0.0036 −0.0239*** 0.0189***  ×ln(Size) (0.0210) (0.0025) (0.0065) Severe damage 0.1184*** 0.0110 0.0633**  ×Prod (0.0122) (0.0318) (0.0312) Mild damage −0.0128 0.0195 −0.0243  ×Prod (0.0454) (0.0158) (0.0449) County FE×sector FE ✓ ✓ ✓ NAICS FE (six-digit) ✓ ✓ ✓ Observationsc 8300 9100 6600 Notes: The sample includes establishments at least 2 years old in the base year, with employees in both the base and end years. The LHS variable is the establishment’s employment growth rate between the 2 years. Robust standard errors in parentheses, clustered by county. aAge of zero occurs only in 2002 sample; later samples are continuers from 2002. bAge of T indicates the firm or one of its original establishments was in operation in 1976. cRounded to the nearest hundred. *Significant at 10%; **significant at 5% and ***significant at 1%. Our first finding is that establishments that were severely damaged have much lower growth rates from 2004 to 2006, even conditional on survival. Although the sign is not surprising, the magnitude is notable: employment growth at severely damaged establishments is 88% log points lower than in undamaged establishments. The difference is likely accounted for by job loss during the rebuilding phase as well as diverted demand toward undamaged businesses, both of which would have suppressed employment.28 Consistent with our survival results, establishments located in areas subject to mild damage grew at rates similar to undamaged establishments. In addition, ceteris paribus, establishments the belonged to larger firms and establishments with higher 2002 productivity grew at higher rates from 2004 to 2006. More interesting from our perspective are the interaction effects. Our prior is that more-productive firms are able to disproportionally take advantage of the new conditions and grow faster. Our findings provide support for this interpretation. We find that establishments subject to severe damage but belonging to larger chains, and those with higher prior productivity, experienced a lesser negative shock to growth than similarly damaged establishments that belonged to smaller firms and had lower prior productivity. Our estimates imply that a severely damaged establishment belonging to a 1000-establishment firm grew on average by 7.8 points more than a single-establishment firm subject to the same damage level between 2004 and 2006; in the undamaged areas, the difference was only 3.8. In other words, the effect of size on growth is doubled for establishments that experienced severe damage. This is true even after controlling for productivity. We interpret this as evidence that establishments of large firms had an advantage rebuilding their capital and operations. Even more strikingly, a one-standard-deviation increase in productivity increased a severely damaged business’s 2-year employment growth by 9.4 points; in the undamaged areas, the difference was only 2.3 points (a 4-fold effect). The next two columns show results, from, first, 2002–2004 regressions (a falsification test) and, second, 2004–2010 regressions (long-run analysis). In the 2002–2004 period, we find that productivity was a strong predictor of growth. The interaction of chain size and future damage is negative, suggesting that large chains may have been relatively ill suited to local demand conditions in the area that later experienced the most severe storm damage. If we are willing to assume that the 2002–2004 results would have been replicated in 2004–2006 but for the storm, the falsification exercise can also be used to provide a triple-difference interpretation of our main results. In this interpretation, the protective impact of chain size on damaged businesses is even greater. The last column shows long-run results up to 5 years after the storm. Results are conditional on surviving to 2010, and as such should be interpreted as a selected sample. We find that initial productivity and size are strong predictors of survival for all businesses. Looking at the damaged areas specifically, however, we find that size no longer provides an advantage. By contrast, productivity does confer an advantage. These results mirror our findings on exits from Table 5. Although size confers an initial advantage, even conditional on productivity, in the long run this advantage disappears and the most productive establishments, regardless of size, are the ones that are able to take advantage of opportunities and grow. 5. Concluding remarks Our analysis uses Hurricane Katrina as a natural experiment to examine the impact of an external cost shock due to capital destruction on business activity.29 Consistent with a ‘cleansing’ hypothesis, we find that less-productive establishments exited disproportionately following the initial shock. But even after controlling for productivity, we find that establishments belonging to small firms were disproportionately affected. Business owners who reported relying, at least in part, on credit-card debt to finance capital projects were particularly vulnerable to exit following severe damage. Focusing on survivors, we find that large firms had an advantage rebuilding their operations quickly after the storm, as did the more productive firms of all sizes. Five years after the storm, size no longer confers an