Penalized logistic regression with low prevalence exposures beyond high dimensional settings

Area of Science:

• Biostatistics
• Epidemiology
• Statistical modeling

Background:

• Estimating risk factors with very low exposure prevalence for binary outcomes presents challenges for standard statistical methods like logistic regression.
• Classical methods often yield unreliable results in these low-prevalence scenarios.

Purpose of the Study:

• To evaluate the effectiveness of penalized regression techniques in improving the estimation of low-prevalence risk factors.
• To compare Firth correction, ridge, lasso, and boosting methods in low-dimensional settings.

Main Methods:

• Utilized a large unmatched case-control study dataset (France, 2005-2008) on prescription medicines and road traffic accidents.
• Conducted an accompanying simulation study to assess method performance.
• Applied Firth correction, ridge, lasso, and boosting regression techniques.

Main Results:

• Penalized regression methods, particularly Firth correction and boosting, significantly enhance the estimation of risk factors with prevalences below 0.1%.
• These improvements are especially pronounced for ultra-low prevalences.
• The study demonstrated the utility of these methods even in low-dimensional settings.

Conclusions:

• Penalized regression techniques offer substantial benefits for estimating low-prevalence risk factors in binary outcome studies.
• Firth correction and boosting are recommended for improving risk factor estimation, especially at extremely low prevalences.
• The findings support the use of penalized techniques when dealing with a moderate number of low-prevalence exposures.