Abstract:
|
Medical studies are frequently afflicted with missing values. Multiple imputation (MI) is one solution for addressing missing data with broad applicability. Certain data structures present unique challenges to implementing MI. One such structure arises in data sets with time-varying covariates. "Proper" imputation should emulate the scientific model driving the analysis. Incorporating correlation among observations within a subject over time can be a challenge if the scientific model assumes observations are independent and then adjusts the standard errors for clustering through robust estimation. We focus on the feasibility of a widely used approach where each cluster's observations are transformed from long to wide format and stacked together prior to imputation. The scientific model is then fit after transforming back to a long format. We explore the effect of several parameters, including the number of observations per cluster, the number of variables, the intra-class correlation coefficient, and the percentage of missing observations on the following: bias, coverage, mean squared error, and computation time. Our study informs guidelines for applying this MI approach.
|
ASA Meetings Department
732 North Washington Street, Alexandria, VA 22314
(703) 684-1221 • meetings@amstat.org
Copyright © American Statistical Association.