advantage on the initial cohort for either survival or growth. Although large firms may have been able to disproportionally survive the storm and rebuild initially, they did not perform better over time. The short-run results suggest that binding constraints other than those captured by an establishment’s productivity serve as a selection mechanism for small businesses following a cost shock. Small businesses survive the initial shock at much lower rates than large ones. They are also not as quick to rebuild their operations. As businesses age and grow, this selection mechanism diminishes in importance and eventually disappears, consistent with the idea that access to finance gives an initial advantage at entry but this advantage dissipates over time to the benefit of productivity.30 But small businesses that face a major cost shock early in their development cannot reach this later phase. Our data do not allow us to identify the precise channel by which financial constraints impact the survival of firms. For example, large firms may have an advantage over small ones due to access to internal capital markets (e.g., Fazzari et al., 1988; Lamont, 1997; Campello et al., 2010); this explanation implies some sort of failure in external capital markets, for example due to an information asymmetry. Alternatively, small and credit-constrained firms may be under-insured compared with larger firms, for example because they are less risk averse, uninformed about their risks, face higher insurance premia or are too illiquid to pay regular insurance premia.31 Although we cannot explicitly test for risk aversion, because small firms are less diversified than large ones they are unlikely to be less risk averse; all the other explanations imply some sort of financial friction. Since the channel determines the policy implications of our findings, more work is needed to determine the relative importance of these and other possible mechanisms. Our findings have important implications for understanding the impact of large cost shocks. First, cost shocks can take a long time to dissipate. In the case of Hurricane Katrina, the recovery took years despite a major effort by the federal government and others to aid in the recovery. Second, there is wide heterogeneity in business response to cost shocks. Small, credit-constrained firms are disproportionately affected and exit immediately. Some of these small firms are highly productive and may have otherwise survived. Third, in the case of Katrina, exiting businesses made up a large share of the local economy, and represent a loss not only of structures but also of entrepreneurial and social capital, which have been difficult to replace and rebuild. Fourth, the impact on the local economy from the high level of small-firm exit may itself have induced a second wave of business exits, further impeding and delaying recovery. Our long-run analysis shows that the first wave of small-business exit was followed by a second wave of large-firm establishment exit, particularly of less-productive large firms. We interpret this as evidence that market mechanisms eventually prevail. These results are particularly relevant when developing strategies for a prompt recovery in the aftermath of an economic shock whether caused by a natural disaster and likely otherwise. The Great Recession hit in 2007, less than 2 years after Katrina; by 2010 the region had yet to return to pre-2005 levels of economic activity. Measures that can directly address these additional risk factors may be able to speed up recovery when the next natural disaster hits. Figure B1 View largeDownload slide Log number of Mississippi stores, restaurants, and hotels by area damage status, relative to 2002. (a) All Mississippi and (b) damaged counties. Figure B1 View largeDownload slide Log number of Mississippi stores, restaurants, and hotels by area damage status, relative to 2002. (a) All Mississippi and (b) damaged counties. Figure B2 View largeDownload slide Log number of Mississippi stores, restaurants, and hotels that existed in 2002 by area damage status, relative to 2002. (a) All Mississippi and (b) damaged counties. Figure B2 View largeDownload slide Log number of Mississippi stores, restaurants, and hotels that existed in 2002 by area damage status, relative to 2002. (a) All Mississippi and (b) damaged counties. Funding Javier Miranda received no direct financial support from any organization for this paper. Emek Basker was partially funded from an ASA/NSF/Census Bureau Fellowship. The funding agencies are the American Statistical Association, the National Science Foundation, and the U.S. Census Bureau. There is no grant number. The fellowship is describe here: https://www.census.gov/srd/www/fellweb.html. Footnotes 1 The debate on whether shocks lead to productive cleansing or counterproductive destruction can be traced back to Schumpeter’s creative destruction hypothesis (Schumpeter, 1939, 1942). 2 For example, Mian and Sufi (2014a, 2014b) argue that a decline in consumer demand, not a decline in credit supply, was responsible for unemployment during the Great Recession. 3 More detail on Katrina’s impact on the Mississippi coast is available in Appendix A. 4 For more information on the LBD, see Jarmin and Miranda (2002). 