Multiple Imputation and Random Survival Forests: Application to the Demographic and Health Survey Child Survival Data

Abstract

Background: Demographic and Health Surveys (DHS) provide data on a wide scope of risk-factors of under-five child survival. Missing covariate data is inevitable in the DHS under-five survival data since data is collected retrospectively and on a large number of covariates. We studied the missing data problem on the risk-factors of under-five child survival in DHS data sets. Methods: Random survival forests model was first used for selecting the highly predictive risk factors from a pool of over 400 covariates, from which a subset of 50 covariates was selected. Multiple imputation by chained equations (MICE) and random forests were applied to handle missing covariate data. Imputed data was then analyzed using random survival forests and Cox-regression models. Results: The results showed that missingness in covariates was more related to the time to event (52%) than the event status (19%) response variables. The ranking of under-five risk factors from imputed data sets was closely related to the ranking from the observed values, albeit, multiple imputation led to increase in the variable importance scores. The unadjusted estimates from the Cox-regression model based on imputed values were closely similar to the estimates from the observed values. However, minimal discrepancies in estimates were observed in covariates with over 30% missing data. Random forests approach shown potential for producing estimates much closer to the true estimates with high level of missing than MICE. Conclusion: Multiple imputation shown potential to produce estimates closely similar to the true estimates even with high missingness. Random forests imputation shown potential to perform better than MICE imputation strategies. The current study results may need to be validated using a larger simulation study and other non-response models for decisive conclusions to be made.