I am working on thesis work for Mphil. I am interested to estimate the average treatment on treated effect (ATT) by using propensity score matcing method. The primary data on 320 respondents, out of which 160 are beneficiaries of a social cash transfer and 160 are non-beneficiaries have been targeted. The propensity scores will be calculated through binary logistic regression. Is this suitalbe to estimate ATT in this situation?
Yes, but there are some issues that may make this procedure slightly challenging. First, matching for the ATT generally involves discarding control units until a sample of control units comparable to your treated units remain. When you don't have more control units than treated units, you will need to match with replacement, meaning that you will discard some control units, but others will be given a larger weight in the effect estimation. Matching with replacement has been shown to often have poor empirical properties.
Second, the theory for estimating marginal treatment effects depends on random sampling from a population of interest. If your 360 units were selected for a specific reason, and that reason could be related to the treatment effect, then your treatment effect estimate will not generalize to a meaningful population. Ideally, you want your treatment effect to generalize to others who would take the treatment in the population, not just others who would take the treatment in your chosen sub-population. The reason I mention this is that it is unusual to have equal numbers in each group if treatment is assigned non-randomly.