Primary Submission Category: Randomized Studies
Safety first: Design-informed inference with `propertee`
Authors: Ben Hansen, Joshua Errickson, Xinhe Wang, Joshua Wasserman,
Presenting Author: Ben Hansen*
In studies with social or medical data, units of analysis may be divisible into “clusters”, for instance students nested within classrooms or patient observations nested within patients. Despite its importance to inference, cluster composition can be hazy, even arbitrary.
In randomized controlled trials (RCTs), as well as those observational studies that model themselves on RCTs (Rubin, 2008; Hernan & Robins, 2016), investigators specify concrete units of assignment in the process of articulating the study’s design (Rosenbaum, 2010). In most analyses of such evaluation studies, the appropriate clusters are simply the units of assignment; in all such studies, appropriate clustering keeps assignment units intact.
Our “propertee” R package aims to facilitate several seemingly straightforward impact estimation tasks that can difficult to execute safely. These include making use of predictions from a model fitted to external or partially external samples, and bringing in design-based inverse probability weights. Its method includes separate elicitation of study design information, identifying the assignment units as a matter of course. This makes it easy and safe to, for example, produce Hajek estimates and associated standard errors for RCTs with large or small blocks — even in the presence of grouped assignment to treatment, irregular repeated measurements, or subgroup-level estimation with subgroups seen only under treatment in some blocks and only under control in others.