Prohibitions on Health Insurance Underwriting:

A Means of Making Health Insurance Available and Affordable

Or a Cause of Market Failure?

 

 

Authors:

 

Mark J. Browne

School of Business

University of Wisconsin – Madison

Madison, WI  53706-1323

 

Edward W. Frees

School of Business

University of Wisconsin – Madison

Madison, WI  53706-1323

 

 

 

Abstract

 

Health insurance underwriting restrictions that prohibit insurers from using disability status, gender, and age to classify risks will in theory result in greater insurance consumption by those positively impacted by the prohibition.  Conversely, the prohibitions are expected to result in less health insurance consumption by those negatively impacted.  Binomial and multinomial logit analysis and data from the Current Population Survey are used to test these hypotheses.  Contrary to the hypotheses, evidence is not found that underwriting prohibitions pertaining to gender and age have effected the consumption of insurance in the small group market.  Some evidence is found that underwriting restrictions on the use of disability status have affected insurance purchases in the small group market.

 

 

A.        Introduction

            This project examines the effects of health insurance underwriting restrictions on insurance coverage.  The study addresses several important questions that have to date received little empirical analysis.  In theory underwriting restrictions result in a lower price of insurance for those who would have been adversely categorized had underwriting been permitted.  Conversely the restrictions result in a higher price of insurance for those who would have been favorably categorized by the underwriting criteria.  While theory suggests that increasing the price of insurance that is paid by those who benefit from underwriting restrictions will encourage them to reduce their coverage, it is unknown whether underwriting prohibitions will result in a significant enough price increase to induce this.  Similarly it is unknown whether those who currently do not purchase coverage and who would benefit by not being underwritten against will be enticed by a lower price to purchase coverage.  The research addresses these empirical questions by examining the effects on health insurance consumption of underwriting restrictions pertaining to disability status, gender, and age passed by the various states. 

            The study extends our knowledge of how insurance markets function when information is incomplete.  Studies of health insurance markets by Browne (1992) and Browne and Doerpinghaus (1993, 1995) have consistently shown that information asymmetries result in reduced insurance purchases by low risks.  The study tests whether lower risks are induced by underwriting restrictions to discontinue their coverage as opposed to reducing the amount of coverage they hold.

 

B.         Background and Significance

            In spite of the potentially dire consequences to one's health, and perhaps life, of not having health insurance, 15 percent of the population below age 65 did not have health insurance according to a 1991 study of the United States General Accounting Office (GAO).^[1]  The Employee Benefits Research Institute (EBRI) reports that the percent of the population uninsured grew to 18.1% in 1993.^[2]  The GAO and EBRI studies found that compared to the general population the uninsured tended to be younger and to have lower incomes.  A disproportionate number of uninsured come from minority groups and are unmarried.

            Reasons why individuals may not have health insurance include a lack of understanding of the benefits of health insurance and financial inability to purchase insurance.  For some individuals in poor health, the cost of insurance may seem prohibitively expensive.  For some individuals in good health, the cost of health insurance may seem too high in relation to the potential costs of being uninsured.

            Efforts to reform the health care system have centered on the goal of reducing the number of uninsured individuals.  Reform proposals range from nationalization of the health care delivery system to modifications in the way insurers are permitted to operate.  One proposal that has received considerable discussion, but scant empirical analysis, is a prohibition on underwriting, the process by which health insurers choose who they will insure.  Recently, the state of New York prohibited most forms of underwriting by mandating that health insurers practice community rating.  Several other states are currently considering similar legislation.  Underwriting restrictions are supported by many ethicists who contend that underwriting is morally indefensible (Daniels, 1990) and by market reformers who view underwriting as a barrier to insurance for those in poor health.  Health insurers, despite their stated commitment to universal coverage, oppose restrictions on underwriting.^[3]

            Those supporting underwriting restrictions emphasize that underwriting has the potential to result in cream skimming in insurance markets.  Cream skimming refers to the selection of good risks by insurers and the refusal to insure high-risks.  Examples provided by Aaron and Bosworth (1994) include insurers who refuse to insure gas station attendants because they face a high risk of violence on the job and male hairdressers who are believed at high risk of AIDS.  Pauly (1984) contends that cream skimming occurs when regulation precludes insurers from charging rates commensurate with risk.  Rate regulation may cap the maximum premium insurers can charge, thus making it an unprofitable proposition to insure high risks.  Pauly’s argument suggests that in the absence of regulation insurers would be willing to provide insurance to high risks, albeit at high premium levels.

