Comorbidity and medical costs among Medicare beneficiaries with HIV in California: A case study in predictive data analysis
Kodi B. Arfer. PhD
Center for HIV Identification, Prevention, and Treatment Services, Semel Institute for Neuroscience and Human Behavior, UCLA
Data analysis typically focuses on inference about the unobservable true parameters of a postulated true data-generating model. An alternative is to focus on the model’s accuracy in predicting values on outcome variables. Predictive data analysis allows evaluating a wide variety of models in a straightforward way. It produces measures of predictive accuracy that show how practically useful a model and a set of variables are in predicting an outcome. Using Medicare claims data from HIV-positive Californians, I investigate how comorbidity information—knowing what conditions other than HIV a patient has been diagnosed with—can be used to predict medical costs. I show how difficulties in the data, from highly skewed distributions to the differences between Medicare and Medicaid, can be addressed with predictive strategies. I find that comorbidities are indeed predictively useful, especially for inpatient costs.
Click here to download the presentation slides. (19)