This paper looks at the effect of technological and organisational changes on the probability for workers in the sec‑ ond part of their careers of transmitting their knowledge to other colleagues in their employing firm. We use matched employer‑ employee data to link changes occurred at the firm level with knowledge transmission behaviours meas‑ ured at the individual‑ level. To control for selection bias based on differences in observable characteristics between workers employed in changing work environments and those employed in non‑ changing ones, we apply propensity score matching techniques. We find that ICT and management changes reduce significantly the probability for work ‑ ers over 45 of transmitting their knowledge to their colleagues. Then, we analyse the role of training in mitigating this negative impact. To address issues of self‑ selection into training, we use propensity score matching methods and a proxy for unobservable productivity. We show that participation in a training program regarding ICT tools may help older workers restore their role of knowledge transmitters. Keywords: Older workers, Knowledge transmission, Skill obsolescence JEL Classification: J14, J24 role of older workers in the process of knowledge transfer 1 Introduction within organisations. The rapid ageing of the population in most developed The goal of this paper is first to understand why the par - countries urges companies to find new Human Resource ticipation to knowledge transfer strongly declines in the management strategies for a successful integration of second part of careers and to address which are the main a more age-diverse workforce. In this setting, European factors behind it. In what follows, we put forward the Union employers and trade unions have negotiated in role of Information and Communication Technologies March 2017 a framework agreement on active ageing as (ICT hereafter) and management changes. Some empiri- well as an inter-generational approach. This agreement cal evidence show that these innovations accelerated the has two main goals: improving the ability of workers of all obsolescence of specific skills acquired by senior workers ages to remain healthy and active in work until the legal (De Grip and Van Loo 2002) and affected them negatively retirement age and facilitating transfers of knowledge through adaptability requirements (Aubert et al. 2006; and experience between generations. However, recent Bartel and Sicherman 1993; Greenan et al. 2014). New studies that have examined the question of intergenera- management practices that have often accompanied the tional knowledge transmission, mainly in terms of men- introduction of ICT over the past three decades signalled toring, highlight a striking fact: the under-representation a move towards multi-skilling, greater autonomy and a of workers aged over 45 among mentors (Masingue 2009; constant redefinition of tasks to be performed (Greenan Molinié and Volkoff 2013). This leads us to investigate the and Mairesse 1999). This may suggest that workers in the second part of their career progressively lost their role of *Correspondence: pierre‑jean.messe@univ‑lemans.fr knowledge and experience transmission in the dynamic Le Mans University, GAINS, TEPP‑ CNRS, LEMNA, Le Mans, France Full list of author information is available at the end of the article © The Author(s) 2018. This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creat iveco mmons .org/licen ses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. Greenan and Messe J Labour Market Res (2018) 52:6 Page 2 of 16 work environment of the most technology advanced workers may become obsolete while others remain valid firms. and valuable for the organisation. Hence, training older We test this assumption using a French matched workers to update their obsolete skills is a way to main- employer–employee survey on organisational changes tain access to those skills that still contribute to the and computerisation (COI) conducted in 2006. Inter- knowledge base of production. In doing so, these older viewed workers declare how frequently they show work workers remain integrated to the process of knowledge practices to other colleagues or help them when they transmission within the organisation. Our empirical encounter problems. Even though respondents do not results support this theory: using propensity score directly report whether they are mentors or not, we can matching techniques and controlling for unobserved identify the workers who transmit their skills informally individual ability by a proxy, we show a positive effect of within the firm. older workers’ participation in a training program regard- In addition, interviewed firm representatives report ing the use of new ICT tools on their probability of trans- about the introduction of modern management tools and mitting their skills to other colleagues. This effect is ICT equipment in their organisation, at the time of the stronger in firms with ICT and management changes. survey and 3 years before from retrospective questions. The remainder of our paper is organised as follows. In This allows us to control for changes that occurred within the next section, we briefly summarize the existing litera - the work environment. To make the workers employed ture on intergenerational skills transmission. We present in changing and non-changing firms comparable in data and descriptive statistics in Sect. 3 and we discuss terms of observable characteristics, we rely on propen- our empirical strategy in Sect. 4. We describe our results sity score matching techniques (Rosenbaum and Rubin in Sect. 5 looking first at effects of technological and 1983). As matched employer–employee survey have a organisational changes on the probability of transmitting complex sampling design, we rely on the recent literature knowledge and then assessing how training may interact regarding the application of propensity score matching to with this effect. Section 6 concludes. complex surveys (Austin et al. 2016; DuGoff et al. 2014; Zanutto 2006). We show that ICT and management 2 The transmission of vocational skills changes have a negative and significant effect on the between generations of workers: a brief probability of transmitting vocational skills among more literature review experienced worker. The management literature has thoroughly studied the In a second part of the paper, we examine whether interactions between workers of different ages or experi - training mitigates this negative effect. Indeed, as the ence levels, particularly in terms of mentor–protégé rela- underlying mechanism is the acceleration of skill obsoles- tionships (Ragins and Kram 2007). The use of the word cence in changing firms, training may contribute to “mentor” has been widely discussed, putting forward the updating older workers’ skills. However, one could won- difference between a sponsor, whose role is only to sup - der why this may increase their probability of transmit- port the career of the protégé, and a mentor who may ting their knowledge. In other words, since ICT and also provide to her protégé a psychosocial assistance management changes have depreciated their skills, work- (Chao 1998). For the employer, setting mentor–pro- ers in the second part of their careers would not have any tégé relationships may help professionals learn technical valid knowledge to transmit. We rely here on the task- knowledge and organisational ropes as well as improve based approach literature, which relates the tasks per- managerial talent (Kram 1988). In addition, for the men- formed on the jobs to the skills needed to carry them out. tor and the protégé, mentorship is a tool of professional A job is synthetized by a bundle of tasks and required development (Kram 1988), and improves work satisfac- skills. As noted by Green (2012), the frontier between tion (Hunt and Michael 1983). tasks and skills affected by the development of ICTs and Unlike the management literature, ergonomics focuses those who are not remains difficult to draw. Jobs rede - on the quality of the transmission and on how this activ- signs depend on how management decides to take advan- ity works out. Lefebvre et al. (2003) show that several tage of the new opportunities that come with ways of transmitting vocational skills exist. The trainer technological progress. In the concomitant tasks recon- may show some practices and provide explanations, she figuration, part of the skills accumulated by experienced may give advice when the trainee encounters a difficult problem or she can also leave her alone, checking the quality of the work done and then giving some feedback. As the trainer has to combine the transmission activity Hereafter, we refer to these workers as internal trainers. with her other daily tasks, time constraints may harm the See Autor (2013) for a survey and Görlitz and