A recent article on risk adjustment by a group of researchers from Dartmouth attracted our attention. Entitled Observational intensity bias associated with illness adjustment: cross sectional analysis of insurance claims , it describes biases that are found in various types of healthcare risk adjustment processes, with the bias is being caused by an increasing number of physician–patient encounters.
Risk adjustment methodologies such as the hierarchical condition categories (HCC) process compute risk scores for groups of patients by utilizing diagnoses reported for those patients. In the HCC model, each patient is given a cumulative risk factor for each diagnosis-based condition recognized by the model, which along with an age/gender factor constitutes that patient’s risk score. While risk scores are not intended to be accurate at a patient level, they are expected to be valid for larger groups of patients, and are used to explain relative costs among different populations of patients. Differences in cost, of course, can be a result of varying numbers of patient–physician encounters, with each encounter contributing to the patient cost. Thus, the risk scores are used to explain, in part, differences in the number of these encounters.
The interesting point of this article, though, was that the risk scores themselves can be affected by the number of encounters, with a larger number of encounters identifying an increasing number of diagnoses. Thus, the number of encounters affects the risk score – which can be used to justify the number of encounters! Obviously this creates a credibility problem, which the researchers suggest can be resolved by modifying the risk score based on the number of encounters. This apparently dampens the "positive feedback loop" in which a larger number of encounters seems to justify a larger number of encounters.
This writer is neither a clinician nor statistician, but as a typical consumer of healthcare services was struck by the idea that a progressively larger number of physician-patient encounters identifies additional diagnoses that were sufficiently significant to increase a risk score, but that were not identified in previous encounters. The diagnoses utilized by the HCC risk scoring model (and probably the other models, with which we are not as familiar) are those such as cerebral hemorrhage, congestive heart failure, multiple sclerosis, diabetes and others, which are so significant that we would hope they would be identified early in the patient diagnosis process. This apparently is not the case.
Doffing our hats as healthcare consumers and reverting to our data analytics role, the takeaway from this article appears to be that users of these risk adjustment models must be sensitive to their biases related to the number patient encounters. Risk adjustment can contribute useful information in understanding differences in care patterns among groups of patients, but like any other statistical tool it contains flaws and biases that must be considered in its use.