Testing for average treatment effects in choice-based samples

It is shown that in causal inference based on choice-based samples, the consistent estimation and t-tests of propensity scores and average treatment effects can be performed only from biased subsamples without external knowledge about the original random samples. Thus, program evaluation becomes more feasible. Empirical analysis is a policy evaluation of subsidies for solar photovoltaics in China. From an economic perspective, subsidies should have a positive effect. However, if the ratio of the treatment and control groups in the subsamples deviates significantly from the true population proportion, the estimate of the average treatment effects may be biased. In such cases, inference based on the conditional likelihood function, which accounts for this changed ratio, is preferable. Using relatively small and biased subsamples, the policy effects based on the resulting estimation method are significant and consistent with a numerical level from possible populations.