Robustness reproducibility of "Improving workplace cimate in large corporations: a clustered randomized intervention"

Alan et al. (2023) carry out a field experiment where they randomly allocate 20 corporations in Turkey to a treatment group or a control group. White-collar employees at the headquarters of the corporations are invited to participate in a training program to improve the workplace environment. They report that the program reduces separation (workers quitting) and improves prosocial behavior, workplace quality and support networks. We test the robustness reproducibility of these results, focusing on the results reported in Table 8 of the original paper. We first successfully reproduce the results in Table 8 computationally based on the posted code and data, and we then carry out five robustness tests. We do not find robust support for an effect of the treatment on any of the four primary outcome variables (separation, prosocial behavior, workplace quality and support networks). The relative effect size of the robustness tests averaged across the primary hypotheses is 0.62, suggesting some inflation in the original effect sizes. The effects reported in the paper are driven by the additional employees added to the sample about one year after the initial baseline data collection and after the randomization of firms to treatment and control (and this sample is not balanced on observables across the treatment and control group). Not having access to the raw data limited the possible robustness tests.