Data
We use information from five nationwide administrative registries and one survey which are all linked by Statistics Netherlands at the individual level
4 . The administrative data would be readily available if risk adjustment were implemented and include (1) health care expenditures in 2000–2004 from the health insurance data collected by Vektis; (2) use of LTC in 2004 and 2005, which includes home care, social assistance, assistance with activities of daily living and inpatient stays in either a residential home or a nursing home and which comes from the Central Administration Office of the LTC insurance scheme (CAK); (3) hospital admissions in 2002, 2003 and 2004 from the hospital discharge register (LMR); (4) demographic information for 2004 from the municipal register (GBA) and (5) mortality from the cause-of-death registry (CBS). In addition, the General Survey of Living Conditions (POLS) held in 2004 provides details on health, disability, and other individual characteristics for a randomly drawn, representative sample of the non-institutionalized population. Prior health care expenditures are registered for sickness fund enrollees only (two-thirds of the population)
5 and LTC use is registered for adults only (
\(\ge \)18 years of age); the other administrative data sets comprise the entire Dutch population.
The sample was further reduced for two reasons. First, the records for one third of those eligible for sickness fund coverage cannot not be linked. Second, 1.7 % of the sample was excluded because of item non-response which always was the result of missing co-residence status. As a result, the final sample consists of individuals who were insured through a sickness fund, did not die in 2004 and whose records could be linked to the municipality register. The total study population was 5,719,934, which is 45 % of the Dutch adult population in 2004. From this subset of the population, 7,790 individuals were included in the 2004 POLS survey; 3,619 of these respondents also completed the more specific health module.
Methods
A good risk adjustment system should reduce insurers’ incentives for risk selection while maintaining their incentives for efficiency. Ideally, after risk adjustment there are no easily identifiable subgroups for which insurers are undercompensated or overcompensated. In addition to an accurate prediction of individual expenditures, good risk adjusters should provide appropriate incentives and should be administratively feasible (Van de Ven and Ellis
2000). Partly following Beck et al. (
2010) and Shen and Ellis (
2002) among others, we identify the extent to which a risk adjustment model can reduce incentives for risk selection in three steps. First, we measure the insurers’ incentives to select against subgroups
6 based on individual characteristics in case of community-rated annual contracts but in the absence of risk adjustment. To quantify the insurers’ incentives for risk selection, we calculate the difference between the average actual expenditures by subgroup and the average expenditures for the entire population in 2005. We consider the incentives for risk selection to be strong when the number of users in the subgroup is substantial
\((>\)300), the predicted loss for a person in this group—the difference between observed expenditures for this subgroup and average expenditures for the entire population—is large (
\(>\)1,000 euro) and significantly
\((p < 0.05)\) different from zero. When these criteria are met, the benefits of risk selection are likely to exceed the costs and therefore the subgroup is included in the risk-adjustment model.
Second, we build the full risk adjustment model in a stepwise manner to examine to what extent each set of individual characteristics contributes to explaining individual variation in LTC use. To this end, we estimate a series of four models. We first test the impact of a basic model based on demographic characteristics on the predicted loss for all subgroups. Next, we add subgroups based on (i) prior LTC use, and (ii) prior health care expenditures and hospital admissions to this basic model variables. The full model includes all subgroups that were identified in the first model. For each risk adjustment model, the remaining predicted loss is the difference between the observed expenditures for these subgroups and the expenditures predicted by the risk adjustment model.
Third, for each subgroup that is included in the full model, we assess the impact of including this subgroup in the risk adjustment formula on the insurers’ incentives for efficiency—a commonly used selection criterion (see e.g. Van Kleef and Van Vliet (
2010), Van de Ven and Ellis (
2000) and Pope et al. (
2000)). Subgroups that are likely to have a negative impact on the insurers’ incentives for efficiency are those for which conditions of eligibility can be easily manipulated by insurers and for which it is highly attractive for them to do so. Manipulation may be financially attractive when the expected benefits exceed the costs, which consist of the required effort and the cost of the additional treatment that the enrollee is required to receive to be eligible for the subgroup. Excluding these subgroups from the full model results in an incentive compatible risk adjustment model. This third step thus sheds light on the tradeoff between creating incentives for efficiency and incentives for risk selection.
All five models described above are estimated by ordinary least squares regression (OLS) in order to facilitate interpretation of the results (Van de Ven and Ellis
2000)
7. Moreover, all current Dutch risk adjustment models use OLS, so using OLS increases the comparability and compatibility with these models.
The POLS sample was very small compared to the population of sickness fund enrollees and therefore the subgroups based on detailed information about health status, disability and socio-economic status from the POLS survey are not included in the risk adjustment model. Instead these subgroups are used as a benchmark to evaluate the impact of the risk adjustment model on incentives for risk selection.
