Comparison of Missing Data Analysis Methods in Cox Proportional Hazard Models

Objective: Data with missing value are common in clinical studies. This study investi - gated to assess the effects of different missing data analysis techniques on the performance of Cox proportional hazard model. Material and Methods: In order to see how sample size and missing rate effect the missing data analysis techniques, we derived the survival data with 25, 50 and 100 sam - ple sizes. Some elements of the survival data with different sample size were deleted in different rates under MAR (Missing at Random) assumption to generate incomplete data sets which had 5%, 10%, 20% and 40% missing value for each data. Data sets with missing values were completed by five missing data analysis techniques (complete case Analysis-CCA, mean imputation, regression im - putation-REG, expectation maximization-EM algorithm, multiple imputation-MI). The new com - pleted data sets were analyzed by Cox proportional hazard model and their results were compared with results of original data. Results: The difference between the techniques grew for increasing missing rate and while the sample size increased the methods were similar to each other. CCA was the most affected from sample size. The estimates from the methods REG, EM and MI were very similar to each other and real value. Conclusion: Multiple imputation method as impute more than one value for each missing value should be preferred instead of single imputation methods as im - pute only one value for each missing value.