5 This is an extremely fine classification. For example, among car dealerships this classification distinguishes between new- and used-car dealerships and between both of those and motorcycle dealerships; in the home-furnishings sector, it distinguishes between stores specializing in floor coverings, window treatments, and other home furnishings; and in the apparel sector, it distinguishes between men’s-, women’s-, children’s-, and family-clothing stores. 6 These business establishments correspond to NAICS 44-45 and 721-722. We exclude from the analysis non-store retailers such as catalog companies and vending-machine operators, NAICS 454, as well as caterers and mobile food-service providers, NAICS 72232 and 72233. 7 In 2004, according to published numbers from County Business Patterns, the four damaged counties, combined, had 2362 retail and accommodation establishments but only 247 manufacturing establishments. 8 The 90% threshold on payroll reduction is arbitrary. We have checked the robustness of our results using various alternative thresholds, as low as 50%, and continue to find qualitatively similar results. 9 The ILBD data allow us to relax this definition in the event that a firm continues to earn revenue in 2006. 10 This threshold was increased to 500 in more recent years. 11 Establishment revenue is not available in annual data sets, but only in the quinquennial (5-year) Economic Censuses. Employment is measured as of the week of 12 March 2002. We drop the top and bottom 1% of our productivity measures to remove the influence of outliers. Our ratio measure is also used in Foster et al. (2002) and Doms et al. (2004). Basker (2012) uses the ratio of revenue to payroll as an alternative measure of productivity. See Foster et al. (2002), Haskel and Sadun (2009), and Betancourt (2005) for further discussion. 12 All establishment counts throughout the paper are rounded to the nearest hundred. This is done to ensure that no confidential information is disclosed in the event that revisions require us to change the sample in small ways. Table 1 County summary statistics, 2004 State County Estabsa Geo-codeda Severe damage Mild damage MS Hancock 200 200 10.7% 68.0% MS Harrison 1000 900 35.2% 16.9% MS Jackson 500 400 6.9% 18.1% MS Stone 100 <100 0.0% MS Rest of state 10,600 8500 0.3% Totalb 12,300 10,000 3.5% 3.5% State County Estabsa Geo-codeda Severe damage Mild damage MS Hancock 200 200 10.7% 68.0% MS Harrison 1000 900 35.2% 16.9% MS Jackson 500 400 6.9% 18.1% MS Stone 100 <100 0.0% MS Rest of state 10,600 8500 0.3% Totalb 12,300 10,000 3.5% 3.5% Damage percentages are of geo-coded establishments. Blank cells indicate fewer than 10 establishments in damage zone. aEstablishment counts represent the retail, restaurant and hotel sectors, rounded to the nearest hundred. bMay not match sum due to rounding. Table 1 County summary statistics, 2004 State County Estabsa Geo-codeda Severe damage Mild damage MS Hancock 200 200 10.7% 68.0% MS Harrison 1000 900 35.2% 16.9% MS Jackson 500 400 6.9% 18.1% MS Stone 100 <100 0.0% MS Rest of state 10,600 8500 0.3% Totalb 12,300 10,000 3.5% 3.5% State County Estabsa Geo-codeda Severe damage Mild damage MS Hancock 200 200 10.7% 68.0% MS Harrison 1000 900 35.2% 16.9% MS Jackson 500 400 6.9% 18.1% MS Stone 100 <100 0.0% MS Rest of state 10,600 8500 0.3% Totalb 12,300 10,000 3.5% 3.5% Damage percentages are of geo-coded establishments. Blank cells indicate fewer than 10 establishments in damage zone. aEstablishment counts represent the retail, restaurant and hotel sectors, rounded to the nearest hundred. bMay not match sum due to rounding. 13 FEMA’s damage classification defines damage categories as follows. ‘Limited Damage: Generally superficial damage to solid structures (e.g., loss of tiles or roof shingles); some mobile homes and light structures are damaged or displaced. Moderate Damage: Solid structures sustain exterior damage (e.g., missing roofs or roof segments); some mobile homes and light structures are destroyed, many are damaged or displaced. Extensive Damage: Some solid structures are destroyed; most sustain exterior and interior damage (roofs missing, interior walls exposed); most mobile homes and light structures are destroyed. Catastrophic Damage: Most solid and all light or mobile home structures destroyed.’ 14 These dots were ‘jittered’ in compliance with Census Bureau disclosure procedures to prevent identification of particular establishments. 15 These figures are reproduced in the Appendix using 2002 as the base year. 16 These figures are reproduced in the Appendix using 2002 as the base year. 17 We need them to have revenue in 2002 so we can control for productivity later on. 18 An alternative to interacting firm size and damage is to interact firm age and damage, or include both interactions. Despite recent evidence that firm age may be a better indicator of a firm’s ability to withstand a serious shock (Haltiwanger et al., 2013), we prefer using firm size in this setting with a limited sample size for two reasons. First, firm size has a technical advantage over firm age, in that it is never censored and can take on any integer value. Firm age is right-censored for about a third of the establishments in our sample, dramatically limiting the explanatory variable of the continuous variable ln(Firm Age). Second, on a conceptual level, since the shock we consider here is destruction of capital stock, the total number of establishments the firm operates provides a measure of the fraction of the capital stock actually destroyed; all single-establishment firms, whether young or old, experienced a 100% capital-stock destruction if Katrina’s winds and storm surge destroyed their one establishment. 