            Aaron and Bosworth (1994) state that the purpose of experience rating is to reduce cream skimming which would occur with community rating.  When community rating is required insurers must charge all risks a similar premium.  Insurers thus have an incentive to select the better risks and not insure the worse risks.  To counter the incentive for cream-skimming, market reforms that call for community rating such as the Clinton plan and the New York plan mentioned earlier entail financial transfers between insurers to offset enrollments of different risk types.  The degree to which these transfers can reduce cream skimming is not known.  Constructing risk-based inter-company transfers is extremely complex.  The best available techniques are not believed to be sufficiently advanced to eliminate the incentive for cream skimming.

            Despite the current national prominence of the debate on underwriting, the issue is not a new one.  Each of the states has passed regulations that restrict the ability of health insurers to use particular types of information for underwriting purposes, including denying coverage and charging differential premiums. In the next section, a model is developed that shows how restrictions on the use of underwriting information may encourage some currently uninsured individuals to purchase health insurance.  The model also shows that if underwriting restrictions are written into law, some individuals who are currently insured may terminate their insurance coverage. 

            Empirical determination of the effect of the underwriting restrictions imposed in the different states on individuals' decisions to insure is our central focus.  In particular the effect of prohibitions on the use of underwriting information pertaining to disability status, gender, and age are studied.  The impact of these restrictions on the different subpopulations that are most likely to be affected will be tested.  The knowledge to be gained from this study has direct implications for public policy focused on health insurance reform.  If underwriting restrictions are associated with reductions in the rate of uninsurance, then this type of market reform holds promise as a way to address in part the problem of uninsurance in the country.  Conversely, if the restrictions result in a greater degree of uninsurance, then they are ill advised.

 

C.        Prior Research:  The Effect of Underwriting Restrictions on the Decision to Insure

            Arrow (1970), Mossin (1968), and Smith (1968) have demonstrated that when insurance is priced at actuarially fair rates insureds prefer policies that offer full coverage.  Since insurance is not a costless business, insurers sell policies above the actuarially fair premium to cover their expenses.  Smith has shown that when insurance is available at a cost that exceeds the actuarially fair value and the probability of loss is greater than zero, the optimal level of insurance coverage will depend on an individual's degree of risk aversion and the cost of insurance. For a given risk‑averse individual, the optimal level of insurance will decrease as the cost of insurance increases.  Depending on the shape of the utility function, the optimal level of health insurance may be zero or exceed the value of the asset, human capital, subject to risk.

            Considerable empirical work has been done on the demand for health insurance.  In part this research has been motivated by the desire to understand why people choose to forego the purchase of insurance.  Van de Ven and Van Praag (1981) and Taylor and Wilensky (1983), as well as others, found that income is positively related to health insurance consumption.  Thomas (1995) demonstrates that the income effect varies by income level.  In particular, demand was found to decrease with increases in income for individuals at or below 125% of the poverty line and to increase with income above this threshold.

            Phelps (1976) provides evidence that price is negatively related to health insurance purchases.  Later studies, including Farly and Monheit (1985) and Browne (1992), support the finding of a negative relationship.

            Demographic characteristics which proxy risk aversion and human capital have been found by various researchers to be important determinants of coverage.  These characteristics include age, gender, industry of employment, degree of physical impairment, marital status, family size, race, and marital status.  Recent studies that have examined the relationship between demographic characteristics and health insurance purchases include Merrill, Jackson, and Reuter (1985), Freeman et al. (1989), Browne (1992), Browne and Doerpinghaus (1993), and Thomas (1995).

            A growing body of literature suggests that information asymmetries in insurance markets affect demand.  Akerlof's (1970) seminal work suggested that "bad" risks will drive "good" risks from an insurance market when information is asymmetric in favor of the buyer of insurance.  This phenomenon is referred to as adverse selection.  Equilibriums obtained in markets when information is asymmetric have been described by Rothschild and Stiglitz (1976), Riley (1979, 1985), and Cho and Kreps (1987). In each of their models the market equilibrium entails different risk types purchasing different policies and no cross-subsidization.  Market equilibriums proposed by Wilson (1977), Grossman (1979), and Hellwig (1987) predict that different risk types will purchase a common policy.  This implicitly entails subsidization of the high risks by the low risks.  Recent work by Young and Browne (1997) suggests that when pooling occurs in a market with adverse selection the pooling policy may have a lower policy limit than if an adverse selection problem did not exist.