Tamm (2016) for a task- quality of knowledge transmission (Thébault et al. 2012). based approach applied to training issues. Greenan and Messe J Labour Market Res (2018) 52:6 Page 3 of 16 In addition, high job rotation may make transmission of where 14,301 workers employed in 6385 firms with more skills harder (Gaudart et al. 2008). than 20 workers in the commercial sector have been Economics has not paid much attention to the role of interviewed. This data set provides detailed information informal transmission of skills in the accumulation of the on demographic and economic characteristics of human capital of workers (Becker 1962). This literature respondents, on their working conditions and on their mainly deals with the optimal amount of investment or firm. As we are interested in knowledge transmission with the distinction between specific and general human within the firm, we exploit information about the interac - capital. Only a few papers have investigated the charac- tions that respondents have with their colleagues to build teristics of the trainers. Yet distinguishing internal (on our dependent variable, i.e. the probability of being an the job) training and off-the-job training (carried out by internal trainer. More precisely, the survey asks workers external training centres) seems to be a key point when the following three questions: “how often do they show studying the learning process. The literature highlights some work practices to their colleagues?”; “how often do heterogeneous returns across both types of training in they help some colleagues when they encounter rela- terms of wages (Kuckulenz and Zwick 2005; Lynch 1992) tional problem with other team members or customers?”; or of firm’s profitability (Bishop 1994; Black and Lynch “how often do they help some colleagues who encounter 1996). Although new workers may acquire skills through technical problems?” From this set of questions, we experience or learning by doing, facilitating knowledge define an internal trainer in the following way: a worker transmission helps to learn how to perform complex who carries out each of these activities at least 2–3 times tasks (learning by watching). In this respect, Bishop a year and at least one activity 2–3 times a month. Using (1991) show that informal training by co-workers or this definition, we consider a multifaceted aspect of training by watching others have positive and significant knowledge transmission. In Sect. 5.3, we test the sensitiv- effects on productivity during the first year of employ - ity of our results to an alternative definition. ment. Liu and Batt (2007) put forward that returns to We use the employer section of the COI survey to informal training depend on whether trainers are super- measure technological or organisational changes in the visors or experienced co-workers. work environment. As shown in Table 1, firm represent - Garicano (2000) and Garicano and Hubbard (2005) atives report about the use of 15 ICTs and 13 manage- have studied the drivers of social learning within an ment tools at the date of the survey and 3 years before. organisation. They show that a knowledge-based hierar - Regarding ICTs, we observe a growing share of firms chy is an optimal way to foster transmission of skills. In that use networking tools. For example, the share of their model, workers are assigned to productive tasks and firms equipped with a website rose from 61.2% in 2003 may ask managers to help them when they encounter a to 73.3% in 2006. In addition, 47.9% of firms had estab - problem they cannot solve. Managers’ role is to provide lished an intranet network in 2003, 57.8% 3 years later. with solutions and to learn how to solve harder problems. Less familiar ICT equipment, like tools for interfacing However, their framework does not account for differ - databases, for automated data archiving or collaborative ences in age or experience between the workers. Rufini tools, also experienced a significant boom between 2003 (2008) filled this gap, considering young workers that and 2006. may either learn on their own or benefit from knowl - Tools for managing external relationships like contrac- edge transmission. In her model, experienced workers tual relationships with suppliers or customers or quality are entrusted with transmitting vocational skills to new workers. Here, experience is a process that builds over time and implies that workers with longer tenure are A list of employers has first been randomly selected and postal question- more likely to transmit their skills. However, her model naire were sent in 2006. Responding firms have then been identified in a does not explain why older workers are actually under- linked employer/employee register from which small random samples of represented among the trainers since it does not account employees (between two and fifteen) have been selected. Then telephone or face to face interviews with employees took place. for technological and organisational changes that may The employer and employee sections of the COI survey have been depreciate the value of specific skills accumulated with matched with other administrative data sources. In this paper, we work with experience. establishment level information used to compute employers’ social contri- bution (Déclaration Annuelles de Données Sociales, DADS) from which we are able to describe the age structure of the company. Furthermore, employ- 3 Data and descriptive statistics ees born in October of an even year (or otherwise in October) have been 3.1 Data and measurement selected within each company, in order to retrieve their work history from To perform our empirical study, we use a matched the DADS administrative panel where employees are sampled according to the same method. employer–employee survey on organisational changes Note that these actions correspond to different transmission strategies and computerisation (COI hereafter) conducted in 2006 described by Lefebvre et al. (2003). Greenan and Messe J Labour Market Res (2018) 52:6 Page 4 of 16 Table 1 Use of ICTs or management tools in 2003 and 2006 % of firms 2003 2006 2006 base metric ICTs Software or firmware for HRM 63.4 65.3 0.064 Website 61.2 73.3 0.065 Local area network 61.3 66.7 0.071 Intranet 47.9 57.8 0.084 Software or firmware for R&D 47.4 49.8 0.041 Tools for data analysis 39.5 47.1 0.065 Electronic data interchange system 36.2 45.8 0.06 Databases for HRM 34.5 38.5 0.082 Enterprise resource planning 26.6 29.6 0.059 Databases for R&D 26.1 28.8 0.075 Extranet 25.0 30.2 0.081 Tools for interfacing databases 21.1 28.6 0.087 Tools for automated data archiving or research 21.4 27.4 0.067 Collaborative tools (groupware) 15.1 21.0 0.099 Tools for process modelling (workflow) 8.8 12.7 0.111 Management tools Contractual commitment to provide a product or service or customer service within a limited time 66.1 68.5 0.087 Long‑term relationships with suppliers 51.7 54.7 0.076 Requirement for suppliers to meet tight deadlines 51.5 53.5 0.090 Quality certifications 36.3 41.4 0.092 Satisfaction surveys for customers 32.9 38.7 0.079 Teams or autonomous work groups 30.7 33.8 0.089 Tools for tracing goods or services 28.3 32.9 0.093 Tools for labelling goods or services 28.3 30.8 0.075 Call or contact centres 25.5 28.0 0.080 Just in time production 22.9 24.3 0.071 Methods of problem solving (FMEA) 17.3 20.9 0.114 Customer relationship management (CRM) 9.7 14.3 0.072 Environmental or ethical certification 9.7 12.9 0.107 Source: COI survey 2006/INSEE-DARES-CEE Coverage: Firms with 20 employees or more in the commercial sector Note: “2006 base metric” refers to the coefficients from the Multiple Correspondence Analyses in 2006. It is the reference metric used to calculate the composite indicators of intensities in use of ICT and management tools in 2003 and 2006 certifications, already well established in 2003, continued at preventing risks and improving design and processes to grow. The introduction of less established management implies a laborious procedure for designers/engineers tools is more likely to depreciate the value of older work- (Stone et al. 2005), that redefines the tasks performed in ers’ knowledge. For example, with the development of such jobs. This may have accelerated the obsolescence of Customer Relationship Management (CRM) tools (9.7% older workers’ specific knowledge about how to perform if firms in 2003, 14.3% in 2006), the information about this kind of activities before the introduction of these customers becomes mainly based on data analysis and new tools. the customer relationship is more and more automated. No tool or equipment alone can summarize the hetero- This reduces the value of information that older work - geneity of observed management strategies. We thus use ers have accumulated about their customers over time. the composite indicators built by Greenan et