A classification tree analysis of selection for discretionary treatment

Abstract

Objectives: To study treatment bias in observational outcomes research, the authors present a nonlinear classification tree model of clinical and psychosocial factors influencing selection for interventional management (lower extremity bypass surgery or angioplasty) for patients with intermittent claudication.

Methods: The study sample includes 532 patients with mild to moderate lower extremity vascular disease, without prior peripheral revascularization procedures or symptoms of disease progression. All patients were enrolled in a prospective outcomes study at the time of an initial referral visit for claudication to one of the 16 Chicago-area vascular surgery offices or clinics in 1993-95. The influence of baseline sociodemographic, clinical, and patient self-reported health status data on subsequent treatment is analyzed. Study variables were derived from lower extremity blood flow records and patient questionnaires. Follow-up home health visits were used to ascertain the frequency of lower extremity revascularization procedures within 6 months of study enrollment. Hierarchically optimal classification tree analysis (CTA) was used to obtain a nonlinear model of treatment selection. The model retains attributes with the highest sensitivity at each node based on cutpoints that maximize classification accuracy. Experimentwise Type I error is ensured at P < 0.05 by the Bonferroni method and jackknife validity analysis is used to assess model stability.

Results: Seventy-one of 532 patients (13.3%) underwent interventional procedures within 6 months. Ten patient attributes were used in the CTA model, which had an overall classification accuracy of 89.5% (67.6% sensitive and 92.9% specific), achieving 57.7% of the theoretical possible improvement in classification accuracy beyond chance. Eleven model prediction endpoints reflected a 33-fold difference in odds of undergoing lower extremity revascularization.

Conclusions: Initial ankle-brachial index (100%), leg symptom status over the previous six months (89%), self-reported community walking distance (74%) and prior willingness to undergo a lower extremity hospital procedure (39%) were used to classify most patients in the sample. These attributes are critical control variables for a valid observational study of treatment effectiveness.