Validation and updating of risk models based on multinomial logistic regression

Area of Science:

• Medical Statistics
• Epidemiology
• Health Services Research

Background:

• Risk prediction models frequently exhibit suboptimal performance during external validation.
• Existing methods for updating multinomial logistic regression models require enhancement for improved risk prediction.

Purpose of the Study:

• To develop and evaluate methods for updating multinomial logistic regression models for enhanced risk prediction.
• To compare recalibration, revision, and extension approaches for model updating.

Main Methods:

• Applied recalibration, revision, and extension techniques to a risk model for pregnancies of unknown location (PUL).
• Utilized a closed testing procedure to guide the selection of updating complexity.
• Validated updated models using temporal and geographical datasets, with internal validation via bootstrap resampling.

Main Results:

• Recalibration can impact discrimination in multinomial models, unlike in dichotomous models.
• Model revision, particularly with functional form adjustments, improved discrimination.
• Penalized estimation enhanced model calibration, and recalibration significantly improved performance in the case study.

Conclusions:

• Novel methods are available to update multinomial risk models for improved external validity.
• A closed testing procedure aids in selecting between recalibration and revision.
• Full model revision is recommended for large sample sizes within each outcome category.