Response to Adjisse, Blimpo, and Castañeda Dower (2024)

Adjisse, Blimpo, and Castañeda Dower (2024) reproduce the tables and figures of Balán et al. (2022) with the exception an incorrect sign on one coefficient in Table 5. Although not central to Balán et al. (2022), we are grateful to have discovered this error and have submitted a corrigendum to the AER accordingly. They then conduct three additional analyses. 1. They construct standard errors with randomization inference, which does not lead to qualitatively different results. 2. When comparing the main three treatment arms, they control for baseline trust in the chief. This covariate is Balánced for these three treatment arms, but imBalánced in a fourth treatment arm we do not study in the paper. This variable is also uncorrelated with the outcome and thus an unlikely source of omitted variable bias. By including this control, the authors restrict the analysis to the baseline survey sample, a sample size reduction of 91%, which not surprisingly increases the standard errors. In this sub-sample, the magnitude of the coefficient on one treatment indicator increases slightly, suggesting if anything stronger evidence of the informational mechanism proposed in the paper. Controlling for the average level of trust in the chief in the full sample leaves our results unchanged. 3. They examine a prediction exercise - one of four auxiliary descriptive analyses regarding the mechanism - in which we compare the properties visited by different types of tax collectors across treatment arms. They add 13 binary neighborhood-level variables to a regression predicting tax compliance at the household level. This unconventional prediction approach fits noise and worsens the prediction, introducing classical measurement error into the ultimate exercise. When we redo their check with LASSO, it drops all but two of these variables and replicates our results. Although we appreciate the interest in our paper, we question the statistical value of the latter two analyses.