Randomization by Survey Date and Evaluation of Treatment Effect Estimators

In some surveys, the survey interview dates are scattered over many months. If a treatment unrelated to the survey purpose takes place during the survey period, then randomized data can obtain because individuals interviewed before and after the treatment timing are homogeneous. This happened in the Korea Labor Institute Panel Study (KLIPS) data for a work hour reduction law in 2004. We verify the random nature of the data and estimate the effects. This 'randomization by survey date' design may hold in other treatment effect analyses based on survey data, if the treatment timing falls in the middle of the survey sampling period. Furthermore, comparing the effect estimate based on the randomized data to those based on observational data, we find that panel data difference-in-differences (DD) performs best, closely followed by repeated cross-section DD. As long as DD is applicable, cross-section estimation with observational data using a post-treatment period had better not be used