A1595
Title: Safety first: Design-informed inference for treatment effects via the propertee package for R
Authors: Joshua Errickson - University of Michigan (United States)
Joshua Wasserman - University of Michigan (United States)
Adam Sales - University of Texas - Austin (United States)
Ben Hansen - University of Michigan (United States) [presenting]
Abstract: When treatments are allocated by cluster, it is vital for correct inference that the clustering structure be tracked and appropriately attended to. In randomized trials and observational studies modeled on RCTs, clustering is determined at the early stage of study design, with subtle but important implications for the later stage of treatment effect estimation. A first contribution of our "propertee" R package is to make analysis safer by providing self-standing functions to record treatment allocations, with the thus-encoded study design informing subsequent calculations of inverse probability weights, if requested, and of standard errors. A second contribution is to facilitate the use of precision-enhancing predictions from models fitted to external or partly external samples. The user experience is kept simple by adapting familiar R mechanisms such as predict(), lm(), offset(), the sandwich package and summary.lm(); it uses stacked estimating equations under the hood. The propertee package makes it easy and safe to produce Hajek- or block fixed effect estimates with appropriate standard errors, even in the presence of grouped assignment to treatment, repeated measures, subgroup-level estimation and/or covariance adjustment.
