Purpose: To develop an empirical model describing decisions related to treatment of anemia in patients with non-dialysis CKD. Method: Incident CKD patients (eGFR<60 cc/min/1.73m2) in the Veterans’ Health Administration system, receiving outpatient care and subsequently developing anemia (Hgb<12 gm/dl) were identified between 2000 and 2005. Exclusion criteria included: 1) chemotherapy in the 6 months prior to index date, 2) <6 months of a window for ESA prescription, resulting in a cohort of 145,173 patients. A person-month panel was constructed from the population. Treatment with ≥1 transfusions or ESA administrations in each month were identified. Each patient contributed monthly data from the index date until date of death or censoring (eGFR<10 ml/min/1.73m2, dialysis initiation, loss to follow up, or the end of the study period) resulting in 4,458,970 person-months of data. Two models of the treatment decision were considered. First, we consider a single regression function model of treatment choice. An ordered logistic regression model was fit with no treatment, ESA treatment, and transfusion as the levels of treatment. Second, a more flexible two-part conditional model was fit with a first stage modeling of decision to treat and a second stage modeling of transfusion vs. ESA conditional on treatment. Covariates used in all models included patient demographics, clinical characteristics, treatment location, and center tendency to use either ESA or transfusion. Models were compared using Cox non-nested hypothesis tests. Standard errors were adjusted for the clustering of observations within patients. Result: Both models are consistent with treatment escalating from no treatment, to ESA treatment, to transfusion with greater anemia severity. Compared to patient demographics and clinical characteristics, center tendency was the strongest predictor of future treatment initiation and choice (p-value<0.0001). Cox non-nested hypothesis tests comparing the two stage model to the ordered model strongly rejected the ordered model in favor of the two stage model (p-value<0.001.) Conclusion: Compared to the ordered model, the two-part model is significantly better and captures the complex relationships between anemia severity and treatment choice in CKD non-dialysis.
Candidate for the Lee B. Lusted Student Prize Competition