            Empirical evidence to date suggests that the pooling equilibrium models better describe the health insurance market than the separating equilibrium models.  Browne (1992) and Browne and Doerpinghaus (1993) both report findings that are consistent with a pooling equilibrium in the market for non-group health insurance.  Browne and Doerpinghaus (1994) report similar findings in their study of adverse selection in the market for Medicare supplemental insurance.

            To illustrate how adverse selection can affect the demand for insurance consider two individuals.  The actuarially fair premium, P_j, for individual j, who has a probability of z_j of having no loss and a probability of (l- z_j ) of having a loss of size S_j  is:

 

P_j = (l- z_j) S_j.

 

For simplicity, assume that there are only two types of individuals, low (L) and high (H) risks, so that j=L or H. High risk individuals have a combined probability of loss and loss severity so that P_H > P_L . As has been previously mentioned, as long as insurance is available at an actuarially fair rate and both types of consumers are risk‑averse, both individuals will purchase complete coverage albeit at different premium rates.  If insurance companies cannot differentiate between risk types, a pooled premium will be charged to both.  The pooled premium, P_p, will be:

 ,

 

where M is the number of low risk types and N is the number of  high risk types.

            The amount by which the pooled premium exceeds the actuarially fair premium for low risk types is:

.

 

The excess E_L is positively related to the number of high risks in the pool, the high risks' probability of loss, and the high risks' severity of loss.  It is negatively related to the number of low risks in the pool, the low risks' probability of loss, and the low risks' severity of loss.

            As previously mentioned low and high risks may purchase a pooling policy if insurance companies are unable to distinguish between risk types.  If individual insureds are better able than insurance companies to perceive whether they are high or low risks, the high risks will have a greater demand for insurance than the low risks.  The possibility exists that the low risks will purchase no coverage.  As Smith (1968) noted, the decision to insure will depend on the shape of one's utility function and the cost of insurance.  Insurers may be unable to differentiate between low and high risks because of the costs of gathering information or because of regulatory restrictions on classification.  Each of the states has different restrictions on the use of information that health insurers may use to classify insureds.

            Underwriting restrictions are passed ostensibly to discourage insurers from discriminating in a way that is contrary to social policy.  In addition to prohibiting socially unacceptable discrimination, underwriting restrictions also have the effect of changing the cost of insurance.  Individuals who would have otherwise been discriminated against had the use of the information not been prohibited will be charged a lower price for insurance.

            Insureds who would have been classified favorably by the criteria will be charged a higher price for insurance.  The models of insurance market equilibrium previously mentioned suggest that there are three possible effects that regulatory restrictions on information may have on the purchase of health insurance.  First, low‑risk insureds, in this case those who would have been favorably categorized had the restriction not been implemented, may choose not to insure.  Second, if low‑risks do purchase insurance coverage, they may purchase insurance policies that offer less coverage than they would have otherwise purchased.  Third, high‑risk insureds may purchase more coverage than they otherwise would have.  Our empirical analysis focuses on the effect underwriting prohibitions have on the demand for health insurance by people of different races and of different ability statuses in the presence of different underwriting regulations.

 

D.        Data, Research Design and Methods

            The sources of data strongly influence our research design and methodology. Thus, we first discuss the data sources. Then, we discuss the variables that will be investigated to test the hypothesis that state regulatory restrictions on the use of underwriting criteria affect individuals' decisions to insure. This is followed by a brief description of the models and estimation techniques.

 

Data

            This study makes use of data from two sources: the Current Population Survey (CPS) conducted by the Bureau of Labor Statistics and state statutes that provide information on state underwriting restrictions.

            The Current Population Survey reports information on health insurance coverage in the March survey.  This information has been collected continuously since 1982.  We use the survey data for the period 1991 through 1995.  Prior to 1991 the survey did not collect information on employment group size in sufficient detail for our analysis.  The survey is the only publicly available source of individual level data on health insurance coverage during the complete study period.

            Long and Marquis (1996) fault the CPS for providing inaccurate estimates of the uninsured population.  Their contention is that the question requesting health insurance status can be misconstrued.  While this adds noise to our analyses, the effect suggested by Long and Marquis should occur equally across individuals and should in no way bias our study.  Another drawback to the CPS is that the question on health insurance coverage has changed slightly during the study period.  The manner in which the questions changed did affect our research design; see the following section.

