Hot Deck Multiple Imputation for Handling Missing Accelerometer Data.

Area of Science:

• Biomedical Engineering
• Epidemiology
• Data Science

Background:

• Missing data due to non-wear is a significant challenge in accelerometer studies.
• Existing methods for handling missing accelerometry data are often ad-hoc or rely on restrictive assumptions.
• Accelerometer data are high-dimensional, episodic, and skewed time-series, complicating imputation.

Purpose of the Study:

• To develop and evaluate a flexible hot deck multiple imputation (MI) procedure for addressing missing accelerometry data.
• To compare the performance of hot deck MI against standard available case (AC) and complete case (CC) analyses.
• To assess imputation performance for both 24-hour and daytime-only accelerometry data.

Main Methods:

• Developed a hot deck MI procedure where missing segments are replaced by observed segments from "donor pools".
• Donor pool selection and imputation weights were based on non-wear and accelerometer-derived variables.
• Compared hot deck MI, AC, and CC analyses using a simulation study with 2,550 women's accelerometry data.

Main Results:

• Hot deck MI demonstrated less bias and better 95% confidence interval (CI) coverage than AC and CC for 24-hour data.
• For daytime-only data, MI showed less bias and better 95% CI coverage compared to AC.
• CC analysis for daytime data yielded similar bias and 95% CI coverage but with longer CIs than MI.

Conclusions:

• Flexible hot deck MI is a robust method for imputing missing accelerometry data, outperforming traditional AC and CC methods.
• This imputation strategy enhances the accuracy and reliability of physical activity and sedentary behavior estimates derived from accelerometers.
• The findings support the adoption of hot deck MI for improving the analysis of accelerometer data, particularly when dealing with significant missingness.