19 To maintain the same age distribution of firms in the 2002 sample, we drop firms with ages 0 or 1 in 2002. 20 In this context, it is interesting to note that Dessaint and Matray (2013), using data from large publicly traded firms, find evidence that managers tend to over-react to hurricane risks. 21 As noted in Appendix A, out-migration was very limited. 22 We are not aware of recent survey data on credit-card usage by small businesses. In the 1993 and 1998 waves of the National Survey of Small Business Finances, respectively, 29% and 34% of business owners reported using a business credit card; over 40% reported using a personal credit card for business expenses (Blanchflower et al., 2003). However, Berger and Udell (1998) report that, after accounting for paid bills, less than 1% of small-business debt in the 1993 wave came from credit-card debt. This finding is consistent with small businesses using credit cards as a payment method but, for the most part, not relying on them for long-term financing. 23 One might be concerned that use of credit card for capital improvements and expansion in the SBO is not a signal of credit constraints. For example, a restaurant that buys a new oven using a credit card could check the box indicating it used a credit card to finance a capital improvement. If this is the case it will cause attenuation bias in our estimated coefficients. 24 Only 30% of SBO establishments, compared with 60% of establishments in the full LBD sample, are single-units. As our previous results show, larger firms were less affected by severe damage in the short run. Table 6 Difference-in-difference survival regressions: auxiliary analysis, 2004–2006 Baseline Baseline Restrictive Expansive Severe damage −0.2923*** −0.2753*** −0.3838*** −0.2967*** (0.0595) (0.0542) (0.1132) (0.0466) Mild damage −0.1309 −0.1006 −0.2307*** −0.1038 (0.1776) (0.1039) (0.0557) (0.1031) ln(FirmSize) 0.0079* 0.0078* 0.0087** 0.0078* (0.0044) (0.0041) (0.0040) (0.0042) ln(FirmAge) 0.0296* 0.0288 −0.0032 0.0279 (0.0176) (0.0174) (0.0193) (0.0169) I(FirmAge=T)a −0.0087 −0.0075 0.0199 −0.0069 (0.0277) (0.0281) (0.0306) (0.0278) Productivity 0.0218 0.0232* 0.0309** 0.0184 (0.0142) (0.0135) (0.0144) (0.0127) Credit cardb −0.0116 0.0015 −0.0194 (0.0366) (0.0365) (0.0378) Severe damage 0.0046 −0.0099** −0.0125 −0.0084*  ×ln(Size) (0.0032) (0.0041) (0.0121) (0.0045) Mild damage −0.0103 −0.0123 −0.0151 −0.0123  ×ln(Size) (0.0074) (0.0077) (0.0137) (0.0078) Severe damage 0.0248 0.0426** 0.0640*** 0.0449***  ×Prod (0.0196) (0.0173) (0.0231) (0.0146) Mild damage 0.0580 0.0544** 0.0702*** 0.0552**  ×Prod (0.0356) (0.0263) (0.0130) (0.0263) Severe damage −0.7316*** −0.7049*** −0.7124***  ×CreditCard (0.0536) (0.0515) (0.0515) Mild damage D D D  ×CreditCarde County FE ✓ ✓ ✓ ✓ NAICS FE (six-digit) ✓ ✓ ✓ ✓ Observationsd 2400 2400 2400 2400 Percent predicted outside [0,1]c 10% 10% 6% 10% Baseline Baseline Restrictive Expansive Severe damage −0.2923*** −0.2753*** −0.3838*** −0.2967*** (0.0595) (0.0542) (0.1132) (0.0466) Mild damage −0.1309 −0.1006 −0.2307*** −0.1038 (0.1776) (0.1039) (0.0557) (0.1031) ln(FirmSize) 0.0079* 0.0078* 0.0087** 0.0078* (0.0044) (0.0041) (0.0040) (0.0042) ln(FirmAge) 0.0296* 0.0288 −0.0032 0.0279 (0.0176) (0.0174) (0.0193) (0.0169) I(FirmAge=T)a −0.0087 −0.0075 0.0199 −0.0069 (0.0277) (0.0281) (0.0306) (0.0278) Productivity 0.0218 0.0232* 0.0309** 0.0184 (0.0142) (0.0135) (0.0144) (0.0127) Credit cardb −0.0116 0.0015 −0.0194 (0.0366) (0.0365) (0.0378) Severe damage 0.0046 −0.0099** −0.0125 −0.0084*  ×ln(Size) (0.0032) (0.0041) (0.0121) (0.0045) Mild damage −0.0103 −0.0123 −0.0151 −0.0123  ×ln(Size) (0.0074) (0.0077) (0.0137) (0.0078) Severe damage 0.0248 0.0426** 0.0640*** 0.0449***  ×Prod (0.0196) (0.0173) (0.0231) (0.0146) Mild damage 0.0580 0.0544** 0.0702*** 0.0552**  ×Prod (0.0356) (0.0263) (0.0130) (0.0263) Severe damage −0.7316*** −0.7049*** −0.7124***  ×CreditCard (0.0536) (0.0515) (0.0515) Mild damage D D D  ×CreditCarde County FE ✓ ✓ ✓ ✓ NAICS FE (six-digit) ✓ ✓ ✓ ✓ Observationsd 2400 2400 2400 2400 Percent predicted outside [0,1]c 10% 10% 6% 10% Notes: The sample includes establishments at least 2 years old in the base year, in the SBO, with employees in 2004. The LHS variable is an indicator for the establishment surviving from 2004 to 2006. Robust standard errors in parentheses, clustered by county. aAge of T indicates the firm or one of its original establishments was in operation in 1976. bUsed a credit card to finance capital improvements or expansion in 2002. cRounded to the nearest percentage point. dRounded to the nearest hundred. eCoefficients on this variable are statistically insignificant and suppressed to comply with disclosure avoidance. *Significant at 10%; **significant at 5% and ***significant at 1%. Table 6 Difference-in-difference survival regressions: auxiliary