            Information on underwriting restrictions is reported in state statutes.  There is significant variation in the restrictions enacted by the different states.  Further, the years of enactment of similar statutes by different states vary.  As mentioned earlier, our primary focus is on underwriting restrictions related to disability status, gender, and age. Other restrictions, such as bans on underwriting on the basis of exposure to DES, infliction with sickle cell anemia, and sexual preference, are not considered because of data limitations.  Underwriting prohibitions can result from either regulation or legislation.  We consider a state to prohibit a particular form of underwriting if either the state has enacted a law or regulation specifically forbidding the use of that criteria for classification purposes for all lines of health insurance or has enacted a community rating law that does not allow for differential pricing based on the criteria.  We also account for prohibitions that were enacted as part of small group reform laws.  That is, some states prohibit the use of specified underwriting criteria only in the small group market.  Appendix D reports the underwriting restrictions in effect in the different states during the period of analysis.

            Using broad cross-sections of the population, Gruber and Currie (1996) found little evidence that state mandates of health insurance coverage affect the purchase of health insurance. In consideration of related legal requirements of underwriting restrictions, our analysis differs in two fundamental ways. First, because different types of underwriting restrictions may be reasonably assumed to affect only certain consumers, our analysis subdivides the population into a smaller, more homogeneous subset. Second, although our initial analysis considers only the purchase of a type of insurance or not (the logistic regressions) as in Gruber and Currie, our subsequent analysis considers purchases of different types of insurance (the multinomial logits, described in Section D).

            The subpopulations we consider consist of only single person households. As described in Appendix A, the CPS is organized at the household level where insurance purchase decision-making is murky. By working with single person households, we clearly identify the decision-maker.

            We are interested in the effects of underwriting restrictions based on disability status, gender, and age; LAWDISA, LAWGEND, and LAWAGE are the corresponding indicator variables. In our first analysis, we focus on the small group market, employer size less than 10. The data are from 1991-1995.  Prior to 1991 the CPS does not distinguish group size less than 25.  For the purposes of our second analysis, using data from the period 1988-1995, we define the small group market as consisting of firms with 25 or fewer employees.

            The individual and state level control variables are described in Appendix C.  Further, because underwriting laws may affect specific subgroups differently, Appendix C describes interaction variables to accommodate this feature.     

Research Design

            To assess an individual's demand for health insurance, we use a variable to describe the type of health insurance purchased.  Specifically, the response variable denoted as y_ist, for time periods t = 1 , ..., 5 (corresponding to years 1991 through 1995), s = 1, ..., 50 (corresponding to each of the 50 states) and i = 1, ..., n_st, where n_st is the number of individuals in state s during time period t.  Thus, for individual i in state s during time period t, we define

           

A detailed description of the choice of response variable is in Appendix B.

            Different underwriting restrictions affect the demand for group and individual insurance in different ways. Further, it is important to account for other forms of coverage that individuals may seek in addition to simply the private health market.  Thus, we have included four categories that an individual may choose from, categories that represent various combinations of private individual and group and public insurance.

            We are primarily interested in testing hypotheses concerning the effect of underwriting restrictions on the demand for health insurance.  To assess these effects, we use indicator variables of the form W_jst.  Here, W_jst indicates whether underwriting restriction of type j is in effect in state s during time period t.  The underwriting restrictions investigated are race and disability status.  Analyzing y_ist in terms of W_jst is in the spirit of a “before/after” analysis, trying to determine whether an intervention has an important impact on the tendency to purchase health insurance.

            Because the CPS data may provide a non-representative sample of state populations, we also consider a number of explanatory variables that may influence an individual's decision to purchase health insurance.  Individual level variables that may be directly related to underwriting restrictions include disability status, gender and age.  These variables are denoted by R_j,ist, j = 1, ..., 5.  Other individual level variables that are used as controls include family income, whether or not the individual is employed, marital status, years of education, and race.  Also used is a variable indicating whether the individual was employed full or part time.  Full time is defined as working twenty or more hours per week.  These variables are denoted by X_j,ist, j = 1, ..., K.  Here, K describes the number of explanatory variables used to define each of the criteria believed to affect the demand for health insurance.  For instance, three categorical variables, Black, White, and Other, define race.  We have split off R from X because we feel it is more important to investigate potential interaction effects with the R variables that may be related to W than the X variables.

