Introduction
Many health insurance systems are based on the model of regulated competition. Competition among health insurers helps to improve efficiency of health insurance systems and regulation helps to protect public objectives like individual affordability of health plans. One element of the regulatory framework is risk equalization, a mechanism that compensates health insurers for predictable spending variation across individuals [
34,
35]. In the presence of premium-rate restrictions, as applied in (almost) all regulated health insurance markets, risk equalization mitigates incentives for risk selection.
Over the past decades, risk equalization systems have evolved from simple demographic models to sophisticated health-based models. An example of the latter is the model applied in the Netherlands, which includes risk adjusters based on an extensive series of demographic, socioeconomic, and morbidity-based variables. Even these sophisticated models, however, do not completely correct for predictable spending variation [
13,
23,
30]. Van Kleef et al. [
31] find that the Dutch risk equalization model of 2016 undercompensates health insurers for the group of consumers who reported a fair or (very) poor health status in the prior year and overcompensates them for the group of consumers who reported a (very) good health status in the prior year. On average, the former group (about 24% of the population) confronts health insurers with a predictable loss of around 500 euros per person per year, while the latter (about 75% of the population) confronts them with a predictable profit of around 180 euros per person per year [
31].
Correlation between consumers’ (self-reported) health and their profitability to health insurers can be problematic for the functioning of health insurance markets. When the unprofitable groups in poor health value (specific features of) health plans differently than the profitable groups in good health, health insurers are confronted with incentives to design their plans in a way that these are more attractive to healthy consumers than to unhealthy consumers. For instance, health insurers might refrain from contracting high-quality care for unprofitable groups with particular chronic medical conditions [
10,
11]. These actions, which we refer to as
selection via plan design, threaten the efficiency of health plans [
12,
15,
17,
25,
27,
36].
This paper seeks to mitigate incentives for selection via plan design by incorporating health survey information in the risk equalization model. However,
direct use of self-reported health measures as a basis for risk adjusters is problematic, because the required survey information is not available for the entire population (which is typically considered a requirement for calculating individual-level risk equalization payments). Collecting this information for the entire population would usually be considered too cumbersome and costly [
34].
Although self-reported health measures are not appropriate as a basis for risk adjusters, they can be used
indirectly in risk equalization models through the method of constrained regression (CR). Conventional risk equalization models are usually estimated by means of ordinary least squares (OLS). Given a set of risk adjusters, OLS results in coefficients that minimize the sum of squared residuals. CR allows for estimating coefficients that minimize the sum of squared residuals
conditional on a pre-specified under- or overcompensation (for instance zero) for specific groups [
29]. Previous research has shown that application of CR can improve payment fit for groups not explicitly flagged by risk adjusters. At the same time, CR typically worsens payment fit for groups explicitly flagged by risk adjusters. Van Kleef et al. [
29] have applied CR in the Dutch context for the risk equalization model 2015 and concluded that the improved payment fit for some groups can potentially outweigh the deteriorated payment fit for other groups.
The aim of this study is to examine and evaluate the use of health survey information in risk equalization through CR. To do so, we use administrative data and health survey information from the Netherlands. The administrative data are from 2013 and contain information on medical spending and risk adjuster variables for the entire Dutch population (N ≈ 17 m). These data are used to replicate the Dutch risk equalization model of 2016. Furthermore, we use health survey data from 2012 based on a large sample of the Dutch population (N ≈ 387 k). We estimate six models, that is, one base model estimated with OLS (i.e., the Dutch risk equalization model 2016) and five models estimated with CR.
Our empirical application comes with two methodological challenges. First, to meaningfully use health survey information as a basis for CR to improve risk equalization, this information must be representative for the population. As with most samples, this is not entirely the case for our survey sample. Prior studies have shown that this sample is somewhat healthier than the population [
29,
37]. We address this by rebalancing the sample using a raking procedure [
1,
18] to correct for mismatches between the sample and the population. Second, a metric is required to evaluate the outcomes of CR relative to OLS. We use a new standardized evaluation metric that summarizes under- and overcompensations for a cross tabulation of two types of groups, i.e., groups explicitly flagged by risk adjusters (for which previous research has demonstrated an
increase in under-/overcompensation with CR compared to OLS) and groups not explicitly flagged by risk adjusters (for which previous research has shown a
decrease in under-/overcompensation with CR compared to OLS). More specifically, we first calculate the total under-/overcompensation per group, take the absolute value of these total under-/overcompensations, and then sum these over the relevant groups.
The structure of this paper is as follows. “
The Dutch health insurance market” section describes relevant aspects of the Dutch health insurance system. “
Literature review” section summarizes the relevant theory and previous research on selection via plan design and CR. “
Data and methods” section describes the data and methods for our empirical application and “
Results” section presents the results. Finally, “
Discussion” section summarizes and discusses the main findings.
The Dutch health insurance market
The Dutch health insurance market has two main components: a basic health insurance and a supplementary health insurance. Supplementary health insurance operates on the basis of free competition and is beyond the scope of this research. The basic health insurance operates on the basis of regulated competition. Regulations implemented by the Dutch government to ensure individual affordability and accessibility of the basic health insurance, include an individual mandate to buy basic health insurance, annual open enrollment, community-rated premiums, risk equalization, and a standardized benefit package. The latter means that health plans have to cover a fixed set of benefits. Insurers are, however, free to selectively contract healthcare providers. Although this is intended to improve the efficiency of health care, health plans can also use this instrument to engage in selection via plan design, e.g., by not contracting good quality health care for specific unprofitable groups of consumers, also known as ‘quality skimping’ [
32,
36].
