ADJUSTMENT OF RISK IN HEALTH
There are three problems in competitive health insurance systems related to the way in which the insurance service is paid. These problems originate in the existence of a heterogeneous population, with large differences in their propensity to claim services and major social challenges such as equity and economic efficiency. Once these problems have been identified and quantified, we propose two concrete ways to mitigate them in the contributive regime and evaluate them quantitatively and in a comparison with the current state of the Colombian health system.
In Colombia, Law 100 of 1994 transformed the public health system into a competitive system of insurance and cross subsidies in line with some of the best international practices (van de Ven, W., & Ellis, R. (1999)). In the case of public systems and, in particular, in the Colombian case, the core of this form of organization is based on three elements: a benefit plan (ie, the POS in Colombia), a set of insurers (ie, health promoting entities or EPSs) that, through their network of health services providers (ie, IPSs), have the responsibility to offer the services, procedures, and drugs in the plan; and, additionally, the definition of a mechanism for the payment of these services. In the absence of a criterion for social equity, if insurers enter into direct negotiation with affiliates, disproportionately different premiums would be charged to each individual due to the heterogeneity of the population and the estimated risk of each affiliate. In general, this disparity in the premiums is considered unfair and that is why the State enters to define a mechanism of cross subsidies in which the affiliates with the lowest cost to the system finance those that have the highest cost.
There are two payment characteristics that are relevant to this study. On the one hand, the payment must be ex-ante, that is, prior to the claims of health services by the affiliate. The main motivation for this is to encourage insurers to curb the system’s expenditure. The second characteristic is the need to adjust the payments to the risk profile of the insured. If the distribution of resources did not depend on the risk profile of the insured, perverse incentives would be generated for insurers to select risks. That is to say, the insurers would seek to affiliate only healthy people, of low risk, and in that way, they would try to maximize their expected benefits by affiliate. For the same reason, they would have incentives not to enroll some individuals, a practice considered illegal, or to use subtle strategies to discourage the affiliation of high risk individuals (strenuous procedures, long queues, poor quality of service, delays in requesting appointments and authorizations, etc.). The value transferred to each insurer adjusted for the risk characteristics of the affiliates was studied in this project.
The theory states that, in principle, except maybe for regulatory reasons, the adjustment mechanism should try to minimize the uncertainty of future spending. In Colombia, this ex-ante adjustment is made based on some age groups, sex, and three areas of residence. This ex-ante adjustment mechanism can be considerably improved using information currently available from the Ministry of Health and Social Protection.
To improve this ex ante risk adjustment it is necessary to do two things: enrich the information used to make the adjustment, taking into account morbidity information, and use a more sophisticated algorithm or formula that improves the prediction of affiliate spending for each EPS. Riascos, Romero, and Serna (2017), using machine learning techniques (Buchner, F., Wasem, J., and Schillo, S. (2015), Bagheri, Li, Goote, S., Hasan, A., and Hazard, G. (2013), compare the Colombian government’s model for risk adjustment against alternative specifications that control for (i) the morbidity of members, characterized in 29 groups of long-term illnesses , (ii) indicators of hospitalization, visit to the specialist, and admission to the intensive care unit, and (iii) all the possible interactions between the sociodemographic variables of the government’s base model. Table 1 shows the results of the comparison in terms of the total expenditure forecast and Table 2 shows the results of the comparison of the expenditure prediction for the highest expenditure quintile.
Table 1. Out-of-sample fitting for the full distribution
Note: This table presents the RMSE (square root of the average squared prediction errors), MAE (average of the absolute errors), PR prediction ratio for the annualized and non-annualized expense, and the R-square in the complete sample. WLS UPC is a linear regression model estimated with weights. WLS UPC + Dx is the same as the previous one plus the diagnostic groups. ANN FS is a neural network with statistical selection of variables. RF FS is a random forest with statistical selection of variables and the GBM FS model is a tree boosting with statistical selection of variables. Calculations of the authors based on the Sufficiency Bases.
Table 2. Out-of-sample fitting in the last quintile
Note: See explanation for Table 1. The RMSE, MAE and PR are defined in the same way except that, in this table, they are estimated only for the most expensive quintile of the expenditure distribution.
As can be seen, GBM FS manages to predict 50% of the total expenditure of this quintile representing an improvement of 7 percentage points with respect to the linear model with all the variables, Model 2, and of 20 percentage points relative to the government’s base model, Model 1. In any case, the results of this table suggest that the current risk adjustment model of the Ministry (WLS UPC) does not predict well the annual health expenditure of the top quintile: it manages to predict only 33.5% of spending in this quintile. The results also show that the inclusion of the diagnostic groups, Model 2, improves the predictive ratio relative to the Ministry’s model by 10 percentage points. This suggests that morbidity is an important risk factor and that adjusting the UPC to these diagnostic groups would improve the description of the risk profile of insurers.
More information in the following video (6 minutes):