            Related studies also suggest the importance of certain control variables. Marquis and Long (1995) used the CPS data to study demand for private health insurance for workers who do not have group insurance. They used an individual’s gender, marital status, income, race, education, age, and number of children. Gruber and Poterba (1994) studied the purchase of health insurance by the self-employed using CPS data. They used an individual’s gender, marital status, race, education, whether or not self-employed and whether an individual worked full-time, part-time or not at all.

            The Gruber and Poterba (1994) study provides empirical evidence that the Tax Reform Act of 1986, which introduced a new tax subsidy for the purchase of health insurance by the self-employed, resulted in increased purchases of health insurance by this group.  During the period of our analysis the tax treatment of the self-employed’s health insurance premiums changed a number of times.  Medical costs from 1980-1982, including insurance premiums, were deductible if they exceeded 3% of adjusted gross income.  The percentage was increased to 5% for 1983-1986 and to 7.5% for 1987 until the present.  However, 25% of the self-employed’s premiums were not subject to this floor from 1987-1994.  This increased to 30% for 1995 and 1996.  We account for this changing tax treatment by including dummy variables for each year in the analysis.

 

Methods

            Because the response, y_ist, is a categorical variable, we use discrete choice models that are now well established in econometrics (see, for example, Greene, 1993, Chapter 21). Let  S_c denote the sum over c = 1,...,4.  With this notation, we will use a multinomial logit model with choice probabilities

 

 ,                                                            (1)

where

V_ist,c =  a_s,c  +  R¢_ist b_1,c  +  X¢_ist b_2,c  +  W¢_st g_c                            (2)

 

Here, R_ist  = (R_1,ist, ..., R_5,ist)¢  represents the vector of individual level control variables that may be related to underwriting restrictions and  b_1,c is the corresponding vector of parameters.  Similarly, X_ist , and W_st are vectors corresponding to the individual variables previously defined and b_2,c, b_3,c and g_c are the corresponding parameters.  The intercept parameters, a_s,c, may vary by state and thus can control for state-specific factors that are not explicitly controlled for.  A convenient normalization for this model is a_s,1 = 0, g₁=0  and  b_j,1 = 0  for j=1,2,3.

            Discrete choice models are desirable because they provide so-called “random utility” interpretations.  Under the random utility model interpretation, we assume that the utility of each choice can be represented as a linear function of explanatory variables plus an error, for example, U_ist,c = V_ist,c + e_ist,c.  We do not observe the utility, only the choice made.  For our application, the choice made is the type of health insurance purchased.  By assuming a special form of the errors, equations (1) and (2) are obtained.

See Judge et al (1985, Chapter 18) for more details.

            To estimate this model, we use maximum likelihood. For the model in equation (2), there are 147 (=49´3) state-specific effects (the a_s,c terms), continuous variables (such as age and income) and categorical variables (such as race and year). Thus, many of the models reported in Section E below have over 200 parameters. We programmed the maximum likelihood routine using the programming language IML procedure in the statistical analysis package SAS. Subsequent reports will include heteroscedasticity corrected t-statistics. Further, we also intend to investigate a semi-parametric specification for the continuous variables age and income.

 

E.         Results

            Analysis of our subpopulation, for which summary data is reported in Tables 1a – 1e, suggests that in general legal underwriting restrictions based on do not significantly affect the demand for health insurance in the small group market.  Evidence is found that underwriting restrictions pertaining to disability status may effect the insurance purchases in the small group market.

            Table 1a shows that in 1995 about 65% were male, 6% were unemployed during the year prior to the survey, 8% had a physical impairment, 7% were black, 89% white and 3% other than black or white. Table 1b shows that in 1995 the mean age of the population was about 40 and the mean income was $27,700, in 1991 dollars.  Table 1c reports that in 1995 approximately 31% of our sample had group health insurance.  In Table 1d, we see that 72% of the uninsured in our sample are males, whereas only 60% of those with group health insurance are male.  The table also reports that the average income of those with group insurance was roughly double that of the uninsured.  Table 1e separates the population on the basis of whether or not an underwriting restriction of a particular type was in effect and reports the portion of the population receiving insurance in each case.  of a particular type was in effect  demonstrates the effects of legal underwriting restrictions on the purchase of health insurance.  For individuals surveyed with underwriting restrictions based on race in effect, 31.6% purchased group health compared to only 29.7% of those where restrictions were not in effect.  Similarly, there was a difference of 2.5% of group health insurance purchases for those surveyed with underwriting restrictions based on physical impairment were in effects compared to those surveyed where the restrictions were not in effect.

