Sensitivity Evaluation of Methods for Estimating Complier Average Causal Mediation Effects to Assumptions

Estimating the effects of randomized experiments and, by extension, their mediating mechanisms, is often complicated by treatment noncompliance. Two estimation methods for causal mediation in the presence of noncompliance have recently been proposed, the instrumental variable method (IV-mediate) and...

Full description

Saved in:
Bibliographic Details
Published in:Journal of educational and behavioral statistics 2020-08, Vol.45 (4), p.475-506
Main Authors: Park, Soojin, Palardy, Gregory J.
Format: Article
Language:English
Subjects:
Compliance (Psychology)
Computation
Depression (Psychology)
Educational tests & measurements
Job Search Methods
Maximum likelihood method
Maximum Likelihood Statistics
Motivation
Professional development
Research Design
Statistical Analysis
Statistical Bias
Training
Citations: Items that this one cites
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Estimating the effects of randomized experiments and, by extension, their mediating mechanisms, is often complicated by treatment noncompliance. Two estimation methods for causal mediation in the presence of noncompliance have recently been proposed, the instrumental variable method (IV-mediate) and maximum likelihood method (ML-mediate). However, little research has examined their performance when certain assumptions are violated and under varying data conditions. This article addresses that gap in the research and compares the performance of the two methods. The results show that the distributional assumption of the compliance behavior plays an important role in estimation. That is, regardless of the estimation method or whether the other assumptions hold, results are biased if the distributional assumption is not met. We also found that the IV-mediate method is more sensitive to exclusion restriction violations, while the ML-mediate method is more sensitive to monotonicity violations. Moreover, estimates depend in part on compliance rate, sample size, and the availability and impact of control covariates. These findings are used to provide guidance on estimator selection.
ISSN:1076-9986
1935-1054
DOI:10.3102/1076998620908599