Population-calibrated multiple imputation for a binary/categorical covariate in categorical regression models

Area of Science:

• Biostatistics
• Medical Informatics
• Epidemiology

Background:

• Multiple imputation (MI) is widely used for handling missing data in medical research.
• Standard MI assumes data are missing at random (MAR), which may be inadequate for missing not at random (MNAR) data.
• External population-level data can offer valuable information for improving imputation models.

Purpose of the Study:

• To introduce and evaluate a novel MI method, the calibrated-δ adjustment, for handling MNAR data.
• To demonstrate how to incorporate population marginal distributions into MI models.
• To assess the performance of the calibrated-δ adjustment MI compared to standard MI under MAR and MNAR scenarios.

Main Methods:

• The calibrated-δ adjustment uses an offset derived from the external population marginal distribution of the incomplete variable.
• Analytic derivations and simulation studies were conducted to evaluate the method.
• The method was applied to impute missing ethnicity data in a UK primary care electronic health records study on type 2 diabetes prevalence.

Main Results:

• The calibrated-δ adjustment MI method yields results comparable to standard MI under the MAR assumption.
• Under two MNAR mechanisms, the proposed method provides more accurate inference than standard MI.
• Application to ethnicity data in a diabetes study revealed scientifically relevant changes in inference for non-White ethnic groups.

Conclusions:

• Calibrated-δ adjustment MI is a pragmatic approach to leverage population-level information for sensitivity analyses.
• This method helps explore potential departures from the MAR assumption in missing data analyses.
• The technique enhances the robustness and accuracy of MI, particularly in complex missing data situations.