al. (2016) by In the same manner, the share of firms using methods multiple correspondence analysis (MCA) to synthesise for problem solving, such as the Failure Mode and Effect the intensity in use of each type of tools in 2006. MCA Analysis (FMEA) methodology (Stamatis, 2003), rose aims at producing a simplified low-dimensional repre - to 17.3% in 2003 to 20.9% in 2006. This method aiming sentation of information in the large frequency table Greenan and Messe J Labour Market Res (2018) 52:6 Page 5 of 16 where each item response, identifying whether the com- employment relationship, we control for full time or part pany uses each of the listed tools, is coded as a dummy time jobs. In addition, the management literature puts variable. The MCA generates quantitative scores, called forward that spatial proximity or job rotations facilitate dimensions, which are linear combinations of the dummy the initiation of mentor–protégé relationships. This is variables that maximise the average correlation between why we include a dummy for employees who change them. The first dimension of the MCA that takes into location frequently to perform their job and another one account the largest part of the observed heterogeneity for employees who have changed colleagues over the last reflects the intensity in use of the selected tools. We 12 months. We also control for time constraints that may interpret the vector of coefficients in the linear combina - impair the quality of transmission (Thébault et al. 2012), tion as a metric determined by the state of tools’ diffusion measured by a pace of work which is imposed by an within and between organisations in 2006. The resulting external demand needing an immediate response. indicator takes higher values when the organisation Regarding the employer, we know the sector, the size and jointly uses a larger number of new tools and/or when the age structure of each firm. Finally, the COI survey these tools are more technologically advanced. As 2006 is provides information about participation in training pro- the implicit reference year in the survey, Greenan et al. grams between 2003 and 2006. We choose to focus on (dir.) (2016) perform the MCA for this year and apply the the training programs regarding the use of new ICT underlying vector of coefficients to the 2003 data in order tools. Indeed, the required adaptation to ICT changes in to obtain two comparable indicators (expressed in base the early 2000s was no more an issue of computer literacy 2006) of the intensity in use of a given set of tools at both but that of being able to complete tasks with a computer dates. They simply compute the indicator of intensity of rather than directly. This has possibly created a situation change for each type of tools as the first difference where the use of some of the skills acquired through between these two composite indicators. We discretize experience has become conditional on mastering a new these indicators by considering as substantial changes ICT tool. that exceed about one standard deviation of their distri- Figure 1 displays the share of internal trainers for work- bution. A large proportion of firms have remained inert ers aged 25–57 by age for the two sub-samples of work- or only experienced marginal changes: 50% for ICT ers employed in changing firms and non-changing firms. changes, 60% for management changes. In the remainder We note first that the share of internal trainers increases of the study, we consider that changing firms have experi - with a peak at 45 years and then falls sharply which is in enced substantial ICT and management changes. This line with the results found by Molinié and Volkoff (2013) choice allows to focus on those structural changes that using another data source. In addition, it appears that the are the most likely to be disruptive in terms of their decrease in the share of internal trainers after age 45 is impact on the work environment of employees. Thus our stronger in changing work environments than in non- category of non-changing firms group together inert changing ones. These simple descriptive statistics may firms, firms with only marginal changes and firms with suggest that ICT and management changes contribute to substantial changes but in one type of tool only. the explanation of the fall in the proportion of internal To select the other covariates that we include in our trainers after age 45. Hereafter, we restrict our sample to model to predict the probability of being an internal We do not control the type of contract (temporary or not) given that the trainer at the end of career, we rely on the previous stud- COI survey has been conducted on a sample of workers with more than ies in management about the determinants of mentor– 1 year of tenure. protégé relationships. In our regressions, we include the 10 We have computed the shares of the workforce aged 29 and less (young gender, age, marital status, educational level, occupa- workers), aged between 30 and 45, and aged 46 and more (older workers). An age pyramid with a high share of a given age group has a share of that tional level, seniority and the quartile groups of the loga- age group which is higher than each of the two other shares. rithm of the net daily wage. To control for a potential We exclude workers aged 58 and more from our sample to attenuate a effect of health, we introduce a dummy indicating the potential selection bias. Indeed, a spike in inflows to unemployment is presence of health limitations. Regarding the observed after 57 (Baguelin and Remillon, 2014). This may be due to the generosity of unemployment insurance system for workers aged 57.5 years as they are entitled to their unemployment insurance rights without any job search obligation. This system may be viewed by both employers and employees as an early retirement scheme. The survey sampling weights are used in the analysis. Running a Probit model in a sample of individuals aged 25–57 years old For more details about the methodology, see Greenan and Mairesse and introducing dummies for different age groups (25–29, 30–34, 35–39, (2006) and the technical annex of Greenan et al. (dir.) (2016). 40–44 and 45–57), we find that being aged 45 and over reduces by 5.29% The intensity of ICT changes have slightly higher dispersion than the points the probability of being an internal trainer compared to workers aged intensity of management changes (their standard deviation is respectively 40-44 years old. This effect is statistically significant at a 5% level. When 0.22 and 0.18). We thus chose 0.20 as the benchmark to discretize both indi- looking only at firms with substantial ICT and management changes this cator. negative effect is still significant at 5% but stronger (− 9.35% points). Greenan and Messe J Labour Market Res (2018) 52:6 Page 6 of 16 yields a causal effect that we will refer to as the Average Treatment Effect on the Treated (ATET hereafter). How - ever, many other confounding factors may influence our dependent variable. To gauge the comparability of treated and non-treated individuals, Table 3 gives the distribution of each of these confounding factors for workers employed in changing firms and for those employed in non-changing ones. It also reports t-stats for differences in means as well as standardized differences . Indeed, Imbens (2015) rec- 25-26 31-32 37-38 45-46 51-52 57 Age ommends some caution when using t-statistics to check All the balance of covariates as sample size may affect them Non-changing firms and they can be non-significant even in the presence of Changing firms covariate imbalance. In Table 3, we see that the propor- changing and non-changing firms tion of internal trainers is lower in changing firms than in Fig. 1 Share of internal trainers by age in changing and non‑ non-changing ones (34.2 and 39.9% respectively) even changing firms. Source: COI survey 2006/INSEE‑DARES‑ CEE. Coverage: though the t-stat for this difference is significant only at Workers aged 45–57 years old with at least 1 year of seniority and the 10% level. t-stats are significant at the 5% level for a employed in firms with 20 workers or more. Changing firms corre ‑ small number of covariates only, mainly the gender, the spond to firms that have experienced substantial ICT and manage ‑ educational level, the age structure of the firm and the ment changes retail trade industry. However, for many other variables, the standardized difference exceeds 10%, which may indi - cate important imbalances. Consequently, as some of workers aged 45–57 to investigate how their probability these variables also affect our dependent variable we have of having a role as an internal trainer is affected by to address a potential confounding bias. changes in the work environment. 