Handling Missing Values in Interrupted Time Series Analysis of Longitudinal Individual-Level Data

Area of Science:

• Epidemiology
• Biostatistics
• Health Services Research

Background:

• Interrupted time series (ITS) analysis commonly uses segmented regression (SR) on averaged data.
• Aggregate-level SR analysis can introduce bias when individual-level data are missing at random (MAR).

Purpose of the Study:

• To illustrate the bias in aggregate-level SR analysis for ITS when data are MAR.
• To propose and evaluate alternative analysis methods using individual-level data.

Main Methods:

• Utilized UK electronic health records to assess weight change post-antipsychotic initiation.
• Contrasted aggregate-level SR with mixed-effects models (with and without multiple imputation).
• Conducted simulation studies to validate findings in a controlled setting.

Main Results:

• Aggregate-level SR indicated greater weight gain (0.799kg/week) than mixed models (0.412kg/week).
• Simulations confirmed aggregate-level SR bias under MAR conditions.
• Mixed models, particularly with multilevel multiple imputation, yielded unbiased estimates.

Conclusions:

• Aggregate-level SR introduces bias in ITS estimates when individual data are MAR due to missingness at the cluster level.
• Recommends avoiding data averaging and employing mixed models, with or without multilevel multiple imputation.
• Mixed models with multiple imputation are suitable for various ITS outcomes, including proportions.