Application of multiple imputation in propensity score methods with partially observed covariates
Abstract/Contents
- Abstract
- Since its introduction in the 1980s, propensity score (PS) methods have been widely used. They have been particularly prevalent in the field of causal inference in order to draw valid inference from observational data. Specifically, PS matching and weighting strategies have been applied to a variety of studies. PS weighting has also been adapted to many other contexts, including the recent methodological developments in generalizing and transporting randomized clinical trial (RCT) results to target populations of interest, referred to as inverse probability of sampling weighting (IPSW). Application of the aforementioned approaches is not straightforward in the presence of missing data, which threatens their statistical validity. Unfortunately, missing data is inevitable in almost all biomedical studies, multiple imputation (MI) is a flexible solution for handling missing data with good statistical properties that is easily accessible in many mainstream computational software. However, one has to make a number of key choices to implement MI in applying PS-based methods that greatly impact the statistical properties of resulting estimators. The choices include which imputation model to use (variables or subpopulations), how to impute PS, how to integrate PS into analysis, and how to estimate the uncertainty of the estimated relationship of interest. We built upon previous work to evaluate novel MI strategies for two key contexts through extensive simulation studies. Specifically, we have designed and implemented Monte Carlo simulations to illustrate the heterogeneity of findings and to develop guidelines for applied methodologists. We additionally illustrated principles using two studies, the Diet Intervention Examining The Factors Interacting with Treatment Success (DIETFITS) and the Frequent Hemodialysis Network (FHN) Daily Network Trial. For both PS matching in the context of classic causal inference and PS weighting in the context of transporting RCT results to target populations of interest, we recommend 1) adopting MI-derPassive and deriving the PS after applying MI 2) implementing INT-within and conducting PS matching or weighting within each imputed dataset before averaging the treatment effects into one summarized quantity 3) estimating the uncertainty of the relationship of interest through a bootstrapped variance estimator for PS matching and a robust variance estimator for IPSW and 4) including key auxiliary variables in the imputation model when possible.
Description
| Type of resource | text |
|---|---|
| Form | electronic resource; remote; computer; online resource |
| Extent | 1 online resource. |
| Place | California |
| Place | [Stanford, California] |
| Publisher | [Stanford University] |
| Copyright date | 2020; ©2020 |
| Publication date | 2020; 2020 |
| Issuance | monographic |
| Language | English |
Creators/Contributors
| Author | Ling, Yun |
|---|---|
| Degree supervisor | Desai, Manisha |
| Thesis advisor | Desai, Manisha |
| Thesis advisor | Baiocchi, Michael |
| Thesis advisor | Chertow, Glenn M |
| Thesis advisor | Tian, Lu |
| Degree committee member | Baiocchi, Michael |
| Degree committee member | Chertow, Glenn M |
| Degree committee member | Tian, Lu |
| Associated with | Stanford University, Department of Biomedical Informatics. |
Subjects
| Genre | Theses |
|---|---|
| Genre | Text |
Bibliographic information
| Statement of responsibility | Yun Ling. |
|---|---|
| Note | Submitted to the Department of Biomedical Informatics. |
| Thesis | Thesis Ph.D. Stanford University 2020. |
| Location | electronic resource |
Access conditions
- Copyright
- © 2020 by Yun Ling
- License
- This work is licensed under a Creative Commons Attribution Non Commercial 3.0 Unported license (CC BY-NC).
Also listed in
Loading usage metrics...