4 Empirical strategy 3.2 Descriptive statistics Propensity score methods are frequently used to address Table 2 provides descriptive statistics to compare the potential confounding in observational studies (Rosen- characteristics of the population of internal trainers baum and Rubin, 1983). The main principle is to create with those of employees who do not take up such a role. groups of treated and non-treated individuals that have Females are under-represented among internal trainers, similar characteristics using matching estimators based as well as workers with low educational or skill level and on the propensity score. In our case, we define the latter seniority, in the first wage quartile group or employed in as the conditional probability of working in a changing part-time jobs. Working conditions also have an influence firm (i.e. being treated) given a set of covariates. More on workers’ participation to knowledge transmission. formally, let X be a vector of potential confounders, P(X) Individuals who change location frequently to perform the propensity score and T the treatment indicator, P(X) their work are less likely to be internal trainers. More is given by: surprisingly, individuals who do not work under strong P(X) = P(T = 1|X) time pressure or who did not change colleagues over the last 12 months are also under-represented among inter- Matching consists in reweighting observed outcomes to nal trainers. Regarding industry-specific correlations, we make the treated and non-treated individuals comparable note an over-representation of workers employed in the in terms of propensity score. There are many types of building industry among internal trainers. matching estimators, among others inverse-probability Our main variable of interest is a dummy T indicating weighting, kernel matching, nearest neighbour substantial ICT and management change at the firm level. Following the potential outcomes literature (Rosenbaum 13 Difference in means between the treated and the control group, divided and Rubin 1983), we will refer to this variable as a treat- by half the sum of standard deviations for each group. Running a simple Probit model regressing this dependent variable on ment. We want to compare the share of internal trainers the same set of covariates and on the indicator of substantial ICT and man- among treated individuals (whose work environment has agement changes, we obtain a negative marginal effect of the latter of 6.8% been hit by substantial changes) and non-treated ones. If points significant at a 5% level. However, regression methods are in general not robust to large differences between treated and non-treated individu- there are no other factors associated with the probabil- als (Imbens 2015). That is why we investigate this effect through alternative ity of being an internal trainer, this comparison in means estimators. Proportion of internal trainers .2 .3 .4 .5 Greenan and Messe J Labour Market Res (2018) 52:6 Page 7 of 16 Table 2 Descriptive statistics of the sample All Internal trainers Not internal trainers Worker’s characteristics Demographic variables Female 0.321 0.276 0.349** Age 45–49 0.472 0.498 0.456 Age 50–54 0.309 0.297 0.317 Age 55–57 0.218 0.205 0.227 Single 0.174 0.125 0.205*** Primary education 0.214 0.140 0.260** Vocational education 0.393 0.385 0.398 High School education 0.152 0.176 0.137 Undergraduate education 0.116 0.146 0.097** Graduate post‑ graduate education 0.126 0.153 0.108** Health limitations 0.112 0.100 0.120 Job’s characteristics High‑skilled occupations 0.494 0.637 0.403*** Low‑skilled occupations 0.506 0.363 0.597*** Seniority < 10 years 0.237 0.189 0.268** Seniority 11–20 years 0.226 0.201 0.243* Seniority 21–30 years 0.330 0.388 0.294** Seniority > 30 years 0.206 0.222 0.195 Log of the daily wage Log of the daily wage first quartile group 0.234 0.140 0.293*** Log of the daily wage second quartile group 0.232 0.202 0.251** Log of the daily wage third quartile group 0.227 0.254 0.210* Log of the daily wage fourth quartile group 0.307 0.404 0.245*** Part‑time work 0.076 0.052 0.091*** Working conditions Has to change location frequently 0.189 0.161 0.207** No external demand needing immediate response 0.524 0.448 0.573*** No change in colleagues over the last 12 months 0.567 0.456 0.638*** Firm’s characteristics High share of young workers (< 30 years) 0.301 0.300 0.302 High share of workers aged 30–45 years 0.500 0.489 0.506 High share of older workers (> 45 years) 0.199 0.211 0.192 Firm size 20‑49 0.148 0.161 0.140 Firm size 50–299 0.263 0.242 0.276* Firm size > 300 0.589 0.597 0.584 Manufacturing 0.400 0.390 0.408 Building 0.065 0.079 0.056** Retail trade 0.171 0.152 0.183* Transports 0.094 0.086 0.099 Housing and finance 0.164 0.196 0.144 Media and services to firms 0.105 0.097 0.110 Observations 4854 1819 3035 Source: COI survey 2006/INSEE-DARES CEE Coverage: Workers aged 45–57 years old with at least 1 year of seniority and employed in firms with 20 workers or more Significance levels for difference in means of characteristics are *** p < 0.01, ** p < 0.05 and * p < 0.1 Greenan and Messe J Labour Market Res (2018) 52:6 Page 8 of 16 Table 3 Descriptive statistics of workers employed in a changing or non-changing firms Non-changing firms Changing firms Standardized difference in means (absolute value in %) Outcome Being an internal trainer 0.399 0.342* 11.86 Demographic variables Female 0.337 0.248*** 19.73 Age 45–49 0.485 0.414* 14.40 Age 50–54 0.303 0.337 7.20 Age 55–57 0.212 0.250 9.00 Single 0.171 0.186 3.97 Primary education 0.215 0.207 2.10 Vocational education 0.392 0.398 1.30 High School education 0.156 0.133 6.49 Undergraduate education 0.122 0.089** 10.68 Graduate post‑ graduate education 0.115 0.173** 16.52 Health limitations 0.115 0.098 5.43 Job’s characteristics High‑skilled occupation 0.485 0.534 9.81 Low‑skilled occupation 0.515 0.466 9.81 Seniority < 10 years 0.235 0.250 3.53 Seniority 11–20 years 0.227 0.226 0.18 Seniority 21–30 years 0.335 0.309 5.60 Seniority > 30 years 0.204 0.215 2.90 Log of daily wage first quartile group 0.241 0.199 10.09 Log of daily wage second quartile group 0.234 0.224 2.44 Log of daily wage third quartile group 0.232 0.208 5.76 Log of daily wage fourth quartile group 0.293 0.369 16.17 Part‑time work 0.077 0.072 2.13 Working conditions Has to change location frequently 0.193 0.172 5.28 No external demand needing immediate response 0.527 0.508 3.96 No change in colleagues over the last 12 months 0.570 0.556 2.75 Firm’s characteristics High share of young workers (< 30) 0.299 0.308 1.86 High share of workers aged 30–45 0.176 0.163 1.00 High share of older workers (> 45) 0.213 0.137*** 20.25 Firm size 20–49 0.153 0.124 8.49 Firm size 50–299 0.258 0.283 5.54 Firm size > 300 0.588 0.593 0.96 Manufacturing 0.394 0.430 7.37 Building 0.063 0.074 4.45 Retail trade 0.182 0.122*** 16.70 Transports 0.097 0.082 5.15 Housing and finance 0.160 0.183 6.00 Media and services to firms 0.104 0.108 1.42 Observations 3993 861 Source: COI survey 2006/INSEE-DARES-CEE Coverage: Workers aged 45–57 years old with at least 1 year of seniority and employed in firms with 20 workers or more Note: Changing firms correspond to firms that have experienced substantial ICT and management changes. Significance levels for t-stats of differences in means are *** p < 0.01, ** p < 0.05 and * p < 0.1 Greenan and Messe J Labour Market Res (2018) 52:6 Page 9 of 16 matching. As the data used here have a complex survey propensity score p(x ) for each non-treated individual design, we rely on the recent literature regarding the over the probability of being employed in a non-changing application of propensity score matching methods to firm conditional on a set of observables. The denomina - complex surveys (Austin et al. 2016; DuGoff et al. 2014; tor is a normalization that ensures that the estimated Zanutto 2006). First we include the weights as a predictor weights add up to one for non-treated individuals. While of the propensity score, without accounting for the com- it is computationally easy, there is evidence that this esti- plex survey design. Second, we estimate the ATET. mator is sensitive to large values of the propensity score More formally, let N and N denote respectively the 0 1 (Frölich 2004). It might also be more affected than other number of observations for the treated and the control estimators in case of small misspecifications of the pro - group, w the sampling weight of non-treated individu- 0,i pensity score. The kernel matching estimator is a non- als and w the sampling weight of treated ones. The 1,i parametric one that reweights non-treated outcomes ATET is equal to: according to the distance between each worker in the control group and the treated observation for which the counterfactual is estimated. Let p(x ) the propensity ATET = T w Y i 1,i i 1,i i=1 score of a treated individual i and p x its counterpart i=1 j (1) N for a non-treated individual j, the weight placed on obser- p−p