Abstract:
Cardiorespiratory fitness (CRF) is often used as an indicator of physical fitness that
has an inverse association with mortality. It is practical to identify subgroups where
CRF exerts positive or negative association with mortality status than on the entire
population. We considered Classification and Regression tree-based algorithms: GUIDE
and Virtual Twins (VT) to find subgroups of participants where CRF exerts positive or
negative association with mortality status from the Ball State Adult Fitness Program
Longitudinal Lifestyle Study data. Results are compared with traditional penalized
Logistic regression with LASSO penalty. The results indicate that VT for classification
and regression tree identified subgroups by considering BMI (Body Mass Index) and age
as important predictors for different thresholds. We observed that when the threshold
was larger than some specific value, VT couldn’t identify any subgroups. GUIDE selected
subgroups with BMI, age, and sex as important predictors. Finally, Logistic regression
with LASSO penalty found the interaction with fitness rank and other predictors: age,
sex, smoking status, and hypertension. Thus, our results suggest that tree-based methods
identified fewer predictors compared to the non-tree based logistic regression method.
