|
Activity Number:
|
427
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Wednesday, August 5, 2009 : 8:30 AM to 10:20 AM
|
|
Sponsor:
|
Section on Bayesian Statistical Science
|
| Abstract - #305392 |
|
Title:
|
A Latent Class Mixture Model to Impute Missing Covariates in Observational Studies
|
|
Author(s):
|
Robin Mitra*+ and Jerome Reiter
|
|
Companies:
|
Southampton Statistical Sciences Research Institute and Duke University
|
|
Address:
|
Southampton Statistical Sciences Research Institut, Southampton, International, SO17 1BJ, United Kingdom
|
|
Keywords:
|
Data augmentation ; Latent variable ; Multiple imputation ; Propensity score
|
|
Abstract:
|
Propensity score matching is a commonly used approach to estimate treatment effects in observational studies. Estimation of the propensity scores is complicated if there is missing covariate data. In this talk we present an approach to estimate propensity scores with missing covariate data in observational studies. We propose to multiply impute the missing data, from which propensity scores can be estimated, using a general location mixture model with two latent classes. One latent class comprises the control units with a similar covariate distribution to the treated units covariates with the other class for all other controls. This formulation reduces the influence of outlying control records on estimation of parameters in the imputation model and can lead to better balanced matched controls. The benefits of this approach are illustrated with simulations and with a breastfeeding study.
|