            Table 2 reports estimation of the logit model with control variables and state specific effects, but without the underwriting variables.  The resulting –2 log likelihood statistic is 10,740.299. 

Table 3 reports the estimation of the multinomial logit model with control variables and state specific effects, but without the underwriting variables.  The –2 log likelihood statistic for this model is 18780.4.  Each of the control variables is highly significant in one of the models.  Most are highly significant in both. 

Table 4 reports the – 2 log likelihood statistic for different estimations of the multinomial logit model that build off of the base model mentioned above. Using a chi-square test, we may state whether these models are statistically different than our base model.  The analysis supports the inclusion of both the year and state effects.  For gender and age including the underwriting restrictions and terms representing their interaction with corresponding demographic characteristics does not appreciably improve the model.  The chi-square statistic for testing the importance of the gender based underwriting prohibitions, 4.7, has a corresponding p-value of 0.5828.  Similarly, for age based underwriting prohibitions the statistic is 4.8 and the p-value is 0.5697.  In both cases, the test of the joint hypothesis that the underwriting restriction and the interaction of the underwriting restriction and the corresponding risk factor are statistically different from zero is not rejected.  In contrast, the chi-square statistic for testing the importance of the disability based underwriting statistic, 13.3, has a p-value of 0.0385.  This provides some support for the hypothesis that underwriting restrictions of this type affect the consumption of health insurance in the small group market.

Table 5 reports parameter estimates and p-values for the multinomial model testing the disability based underwriting restrictions.  The positive sign on the term that interacts the presence of a disability statute and the indicator variable for being non-disabled provides support for the hypothesis that the non-disabled reduce their insurance consumption as a result of prohibitions on the use of disability status by insurers.  The results suggest that the non-disabled are more likely to be uninsured as a result of prohibitions on the use of disability status as an underwriting criteria.  The analysis also suggests that the disabled are less likely to have group insurance and more likely to have government insurance as a result of this underwriting prohibition.

            The second analysis, for which we define the small group market as consisting of firms with 25 or fewer employees, uses data from the period 1988-1995.  The logit and multinomial logit estimations are reported in Tables 6 and 7.  Table 8 provides a comparison of  different multinomial logit models for this population.  The findings mirror those of the first analysis. Underwriting restrictions pertaining to gender and age are not found to have a statistically significant impact on insurance consumption.  Evidence consistent with disability based underwriting restrictions affecting consumption of health insurance in the small group market is found.

            Table 9 is similar to Table 5 in that it reports parameter estimates and p-values for the multinomial model testing the disability based underwriting restrictions.  This analysis provides additional support for the earlier finding that the disabled are less likely to have group insurance and more likely to have government insurance as a result of underwriting prohibitions on the use of disability status.  Unlike Table 5 there is no support for the argument that the non-disabled reduce their insurance consumption in the group market in response to these regulations.  Because this analysis defines small groups as those with 25 or fewer employees and the earlier one defined small groups as those with 10 or fewer employees, the discrepancy in results is not surprising.  In general, underwriting of individuals is more common the smaller the group size.  With underwriting decreasing as group size increases, the effect on of underwriting prohibitions would be expected to be less pronounced as group size increases.

 

F.         Conclusion

            Our preliminary results suggest that underwriting regulations pertaining to gender and age do not impact the consumption of health insurance in the small group market. Some evidence is found that prohibitions on the use of disability status may be related to insurance consumption.  These results are preliminary.  We anticipate expanding the data set to include the years 1983-1988.  The analyses reported in this paper employ data from 1988-1995. 

Future analysis will include a study of additional underwriting prohibitions.  Future research will also focus on the market effects of laws of different stringency.  In some states underwriting prohibitions apply only when insurers can not provide actuarial justification for using the prohibited criteria.  In others the prohibitions can not be overridden with statistical arguments.  In the current study we do not consider states with the latter form of prohibition to have a restriction.  Finally, we intend to undertake an analysis of the individual market for health insurance.  This analysis will be similar in nature to the analysis of the small group insurance market reported in this paper.

