No previously published research assesses the usefulness of incorporating pharmacy data into prospective risk adjustment techniques in any national health system. To date, research in the European and Spanish contexts has focused on using diagnoses-based risk assessment variables [3
]. However, our study, as well as others carried out within the Spanish National Health System [20
], determined that the accuracy of the diagnostic codes allocated by primary health care physicians in their computerised medical records could be improved.
The results of this study confirm that much can be learned by looking at pharmacy data, especially when forecasting drug expenditures. Studies carried out in the U.S [21
] and Europe [23
] have outlined the potential of pharmacy data to improve the system of risk adjustment for both care management program capitation payments and pharmacy budget planning. This is of particular interest in a situation in which the information related to drug consumption is routinely recorded and automated, as it is in Spain. Moreover, the fact that the applied drug classification system (ATC classification) is uniform and used all over the country makes the use of pharmacy data very feasible--even more so if we consider the regular updating of the national ATC code databases issued by the Ministry of Health, according to the Guidelines of the World Health Organisation [24
Nevertheless, several challenges are posed by medication-only PMs. First, using pharmacy data as a risk adjuster for resource allocation could create perverse incentives, encouraging inappropriately prescribed drugs to be given higher budgets and promoting the inappropriate use of these drugs in the future. In consequence, if pharmacy data are to be used for budget allocation purposes, intensive monitoring activity will be required to prevent the inappropriate prescribing of drugs. As it happens in other European countries [23
], in Spain, social security funds are not care providers themselves unlike Health Maintenance Organizations (HMO) in the U.S.; this may reduce the danger of inappropriate prescription behaviour.
Second, pharmacy claims data may not always portray an accurate clinical picture, because some prescribed medications have multiple indications from which a patient's disease status must be inferred and because one disease may have many medication options in terms of its management [25
]. One of the most important innovations presented by the Rx-PM from the ACG system is its clinically oriented approach, which captures the unique clinical information embedded in medication-use data instead of attempting to identify diseases/conditions based on medications [19
The benefits of local calibrations have become evident. As was the case with other risk adjustment tools, such as the Diagnosis-Related Groups used in acute care hospitals, adaptation processes have been developed by other countries on the grounds that the philosophy of health care, resource consumption patterns and funding approaches differ from those in the country where the tool was originally created [26
]. Indeed, importing clinical predictors related to weights resulting from empirical evaluations carried out with U.S. cost data could lead to the incorporation of U.S. funding incentives and disincentives into the Spanish health care system. The results of our work show that the statistical performance of PMs was optimised using Spanish weights. This is due to the fact that local weights were calculated by regressing pharmacy cost data from our own health system on the explanatory variables for ACG-PMs.
Moreover, our sensitivity analysis demonstrated that these results were fairly robust even when using U.S. weights derived from a Medicare managed care population for individuals over the age of 65.
The main limitation of this study is related to the infeasibility of applying a split half method, which involves dividing the total sample in two and correlating the results, as a way of assessing the reliability of a test. In order to apply this technique to our particular study, the sample size would need to be larger than 80,000 patients so that each of the split groups would surpass the minimum number of individuals required for predictive modelling. On the other hand, the study population is not representative of either the population of the Aragon region or that of Spain, so results must be cautiously interpreted. Moreover, given the applied patient inclusion criteria (i.e. enrolees seen at least once by a public general practitioner during both Year 1 and Year 2), 7.9% of the enrolees that went to the health centre in 2006 (Year-1) and 14.3% of the enrolees that went in 2007 (Year-2) were excluded from the study sample due to their non-attendance in 2007 and 2006 respectively. Future studies may consider including patients with a discontinuous utilization of health care services when measuring the performance of PMs.
Nevertheless, the validity of the sample is backed up by the fact that the proportion of women, the age distribution, the prevalence of chronic conditions and the behaviour of patients with regard to pharmacy expenditure are consistent with those of previous studies carried out in the Spanish primary care setting [3
Another potential limiting factor could be related to the relatively recent incorporation of electronic medical records into the primary care setting. Even if a series of inclusion criteria were applied during the health centre selection process to guarantee the quality and reliability of the clinical data, a three-year period of experience in the use of electronic records is still short enough that we might question the maturity of the information systems. This could lead to an overestimation of the clinical under-coding effect highlighted in this study. Thus, using even slightly more recent data could substantially boost model performance measures.
An additional reason for overestimating the under-coding phenomenon arises from the lack of connection between databases for primary and specialised care. Although the primary care general practitioners are considered the gatekeepers of the health system and would therefore need to have a recompilation of all diagnostic episodes of a patient, certain codes could be missing when these disease are followed by specialised physicians. In the study region, specialised physicians have poorer information systems than primary care physicians and, consequently, data are less available. Still, drug claims collect prescriptions carried out by both specialised and general practitioners. This situation could explain cases in which patients consume drugs for diseases that are not registered in general practitioners' office records, as reported in this paper.
Even if these two previous statements have been considered as potential limitations of the study due to their association with a poor quality of the data, they support the need to incorporate pharmacy data when carrying out risk adjustment.
Implications for the Spanish National Health System
When the target of a health care organisation is the management of an individual's medication use, predictive models based on pharmacy data are particularly useful. Adding diagnostic markers to medication data does not appear to improve predictions for pharmacy costs [21
], which tend to show a pronounced degree of persistence from year to year, particularly among the heaviest users [27
]. This has long been the rationale for using prior costs in themselves for budget planning by hospitals and the primary care setting in Spain [2
]. In terms of statistical performance, prior cost is a fairly good predictor of future cost--even better than diagnostic or pharmacy-related variables [3
]--, but it has some limitations. First, prior cost has no inherent clinical meaning, and is therefore of low relevance to clinicians who wish to intervene. It is not tied to morbidity and, thereby, cannot be translated into clinical action. Second, prior cost is subject to the phenomenon of regression to the mean (i.e., the natural tendency of groups of individuals who are high cost one year to move towards mean costs in the following years). Third, prior-use measures are not entirely appropriate as risk factors for risk-adjusted rate setting or profiling as they potentially could provide incentives to excessive and inappropriate pharmacy use.
Screening tools based on diagnostic or medication data can identify reliable "early warning signs" of future expenses that can then promote secondary prevention through patient care management [4
]. Although the beneficial effects of care management have not been consistently demonstrated[28
], preliminary evidence from an intensive nurse-based intervention for high-risk elderly individuals appears to show that it holds great promise in terms of cost reduction[29
] and better quality of care [30
]. A randomised clinical trial carried out recently in the Spanish primary care setting has confirmed the effectiveness of intervention strategies in decreasing the number of consultations of frequent attenders[31
], paving the way for the implementation of further cost efficiency-focused strategies.
Moreover, the optimal predictive capacity of the ACG-PM proves its usefulness for future budget planning. This has been demonstrated to entail the largest impact on pharmaceutical revenues among existing pharmacy regulatory measures[32
Last, PMs provide a means of determining physician prescription profiles while adjusting for patient case-mix, so decisions about incentives, efficiency improvement efforts or even sanctions can be targeted towards the "right" physicians [33