Approaches to Missing Data Inference Results from CaPSURE

Abstract

Objective: There are multiple reasons for missing data in observational studies; excluding patients with missing data can lead to significant bias. In this study, we evaluated several methods for assigning missing values to health service utilisation.

Design and setting: Cancer of the Prostate Strategic Urologic Research Endeavor (CaPSURE) is a US national database of men with prostate cancer. Physician visits and diagnostic tests for 342 patients newly diagnosed with prostate cancer were evaluated.

Patients and participants: Patients were followed for a full year (observed data, n = 228) and patients with incom plete data (predicted data, n = 114) were included.

Interventions: We used the following approaches for imputing missing data: assigning the group mean, a time-specific mean, a patient-specific mean, a stratified mean (by age, localised disease and insurance status) and carrying the last observation forward and/or backward.

Main outcome measures and results: All prediction strategies resulted in higher estimates (19.3 to 23.1) for annual physician visits than was observed (17.1 ± 15.5), and differences were statistically significant for both the last observation carried forward (23.1 ± 15.5) and the patient’s individual mean (22.7 ± 36.1) when predicting physician visits. The same strategies had higher predicted values for x-rays (1.8 ± 5.1 and 1.8 ± 4.4 vs 1.1 ± 1.9 for the observed group), although the last observation carried forward was not statistically different from the observed value.

Conclusions: We were unable to identify a single optimal strategy. However, imputation from individual means and the last observation carried forward methods did not perform as well as the other strategies. While the differences observed in this study were small, we anticipate that with increased length of follow-up and more dropouts, there would be greater differences among strategies.