CMStatistics 2017: Start Registration
View Submission - CMStatistics
Title: Tests for the unconfoundedness assumption using quasi-instruments Authors:  Emma Persson - Umeå University (Sweden) [presenting]
Xavier de Luna - Umea University (Sweden)
Per Johansson - Uppsala University (Sweden)
Abstract: In observational studies, the identification of an average causal effect of a treatment on an outcome of interest commonly relies on the unconfoundedness assumption. An alternative approach to identification is to use an instrument. When an instrument is available, it may also be used to test for the unconfoundedness assumption even in situations where the instrument does not fulfill all necessary conditions to yield nonparametric identification of the average causal effect. We propose tests for the unconfoundedness assumption using such quasi-instrumental variables, both using matching and parametric models. Our approach allows for discrete as well as continuous instruments, and is applicable to situations where the targeted parameter is the average treatments effect on the treated. We perform a simulation study to evaluate the finite sample performance and compare power with Durbin-Wu-Hausman tests. Finally we apply the results to a case study where the interest lies in evaluating the effect of cognitive behavioral therapy on three outcomes; outpatient care, sickness leave and medical drug prescriptions.