Cohen and Siegelman (2009) document empirical research on adverse selection in 5 markets: i) automobile insurance, ii) annuities and life insurance, iii) long term care, iv) crop insurance, and v) health insurance. The presence of adverse selection varies not only across markets, but also within markets depending on the product sold and the type of individuals who buy the product.
Adverse selection occurs when high risk individuals are the ones more likely to purchase insurance. However, measuring adverse selection may not be asstraightforward empirically as it seems. Typically, economists conclude that there is adverse selection in a market if a correlation exists between risk levels and whether or not the individual buys insurance.
If adverse selection is at work, however, this correlation may not necessarily show up in the data. For instance, high risk individuals may be risk loving while low risk individuals are risk averse. Thus, adverse selection may be occurring, but due to the correlation of risk preferences, this may not be born out in the data. Further, a correlation between risk and insurance status does not necessarily imply the existence of adverse selection. In health insurance markets, insured individuals may incur more cost due to moral hazard. If the insured and uninsured have equal risk levels, the person with health insurance may still incur more medical costs, because these services are generally free to them.
Further, detecting adverse selection econometrically, is not simple as well. The researcher must have full access to the insurer’s information to reliably estimate the level of private information in a market. Further, all accidents do not result in claims. Individuals with high deductible health insurance may fall ill, but not go to the doctor. Drivers who get in fender benders may not report the incident. There may also be unobservable differences among policyholders. Finally, one may use a variety of econometric techniques to estimate the presence of adverse selection. Many researchers use a simple OLS structures as follows:
- Riski=α + βCoveragei + γXi + εi
Others use a dual regression method.
- Riski=f(Xi) + εi
- Coveragei =g(Xi)+ ηi
In this case, the regressions are estimated simultaneously. If the correlation between the residuals is positive and statistically significant, then there is evidence of coverage-risk correlation.
The authors conclude that the presence of adverse selection will depend on the following factors:
- an absence of useful private information,
- the existence of private information for some but not all policymakers in a market,
- policyholders’ inability or failure to use the private information they have,
- the presence of superior information or predictive power on the part of the insurer,
- propitious selection resulting from the interaction between risk and risk aversion or other policyholder characteristics associated with an increased tendency to purchase insurance, and
- institutional arrangements.
Source: Cohen, Alma and Siegelman, Peter (2009) “Testing for Adverse Selection in Insurance Markets,” NBER Working Paper # 15586.