Online Program

Selective and future ignorability in causal inference

*Marshall Joffe, University of Pennsylvania 
Wei Peter Yang, University of Pennsylvania 

Keywords: causal inference, confounding, instrumental variables

Most attempts at causal inference in observational studies are based on assumptions that treatment assignment is ignorable, or on extensions of these assumptions; ignorability involves conditional independence of treatment and the potential outcomes given a set of measured covariates. Such assumptions are usually made casually, largely because they justify the use of available statistical methods and not because they are truly believed. These assumptions will often be implausible, because adequate information on confounders is not available in the data. The inadequacy of information on confounders may result from several sources, including the failure to collect information altogether on given important confounders, and the failure to collect information on confounders adequately in subsets of the data.In recent work, we have formalized variants of ignorability assumptions, which we term selective and future ignorability, which can more correctly represent the situation obtaining in many studies. Under selective ignorability, conditional independence obtains in a known subset of the data; under future ignorability, independence of treatment and potential outcomes holds conditionally on a combination of measured covariate history and future potential outcomes. We have developed initial approaches to inference which are more appropriate than standard methods when selective and/or future ignorability conditions obtain. We outline these assumptions and sketch how they may be used along with structural nested models and semiparametric methods to obtain valid inference. We motivate and illustrate our development by considering an analysis of an observational database to estimate the effect of erythropoietin use on mortality among hemodialysis patients.