            The role that underwriting plays in individuals’ decisions to purchase insurance will affect whether various forms of health insurance regulation, including mandated community rating and prohibitions on denying coverage to those with preexisting conditions, will be successful in reducing the rate of uninsurance or will result in greater uninsurance.  The need for research on health insurance underwriting is critical not only to understand how insurance markets operate but also to aid in the making of wise public policy.

 

 

References

 

Aaron, Henry J. and Barry P. Bosworth, “Economic Issues in Reform of Health Care Financing.”  In Brookings Paper on Economic Activity:  ed.  Martin Neil Baily, Peter C.  Reiss, and Clifford Winston, Washington, D.C.:  Brookings Institution (1994).

Akerlof, G., “The Market for "Lemons": Quality Uncertainty and the Market Mechanism,” 84, Quarterly Journal of Economics, 488 (1970).

Arrow, Theory of Risk Bearing (1970).

Browne, M., “Evidence of Adverse Selection in the Individual Health Insurance Market,” 59, Journal of Risk and Insurance, 13 (1992).

Browne, M. J., “Health Insurance for the “Uninsurable”:  State High-Risk Pools,” Journal of Insurance Regulation, (1997).

Browne, M. and H. Doerpinghaus, “Information Asymmetries and Adverse Selection in the Market for Individual Medical Expense Insurance,” 60, Journal  of Risk and Insurance, 300 (1993).

Browne, M. and H. Doerpinghaus, “Asymmetric Information and the Demand for Medigap Insurance,” 31, Inquiry,  445 (1995).

Cho, In-Koo and D. Kreps, “Signaling Games and Stable Equilibria,” 102, Quarterly Journal of Economics, 179 (1987).

Cutler, D.  and J. Gruber, “Does Public Insurance Crowd Out Private Insurance?” Quarterly Journal of Economics, 391 (1996).

Daniels, “Insurability and the HIV Epidemic:  Ethical Issues in Underwriting,” 68, The Milbank Quarterly, 497 (1990).

Dubay, L. and G. Kenney, “Did Medicaid Expansions for Pregnant Women Crowd Out Private Coverage?” 16, Health Affairs, 185 (1997).

Farley, P. and A. Monheit, “Selectivity in the Demand for Health Insurance and Health Care,” 6, Advances in Health Economics and Health Services Research, 231 (1985).

Freeman, H., L. Aiken, R. Blendon, and C. Corey, “Uninsured Working Age Adults:  Characteristics and Consequences,” 24, Health Services Research, 811 (1990).

Greene, W. H., Econometric Analysis, Second Edition, Macmillan:  New York (1993).

Grossman, H., “Adverse Selection, Dissembling, and Competitive Equilibrium,” 10, Bell Journal of Economics, 336 (1979).

Gruber, J.  and J. Poterba, “Tax Incentives and the Decision to Purchase Health Insurance:  Evidence from the Self-Employed,” Quarterly Journal of Economics, 701 (1994).

Hausman, J. A. and W.  E. Taylor, “Panel data and unobservable individual effects,” Econometrica, 1377 (1981).

Hellwig, M., “Some Recent Developments in the Theory of Competition in Markets with Adverse Selection,” 31, European Economic Review, 319 (1987).

Judge, G. G., W.  E. Griffiths, R.  C. Hill, H. Lütkepohl, and T.  C. Lee, The Theory and Practice of Econometrics, Wiley:  New York (1985).

Long, S. H. and M.S. Marquis (1996), “Some pitfalls in making cost estimates of state health insurance coverage expansions.”  Inquiry 33: 85-91 (Spring).

Marquis, M.S. and M.R. Holmer (1996), “Alternative models of choice under uncertainty and demand for health insurance.” The Review of Economics and Statistics, 421-427.

Marquis, M.S. and S.H. Long (1995), “Worker demand for health insurance in the non-group market.” Journal of Health Economics 14, 47-63.

Merrill, J., C. Jackson, and J. Reuter, “Factors that Affect the HMO Enrollment Decision:  A Tale of Two Cities,” 22, Inquiry, 388 (1985).

Mossin, “Aspects of Rational Insurance Purchasing,” 76, Journal Political Economics, 553 (1968).

Pauly, Mark V., “Is Cream-Skimming a Problem for the Competitive Medical Market?”  3, Journal of Health Economics, 88 (1984).

