Variation in pricing, treatment, coding intensity and patient severity in spells of health care treatment

Existing theoretical and empirical models of risk adjustment generally presume exogenous health status information is available to calculate risk adjusted payments. The norm is that only one signal per disease per year is used for predictions, even though in practice multiple signals are generally available. The literature suggests that consumers dominate the initialization of care while providers largely determine the intensity of treatment once care is initialized. The central insight of this paper is that the number and severity of health status signals over multiple years can be used to distinguish consumer taste heterogeneity for initiating treatment from provider practice style variation in the intensity of treatment conditional on an initial patient contact.

The structure of the analytical model is as follows. Each period a representative consumer is either healthy or sick but does not know what illness she has unless treatment is sought. She decides whether to seek treatment before knowing whether she has illness A (acute, e.g., the flu) or C (chronic, e.g. cancer). Her decision to seek treatment depends on her own taste parameter for treatment. The consumer’s expected cost depends on her severity of illness as measured by observed diagnoses, but also on consumer taste and provider treatment style. Assuming that treatment taste is independent of illness probabilities and patients do not sort by provider intensity, then higher or lower rates of consumer initiation of treatment are not associated with being sicker or more appropriate to treat, only pure taste variation. Geographic spending variation is not necessarily inefficient: it depends on correlations and the value of treatment.  

This analytical model is operationalized empirically using MarketScan commercial claims data (~3 million). We use adults age 21-64 from 2007 to 2011, enrolled in identifiable plans with high rates of provider specialty coded. The week is used as the unit of analysis. Rates of diagnoses (binary flags for each of 394 Verisk Health’s DxCG HCCs) and visits (measured as quantity of distinct procedures provided in a one week period) are evaluated both between weeks and within each week. The number of weeks in which a patient is treated for disease k is a signal of severity of illness and patient taste for treatment, while the per week number of visits (or diagnoses) for a disease k to a provider conditional on at least one is a signal of provider intensity style.  Separate fixed effects capture consumer taste and provider style for each disease, while diagnoses and plan fixed effects control for patient severity and plan generosity. For exploratory work we have been focusing on two conditions: Influenza (acute) and Asthma/COPD (chronic), but the final paper will explore multiple conditions.

Fixed effects suggest that the majority of variation in coding is due to heterogeneous consumer tastes rather than provider style, and neither dimension is explained by county level demand or supply side variables. There is meaningful correlation between patient fixed effects across diseases.