Multivariate matching with non-normal covariates in observational studies
A new family of distances is generated by relating the estimated treatment effect to the squared bias of such an estimate. This family arises by incorporating prior beliefs about the relationship between the matching covariates and their importance in predicting the treatment effect. The resulting f...
Saved in:
| Main Author: | |
|---|---|
| Corporate Author: | |
| Other Authors: | |
| Format: | Thesis Book |
| Language: | English |
| Subjects: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | A new family of distances is generated by relating the estimated treatment effect to the squared bias of such an estimate. This family arises by incorporating prior beliefs about the relationship between the matching covariates and their importance in predicting the treatment effect. The resulting family of distances is broad enough to include both Mahalanobis distance and propensity score distance, as well as new distances that address the specific problems studied in the thesis A new matching construction, called Rank Propensity plus Variable matching, is derived from this distance family. A large computer simulation examines the behavior of this method relative to conventional matching methods. In low dimensions, this method is superior to Mahalanobis distance matching and propensity score matching, while in high dimensions this method is superior when the correlation among the covariates is high. In high dimensions with low or moderate correlation, propensity score matching based on ranks (instead of the conventional maximum likelihood estimate) is found to be the superior method. In summary, large improvements over conventional matching methods are found for the sampling situations studied A new measure of multivariate dispersion, the Normal/T Dispersion MLE, is derived to address the outlier problem. This estimator, based on a joint multivariate normal/multivariate t distribution, is the maximum likelihood estimator for the covariance matrix of the underlying normal distribution and is found through an implementation of the EM Algorithm. Unlike other robust dispersion estimates, this estimator does not require the underlying data distribution to be elliptically symmetric Multivariate matching is used to remove bias between treatment and control groups in observational studies. The conventional matching distances used, Mahalanobis distance and propensity score distance, are shown to behave poorly in the presence of outliers, rare binary variables, and collinearity. These three problems are specifically addressed in the thesis, and new matching techniques are developed which are superior to the conventional methods in these settings |
|---|---|
| Item Description: | Source: Dissertation Abstracts International, Volume: 58-11, Section: B, page: 6046 Supervisor: Paul Rosenbaum |
| Physical Description: | 322 pages Also available in print |
| Format: | Mode of access: World Wide Web |
| ISBN: | 9780591660319 |
| Access: | Restricted for use by site license |