5.4. Robustness checks
To check the robustness of our results, we estimate the models with different specifications, samples, and estimation methods. First, we examine different specifications with smaller sets of control variables: Model 1, where neither the influence of groups or organizations nor the motivations for environmental practices are controlled for, and Model 2, where we do not control for the motivations for environmental practices. Model 3 includes all the control variables.
The results are presented in columns 3 and 4 of Table 4. In both models, the effects of ISO 14001 are found to be positive and significant. For example, in Model 2, the probability of assess (require) increases by 41.7% (51.9%) when the facility adopts ISO 14001. Actually, the sample we have used is part of the full sample; the full sample contains facilities that do not report their motivations for environmental practices. For this reason, we have more observations available for Models 1 and 2. The estimation results of Models 1 and 2 using the full sample are presented in columns 1 and 2 of Table 4, respectively. Although
the estimated effects become slightly smaller in most cases, our main results clearly do not change.�Our second robustness check involves examining whether our main findings are driven by the joint estimation. On the one hand, estimating Eqs. (1)–(3) jointly is more efficient than that of each subsystem of two equations—that is, Eqs. (1) and (3) and Eqs. (2) and (3). On the other hand, the three-equation system is less robust than each two-equation system because if there is any misspecification in the assess equation, the misspecification bias will spillover to the require equation and vice versa. This motivates us to estimate two-equation systems: a bivariate probit of Eqs. (1) and (3) and a bivariate probit of Eqs. (2) and (3). For each bivariate probit model, we examine all combinations of the same control variables described in Models 1–3 for the full sample and the subsamples. The outcome is ten bivariate probit model estimations. As Table 4 indicates, our main results appear to be robust; the estimated effects are very similar in size to the corresponding ones from the multivariate probit.
As a third robustness check, we estimate the following linear probability models: ASSESSi 1⁄4 yAISOi þdAu X1i þeiA,
REQUIREi 1⁄4yRISOiþdRu X1iþeiR,
where we treat ISOi as an endogenous variable. That is, we estimate each equation by two-stage least squares (2SLS) to determine whether the normality assumption is a reason for our main results. As Table 4 shows, the 2SLS estimates of the effects are found to be larger in all cases. These findings suggest that our main results are not driven by the normality assumption.
As discussed previously, ISOi is found to be exogenous in the require equation, and thus a univariate probit model can provide consistent estimates. The univariate results are presented in Appendix A. Although the estimated effects are smaller than those in the multivariate probit models, they are all positive and significant. The effect is estimated to be smallest when we use the subsample for Model 2. However, our results still suggest that the adoption of ISO 14001 increases the probability that facilities require their suppliers to undertake specific environmental measures, by 37.7%. Overall, our main results appear to be quite robust to different specifications, samples, and estimation methods.