Predicting absolute risk for a person with missing risk factors.

Area of Science:

• Biostatistics
• Epidemiology
• Medical Informatics

Background:

• Accurate absolute risk projection is vital for clinical decision-making.
• Missing predictor data in target populations complicates absolute risk estimation.
• Existing methods may introduce bias when predictor distributions differ between reference and target datasets.

Purpose of the Study:

• To compare the performance of seven methods for projecting absolute risk when predictors are missing.
• To evaluate bias and mean squared error of different imputation strategies.
• To identify optimal methods for absolute risk prediction in the presence of missing data.

Main Methods:

• Simulations using real breast cancer predictor distributions and outcome data.
• Comparison of methods imputing individual predictors versus risk scores.
• Analysis of a real-world breast cancer dataset.

Main Results:

• The greatest bias stemmed from differing predictor distributions between reference and target populations.
• No single method achieved unbiased predictions when distributions varied.
• Multiple imputation methods showed comparable performance, with less bias but higher variability than single risk score methods.
• Violation of the missing at random (MAR) assumption did not lead to severe bias.

Conclusions:

• Selecting a reference dataset that closely matches the target population's predictor distribution is essential for reducing bias in absolute risk predictions.
• Careful consideration of reference data is paramount when dealing with missing risk factors.
• Multiple imputation techniques offer a robust approach for absolute risk projection with missing data.