Imputation and missing indicators for handling missing data in the development and deployment of clinical prediction...

Area of Science:

• Clinical Epidemiology
• Biostatistics
• Health Informatics

Background:

• Missing data is a common challenge in clinical prediction model development, validation, and deployment.
• Multiple imputation is standard but difficult at deployment; regression imputation offers a pragmatic alternative.
• The utility of missing indicators for informative missingness in prediction models is unclear.

Purpose of the Study:

• To compare the predictive performance of clinical prediction models using multiple imputation versus regression imputation.
• To evaluate strategies for handling missing data under different deployment scenarios (missing data permitted or prohibited).
• To assess the impact of including or omitting the outcome in imputation models and using missing indicators.

Main Methods:

• Simulated data under various missing data mechanisms.
• Compared predictive performance of models developed with multiple imputation and regression imputation.
• Evaluated scenarios with and without missing data at deployment, with/without outcome in imputation models, and with/without missing indicators.
• Applied methods to critical care data.

Main Results:

• When complete data is available at deployment, standard multiple imputation (using outcome) and regression imputation (omitting outcome) performed well.
• If missing data is allowed at deployment, omitting the outcome from the imputation model during development was preferred.
• Missing indicators generally improved performance but could be detrimental under outcome-dependent missingness.

Conclusions:

• Standard multiple imputation principles may not directly apply to clinical prediction models, especially with missing data at deployment.
• Multiple imputation and regression imputation showed comparable predictive performance.
• The optimal missing data strategy depends on the specific study and whether missing data is permitted at deployment.