Survival Analysis with PHREG: Using MI and MIANALYZE to Accommodate Missing Data
Survival analyses based on a data collection process which the researcher has little control over are often plagued by problems of missing data. Deleting cases with any missing data will result in information loss and usually results in bias, while many analytic procedures that retain this information in some form underestimate the resulting uncertainty in parameter estimates and other output. SAS Version 8 includes two new procedures that allow the researcher to generate "complete" data sets from incomplete data by multiple imputation and to analyze the resulting data in ways which adequately account for the uncertainty involved. This paper presents suggestions for optimal use of PROC MI to perform such multiple imputation and PROC MIANALYZE to conduct various statistical analyses of modeling output, in this case from PROC PHREG, including design of control macros, structure of multiply imputed datasets, generation of binary from non-binary categorical variables, and options for presentation of results.
View the pdf for this paper.