analysis, 2004–2006 Baseline Baseline Restrictive Expansive Severe damage −0.2923*** −0.2753*** −0.3838*** −0.2967*** (0.0595) (0.0542) (0.1132) (0.0466) Mild damage −0.1309 −0.1006 −0.2307*** −0.1038 (0.1776) (0.1039) (0.0557) (0.1031) ln(FirmSize) 0.0079* 0.0078* 0.0087** 0.0078* (0.0044) (0.0041) (0.0040) (0.0042) ln(FirmAge) 0.0296* 0.0288 −0.0032 0.0279 (0.0176) (0.0174) (0.0193) (0.0169) I(FirmAge=T)a −0.0087 −0.0075 0.0199 −0.0069 (0.0277) (0.0281) (0.0306) (0.0278) Productivity 0.0218 0.0232* 0.0309** 0.0184 (0.0142) (0.0135) (0.0144) (0.0127) Credit cardb −0.0116 0.0015 −0.0194 (0.0366) (0.0365) (0.0378) Severe damage 0.0046 −0.0099** −0.0125 −0.0084*  ×ln(Size) (0.0032) (0.0041) (0.0121) (0.0045) Mild damage −0.0103 −0.0123 −0.0151 −0.0123  ×ln(Size) (0.0074) (0.0077) (0.0137) (0.0078) Severe damage 0.0248 0.0426** 0.0640*** 0.0449***  ×Prod (0.0196) (0.0173) (0.0231) (0.0146) Mild damage 0.0580 0.0544** 0.0702*** 0.0552**  ×Prod (0.0356) (0.0263) (0.0130) (0.0263) Severe damage −0.7316*** −0.7049*** −0.7124***  ×CreditCard (0.0536) (0.0515) (0.0515) Mild damage D D D  ×CreditCarde County FE ✓ ✓ ✓ ✓ NAICS FE (six-digit) ✓ ✓ ✓ ✓ Observationsd 2400 2400 2400 2400 Percent predicted outside [0,1]c 10% 10% 6% 10% Baseline Baseline Restrictive Expansive Severe damage −0.2923*** −0.2753*** −0.3838*** −0.2967*** (0.0595) (0.0542) (0.1132) (0.0466) Mild damage −0.1309 −0.1006 −0.2307*** −0.1038 (0.1776) (0.1039) (0.0557) (0.1031) ln(FirmSize) 0.0079* 0.0078* 0.0087** 0.0078* (0.0044) (0.0041) (0.0040) (0.0042) ln(FirmAge) 0.0296* 0.0288 −0.0032 0.0279 (0.0176) (0.0174) (0.0193) (0.0169) I(FirmAge=T)a −0.0087 −0.0075 0.0199 −0.0069 (0.0277) (0.0281) (0.0306) (0.0278) Productivity 0.0218 0.0232* 0.0309** 0.0184 (0.0142) (0.0135) (0.0144) (0.0127) Credit cardb −0.0116 0.0015 −0.0194 (0.0366) (0.0365) (0.0378) Severe damage 0.0046 −0.0099** −0.0125 −0.0084*  ×ln(Size) (0.0032) (0.0041) (0.0121) (0.0045) Mild damage −0.0103 −0.0123 −0.0151 −0.0123  ×ln(Size) (0.0074) (0.0077) (0.0137) (0.0078) Severe damage 0.0248 0.0426** 0.0640*** 0.0449***  ×Prod (0.0196) (0.0173) (0.0231) (0.0146) Mild damage 0.0580 0.0544** 0.0702*** 0.0552**  ×Prod (0.0356) (0.0263) (0.0130) (0.0263) Severe damage −0.7316*** −0.7049*** −0.7124***  ×CreditCard (0.0536) (0.0515) (0.0515) Mild damage D D D  ×CreditCarde County FE ✓ ✓ ✓ ✓ NAICS FE (six-digit) ✓ ✓ ✓ ✓ Observationsd 2400 2400 2400 2400 Percent predicted outside [0,1]c 10% 10% 6% 10% Notes: The sample includes establishments at least 2 years old in the base year, in the SBO, with employees in 2004. The LHS variable is an indicator for the establishment surviving from 2004 to 2006. Robust standard errors in parentheses, clustered by county. aAge of T indicates the firm or one of its original establishments was in operation in 1976. bUsed a credit card to finance capital improvements or expansion in 2002. cRounded to the nearest percentage point. dRounded to the nearest hundred. eCoefficients on this variable are statistically insignificant and suppressed to comply with disclosure avoidance. *Significant at 10%; **significant at 5% and ***significant at 1%. 25 We have also estimated this regression on the sample selected to have survived from 2004 to 2006, and find statistically and economically lower survival rates by damaged businesses between 2006 and 2010: survival rates of continuing business in these areas were 22 percentage points lower than those in undamaged areas. 26 Unfortunately, we cannot test this with the data we currently have available to us. 27 The idea is similar to the contagion effect of home foreclosures (Harding et al., 2009; Towe and Lawley, 2013), which may create additional externalities, like higher crime rates (Cui and Walsh, 2015). For a general discussion of agglomeration externalities, see Rosenthal and Strange (2003). 28 Katrina hit in August of 2005 and our employment measure in 2006 is for the week of March 12. This gives establishments seven months to rebuild their structures and restore staffing levels. 29 There are several recent studies of the effects of Katrina on population and labor-market outcomes. Among them, Deryugina et al. (forthcoming) study the effects of Katrina on Louisiana residents, and Groen et al. (2013) study Katrina survivors from a broader geographic area. 30 Foster et al. (2016) show that, as a rule, single-unit retailers that contract almost always exit entirely, whereas establishments in large firms are much more likely to contract without exiting, implying they have additional resources or other margins on which to adjust. 31 Kunreuther (1996, 2006) explains households’ low rates of disaster-insurance coverage with a combination of underestimation of the probability of disaster, above-market discount rates and binding budget constraints. These explanations may also apply to small businesses. 