i j − (1 − T )w wY i 0,i 1 i N vation j is defined by K , where K is the Kernel 0,i i=1 i=1 estimator and h is the bandwidth. The higher is the band - where w corresponds to the reweighting of non-treated width, the lower is the variance but the higher is the bias. outcomes to control for differences in propensity scores In addition, to improve the balance of covariates between treated and non-treated observations. Unbiased between treated and non-treated individuals, we com- estimate of the ATET through propensity score match- bine the IPW technique with exact matching by gender. ing methods can be obtained only if the conditional Indeed, as this variable strongly affects both the prob - independence assumption (CIA hereafter) is satisfied, ability of being an internal trainer and the probability which implies that the probability of being employed in a of working in a changing firm, and given that males and changing firm (treatment assignment) is independent of females may differ in terms of background characteris - the fact of being an internal trainer conditionally on a set tics, we match treated and non-treated individuals within of observed covariates. each stratum defined by gender and then we compute For this study, we use two different estimators: the the ATET on average. More formally, let ATET be the inverse probability weighting (IPW hereafter) and the ATET as defined by Eq. (1) but for each stratum s , where kernel matching. The former consists in replacing w for s = 1 and s = 2 correspond respectively to the males and p(x ) i females. Following Zanutto (2006) the following expres- 1−p x ( ) each non-treated individual by . The sion summarizes the estimated ATET for the whole 1−T p x j j j=1 sample: 1−p x numerator corresponds to the ratio of the estimated w i∈S Ts ATET s=1 i∈S s=1 Ts where w denotes the sampling weight for individual i and S corresponds to the set of treated individuals in stra- Ts For a discussion around the performances of these different estimators, tum s. see Huber et al. (2013). Empirical studies often pass over the survey design and the sampling weights. There are however two ways of incorporating sampling weights 5 Results when estimating the propensity score. The first one is to fit a weighted 5.1 The role of ICT and management changes regression model and the second one is to include sampling weights as an We have first to ensure that the estimated propensity additional covariate in the propensity score model. However, Austin (2016) shows that this does not significantly alter the performance of the matching scores of workers employed in changing and non-chang- estimator in terms of balancing covariates. ing firms overlap sufficiently. The most straightforward We do not use direct matching estimators, such as one-to-one or near- way to check the common support between treated and est neighbor matching. Indeed, while it is easy to perform inference for non-treated individuals is to plot the density distribu- IPW and kernel matching by bootstrap techniques that yield valid standard errors, it is not the case for direct matching estimators (Abadie and Imbens tions of the propensity scores in both groups. Figure 2 2008). Even though they suggest an alternative procedure to obtain valid shows that the overlapped region covers all the sam- standard errors for the latter group of estimators, the way of implementing ple of treated. Hence, adopting the Min–Max method such correction method in complex surveys is not clear as far as we know. Greenan and Messe J Labour Market Res (2018) 52:6 Page 10 of 16 suggested by Dehejia and Wahba 1999 will not imply any loss of observations. Then we check whether the implementation of the different matching procedures improves the balance of covariates between the treated and non-treated individu- als. Figure 3 depicts the standardized bias for each covar- iate without matching (plotted by black circles) and after applying each matching technique (crosses for IPW with exact matching by gender, diamonds for kernel match- ing and triangles for IPW matching). For easier reading, we show only the bias for explanatory variables that pre- 0 .1 .2 .3 .4 Propensity score sent a high standardized difference (10 or higher) in the Non-treated unmatched sample. Simple IPW matching and IPW com- Treated bined with exact matching by gender perform better in Fig. 2 Density distributions of the propensity scores for treated terms of balancing covariates and reducing bias. and non‑treated individuals. Source: COI survey 2006/INSEE‑DARES‑ Table 4 reports the estimated Average Treatment Effect CEE. Coverage: Workers aged 45–57 years old with at least 1 year on the Treated resulting from the different matching of seniority and employed in firms with 20 workers or more. Note: Treated individuals are workers employed in changing firms, i.e. firms techniques. The estimates range from − 0.056 using Ker- that have experienced substantial ICT and management changes. nel matching to -0.098 using IPW with exact matching Propensity score is the conditional probability of being employed in a by gender. All these effects are strongly significant. These changing firm given a set of observable characteristics results indicate that when companies implement sub- stantial ICT and management changes, the probability for older workers to transmit their skills falls by 5.6–9.8% points. Note that these results are similar to the marginal firm’s dummy is negative but small (− 0.019) and non- effect of − 6.8% points obtained with a simple Probit significant (t-stat = − 0.34). In the second sub-sample regression and to the − 5.7% points difference in means (non-participants), this effect is stronger (− 0.074) and obtained with simple descriptive statistics (Table 3). Thus significant at a 5% level (t-stat = − 2.13). This would sug - the estimated effect is quite the same whether we control gest that ICT training might help to mitigate the negative or not for the selection bias on observables. effect of ICT and management changes on the probabil - We can interpret these results as causal effects only ity of being an internal trainer. However, as training is if the conditional independence assumption (CIA) is endogenous, we have to control for selection bias both on satisfied, which implies that once controlling for selec - observables and on unobservables. This is the aim of the tion bias on observable variables, there does not remain next section. any bias on unobservable ones. In that case, this critical assumption appears to be plausible. Indeed, we measure 5.2 The role of training the changes at the firm level and the outcome variable at In this section, we examine whether participation in a the individual level. This may ensure us to address any training session regarding the use of new ICT tools may selection bias that may come from unobservable differ - help workers in the second part of their careers to keep ences between individuals. These estimates suggest that their role as internal trainers in changing work environ- substantial ICT and management changes can explain ments. We follow the same strategy as before, introduc- why older workers are less often engaged in the knowl- ing a new treatment variable equal to one if individuals edge transmission process. have benefited from a training session regarding the use Even though we cannot directly test the mechanism at of new ICT tools in 2006. We look at the ATET for the stake, we can assume that it passes through skill obsoles- whole sample and then distinguishing workers employed cence. If this is the case, training could mitigate the nega- in changing firms and those employed in non-changing tive effect of changing work environments. As a first test, ones. This distinction is important. Since we assume that we run a simple Probit model regressing the dependent training may allow older workers to remain integrated to variable on the same set of covariates, on the indicator of the knowledge transfer process through a mechanism of ICT and management changes and distinguishing work- skill updating, we expect that this effect will be higher ers who participated in a training session regarding the in firms that have changed their organisation and work use of new ICT tools and those who did not benefit from processes. One can argue that a more suitable approach this training session in 2006. In the first subsample (par - would be first to apply propensity score matching to ticipants in training), the marginal effect of the changing identify similar workers in changing and non-changing Density 0 2 4 6 8 10 Greenan and Messe J Labour Market Res (2018) 52:6 Page 11 of 16 self-selection into training may also result from some unobservable characteristics. To remove at least partially Aged 45-49 Undergraduate this bias, following the method developed by Behaghel Graduate and + and Greenan (2010), we introduce as a potential con- Female founder a proxy of unobserved individual productiv- 1st quartile of net daily wage ity, using social security records of the employees’ work 4th quartile of net daily wage history (the DADS administrative panel). Starting from High share of older workers 1976, we estimate a wage fixed effect from a Mincerian Retail trade wage regression. We estimate it in a covariance analysis of log wages controlling for education, gender, experi- 0 5 10 15 20 Standardized % bias across covariates ence, industry and time effects for the period before the Unmatched worker enters the firm that employs her/him in 2006 and Matched (IPW with exact matching) Matched (Kernel matching) that answers the firm section of the COI survey. Matched (IPW matching) The distributions of the observable characteristics of Fig. 3 Standardized differences in means of covariates between workers who participated in a training session and of individuals employed in changing firms and those employed in those who have not inform about the potential confound- non‑ changing without and with different matching techniques. ing bias when estimating the ATET for this new treat- Source: COI survey 2006/INSEE‑DARES‑ CEE. Coverage: Workers aged ment variable. Table 5 presents these distributions and 45–57 years old with at least 1 year of seniority and employed in firms with 20 workers or more. Note: Treated individuals are work ‑ reports the t-stats for differences in means and standard - ers employed in changing firms, i.e. firms that have experienced ized differences as in Table 3. The proportion of internal substantial ICT and management changes. We plot standardized trainers is sharply higher among treated workers than biases without matching (black circles) and with different matching among non-treated ones (53.8 and 37.3% respectively, i.e. techniques (black crosses for IPW with exact matching by gender, a difference in means of 16.5% points). However, stand - blue diamonds for Kernel matching and red triangles for IPW match‑ ing). We report only covariates for which standardized bias without ardized bias reaches a very high level for many variables, matching is greater than 10 especially occupational level, educational levels and wage quartile groups. This may show that training participa - tion is very selective and targeted mostly on high-skilled and more productive workers. Furthermore, the Mince- Table 4 Average effect of working in a changing firm rian wage fixed effect is only 0.016 for untrained workers on the probability of being an internal trainer and 0.155 for trained ones. This shows clearly that high- Propensity score method Average treatment ability workers have a higher probability of participating effect on the treated in training. Therefore, we cannot interpret the simple raw difference in means as a causal effect and we have to IPW matching with exact matching by − 0.098** (0.019) gender address again selection bias based at least on observables IPW matching − 0.058** (0.019) and on our proxy for unobservable productivity. Kernel matching − 0.056** (0.017) We adopt the same matching procedures as in the previous section and we display the standardized bias Number of observations 4854 without and with matching in Fig. 4 using the identical Source: COI survey 2006/INSEE-DARES-CEE symbols for each matching method as in Fig. 3. Once Coverage: Workers aged 45–57 years old with at least 1 year of seniority and employed in firms with 20 workers or more again, we see that implementing a technique of Inverse Note Treated individuals are workers employed in changing firms, i.e. firms that Probability Weighting procedure yields the best perfor- have experienced substantial ICT and management changes. Standard errors mance in terms of reducing the bias. Using this tech- (in parentheses) have been estimated using bootstrap procedures with 500 replications. Significance levels are *** p < 0.01, ** p < 0.05 and * p < 0.1 nique, the bias concerning the occupational level shifts from 69.38% before matching to 1.63% after matching and the one concerning the Mincerian individual wage firms and second to apply it to those who participate or fixed effect shifts from 44.79 to 1.04%. do not participate in training. However, the previous sec- Table 6 shows the estimates of the ATET for dif- tion shows that selection bias into changing/non chang- ferent matching techniques for the whole sample ing firms appears to be negligible. Therefore, we address only the self-selection of workers into training based on their observable characteristics. Running a simple Probit model regressing the dependent variable on the In addition, when considering training as the treat- same set of covariates and on the indicator of participation in training, we obtain a positive marginal effect of the latter of 9.3% points significant at a ment, we could question the validity of the CIA. Indeed, 10% level. Greenan and Messe J Labour Market Res (2018) 52:6 Page 12 of 16 Table 5 Descriptive statistics of workers according to their participation in a training program regarding the use of new ICT tools Non-trained workers Trained workers Standardized difference in means (absolute value in %) Outcome Being an internal trainer 0.373 0.537*** 34.27 Demographic variables Female 0.309 0.430*** 25.93 Age 45–49 0.475 0.440 7.36 Age 50–54 0.310 0.300 2.21 Age 55–57 0.213 0.259 11.05 Single 0.176 0.150 7.41 Primary education 0.225 0.101*** 34.68 Vocational education 0.396 0.361 7.42 High School education 0.144 0.220** 20.22 Undergraduate education 0.108 0.186*** 22.85 Graduate post‑ graduate education 0.125 0.131 1.87 Health limitations 0.115 0.079* 12.69 Mincerian wage fixed effect 0.016 0.155*** 44.79 Job’s characteristics High‑skilled occupations 0.463 0.777*** 69.38 Low‑skilled occupations 0.536 0.223*** 69.38 Seniority < 10 years 0.244 0.176*** 17.08 Seniority 11–20 years 0.228 0.209 4.68 Seniority 21–30 years 0.325 0.371 9.67 Seniority > 30 years 0.201 0.244 10.37 Log of daily wage first quartile group 0.247 0.105*** 38.72 Log of daily wage second quartile group 0.231 0.239 1.76 Log of daily wage third quartile group 0.225 0.248 5.49 Log of daily wage fourth quartile group 0.295 0.409*** 24.41 Part‑time work 0.077 0.067 4.04 Working conditions Has to change location frequently 0.194 0.134* 16.83 No external demand needing immediate response 0.531 0.455 15.76 No change in colleagues over the last 12 months 0.573 0.508 13.59 Firm’s characteristics Age structure High share of young workers (< 30 years) 0.296 0.339 9.32 High share of workers aged 30–45 years 0.016 0.034 12.52 High share of older workers (> 45 years) 0.208 0.123*** 23.23 Firm size 20–49 0.148 0.150 0.72 Firm size 50–299 0.267 0.221 11.13 Firm size > 300 0.585 0.629 9.27 Manufacturing 0.404 0.368 7.71 Building 0.069 0.024** 21.89 (second column) and distinguishing workers employed in range from 0.082 to 0.091. When focusing on workers a changing firm (third column) from those employed in employed in changing firms, these effects are stronger, a non-changing work environment (fourth column). We ranging from 0.109 to 0.191. In contrast, in non-changing see first that for the whole sample, the different matching work environment, the ATET is almost three times lower techniques yield highly significant ATET estimates that ranging from 0.041 to 0.067. Overall, this means that Greenan and Messe J Labour Market Res (2018) 52:6 Page 13 of 16 Table 5 continued Non-trained workers Trained workers Standardized difference in means (absolute value in %) Retail trade 0.168 0.202 8.96 Transports 0.101 0.034*** 27.09 Housing and finance 0.153 0.265 28.71 Media and services to firms 0.105 0.107 0.85 Observations 4429 425 Source: COI survey 2006/INSEE-DARES-CEE Coverage: Workers aged 45–57 years old with at least 1 year of seniority and employed in firms with 20 workers or more Significance levels for t-stats of differences in means are *** p < 0.01, ** p < 0.05 and * p < 0.1 Table 6 Average effect of the participation in a training session regarding the use of new ICT tools on probability of being an internal trainer Average Treatment Eec ff t on the Treated All Workers employed Workers employed In changing firms In non-changing firms IPW matching with exact matching by gender 0.091*** (0.026) 0.191*** (0.066) 0.067*** (0.027) IPW matching 0.082*** (0.026) 0.168*** (0.062) 0.041*** (0.026) Kernel matching 0.087*** (0.025) 0.109*** (0.029) 0.066*** (0.028) Number of observations 4854 861 3993 Source: COI survey 2006/INSEE-DARES-CEE Coverage: Workers aged 45–57 years old with at least 1 year of seniority and employed in firms with 20 workers or more Treated individuals are workers who participate in a training session on new ICT tools in 2006. Changing firms correspond to firms that have experienced substantial ICT and management changes. Standard errors (in parentheses) have been estimated using bootstrap procedures with 500 replications. Significance levels are *** p < 0.01, ** p < 0.05 and * p < 0.1 being an internal trainer is more likely for individuals who have participated in training compared to non-par- Aged 55-57 0-10 years of tenure More than 30 years of tenure ticipants. This correlation is stronger in firms that have High-skilled Low-skilled Primary experienced substantial ICT and management changes. High-school Undergraduate Fixed mincerian effect Female Change workplace frequently Health limitations 5.3 A lternative definition of knowledge transmission 1st wage quartile 4th wage quartile No frequent change in colleagues In this study, we consider a multi-faceted definition of No time constraint High-share of mid-age workers High-share of older workers knowledge transmission. Following previous works in Construction Transport Housing and finance ergonomics (Lefebvre et al. 2003), the trainer may show 50-299 workers some practices and provide some explanations or she 0 20 40 60 80 Standardized % bias across covariates may give advice when the trainee encounters a difficult Unmatched problem. In our baseline definition, an internal trainer is Matched (IPW with exact matching by gender) a worker who shows some work practices to colleagues Matched (Kernel) Matched (IPW) and who helps them when they encounter either techni- Fig. 4 Standardized differences in means of covariates between cal or relational problems. In this sub-section, we con- trained and non‑trained individuals without and with different sider only one dimension of transmission, i.e. showing matching techniques. Source: COI survey 2006/INSEE‑DARES‑ CEE. some work practices to colleagues at least 2–3 times a Coverage: Workers aged 45–57 years old with at least 1 year of senior‑ month. This narrows down our definition of an internal ity and employed in firms with 20 workers or more. Note: Treated individuals are workers who participate in a training session on new trainer. Among individuals who correspond to our base- ICT tools in 2006. We plot standardized biases without matching line definition of internal trainer, only 76.95% show work (black circles) and with different matching techniques (black crosses practices at least 2–3 times a month. For the remain- for IPW with exact matching by gender, blue diamonds for Kernel ing 23.05% of these individuals, internal training would matching and red triangles for IPW matching). We report only covari‑ mean rather being helpful for colleagues who encounter ates for which standardized bias without matching is > 10 Greenan and Messe J Labour Market Res (2018) 52:6 Page 14 of 16 Table 7 Average treatment effects on the treated for different definitions of internal trainer Definition of an internal trainer Average treatment effect on the Average treatment effect on the treated of the participa- treated of substantial ICT and manage- tion in a training session regarding the use of new ICT ment changes tools All workers In changing firms In non-changing firms Baseline definition − 0.098*** (0.019) 0.091*** (0.026) 0.181*** (0.066) 0.067** (0.027) Only shows work practices 2–3 times a − 0.080*** (0.019) 0.091*** (0.025) 0.195*** (0.063) 0.051* (0.028) month Number of observations 4854 4854 861 3993 Source: COI survey 2006/INSEE-DARES-CEE Coverage: Workers aged 45–57 years old with at least 1 year of seniority and employed in firms with 20 workers or more In the second column, the treatment is to be employed in a changing firm i.e. a firm that has experienced substantial ICT and management changes. In the other columns, the treatment is the participation in a training session about new ICT tools in 2006. Standard errors (in parentheses) have been estimated using bootstrap procedures with 500 replications. Significance levels are *** p < 0.01, ** p < 0.05 and *p < 0.1 technical or relational problems. In addition, among probability of showing frequently some work practices to workers who do not belong to the category of internal colleagues is non-significant at a 5% level. trainers according to our baseline definition, 11.02% can be considered as internal trainers with our new defini -6 Concluding remarks tion. Their role is purely to show their work to other col - In this paper, we analyse the factors that affect the prob - leagues. Among them, around 44% do not give advice in ability for workers in the second part of their careers of case of technical problems and around 67% do not help transmitting their knowledge (being an internal trainer) colleagues who encounter relational problems with other within their employing firm. We motivate this study by team members or customers. the need for European Union social partners to iden- We check whether our main results are affected or not tify the main barriers that hamper the intergenerational when changing our dependent variable. Table 7 reports cooperation within organisations and the good prac- the ATET estimated using IPW combined with exact tices that may help facilitating it. We focus on the group matching by gender for the baseline multi-faceted defi - of workers over 45 because even though they are more nition of an internal trainer as well as the alternative sim- experienced, they are actually under-represented among pler, definition. In the second column, the treatment those who actively participate to the knowledge trans- considered is the implementation of substantial ICT and mission process. management tools when in the other columns the treat- We find that the introduction of new ICT and manage - ment is the participation in a training session for the ment tools strongly contribute to the reduction of the whole sample (column 3) and for a breakdown of workers probability of being an internal trainer after age 45. After according to whether their work environment has been applying propensity score matching techniques to make changing or not (columns 4 and 5 respectively). workers employed in changing firms and those employed Even when using the alternative definition, substantial in non-changing ones more comparable in terms of ICT and management changes reduce the probability of characteristics, this negative effect ranges from − 5.6 to showing some work practices frequently by 8% points. It − 9.8% points. This suggests that ICT and management is lower than the baseline effect (9.8 p.p.) but it remains changes have accelerated older workers’ skills obsoles- highly significant. The effect of participation in training cence. As a result, in the most dynamic work environ- for the whole sample is + 9.1 p.p. regardless of the defi - ments, older workers lose their role of knowledge and nition of an internal trainer. When distinguishing chang- experience transmitters. ing and non-changing firms and using the alternative In addition, we show that training may help mitigate definition of a trainer, the effect is slightly stronger in this negative effect. After addressing selection bias on the former case (19.5 p.p., relative to the baseline effect observables by matching techniques and using a proxy of 18.1 p.p.) but slightly weaker in the latter case (5.1 p.p. for unobservable individual productivity, we find that relative to the baseline effect of 6.7 p.p.). Note that in the the probability of being an internal trainer after age 45 case of non-changing firms, the impact of training on the is higher for individuals who have participated in train- ing session regarding the use of new ICT tools compared to non-participants. Not surprisingly, this correlation is stronger in firms that experienced substantial ICT and This estimator presents the best balancing properties. Greenan and Messe J Labour Market Res (2018) 52:6 Page 15 of 16 management changes and ranges from + 10.9 to + 19.1% Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in pub‑ points. If we could interpret these results as causal lished maps and institutional affiliations. effects, this would bring new insights on the gain that organisations would derive from training older workers Received: 9 March 2017 Accepted: 9 May 2018 after having changed their work processes and manage- ment practices. Beyond skills’ updating, it would allow maintaining the access to skills and knowledge acquired through experience and that are still valuable for the References organization. This happens when ICT and management Abadie, A., Imbens, G.W.: On the failure of the bootstrap for matching estima‑ tors. Econometrica 76(6), 1537–1557 (2008) changes affect a sub-group of tasks within those that are Aubert, P., Caroli, E., Roger, M.: New technologies, organisation and age: firm bundled into older workers’ jobs, creating a situation level evidence. Econ. J. 116(509), F73–F93 (2006) where the use of the non-affected skills becomes condi - Austin, P.C., Jembere, N., Chiu, M.: Propensity score matching and complex surveys. Stat. Methods Med. Res. 27, 1240–1257 (2016) tional on mastering a new tool or method. Autor, D.H. (2013). The” task approach” to labor markets: an overview, NBER As it stands, we have to be cautious in the interpreta- working paper No. w18711, National Bureau of Economic Research tion of these effects. Causal interpretation would need Baguelin, O., Remillon, D.: Unemployment insurance and management of the older workforce in a dual labor market: evidence from France. Labour to address more in depth the main issue of self-selection Econ. 30, 245–264 (2014) into training. Even though we use matching techniques Bartel, A.P., Sicherman, N.: Technological change and retirement decisions of and a proxy for unobserved individual productivity, the older workers. J. Labor Econ. 11(1, Part 1), 162–183 (1993) Becker, G.S.: Investment in human capital: a theoretical analysis. J Polit. latter controls only imperfectly capture workers’ unob- Economy. 