1075 — Comparison of Methods for Modeling Comorbidities in Health Outcomes Research
Echols C, Axon RN, Hunt KJ, Gebregziabher M, and Egede LE, Center for Disease Prevention and Health Interventions for Diverse Populations;
Adjustment for disease burden or comorbidity is essential in health outcomes research. Outcomes models that use administrative data sets that are not collected for research purposes could be biased if they do not properly account for comorbidity. Some researchers adjust for comorbidity using either Deyo-Charlson, Elixhauser, or their categorical form, while others include diagnostic indicators of individual co-morbidities in their model (Quan’s approach). However, it is unclear which approach is optimal. Therefore, we examined three approaches to determine the most optimal risk-adjustment with applications to data from a national cohort of Veterans with diabetes.
We investigated Deyo-Charlson, Elixhauser, and Quan’s comorbidity measures using three modeling approaches: as a continuous predictor, as a categorical form (0 = none, 1 = one, 2 = two, 3 = three or more), and as a list of all diagnostic indicators of disease. We considered three types of outcomes (binary, continuous, time to event). Generalized linear mixed models (GLMM) were fitted for HbA1c (binary and continuous) and Cox regression was used for modeling mortality with comorbidity as the main covariate. We examined area under the receiver operating characteristics curve (AUC) and goodness of fit (GoF) criteria to determine the optimal comorbidity modeling approach. AIC values were standardized by the total sample size used to fit each model.
Application to a national cohort of 892,223 Veterans with diabetes (followed 2002-2006) showed that modeling comorbidity, irrespective of index, in a categorical form (AIC = 8.77, R2 = 0.42, AUC = 0.95) leads to greater risk adjustment than modeling it as a continuous predictor (AIC = 8.79, R2 = 0.41, AUC = 0.92). In GLMM continuous HbA1c, a list of diagnostic indicators of disease consistently exhibited the best GoF. In GLMM binary HbA1c, a better fit was observed when a list of diagnostic indicators (AIC = 12.5, R2 = 0.95, AUC = 0.61) or four categories (AIC = 12.4, R2 = 0.61, AUC = 0.95) modeled comorbidity compared to including a score (AIC = 12.9, R2 = 0.61, AUC = 0.95). Results were similar for mortality.
How we model comorbidity leads to different levels of risk adjustment irrespective of the type of comorbidity index used. A robust approach seems to categorize the count of diagnostic indicators from ICD-9 codes.
Researchers need to consider alternative scenarios of modeling previously validated comorbidity scores when adjusting for risk in outcomes research.