32 The SBA approved over 13,400 disaster loans for businesses of all sizes affected by the hurricanes from fiscal years 2005 to 2009, and more than 10,700 of these loans were identified as having assisted small businesses. A total of 2362 of these small-business loans went to Mississippi. Most of these loans were specifically directed to small businesses that were not able to obtain credit elsewhere (Small Business Administration, 2008). Unfortunately, we are unable to link information about loans to the LBD because the loans were issued in the name of the owner, not the business; moreover, many owners provided out-of-state addresses to the SBA. Acknowledgements Any opinions and conclusions expressed herein are those of the authors and do not necessarily represent the views of the U.S. Census Bureau. All results have been reviewed to ensure that no confidential information is disclosed. We thank Manuel Adelino, Saku Aura, Susanto Basu, Randy Becker, David Brown, Jeff Brown, Jeff Czajkowski, Tatyana Deryugina, Tim Dunne, Steve Fazzari, Teresa Fort, Lucia Foster, Etienne Gagnon, Jeff Groen, Hanna Halaburda, John Haltiwanger, Ron Jarmin, Bill Kerr, Shawn Klimek, Mark Kutzbach, Traci Mach, David Matsa, Erika McEntarfer, Guy Michaels, Peter Mueser, Justin Pierce, Allison Plyer, Anne Polivka, Andrea Pozzi, Melissa Schigoda, Antoinette Schoar, Chad Syverson and seminar and conference participants for helpful comments and conversations. This research was started while Basker was an ASA/NSF/Census Bureau Fellow visiting the Center for Economic Studies (CES) at the U.S. Census Bureau. E.B. thanks the funding agencies for their generous support and the economists at CES for their hospitality. References Adelino M. , Schoar A. , Severino F. ( 2015 ) House prices, collateral and self-employment . Journal of Financial Economics , 117 : 288 – 306 . Google Scholar CrossRef Search ADS Aghion P. , Fally T. , Scarpetta S. ( 2007 ) Credit constraints as a barrier to the entry and post-entry growth of firms . Economic Policy , 22 : 731 – 779 . Google Scholar CrossRef Search ADS Barlevy G. ( 2003 ) Credit market frictions and the allocation of resources over the business cycle . Journal of Economics and Management Strategy , 50 : 1795 – 1818 . Basker E. ( 2012 ) Raising the barcode scanner: technology and productivity in the retail sector . American Economic Journal: Applied Economics , 4 : 1 – 29 . Google Scholar CrossRef Search ADS Berger A. N. , Udell G. F. ( 1998 ) The economics of small business finance: the roles of private equity and debt markets in the financial growth cycle . Journal of Banking & Finance , 22 : 613 – 673 . Google Scholar CrossRef Search ADS Bertrand M. , Schoar A. , Thesmar D. ( 2007 ) Banking deregulation and industry structure: evidence from the French banking reforms of 1985 . Journal of Finance , 62 : 597 – 628 . Google Scholar CrossRef Search ADS Betancourt R. R. ( 2005 ) The Economics of Retailing and Distribution . Cheltenham, UK : Edward Elgar . Blanchflower D. G. , Levine P. B. , Zimmerman D. J. ( 2003 ) Discrimination in the small business credit market . Review of Economics and Statistics , 85 : 930 – 943 . Google Scholar CrossRef Search ADS Caballero R. J. , Hammour M. L. ( 1994 ) The cleansing effect of recessions . American Economic Review , 84 : 1350 – 1368 . Campello M. , Graham J. R. , Harvey C. R. ( 2010 ) The real effects of financial constraints: evidence from a financial crisis . Journal of Financial Economics , 97 : 470 – 487 . Google Scholar CrossRef Search ADS Cui L. , Walsh R. ( 2015 ) Foreclosure, vacancy and crime . Journal of Urban Economics , 87 : 72 – 84 . Google Scholar CrossRef Search ADS Deryugina T. , Kawano L. , Levitt S. ( forthcoming ) The economic impact of Hurricane Katrina on its victims: evidence from individual tax returns. American Economic Journal: Applied Economics . Dessaint O. , Matray A. ( 2013 ) Do managers overreact to salient risks? Evidence from Hurricane Strikes. HEC Paris Research Paper FIN-2013-1026. Doms M. E. , Jarmin R. S. , Klimek S. D. ( 2004 ) Information technology investment and firm performance in U.S. retail trade . Economics of Innovation and New Technology , 13 : 595 – 613 . Google Scholar CrossRef Search ADS Fazzari S. M. , Hubbard R. G. , Petersen B. C. ( 1988 ) Financing constraints and corporate investment . Brookings Papers on Economic Activity , 1988 : 141 – 206 . Google Scholar CrossRef Search ADS Federal Emergency Management Agency . ( 2011 ) Federal interagency geospatial concept of operations (GeoCONOPS), version 3.0. Fort T. , Haltiwanger J. , Jarmin R. , Miranda J. ( 2013 ) How firms respond to business cycles: the role of firm age and firm size . IMF Economic Review , 61 : 520 – 559 . Google Scholar CrossRef Search ADS Foster L. , Haltiwanger J. , Klimek S. , Krizan C. J. , Ohlmacher S. ( 2016 ) The evolution of national retail chains: how we got here. In Basker E. (ed.) Handbook on the Economics of Retailing and Distribution , pp. 7 – 37 . Cheltenham, UK : Edward Elgar . Foster L. , Haltiwanger J. , Krizan C. J. ( 2002 ) The link between aggregate and micro productivity growth: evidence from retail trade. National Bureau of Economic Research Working Paper 9120. Foster L. , Haltiwanger J. , Krizan C. J. ( 2006 ) Market selection, reallocation and restructuring in the U.S. retail trade sector in the 1990s . Review of Economics and Statistics , 88 : 748 – 758 . Google Scholar CrossRef Search ADS Gertler M. , Gilchrist S. ( 1994 ) Monetary policy, business cycles, and the behavior of small manufacturing firms . Quarterly Journal of Economics , 109 : 309 – 340 . Google Scholar CrossRef Search ADS Gould E. D. , Pashigian B. P. , Prendergast C. J. ( 2005 ) Contracts, externalities, and incentives in shopping malls . Review of Economics and Statistics , 87 : 411 – 422 . Google Scholar CrossRef Search ADS Greenstone M. , Mas A. , Nguyen H.