ABSTRACT
BACKGROUND: Diagnostic models are frequently used to assess the role of risk factors on disease complications, and therefore to avoid them. Missing data is an issue that challenges the model making. The aim of this study was to develop a diagnostic model to predict death in HIV/AIDS patients when missing data exist. METHODS: HIV patients (n=1460) referred to Voluntary Consoling and Testing Center (VCT) of Shiraz southern Iran during 2004-2009 were recruited. Univariate association between variables and death was assessed. Only variables which had univariate P< 0.25 were selected to be offered to the Multifactorial models. First, patients with missing data on candidate variables were deleted (C-C model). Then, applying Multivariable Imputation via Chained Equations (MICE), missing data were imputed. Logistic regression was fitted to C-C and imputed data sets (MICE model). Models were compared in terms of number of variables retained in the final model, width of confidence intervals, and discrimination ability. RESULT: About 22% of data were lost in C-C model. Number of variables retained in the C-C and MICE models was 2 and 6 respectively. Confidence Intervals (C.I.) corresponding to C-C model was wider than that of MICE. The MICE model showed greater discrimination ability than C-C model (70% versus 64%). CONCLUSION: The C-C analysis resulted to loss of power and wide CI's. Once missing data were imputed, more variables reached significance level and C.I.'s were narrower. Therefore, we do recommend the application of the imputation method for handling missing data.
ABSTRACT
BACKGROUND: Prognostic models have clinical appeal to aid therapeutic decision making. Two main practical challenges in development of such models are assessment of validity of models and imputation of missing data. In this study, importance of imputation of missing data and application of bootstrap technique in development, simplification, and assessment of internal validity of a prognostic model is highlighted. METHODS: Overall, 310 breast cancer patients were recruited. Missing data were imputed 10 times. Then to deal with sensitivity of the model due to small changes in the data (internal validity), 100 bootstrap samples were drawn from each of 10 imputed data sets leading to 1000 samples. A Cox regression model was fitted to each of 1000 samples. Only variables retained in more than 50% of samples were used in development of final model. RESULTS: Four variables retained significant in more than 50% (i.e. 500 samples) of bootstrap samples; tumour size (91%), tumour grade (64%), history of benign breast disease (77%), and age at diagnosis (59%). Tumour size was the strongest predictor with inclusion frequency exceeding 90%. Number of deliveries was correlated with age at diagnosis (r=0.35, P<0.001). These two variables together retained significant in more than 90% of samples. CONCLUSION: We addressed two important methodological issues using a cohort of breast cancer patients. The algorithm combines multiple imputation of missing data and bootstrapping and has the potential to be applied in all kind of regression modelling exercises so as to address internal validity of models.
