Multiple imputation in survival models: applied on breast cancer data

Abstract

Background: Missing data is a common problem in cancer research. While simple methods such as completecase (C-C) analysis are commonly employed for handling this problem, several studies have shown that these methods led to biased estimates. We aim to address the methodological issues in development of a prognostic model with missing data.

Methods: Three hundred and ten breast cancer patients were enrolled. At first, patients with missing data on any of four candidate variables were omitted. Secondly, missing data were imputed 10 times. Cox regression model was fitted to the C-C and imputed data. Results were compared in terms of variables retained in the model, discrimination ability, and goodness of fit.

Results: Some variables lost their effect in complete-case analysis, due to loss in power, but reached significance level after imputation of missing data. Discrimination ability and goodness of fit of imputed data sets model was higher than that of complete-case model (C-index 76% versus 72%; Likelihood Ratio Test 51.19 versus 32.44).

Conclusion: Our findings showed inappropriateness of ad hoc complete-case analysis. This approach led to loss in power and imprecise estimates. Application of multiple imputation techniques to avid such problems is recommended.

Keywords: Breast cancer; Missing data; Multiple imputation; Prognostic model.