Selection on Observables
Do We Need Controls?
The question that has been lurking in the back of your head up to this point is whether it would be appropriate to use a difference-in-means estimator when the selection on observable assumption holds. That is, if we can think of our data as coming from a "local" randomized control trial, does the difference in means estimator recover the average treatment effect?
Let's think through this. Applying the Law of Total Probability, we can express the first term in the estimator as follows:
The immediate issue that jumps out to us is that the conditional distribution \(p_{X \vert D=1}(x)\) may not equal the unconditional distribution, \(p_X(x)\), in which case our estimator would be biased. Another way to frame this problem is as follows: We observe a random variable following a distribution \(q(x)\), but we are interested in taking the expected value with respect to a different distribution \(v(x)\). That is:
One approach in this context is to transform the random variable via a correction term. That is, we introduce a new random variable which is a scaled version of our original random variable \(h(x)= f(x)\frac{v(x)}{q(x)}\), and take the average of this random variable with respect to \(q(x)\).
In our context, \(q(x) = p_{X \vert D=1}(x)\) and \(v(x) = p(x)\). Therefore our "correction term" is the ratio of these two values: the unconditional probability of treatment and the propensity score.
Let's now check that if we use this correction term, the difference-in-means estimator will return the average treatment effect. To do so, let \(W_i = \frac{Y_i p_D(1)}{p_{D\vert X_i}(1)}\)
Can Linear Regression be Wrong in this Context?
The above section highlights that even if we can think of our data as generated by local randomized control trails, we need to account for the relative distribution over the covariates. The most immediate question that follows is: does linear regression adjust the relative distribution?