|
Activity Number:
|
216
|
|
Type:
|
Topic Contributed
|
|
Date/Time:
|
Tuesday, August 10, 2004 : 10:30 AM to 12:20 PM
|
|
Sponsor:
|
Social Statistics Section
|
| Abstract - #301292 |
|
Title:
|
Imputation by Propensity Matching
|
|
Author(s):
|
Murthy N. Mittinty*+ and Easaw Chacko
|
|
Companies:
|
University of Canterbury and University of Canterbury
|
|
Address:
|
Department of Mathematics and Statistics, Christchurch, P.Bag 4800, New Zealand
|
|
Keywords:
|
propensity ; matching ; nearest neighbour ; dissimiliarity ; missing at random ; imputation
|
|
Abstract:
|
Missing data is a common phenomenon. Many survey organizations use single imputation methods such as nearest neighbour imputation (NNI) to deal with missing data. Advantages of NNI are it makes use of covariate information and the point estimators obtained from the data imputed by NNI are less biased. In multivariate covariate situation finding NN is complicated because every variable needs to be matched. We investigate the use of propensity matching, introduced by Rosenbaum and Rubin (1983) for observational studies, to find the donor when data is multivariate. NNI by propensity score (NNPS) is investigated using simulations with data missing at random, both linear and convex. We use NNPS as it assures that the conditional distribution of the covariates given the propensity score is same for respondents and nonrespondents. We compare NNPS with NNI by dissimilarity matrix (NNDM) given by Murthy et al. (2003). We use NNDM as it preserves the marginal distributions. The results indicate that estimates from data with imputation by NNPS are often similar to those of NNDM and it reduces the curse of dimensionality.
|
