Influence of Pattern of Missing Data on Performance of Imputation Methods: An Example from National Data on Drug Injection in Prisons

Background: Policy makers need models to be able to detect groups at high risk of HIV infection. Incomplete records and dirty data are frequently seen in national data sets. Presence of missing data challenges the practice of model development. Several studies suggested that performance of imputation methods is acceptable when missing rate is moderate. One of the issues which was of less concern, to be addressed here, is the role of the pattern of missing data. Methods: We used information of 2720 prisoners. Results derived from fitting regression model to whole data were served as gold standard. Missing data were then generated so that 10%, 20% and 50% of data were lost. In scenario 1, we generated missing values, at above rates, in one variable which was significant in gold model (age). In scenario 2, a small proportion of each of independent variable was dropped out. Four imputation methods, under different Event Per Variable (EPV) values, were compared in terms of selection of important variables and parameter estimation. Results: In scenario 2, bias in estimates was low and performances of all methods for handing missing data were similar. All methods at all missing rates were able to detect significance of age. In scenario 1, biases in estimations were increased, in particular at 50% missing rate. Here at EPVs of 10 and 5, imputation methods failed to capture effect of age. Conclusion: In scenario 2, all imputation methods at all missing rates, were able to detect age as being significant. This was not the case in scenario 1. Our results showed that performance of imputation methods depends on the pattern of missing data