Project C: Missing data in prognostic research: can imputation solve all problems?

Department of Epidemiology Unit Medical Technology Assessment
Coordinators: H. Groen, M.D., PhD and Prof. E. Buskens PhD

Prognosis is an important aspect to be taken into account when making treatment choices for patients. For several chronic diseases, prognostic models exist to determine the future risks of patients, depending on absence or presence of one or more risk factors. Treatment aimed at these risk factors may reduce risk.

The construction of prognostic models is an elaborate process, starting with collection of data from the patient population who have developed the disease under study. Data collection is often incomplete, and statistical analyses of the variables that influence prognosis become less reliable if many patients have to be excluded because of missing data.

There are many ways to supplement or impute missing data, but for prognostic models multiple imputation is becoming increasingly popular. This technique involves replicated imputation of missing values in the original data. The resulting datasets vary from each other and reflect the uncertainty surrounding the imputed values. Statistical software such as SPSS allows estimation of pooled risk measures from the multiple datasets. However, the methodology for calculation of measures of performance of a prognostic model, such as the c-statistic and calibration, is less defined.

The proposed project comprises of a literature study on the use of multiple imputation methods in prognostic research. The specific question will be to identify the methods used for determining discriminative power and calibration in models based on multiple imputation.