| Risk adjustment has become an essential tool in modern health services research and management. Specifically, it can be used to characterize populations, fairly profile providers and set their payment rates, and to predict future health care events. The extent to which risk adjustment methods can be applied to different data sources, settings, purposes and populations is not well understood. The ACG risk adjustment methodology has been tested internationally for financial applications, but only on a limited basis for predictive modeling. The utilization of different pharmacy codes as ACG inputs (e.g., ATC versus NDC codes) is also not well understood and might have an impact on model performance. Furthermore, explanatory (i.e., risk) variables vary significantly between population sub groups, and this fact might have an impact on model performance when applied to specific sub-populations. The purpose of this study is threefold: (1) to assess the extent to which the Adjusted Clinical Groups (ACG) system, a widely used tool, performs well in a European primary care setting, (2) to assess the extent to which the ACG performs well using a new source of risk data (ATC pharmacy codes), and (3) to assess the extent to which the ACG performs well for predictive modeling, overall and in two specific sub-populations (children and adults with diabetes). The analysis is based on a dataset from a group of primary care providers in the city of Badalona, Spain, derived from approximately 66,000 people of all ages who utilized services in two consecutive years (2006 and 2007). The statistical approach was based primarily on linear and logistic regression models, to predict health services utilization (in terms of visits and costs). Results show that the tool, as far as predictive modeling is concerned, works as well as it does with populations in the US, where it was developed. Adaptation to ATC codes seems to work equally well. In terms of the two specific sub-groups studied, children are considerably different from adults. That is revealed in both their morbidity patterns and drug prescription rates, the most important predictors utilized in the models. Probably as a consequence of that difference, the models, although still an improvement on simpler demographic models, do not do as well among children as in adults. Differences between adults with and without diabetes are evident, but the two groups seem to be far more similar, since the performance of models in this sub-group approximates that seen in the general adult population. The major conclusion is that risk adjustment models such as ACG can be utilized for predictive modeling in a European primary care system, and that the transition from NDC to ATC codes is feasible. When applying the tool to specific populations, caution should be taken. |
