ATE vs ATET: when should be care about them?

In Chapters 19, 21, and 22 we talk about both the average treatment effect (ATE) and the average treatment effect on the treated (ATET).  We discuss how to compute it in some cases and how to interpret them. In this short piece let us discuss two points scattered in the book: which one should we care about and what is the difference in terms of computation. 

Which one shall we care about?

 Whether ATE or ATET is the more policy relevant one depends on the situation. Who the non-treated are who are included in ATE but not ATET, and will the implementation you really care about have the same kind of individuals left out from ATET. Your slides say “gains from the intervention for those subject it was intended to, and can be compared with its costs” well that may not be true, ATET is different when there is selection, maybe that selection is different in the data than what you want it to be in the implementation you care about (e.g., want to introduce a compulsory program, data compares implementation for some subjects comparing other subjects for whom it was not implemented yet). You may also care about spillovers on non-treated ones.


This is explained in detail in Chapter 21. Basically, (just copy text) “When unweighted, this is an estimate of the average treatment effect on the treated (ATET), because each difference within the average is computed for each and every observation with x = 1. If the weights include the number of x = 0 observations, too, it’s an estimate of the average treatment effect (ATE).”


Plus, actually, you can estimate ATET by OLS in a way that is analogous to ATET vs ATE with matching: reweight observations by the distribution of confounders in the treated group. OLS estimates ATE for the population represented by the data. Weights can make that data represent the treated ones. With many confounder variables that can be cumbersome in practice just like exact matching would be. But you can approximate it by estimating the pscore and using it as a weight. BTW you can also have the pscore as your only rhs variable besides the causal variable instead of all the confounders.


As explained in Chapter 22,  “Diff-in-diffs gives a good estimate of the average treatment effect on the treated (ATET) if the averagechange among untreated subjects is a good counterfactual to the average change among treated subjects. That happens if the outcome of the treated subjects would have changed, on average, in the same way as it changed among the untreated subjects, had the intervention not taken place. In addition, diff-in-diffs gives a good estimate of the overall average treatment effect (ATE) under two conditions. First, the previously described condition for the ATET is satisfied. Second, if the average change among treated subjects is a good counterfactual to the average change among untreated subjects. This second condition is met if the outcome of the untreated subjects would have changed, on average, the same way as the outcome would have changed for the treated subjects, had they, too, been treated”

Read more

We have not dwelled on details a great deal. You may turn to Imbens for more ideas.