To test our hypotheses, we used data from the French Organizational Changes and Computerization's (COI) 2006 survey.4 The COI survey is a matched employer–employee dataset on organizational change and computerization from the National Institute for Statistics and Economic Studies (INSEE), the Ministry of Labor, and the Center for Labor Studies (CEE). The survey contains 7700 firms, with at least 20 employees belonging to the private sector. It is a representative population of French firms from all industries except agriculture, forestry, and fishing. Each firm fills in a self-administered questionnaire concerning the utilization of information technologies and work organizational practices in 2006, and changes that have occurred in those areas since 2003.5 Firms were also interviewed on the economic goals driving the decision to implement organizational changes and the economic context in which those decisions were made.
Within each surveyed firm, employees were randomly selected and asked about their personal socio-economic characteristics, as well as information about their job and position within the organization. The labor force survey defines the employee's job duties and responsibilities at the time of the survey and provides only a few elements dealing with actual changes. In our sample, the respondents are associated with the following departments: 46 per cent to general management; 32 per cent to finance and accounting; 7 per cent to human resources; 2 per cent to manufacturing, logistics, and quality; 7 per cent to information technology; and 6 per cent are classified as others.
The original dataset includes 14 369 employees. In order to obtain information on business export volumes, employee value-added activities, and earnings and wage information, the COI survey was merged with two other databases: the Annual Enterprise Survey (EAE) and the Annual Statement of Social Data (DADS). As a result of these merges, our sample includes 10 663 employees from 5220 firms.
These databases offer a propitious opportunity to examine three relationships: (i) between the firm's environmental orientation, employee training, and interpersonal contacts; (ii) between employee training and interpersonal contacts and labor productivity; and (iii) between environmental standards and labor productivity. By controlling for the organizational changes associated with the adoption of environmental standards, we sought to isolate the positive social identity effect, which implementation of environmental standards may bring about and which might lead to improved labor productivity.
We hypothesize a direct effect of the adoption of environmental standards on labor productivity, as well as mediating effects of training and interpersonal contacts. Hence, in our model, employee training and interpersonal contacts are determined by the adoption of environmental standards. We further argue that the adoption of environmental standards and the degree of training and interpersonal contact within an organization determine labor productivity.
However, the adoption of environmental standards, training, interpersonal contacts, and labor productivity can be influenced by the same variables (e.g., size, sector of activity, firm's strategy), and this may cause a spurious relationship. Thus, an OLS regression is inappropriate because it considers environmental standards adoption, training, and interpersonal contacts as exogenous.
In light of such endogeneity, we used a three-stage least square (3SLS) model (Aerts et al., 2008; Anton et al., 2004) that considers environmental standards, training, and interpersonal contacts as endogenous variables. The model relies on a simultaneous estimation approach (Pindyck & Rubinfeld, 1991), in which (i) the factors that determine environmental standards are estimated simultaneously with (ii) the factors that explain employee training or interpersonal contacts, and (iii) the factors that define labor productivity. We estimated jointly the three equations for each explanatory variable using maximum likelihood.
, , and are latent variables influencing the probability that the firm implements environmental standards; improves employee training or interpersonal contacts; and improves labor productivity, respectively. We consider the following 3SLS model:
where X1 are the vectors of exogenous variables including firm characteristics, such as export level, being a part of a holding company, size, and sector activity. In addition, we control for employee characteristics, including gender, age, education, and wage.
The vector of variable Z1 represents the vectors of instrumental variables that guarantee the identification of the model and help estimate correlation coefficients (Maddala, 1983). Hence, in order to identify the three-stage least square model, we needed additional variables that explain the probability of adopting environmental standards, but are not correlated to the error term of the labor productivity equation. In our case, Z1 indicates that the firm assured timely delivery to its customers and had a client call center in 2003.
Several rationales can explain why the client supply center and client call center variables affect environmental practices. With environmental standards, it is essential to maintain close links with customers in order to identify their needs, to receive feedback necessary for understanding if customer requirements are successfully met, and to determine whether to initiate relevant improvement activities. Hence, firms that have a close link with their customers also have strong incentives to demonstrate goodwill to their customers by implementing successful environmental management systems (Nishitani, 2009). Moreover, the literature argues that a firm that wants to deliver their products or services on time should adopt management practices, because the implementation of such practices improves delivery performance, mainly through reduction in time spent on non-value-added activities (Pekovic, 2010). We presumed that a firm's relationships with clients would not positively influence labor productivity, because scholars have identified potential tradeoffs between customer satisfaction and productivity (Anderson, Fornell, & Rust, 1997). It is worth noting that our proposed instrumental variables do not appear to be a significant determinant of training, interpersonal contacts, and labor productivity in a single equation logit or probit model.
X2 includes two sets of variables: (i) firm characteristics (export level, being a part of a holding, size, and sector activity) and (ii) socio-demographic characteristics (gender, age, age square, education, wage, seniority, occupation, and working hours).
As in the previous case, the vector of variable Z2 represents the vector of the instrumental variable that explains the probability of employee training improvement or interpersonal contacts, but is not correlated to the error term of the labor productivity equation. For employee training and interpersonal contacts, the vector Z2 includes whether the employee uses the informal pronoun “tu” when speaking to his or her superior. The choice of this variable as an instrument seems to be reasonable, because supervisors play a central role in employee work empowerment and integration (Hopkins, 2005) and in developing opportunities for employees to practice their skills (Noe, 1986).
X3 also includes two sets of variables: (i) firm characteristics (export level, being a part of a holding, size, and sector activity) and (ii) socio-demographic characteristics (gender, age, age square, education, wage, seniority, occupation, and working hours).
β1, β2, β3, γ1, γ2, γ3, δ1, δ2, and δ3 are slope coefficients to be estimated.
Finally, α1, α2, α3, μ1, μ2, and μ3 are the intercepts and the disturbance terms for the three equations, respectively.
Because our data provide information on multiple individuals within each organization, there is the potential for correlation of errors across individuals within each organization. We therefore trimmed our sample and used only a single individual respondent per firm in our estimations. As a robustness test, we conducted the analysis with all the 10 663 observations. There is no significant difference in the results between the two samples.7