Umeå University's logo

Background: Variable selection for cardiovascular disease (CVD) risk is important in reducing cost of data collection and thus promote uptake of risk screening among practitioners and general public. Apart from expert knowledge, statistical model is another major component in variable selection. Variable selection determined by p-value in Cox proportional hazard regression after expert selection is the most common approach in constructing CVD risk prediction model. Under the Cox model, there are many other available methods, such as, step-wise Cox selection, LASSO, and survival boosted regression tree. These methods can be applied when the researcher is facing to large amount of potential variables.

Objective: This study aims at 1) identifying the best subset of predictors for the first CVD event in 10 years based on different statistical methods, and, 2) comparing the predictive performance of Cox proportional hazard regression, step-wise Cox selection, LASSO, and survival boosted regression tree with the corresponding set of predictors.

Method: Empirical data was extracted from the Vasterbotten Intervention Programme (VIP), death registry, and hospitalisation record. The cohort selected for analysis includes individuals examined in primary care centre from 1990 to 2005 and were followed up for ten years based on the registry data. Data from 1990 to 1999 are used to fit models. The models are then tested on data from 2000 to 2005. Statistical predictor selection was performed with Cox backward selection, LASSO in Cox model, and survival boosted regression tree model (sBRT). Predictor set based on expert knowledge and p-value was derived from Framingham risk score. Predictor subsets from LASSO and sBRT were passed to Cox regression. All models were compared based on the predictive performance, specifically graphical calibration and AUC.

Result: Variables different from the expert variable subset are found by Cox backward selection, LASSO, and sBRT, such as, education, long sick leave, relationship with colleagues, berry picking, and, alcohol consumption. In terms of predictive performance, the Cox proportional hazard regression model with the expert variable subset performs best. The differences in AUC between models with various variable subsets is little. In our case, LASSO and sBRT do not show better prediction than Cox regression.

Conclusion: Different statistical models may result in different variables subsets. Variable selected based on expert knowledge is of high importance in constructing prediction models. In the case of predicting CVD event in 10 in Northern Sweden, Cox proportional hazard regression shows similar predictive performance as LASSO and sBRT.