Applied comparison of stepwise Cox and Cox-LASSO regression in selecting determinants of joint replacement: insights from the Geelong Osteoporosis Study

Joint replacement (JR) is influenced by multiple determinants. Identifying an appropriate variable selection method is essential for accurate and interpretable models in musculoskeletal epidemiology. Although penalized methods such as LASSO are increasingly recommended, stepwise regression remains common in applied research. We compared these approaches for predicting JR in a large musculoskeletal cohort. Data from 2931 participants aged 20-96y from the Geelong Osteoporosis Study were analysed. Potential determinants included sociodemographic, lifestyle, body composition, and comorbidity variables. Stepwise Cox regression was implemented using the stepAIC() function in the ‘MASS’ package, and Cox-LASSO regression was performed using cv.glmnet() in the ‘glmnet’ package, using R version 4.3.1. Model performance and generalizability were assessed through independent test validation and 5-fold cross-validation. Bootstrap resampling was used to evaluate performance differences between models. Over a median follow-up of 16.7y (IQR:9.7–23.2), both models consistently identified age, spine bone mineral density, and prior fracture as key determinants of JR. Stepwise Cox regression using forward selection achieved slightly higher C-index scores across independent train–test splits and 5-fold cross-validation. All models demonstrated good discriminatory ability (C-index > 0.7). Bootstrap resampling confirmed that the performance difference was unlikely due to random variation (95%CI: 0.0004–0.0190). This case study revealed the practical implications of variable selection in musculoskeletal epidemiology. Stepwise Cox regression may remain useful for moderate-sized, low-dimensional datasets, while penalized methods are better suited to larger, more complex data. These findings guide researchers and clinicians aiming to balance methodological rigor with practical utility in predicting JR.