EMLasso: logistic lasso with missing data

Area of Science:

• Biostatistics
• Clinical Research
• Data Science

Background:

• Missing data in clinical covariates is frequent due to measurement costs.
• Complete case analysis and single imputation methods compromise model selection goals.
• Previous heuristic approaches combined logistic Lasso with multiple imputation.

Purpose of the Study:

• To develop a statistically rigorous method for handling missing covariate data in model selection.
• To improve the accuracy and reliability of predictive models in clinical settings.
• To address missing data in both categorical and continuous predictors.

Main Methods:

• The proposed method is based on the stochastic expectation-maximisation algorithm.
• It extends the logistic Lasso with multiple imputation for robust model selection.
• The approach handles missing data in penalized and non-penalized regression models.

Main Results:

• The method effectively handles missing data in categorical and continuous predictors.
• It provides a statistically sound alternative to incomplete case analysis and single imputation.
• Demonstrated application on lung cancer patient data for acute dysphagia prediction.

Conclusions:

• The new method offers a statistically grounded approach to missing covariate data in model selection.
• It is applicable to various data types and regression models.
• This technique enhances the reliability of clinical predictive models, as shown in lung cancer research.