ABSTRACT:Many social experiments are run in multiple waves, making it possible
to adapt the sampling design in later stages to allow for more
efficient estimation of causal effects. We consider the design of a
two-stage experiment for estimating an average treatment effect, when
covariate information is available for experimental subjects.
Treatment assignment probabilities can be chosen in the second stage
based on data from the first stage. This amounts to choosing the
``propensity score,'' the conditional probability of treatment given
covariates. We propose to select the propensity score in a way that
minimizes the asymptotic variance bound for estimating the average
treatment effect, show how to implement this numerically using
standard statistical software, and derive large-sample properties of
our resulting estimator. We also extend our results to setting with
noncompliance. In this case, the goal is to estimate a local average
treatment effect, and we show how to construct an adaptive procedure
for estimating the LATE.