Investigating Treatment Effects in a Domestic Violence Experiment with Partially Missing Outcome Data

This paper considers the problem of estimating the magnitude of a treatmenteffect in a randomized experiment where the outcome is missing for somecases. The primary concern in such situations is that the distribution ofthe outcome variable may vary in important ways between individuals whoseoutcomes are observed and individuals whose outcomes are missing. Since thedata cannot be used to resolve this concern, it is necessary to take theuncertainty that is created by the missing data into account when developinginferences about the magnitude of the treatment effect. This paper considersa modeling framework that accomplishes this objective. Then, the proposedframework is applied to a study of the effectiveness of different types ofpolice responses to spouse assault incidents in Charlotte, North Carolina.