Risk equalization mitigates incentives for selection via plan design, given premium-rate restrictions. The risk equalization model is used to calculate risk-adjusted payments to health plans, based on the characteristics of their insured population. The Dutch risk equalization model is comprised of three separate models: one for somatic health care, one for mental health care, and one model for copayments due to a mandatory deductible [
32]. This research focuses on the model for somatic health care, which contains the following indirect indicators of health: age, gender, region, socioeconomic status, and source of income. In addition, the model includes the following series of more direct health indicators: pharmacy-based cost groups (PCGs), diagnosis-based cost groups (DCGs), multiple-year high cost groups (MHCGs), durable medical equipment cost groups (DMECGs), physiotherapy spending in the previous year, home care spending in the previous year, and geriatric rehabilitation care spending in the previous year [
32]. In this paper, we refer to these direct health indicators as ‘morbidity-based risk adjusters’.
Discussion
Most health insurance markets with premium-rate restrictions include a risk equalization system to compensate health insurers for predictable variation in spending. Recent research has shown, however, that even the most sophisticated risk equalization systems tend to undercompensate (overcompensate) people with poor (good) self-reported health, which confronts insurers with selection incentives. Self-reported health measures are generally considered infeasible for use as ‘risk adjusters’ in the risk equalization model. The aim of this paper was to examine and evaluate an alternative way of including self-reported health measures in risk equalization, namely through constrained regression (CR). To do so, we estimated five CR models and compared these with the actual Dutch risk equalization model of 2016 estimated with ordinary least squares (OLS). In the CR models, coefficients were estimated by least-squares regression given that the under-/overcompensation for two groups based on self-reported general health are reduced by 20, 40, 60, 80, or 100%.
We first calculated the under- and overcompensations for selected survey groups and groups flagged by the morbidity-based risk adjusters included in the risk equalization model. For the survey groups, the results showed that the chronically ill receive more compensation under CR compared to OLS, while the opposite is true for the complementary groups of healthy people. We observed a similar pattern for the groups (not) explicitly flagged by a morbidity-based risk adjuster; the groups that were explicitly flagged by such a risk adjuster receive more compensation under CR compared to OLS and the groups not explicitly flagged receive less. Next, we researched subsamples of these groups by cross tabulating the groups yes/no (very) good self-reported general health with the groups yes/no explicitly flagged by at least one morbidity-based risk adjuster. The results showed that—compared to OLS—also within the groups of self-reported general health, the CR models move money from the individuals not flagged by a morbidity-based risk adjuster to those flagged by such a risk adjuster. Consequently, we found that payment fit improves for some groups but worsens for others. Van Kleef et al. [
33] reported similar findings.
To evaluate the outcomes under all six models, we constructed a standardized metric that summarizes the absolute under-/overcompensations for relevant subgroups. We evaluated the four groups resulting from the cross tabulation of yes/no (very) good self-reported general health with the groups yes/no explicitly flagged by at least one morbidity-based risk adjuster. In this metric, we take the absolute values of the total under-/overcompensations and sum these over the four groups. The metric then compares the outcomes of a CR model relative to OLS. We find that the CR-20%, CR-40%, and CR-60% models yield more preferable outcomes than OLS, with the CR-40% model yielding the best results (i.e., for the groups analyzed here). This finding shows that a relatively small constraint could already improve conventional risk equalization. This is in line with the conclusions drawn in the paper by Van Kleef et al. [
29] and with the findings of the work by Glazer & McGuire [
15,
16]. Glazer and McGuire [
15,
16] argued that conventional risk equalization estimated with OLS might not be optimal and that overpaying groups flagged as ‘high risk’ and underpaying groups of ‘low risk’ could improve the outcomes of risk equalization. However, our results also show that CR in risk equalization can be pushed too far, since the metric increases sharply as the constraint becomes heavier, with the CR-80% and CR-100% models performing worse than OLS.
Our primary simulations assume equal weighting of (the under-/overcompensations of) subgroups. Acknowledging that regulators might consider the effects of some selection actions to be more harmful than others, we also examined how differentiated weighting could influence the model outcomes. We found that a specific form of weighting (based on assumptions about the effects of quality skimping and underserving versus overserving) substantially affects the outcomes of the CR models relative to OLS. These results demonstrate the relevance of carefully defining the policy objectives which regulators want to include in the evaluation.
The results of this study indicate that the use of health survey information in risk equalization through CR can be promising in reducing incentives for selection via plan design. Practical implementation of survey information in risk equalization through CR, however, needs more work. First, evaluation can be more refined, for example by evaluating the outcomes using other and more groups than analyzed here. In addition, more refined evaluation of risk equalization models could require a welfare approach that incorporates how incentives affect the behavior of insurers, how this behavior of insurers interacts with the behavior of consumers, and how this affects social welfare. Although such a welfare approach is beyond the scope of this paper, we believe that further research into this direction can help to improve the evaluation of risk equalization systems. Second, the choice of groups on which the constraints are based can differ from the groups used is this research. This choice is, however, not ours to make. The method of CR offers regulators an effective tool for protecting specific groups of interest against selection via plan design [
29]. An important insight in this respect is that these groups can also be determined on subsamples of the population, as long as these subsamples are representative for the population.
Acknowledgements
The authors are grateful for the thorough comments on a previous version of this article by two anonymous reviewers. They are also grateful to René van Vliet, Wynand van de Ven, Erik Schut, and Thomas McGuire for valuable feedback during this project. Furthermore, the authors would like to thank the participants of the HSI seminar series and the attendees of the RAN meeting 2018 for helpful discussions. The authors also gratefully acknowledge the Dutch Ministry of Health, the Dutch Association of Health Insurers and Statistics Netherlands for providing the administrative and survey data. Finally, the authors thank the members of the supervisory committee for their valuable comments. Remaining errors are the responsibility of the authors.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.