-L. ( 2014 ) Do credit market shocks affect the real economy? Quasi-experimental evidence from the great recession and ‘normal’ economic times. National Bureau of Economic Research Working Paper 20704. Groen J. , Kutzbach M. , Polivka A. ( 2013 ) Storms and Jobs: The Effect of Hurricanes on Individuals’ Employment and Earnings over the Long Term. Unpublished paper, U.S. Census Bureau. Haltiwanger J. , Jarmin R. , Miranda J. ( 2013 ) Who creates jobs? Small vs. large vs. young . Review of Economics and Statistics , 95 : 347 – 361 . Google Scholar CrossRef Search ADS Harding J. P. , Rosenblatt E. , Yao V. W. ( 2009 ) The contagion effect of foreclosed properties . Journal of Urban Economics , 66 : 164 – 178 . Google Scholar CrossRef Search ADS Haskel J. , Sadun R. ( 2009 ) Entry, exit and labour productivity in UK retailing: evidence from micro data. In Jensen J. B. , Dunne T. , Roberts M. J. (eds) Producer Dynamics: New Evidence from Micro Data . University of Chicago Press , Chicago. Hosono K. , Miyakawa D. , Uchino T. , Hazama M. , Ono A. , Uchida H. , Uesugi I. ( 2012 ) Natural Disasters, Damage to Banks, and Firm Investment, Gakushuin University, Tokyo. Unpublished Paper. Jarmin R. S. , Miranda J. ( 2002 ) The Longitudinal Business Database. Unpublished Paper, U.S. Census Bureau. Jarmin R. S. , Miranda J. ( 2009 ) The Impact of Hurricanes Katrina, Rita and Wilma on Business Establishments . Journal of Business Valuation and Economic Loss Analysis , 4 : article 7. Kast S. ( 2005 ) Disaster bridge loan deadline extended to Jan. 31 for southernmost counties. US Fed News. Kunreuther H. ( 1996 ) Mitigating disaster losses through insurance . Journal of Risk and Uncertainty , 12 : 171 – 187 . Google Scholar CrossRef Search ADS Kunreuther H. ( 2006 ) Disaster mitigation and insurance: learning from Katrina . Annals of the American Academy of Political and Social Science , 604 : 208 – 227 . Google Scholar CrossRef Search ADS Lamont O. ( 1997 ) Cash flow and investment: evidence from internal capital markets . Journal of Finance , 52 : 83 – 109 . Google Scholar CrossRef Search ADS Mian A. , Sufi A. ( 2014a ) House of Debt . Chicago, IL : University of Chicago Press . Mian A. , Sufi A. ( 2014b ) What explains the 2007–2009 drop in employment? Econometrica , 82 : 2197 – 2223 . Google Scholar CrossRef Search ADS Pashigian B. P. , Gould E. D. ( 1998 ) Internalizing externalities: the pricing of space in shopping malls . Journal of Law and Economics , 41 : 115 – 142 . Google Scholar CrossRef Search ADS Rosenthal S. S. , Strange W. C. ( 2003 ) Geography, industrial organization, and agglomeration . Review of Economics and Statistics , 85 : 377 – 393 . Google Scholar CrossRef Search ADS Sayre E. A. , Butler D. ( 2011 ) The Geography of Recovery: An Analysis of the Mississippi Gulf Coast after Hurricane Katrina. Unpublished paper, University of Southern Mississippi. Schumpeter J. A. ( 1939 ) Business Cycles: A Theoretical, Historical and Statistical Analysis of the Capitalist Process . New York, NY : McGraw-Hill . Schumpeter J. A. ( 1942 ) Capitalism, Socialism and Democracy . New York, NY : Harper . Sharpe S. ( 1994 ) Financial market imperfections, firm leverage, and the cyclicality of employment . American Economic Review , 84 : 1060 – 1074 . Small Business Administration . ( 2008 ) Annual performance report. Discussion Paper GAO-041-076, SBA. Smith J. P. ( 2012 ) Hurricane Katrina: The Mississippi Story . Jackson, MS : University Press of Mississippi . Google Scholar CrossRef Search ADS Towe C. , Lawley C. ( 2013 ) The contagion effect of neighboring foreclosures . American Economic Journal: Economic Policy , 5 : 313 – 335 . Google Scholar CrossRef Search ADS Uchida H. , Miyakawa D. , Hosono K. , Ono A. , Uchino T. , Uesugi I. ( 2013 ) Natural disaster and natural selection. Research Institute of Economy, Trade, and Industry (RIETI) Discussion Paper 12-E-062. Zhu T. , Singh V. , Dukes A. ( 2011 ) Local competition, entry, and agglomeration . Quantitative Marketing and Economics , 9 : 129 – 154 . Google Scholar CrossRef Search ADS Appendix A. Katrina’s effect on the Mississippi Coast Hurricane Katrina was the most damaging hurricane of a particularly active 2005 hurricane season. Katrina struck several locations in Florida before veering into the Gulf of Mexico and making landfall again in New Orleans on 29 August 2005 as a Category 3 hurricane. Katrina caused damage in several states, including Alabama and Florida, but the most severe damage to businesses was in Louisiana (primarily due to flooding) and along the Mississippi coast (primarily due to high winds and storm surge). In Louisiana, flood waters did not completely recede for several weeks. The quick recovery of the Mississippi coast depended heavily on two sectors, military and casinos. Uncertainty about whether Keesler Air Force Base, which was heavily damaged, would be rebuilt was resolved within three weeks of the hurricane, when Air Force Secretary Pete Geren visited the base and promised to spend a billion dollars to fully restore it. The casinos, which had been barred from land and therefore