Validation of prediction models based on lasso regression with multiply imputed data.

Area of Science:

• Statistics
• Biostatistics
• Medical Informatics

Background:

• The lasso technique offers parsimony and improved prediction in prognostic studies by shrinking regression coefficients.
• Quantifying optimism and generalizing model performance often involves bootstrap methods.
• The optimal resampling strategy for multiply imputed data in lasso models remains unclear.

Purpose of the Study:

• To investigate the optimism of lasso models in prognostic studies.
• To compare different bootstrap resampling approaches for multiply imputed data.
• To assess the impact of these approaches on the estimated model performance.

Main Methods:

• A cohort of Chronic Obstructive Pulmonary Disease patients was used to predict dyspnea.
• Four methods of handling multiply imputed data within bootstrap resampling were compared.
• Simulated datasets were used alongside study data to evaluate the approaches.

Main Results:

• Lasso model performance showed optimism and suboptimal calibration due to over-shrinkage.
• Estimates of optimism varied significantly based on the chosen data handling method.
• Resampling completed datasets underestimated optimism, while incorporating MI in validation yielded more accurate, though slightly overestimated, results.

Conclusions:

• Prognostic models using the lasso technique can exhibit optimism.
• Internal validation results are highly dependent on the bootstrap resampling strategy employed for multiply imputed data.