Comparison of Approaches for Handling Missingness in Covariates for Propensity Score Models
Author | : Jiangxiu Zhou |
Publisher | : |
Total Pages | : |
Release | : 2015 |
ISBN-10 | : OCLC:927775580 |
ISBN-13 | : |
Rating | : 4/5 (80 Downloads) |
Download or read book Comparison of Approaches for Handling Missingness in Covariates for Propensity Score Models written by Jiangxiu Zhou and published by . This book was released on 2015 with total page pages. Available in PDF, EPUB and Kindle. Book excerpt: Causal effect estimation with observational data is subject to bias due to confounding. Although potential confounders could be adjusted for by fitting a multiple regression model, a more effective way to control for confounding is to use propensity score methods. Propensity scores are most commonly estimated from logistic regression with a binary exposure; generalized propensity scores could be estimated instead using linear regression if the exposure is continuous. One unresolved issue in propensity score estimation is handling of missing values in covariates. As covariates are only used for propensity score estimation but not for later outcome analysis, missing values in covariates may need to be handled differently from missing values in outcome analysis. Several approaches have been proposed for handling covariate missingness, including multiple imputation (MI), multiple imputation with missingness pattern (MIMP) and treatment mean imputation. There are other potentially useful approaches that have not been evaluated, including single imputation, single conditional mean imputation and Generalized Boosted Modeling (GBM), which is a nonparametric approach of estimating propensity scores and missing values are automatically accounted for in the estimation.To evaluate the performance of single imputation, single conditional mean imputation and GBM in comparison to the previously proposed approaches including treatment mean imputation, MI and MIMP, a simulation study is conducted with a binary exposure. Results suggest that when all confounders are included for propensity score estimation, single imputation, single conditional mean imputation, MI and MIMP perform almost equally well and better than treatment mean imputation and GBM. To examine whether the finding could be extended to a continuous exposure setting, another simulation study is conducted. Results suggest that single imputation, single conditional imputation, MI, MIMP and GBM with single conditional mean imputation have equally good and better performance than treatment mean imputation and GBM with incomplete data under scenario A (linearity and additivity). None of the approaches perform well under scenario G (nonlinearity and nonadditivity). These approaches are further demonstrated and compared through an empirical analysis with the Adolescent Alcohol Prevention Trial (AAPT). A similar pattern of results is observed as in the simulation study. It is recommended to impute missing covariates using different approaches and similar estimates help provide more confidence in the estimates.