70(5, Part 2), 9–49 (1962) served ability. As highlighted by Leuven and Ooster- Behaghel, L., Greenan, N.: Training and age‑biased technical change. Ann. beek (2008), one more convincing way of correcting this Econ. Stat. 99–100, 317–342 (2010). https ://doi.org/10.2307/41219 169 Bishop, J.H.: On‑the ‑job training of new hires. Market failure in training?, pp. potential bias is to narrow down the non-treated group 61–98. Springer, Berlin (1991) to non-participants who did not participate in training Bishop, J.: The impact of previous training on productivity and wages. Training due to some random event. Even though we do not have and the private sector: International comparisons, pp. 161–200. University of Chicago Press, Chicago (1994) this information in the data we use, some recent surveys Black, S.E., Lynch, L.M.: Human‑ capital investments and productivity. Am. Econ. include questions regarding the reason of non-partici- Rev. 86(2), 263–267 (1996) pation in a training session. Using surveys that combine Chao, G.T.: Invited reaction: challenging research in mentoring. Hum. Resour. Dev. Q. 9(4), 333–338 (1998) this kind of questions with information on changes that De Grip, A., Van Loo, J.: The economics of skills obsolescence: a review. The occurred in the work environment and knowledge trans- economics of skills obsolescence, pp. 1–26. Bingley, Emerald Group mission practices should help us to check whether train- Publishing Limited (2002) Dehejia, R.H., Wahba, S.: Causal effects in nonexperimental studies: reevalu‑ ing older workers really helps them to remain integrated ating the evaluation of training programs. J. Am. Stat. Assoc. 94(448), in their firm’s knowledge transfer process. We leave this 1053–1062 (1999) issue for further investigation. DuGoff, E.H., Schuler, M., Stuart, E.A.: Generalizing observational study results: applying propensity score methods to complex surveys. Health Serv. Res. Authors’ contributions 49(1), 284–303 (2014) PJM and NG have carried out the literature review, devised the empirical Frölich, M.: Finite‑sample properties of propensity‑score matching and weight ‑ strategy, discussed the results and revised the manuscript together. NG has ing estimators. Rev. Econ. Stat. 86(1), 77–90 (2004) assembled the dataset and built the measurement frame for ICT and manage‑ Garicano, L.: Hierarchies and the organization of knowledge in production. J. ment changes and for the wage fixed effect. PJM has built the measurement Political Economy 108(5), 874–904 (2000) frame for internal trainers, worked out how to apply matching methods to Garicano, L., Hubbard, T.N.: Hierarchical sorting and learning costs: theory and complex survey data and run the regressions. All authors read and approved evidence from the law. J. Econ. Behav. Organ. 58(2), 349–369 (2005) the final manuscript. Gaudart, C., Delgoulet, C., Chassaing, K.: La fidélisation de nouveaux dans une entreprise du BTP. Approche ergonomique des enjeux et des détermi‑ Author details nants. Activités 5(5–2), 2–24 (2008) 1 2 CNAM Lirsa CEET and TEPP‑ CNRS, Noisy‑le‑Grand, France. Le Mans Univer‑ Green, F.: Employee involvement, technology and evolution in job skills: a task‑ sity, GAINS, TEPP‑ CNRS, LEMNA, Le Mans, France. based analysis. ILR Rev. 65(1), 36–67 (2012) Görlitz, K., Tamm, M.: The returns to voucher‑financed training on wages, Acknowledgements employment and job tasks. Econ. Educ. Rev. 52, 51–62 (2016) We would like to thank conference participants at JMA (2014) and EALE (2014) Greenan N., Hamon‑ Cholet S., Ughetto P. (dir.) (2016) Salariés du public, salaries and two anonymous referees for their helpful comments and suggestions. du privé face aux changements, Collection Conception et dynamique des Any remaining errors are ours. organisations, L’Harmattan Greenan, N., Mairesse, J. (1999). Organizational change in French manufactur‑ Competing interests ing: what do we learn from firm representatives and from their employ‑ The authors declare that they have no competing interests. ees? NBER working paper No. w7285, National Bureau of Economic Research Ethics approval and consent to participate Not applicable. Greenan and Messe J Labour Market Res (2018) 52:6 Page 16 of 16 Greenan, N., Mairesse, J.: Les changements organisationnels, l’informatisation Lynch, L.M.: Private‑sector training and the earnings of young workers. Am. des entreprises et le travail des salariés. Revue Économique 57(6), Econ. Rev. 82(1), 299–312 (1992) 1137–1175 (2006) Masingue, B. (2009). Seniors tuteurs: comment faire mieux?. Rapport au Secré- Greenan, N., Narcy, M., Volkoff, S.: Aging, changes and quality of working life. In: taire d’État chargé de l’emploi Korunka, C., Hoonakker, P. (eds.) The Impact of ICT on quality of working Molinié, A.F., Volkoff, S. (2013). “ Avoir un rôle de tuteur…”: qui et dans quel life, Chapter 10, pp. 163–175. Springer, Berlin (2014) travail?. Connaissance de l’emploi no 101 Huber, M., Lechner, M., Wunsch, C.: The performance of estimators based on Ragins, B.R., Kram, K.E.: The handbook of mentoring at work: theory, research, the propensity score. J. Econom. 175(1), 1–21 (2013) and practice. Sage, Thousand Oaks (2007) Hunt, D.M., Michael, C.: Mentorship: a career training and development tool. Rosenbaum, P.R., Rubin, D.B.: The central role of the propensity score in obser‑ Acad. Manag. Rev. 8(3), 475–485 (1983) vational studies for causal effects. Biometrika 70(1), 41–55 (1983) Imbens, G.W.: Matching methods in practice: three examples. J. Hum. Resour. Rufini, A. The organization of social learning in firms: should it be formal or 50(2), 373–419 (2015) informal? Mimeo GREDEG (2008) Kram, K.E.: Mentoring at work: developmental relationships in organizational Stamatis, D.H.: Failure mode and effect analysis: FMEA from theory to execu‑ life. University Press of America, Lanham (1988) tion. ASQ Quality Press, Milwaukee (2003) Kuckulenz, A., Zwick, T.: Heterogeneous returns to training in personal services. Stone, R.B., Tumer, I.Y., Stock, M.E.: Linking product functionality to historic Job quality and employer behaviour, pp. 216–234. Palgrave Macmillan, failures to improve failure analysis in design. Res. Eng. Design 16(1–2), London (2005) 96–108 (2005) Lefebvre, S., Cloutier, E., Ledoux, É., Chatigny, C., Saint‑ Jacques, Y.: Transmission Thébault, J., Gaudart, C., Cloutier, E., Volkoff, S.: Transmission of vocational skills et vieillissement au travail. Vie et vieillissement 2(1–2), 67–76 (2003) between experienced and new hospital workers. Work 41(2), 195–204 Leuven, E., Oosterbeek, H.: An alternative approach to estimate the wage (2012) returns to private‑sector training. J. Appl. Econom 23(4), 423–434 (2008) Zanutto, E.L.: A comparison of propensity score and linear regression analysis Liu, X., Batt, R.: The economic pay‑ offs to informal training: evidence from of complex survey data. J. Data Sci. 4(1), 67–91 (2006) routine service work. ILR Rev. 61(1), 75–89 (2007)
Journal for Labour Market Research – Springer Journals
Published: May 28, 2018
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
Query the DeepDyve database, plus search all of PubMed and Google Scholar seamlessly
Save any article or search result from DeepDyve, PubMed, and Google Scholar... all in one place.
Get unlimited, online access to over 18 million full-text articles from more than 15,000 scientific journals.
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.
“Hi guys, I cannot tell you how much I love this resource. Incredible. I really believe you've hit the nail on the head with this site in regards to solving the research-purchase issue.”Daniel C.
“Whoa! It’s like Spotify but for academic articles.”@Phil_Robichaud
“I must say, @deepdyve is a fabulous solution to the independent researcher's problem of #access to #information.”@deepthiw
“My last article couldn't be possible without the platform @deepdyve that makes journal papers cheaper.”@JoseServera