Abstract
Researchers are often faced with evaluating the effect of a policy or program that was simultaneously initiated across an entire population of units at a single point in time, and its effects over the targeted population can manifest at any time period afterwards. In the presence of data measured over time, Bayesian time series models have been used to impute what would have happened after the policy was initiated, had the policy not taken place, in order to estimate causal effects. However, the considerations regarding the definition of the target estimands, the underlying assumptions, the plausibility of such assumptions, and the choice of an appropriate model have not been thoroughly investigated. In this paper, we establish useful estimands for the evaluation of large-scale policies. We discuss that imputation of missing potential outcomes relies on an assumption which, even though untestable, can be partially evaluated using observed data. We illustrate an approach to evaluate this key causal assumption and facilitate model elicitation based on data from the time interval before policy initiation and using classic statistical techniques. As an illustration, we study the Hospital Readmissions Reduction Program (HRRP), a US federal intervention aiming to improve health outcomes for patients with pneumonia, acute myocardial infraction, or congestive failure admitted to a hospital. We evaluate the effect of the HRRP on population mortality among the elderly across the US and in four geographic subregions, and at different time windows. We find that the HRRP increased mortality from pneumonia and acute myocardial infraction across at least one geographical region and time horizon, and is likely to have had a detrimental effect on public health.