Variables
In each of the models, the dependent variable measures public LTC expenditures in 2005. In case the individual dies in 2005, expenditures are annualized by dividing expenditures by the share of the year the individual was alive. The data set provides information on the quantity of LTC that was provided in kind, which was 95 % of the publicly financed LTC in the Netherlands in 2006
8 (CVZ
2011). The quantities provided, i.e. days institutionalized or hours of home care, are multiplied by the maximum prices as set by the government in order to calculate expenditures; co-payments are not taken into account. The data contains information about institutional care use in 2004 and 2005 and about all use of six types of home care in 2004. For 2005, the data contained information about use of only four out of six types of home care
9.
The set of subgroups that make up the basic model are based on three demographic characteristics: age, gender and co-residence, i.e. whether someone lived in a single-person household. Age and gender are the backbone of any risk-adjustment model, while co-residence proxies informal care availability. Informal care availability is an element of the eligibility assessment procedure for homecare (CIZ
2005) and formal LTC use is known to be correlated with informal LTC use (Bonsang
2009; Van Houtven and Norton
2004).
The subgroups of LTC users are based on prior LTC
use rather than
expenditures because using prior LTC use as a risk adjuster rewards insurers for negotiating lower prices with providers. Subgroups are created for each type of home care and each type of institutional care separately. Each of the subgroups of home care users consists of individuals who used this specific type of home care at least 1 hour per week on average. In selecting subgroups of institutional care users, we aim at balancing responsiveness to changes in LTC use against incentives for overreporting and oversupply resulting from the (partial) reimbursement of additional expenditures in the future. Therefore, for each of the four types of institutional care, four subgroups are generated consisting of individuals who stayed in an LTC institution for
\(\ge \)1 day, 91–180 days, 181–365 days, and the entire year (366 days), respectively. These subgroups reflect differences in expected future expenditures between long-term and short-term residents: future expenditures are positively correlated with the number of days that the individual is institutionalized. Furthermore, following Van Barneveld et al. (
1997), two subgroups are created consisting of patients who received home care and institutional care, respectively, on the last day of 2004, which shows the size of the predictable loss for enrollees who only use a very small amount of LTC in the prior year.
We also include subgroups based on prior HCE. Each of these subgroups measures health care expenditures
10 that are associated with LTC use: expenditures on hospital and outpatient care, prescription drugs, paramedical care, transportation, and durable medical equipment. For each of these categories, three subgroups are constructed that consist of persons who are among the 15 % who had the highest expenditures during the last year (omitted for hospital and outpatient expenditures), during each of the last 3 years, and during each of the last 5 years. Because the data only includes HCE covered by sickness funds, we also include a variable indicating which persons were not insured through a sickness fund in one of the 4 years preceding 2004. If someone was no longer registered with a sickness fund during a year, e.g. because of losing his/her eligibility status due to exceeding the income threshold, and hence is not in the data set for the entire year, expenditures are annualized.
In addition to the subgroups based on prior HCE, we also create subgroups based on hospital admissions because information on hospitalization and diagnosis information may help to predict LTC use (Wong et al.
2010). Subgroups are based on 94 diagnoses (based on a grouping algorithm of ICD-9 codes, see Polder et al.
2002) and on 48 types of treatments (based on ICD-9-CM volume 3 codes) using hospital admission data from 2002–2004. In addition, we create 12 Diagnostic Cost Groups (DCGs). DCGs are used for risk adjustment in the Dutch health insurance scheme and consist of clinically homogenous inpatient diagnoses for chronic health problems that have similar future HCE (Van de Ven and Ellis
2000). Using the ICD-code of the main diagnosis and the medical specialty that set this diagnosis, each individual is assigned to either the reference group (DCG 0)—people with no hospital admission or an incidental admission (e.g. fractures)—or the highest DCG they are eligible for (Rijksoverheid
2005; Prinsze and van Vliet
2007)
11. We include the DCGs but not the separate subgroups based on diagnoses and treatments in the risk adjustment model because the subgroups based on diagnosis and treatments and the DCGs overlap. Furthermore, the impact of the DCGs on the incentives for efficiency is known to be limited in the context of health insurance (Lamers
1998) while including all subgroups separately will increase incentives for oversupply and over-reporting.
As the administrative data do not provide detailed information on personal characteristics, subgroups based on health, disability and socio-economic characteristics could only be created using the smaller set of respondents that completed the POLS survey. Although it is much smaller and persons in nursing homes are not sampled, this survey allows investigating incentives for insurers to use such questionnaires for risk selection purposes. The same subgroups are used as in De Meijer et al. (
2011), who study determinants of LTC expenditures among the elderly, and in Stam and van de Ven (
2008), who identify subgroups that generate losses for health insurers. Of these subgroups, only those are selected for which the predicted loss deviates significantly from zero in the absence of risk adjustment. Because the average predicted profit without risk adjustment for the POLS sample and the subsample answering the health module are positive, the predictions for these samples are adjusted by subtracting the mean deviation from zero for the relevant sample multiplied by the ratio of the individual’s observed expenditures to the sample mean observed expenditures in order to ensure that the average predicted profit was zero for this subsample.