operated on floating barges, presented a bigger challenge when they threatened not to rebuild unless they were allowed on land. Their threat was heeded: a month after the storm, on 30 September 2005, a land-based casino bill made it through the Mississippi legislature (Smith, 2012, 218–231). The combination of the Federal government’s explicit commitment to rebuilding Keesler and the casino bill seemed to seal the return of the Mississippi Gulf Coast. Figure A1 shows air travel to/from Gulfport-Biloxi International Airport relative to its January 2005 level, using monthly data from the Bureau of Transportation Statistics. There was a large negative shock of approximately 120 log points in September 2005, after which the air travel recovered relatively quickly; by early 2007 it was back to national trend. Figure A1 View largeDownload slide Air travel to/from Gulfport–Biloxi vs. all other domestic travel. Figure A1 View largeDownload slide Air travel to/from Gulfport–Biloxi vs. all other domestic travel. Business recovery was also aided by a web of government programs that provided post-storm support to residents and business owners. The most substantial program directed at business owners was a loan program administered by the Small Business Administration (SBA). Access to this program was not restricted to small business and it offered lower interest rates and longer terms than conventional loans.32 In addition, Mississippi offered small businesses in the worst-hit areas a 180-day, no-interest loan program; by the end of 2005, 392 small businesses had taken loans totaling over $9 million under this program (Kast, 2005). Smith (2012) reports that ‘by January 2006, three of the 13 destroyed casinos were back in business [….] Seven more casinos were scheduled to reopen by the end of 2006’ (231). The large investments these casinos made clearly signals that they, like the rest of Mississippi, expected a swift return to pre-Katrina conditions. Mississippi’s population remained largely in place in the aftermath of Katrina. Table A1 lists the 2000 and 2010 population in the affected counties and the rest of the state. Population changes between 2000 and 2010 were generally modest in Mississippi. The only exception is for one of the damaged counties, Stone County, which saw a population gain of nearly 27%. Stone County is very small, however, and accounts for little economic activity; fewer than 0.5% of our observations are from Stone County. In addition, the local unemployment rate, which rose in Hancock, Harrison and Jackson counties in 2005 and 2006, had returned to its pre-storm levels by 2007 (Sayre and Butler, 2011). Table A1 Population of selected Mississippi counties 2000–2010 2000 2010 Log County Population Population Change Hancock 42,967 43,929 +2.2% Harrison 189,601 187,105 –1.3% Jackson 131,420 139,668 +6.1% Stone 13,622 17,786 +26.7% Rest of state 2,467,048 2,578,809 +4.4% 2000 2010 Log County Population Population Change Hancock 42,967 43,929 +2.2% Harrison 189,601 187,105 –1.3% Jackson 131,420 139,668 +6.1% Stone 13,622 17,786 +26.7% Rest of state 2,467,048 2,578,809 +4.4% Source: Authors’ calculations from population census, 2000 and 2010. Table A1 Population of selected Mississippi counties 2000–2010 2000 2010 Log County Population Population Change Hancock 42,967 43,929 +2.2% Harrison 189,601 187,105 –1.3% Jackson 131,420 139,668 +6.1% Stone 13,622 17,786 +26.7% Rest of state 2,467,048 2,578,809 +4.4% 2000 2010 Log County Population Population Change Hancock 42,967 43,929 +2.2% Harrison 189,601 187,105 –1.3% Jackson 131,420 139,668 +6.1% Stone 13,622 17,786 +26.7% Rest of state 2,467,048 2,578,809 +4.4% Source: Authors’ calculations from population census, 2000 and 2010. Appendix B. Additional figures Published by Oxford University Press 2017. This work is written by US Government employees and is in the public domain in the US.

Journal

Journal of Economic GeographyOxford University Press

Published: Aug 24, 2017

There are no references for this article.

You’re reading a free preview. Subscribe to read the entire article.


DeepDyve is your
personal research library

It’s your single place to instantly
discover and read the research
that matters to you.

Enjoy affordable access to
over 18 million articles from more than
15,000 peer-reviewed journals.

All for just $49/month

Explore the DeepDyve Library

Search

Query the DeepDyve database, plus search all of PubMed and Google Scholar seamlessly

Organize

Save any article or search result from DeepDyve, PubMed, and Google Scholar... all in one place.

Access

Get unlimited, online access to over 18 million full-text articles from more than 15,000 scientific journals.

Your journals are on DeepDyve

Read from thousands of the leading scholarly journals from SpringerNature, Elsevier, Wiley-Blackwell, Oxford University Press and more.

All the latest content is available, no embargo periods.

See the journals in your area

DeepDyve

Freelancer

DeepDyve

Pro

Price

FREE

$49/month
$360/year

Save searches from
Google Scholar,
PubMed

Create lists to
organize your research

Export lists, citations

Read DeepDyve articles

Abstract access only

Unlimited access to over
18 million full-text articles

Print

20 pages / month

PDF Discount

20% off