The performance of multiple imputation for missing data in the context of Cox regression model

Missing data is a prevalent problem in data analysis. In the present dissertation I investigated the performance of multiple imputation (MI) techniques using a real data set as well as through a Monte Carlo study in the context of a Cox regression model. For the Monte Carlo study, I compare the MI methods for handling missing data between a restrictive and an inclusive strategy. Meanwhile, the implementation of MI methods in handling missing data in surveys with multiple-item scales is also investigated in the present dissertation. In the real data setting, the effect of participation in interracial dating on relationship stability is examined by applying a restrictive MI strategy to impute the missing multiple-item scales at the scale level. The Monte Carlo analysis shows that the MI techniques outperform the complete case analysis across all the experimental conditions under MAR or MCAR in the sense that the MI estimates have less cross-sample fluctuations and are closer, on average, to the true population values. Among all the MI techniques, there is no major difference in the performance of MI with either a restrictive or an inclusive strategy under a various experimental conditions. The Monte Carlo studies also shows that missing items in multiple-item scales can be imputed at the scale level on the condition that the items in a given scale have more than five response categories and the inter-item correlations are high. We should impute the missing items at the item level if the items in a given scale have only two response categories. Lastly, in the real data setting, I impute for the missing multiple-item scales and other missing single-item variables at the scale level with a restrictive strategy to investigate the effect of participation in interracial dating on relationship stability. The MI results show that adolescents who involved in the interracial dating are more likely to experience relationship disruption than their in-group dating counterparts while the complete case analysis shows that participation in interracial dating has no effect on relationship stability.