Pay for performance (P4P) is increasingly being used to stimulate healthcare providers to improve their performance. However, evidence on P4P effectiveness remains inconclusive. Flaws in program design may have contributed to this limited success. Based on a synthesis of relevant theoretical and empirical literature, this paper discusses key issues in P4P-program design. The analysis reveals that designing a fair and effective program is a complex undertaking. The following tentative conclusions are made: (1) performance is ideally defined broadly, provided that the set of measures remains comprehensible, (2) concerns that P4P encourages “selection” and “teaching to the test” should not be dismissed, (3) sophisticated risk adjustment is important, especially in outcome and resource use measures, (4) involving providers in program design is vital, (5) on balance, group incentives are preferred over individual incentives, (6) whether to use rewards or penalties is context-dependent, (7) payouts should be frequent and low-powered, (8) absolute targets are generally preferred over relative targets, (9) multiple targets are preferred over single targets, and (10) P4P should be a permanent component of provider compensation and is ideally “decoupled” form base payments. However, the design of P4P programs should be tailored to the specific setting of implementation, and empirical research is needed to confirm the conclusions.

Pay for performance (P4P) has become a popular approach to performance improvement in health care. Most of the P4P literature has focused on the United States and there is limited insight in the characteristics of major programs initiated in other countries. This article systematically describes and reviews P4P programs outside the United States. Our literature search identified 13 programs initiated in 9 countries. Although the programs share many similarities, they differ in several important respects, also when compared with the typical P4P program in the United States. In addition, there are clearly possibilities to increase incentive strength and minimize incentives for undesired behavior. In part, observed heterogeneity will be a consequence of contextual differences, but design choices often also seem to be made arbitrarily. In designing their programs, purchasers are hampered by limited knowledge of the influence of specific design choices and effective strategies to mitigate undesired behavior.

Most health insurance markets with premium-rate restrictions include a risk equalization system to compensate insurers for predictable variation in spending. Recent research has shown, however, that even the most sophisticated risk equalization systems tend to undercompensate (overcompensate) groups of people with poor (good) self-reported health, confronting insurers with incentives for risk selection. Self-reported health measures are generally considered infeasible for use as an explicit ‘risk adjuster’ in risk equalization models. This study examines an alternative way to exploit this information, namely through ‘constrained regression’ (CR). To do so, we use administrative data (N = 17 m) and health survey information (N = 380 k) from the Netherlands. We estimate five CR models and compare these models with the actual Dutch risk equalization model of 2016 which was estimated by ordinary least squares (OLS). In the CR models, the estimated coefficients are restricted, such that the under-/overcompensation for groups based on self-reported general health is reduced by 20, 40, 60, 80, or 100%. Our results show that CR can improve outcomes for groups that are not explicitly flagged by risk adjuster variables, but worsens outcomes for groups that are explicitly flagged by risk adjuster variables. Using a new standardized metric that summarizes under-/overcompensation for both types of groups, we find that the lighter constraints can lead to better outcomes than OLS.

Background. Profiling is increasingly being used to generate input for improvement efforts in health care. For these efforts to be successful, profiles must reflect true provider performance, requiring an appropriate statistical model. Sophisticated models are available to account for the specific features of performance data, but they may be difficult to use and explain to providers. Objective. To assess the influence of the statistical model on the performance profiles of primary care providers. Data Source. Administrative data (2006–2008) on 2.8 million members of a Dutch health insurer who were registered with 1 of 4396 general practitioners. Methods. Profiles are constructed for 6 quality measures and 5 resource use measures, controlling for differences in case mix. Models include ordinary least squares, generalized linear models, and multilevel models. Separately for each model, providers are ranked on z scores and classified as outlier if belonging to the 10% with the worst or best performance. The impact of the model is evaluated using the weighted kappa for rankings overall, percentage agreement on outlier designation, and changes in rankings over time. Results. Agreement among models was relatively high overall (kappa typically >0.85). Agreement on outlier designation was more variable and often below 80%, especially for high outliers. Rankings were more similar for processes than for outcomes and expenses. Agreement among annual rankings per model was low for all models. Conclusions. Differences among models were relatively small, but the choice of statistical model did affect the rankings. In addition, most measures appear to be driven largely by chance, regardless of the model that is used. Profilers should pay careful attention to the choice of both the statistical model and the performance measures.

Background: The risk-equalization (RE) model in the Dutch health insurance market has evolved to a sophisticated model containing direct proxies for health. However, it still has important imperfections, leaving incentives for risk selection. This paper focuses on refining an important health-based risk-adjuster in this model: the diagnosis-based costs groups (DCGs). The current (2017) DCGs are calibrated on “old” data of 2011/2012, are mutually exclusive, and are essentially clusters of about 200 diagnosis-groups (“dxgroups”). Methods: Hospital claims data (2013), administrative data (2014) on costs and risk-characteristics for the entire Dutch population (N≈16.9 million), and health survey data (2012, N≈387,000) are used. The survey data are used to identify subgroups of individuals in poor or in good health. The claims and administrative data are used to develop alternative DCG-modalities to examine the impact on individual-level and group-level fit of recalibrating the DCGs based on new data, of allowing patients to be classified in multiple DCGs, and of refraining from clustering. Results: Recalibrating the DCGs and allowing enrolees to be classified into multiple DCGs lead to nontrivial improvements in individual-level and group-level fit (especially for cancer patients and people with comorbid conditions). The improvement resulting from refraining from clustering does not seem to justify the increase in model complexity this would entail. Conclusions: The performance of the sophisticated Dutch RE-model can be improved by allowing classification in multiple (clustered) DCGs and using new data. Irrespective of the modality used, however, various subgroups remain significantly undercompensated. Further improvement of the RE-model merits high priority.