Will the true causal effect please stand up? A critique of using fixed-effects regression to estimate the effects of personal contacts on wages

Abstract

<p>There is an ongoing controversy about whether the correlation between job finding via personal contacts and wages reflects a causal effect. Critics such as Mouw (2003) argue that controlling for unobserved confounders, preferably by fixed-effects regressions, removes spurious correlations and reveals the actual null effect of personal contacts. More recently, however, McDonald (2015) applied fixed effects regressions and found a significantly positive effect. In this paper, we argue that both the Mouw (2003) and McDonald (2015) results are subject to sample selection. Results are potentially biased because their fixed effects estimators exclude those persons who use only one job search method. We propose difference-in-differences matching as an alternative estimator that does not induce the same sample selection bias. Re-analyzing the data used in Mouw (2003) and McDonald (2015), we find that in both cases, this alternative estimator is unbiased by unobserved confounders as well as by sample selection and gives us a causal null effect, which supports Mouw`s (2003) original argument.</p>