Phelps, C., “Demand for Reimbursement Insurance.”  In The Role of Health Insurance in the Health Services Sector, ed. R. Rosett, New York: National Bureau of Economic Research (1976).

Riley, J., “Informational Equilibrium,” 47, Econometrica, 331 (1979).

Riley, J., “Competition with Hidden Knowledge,” 93, Journal of Political Economics, 958 (1985).

Rothschild, M. and J. Stiglitz, “Equilibrium in Competitive Insurance Markets:  An Essay on the Economics of Imperfect Information,” 90, Quarterly Journal of Economics, 629 (1976).

Smith, “Optimal Insurance Coverage,” 76, Journal Political Economics, 68 (1968).

Taylor, A. and G. Wilensky, “The Effect of Tax Policies on Expenditures for Private Health Insurance.”  In Market Reforms in Health Care: Current Issues, New Directions, Strategic Decisions.  J. Meyer, ed. Washington, D.C.: American Enterprise Institute.

Thomas, K., “Are Subsidies Enough to Encourage the Uninsured to Purchase Health Insurance?  An Analysis of Underlying Behavior,” 31, Inquiry, 415 (1995).

Van de Ven, W. and B. Van Praag, “The Demand for Deductibles in Private Health Insurance,” 17, Journal of Econometrics, 229 (1981).

Young, V. and Browne, M., “Explaining insurance policy provisions via adverse selection,” Geneva Papers on Risk and Insurance Theory, 22: 121-134 (1997).

Wilson, C., “A Model of Insurance Markets with Incomplete Information,” 16, Journal of Economic Theory, 167 (1977).

“Medicaid and Private Insurance: Evidence and Implications,” 16, Health Affairs, 194 (1997).

 

 

 

APPENDIX A

 Unit of Analysis

 

 

            This article uses the family as the level of decision-making for purchasing insurance. The work of Marquis and Long (1995) and Long and Marquis (1996) distinguishes between the concept of a census family and an insurance family.  As described by Long and Marquis (1996), a census family includes all people related to the head of the household by either blood or marriage. This is the customary definition employed by government statistics programs such as the Current Population Survey (CPS). In contrast, an insurance family is defined to include the household’s head, spouse and dependent children up to age 18 (23 if in school). This definition corresponds to one used by private insurers and government for coverage purposes. In general, a census family is at least as large as an insurance family. Long and Marquis (1996) give examples of individuals that may be included in a census family yet not an insurance family. These examples are adult children (and their families) living in homes of their parents, adult siblings co-habitating and parents living at homes of their adult children.

            As in Marquis and Long (1995), it is possible to analyze observations organized by insurance family from the CPS data. However, Marquis and Long (1995) analyze only the non-government and non-group markets. That is, all individuals whose census family members answer positively to either the group insurance or government market questions were excluded from their analysis. When considering group and government markets, the difficulty in analyzing observations by insurance family lies in the nature of the CPS questions. The CPS questions ask only whether any member of the (census) family has group, or government, health insurance. For example, if one member of a (census) family receives military health benefits, this means that all members of the family are coded as having received government insurance. Thus, by including several members of a family, responses for the group and government variables become inextricably intertwined.

            To handle these difficulties, we examine two sub-populations. The smaller sub-population consists of single person households as the unit of analysis. For people within this sub-population, questions regarding group and government health insurance pertain to the individual. There is no ambiguity regarding who is making the purchasing decision. The drawback of this smaller sub-population is that single person households comprise a small portion of the market place.

The larger sub-population which we be examined in a future study consists of household (census family) heads as the unit of analysis. The advantage of this larger sub-population is that results are more broadly applicable to the entire population when compared to the smaller sub-population. By restricting consideration to only one individual from each family, we limit the problem of a household member driving the several responses within a household. However, there remains the drawback of a distant family member potentially influencing the response of a household head. We eliminate this difficulty by considering the smaller sub-population consisting of single person households.

 

 

 

 

APPENDIX B

 The Response Variable

 

            The goal of this article is to assess the effect of underwriting laws on an insurance family’s choice of health insurance. To achieve this goal, we must work with the timing of the laws and the constraints imposed by the Current Population Survey (CPS) data. Because health insurance regulations are at the state level, there are many different underwriting laws passed over a long period of time. Our interest is in comparing health insurance purchases before and after the passage of the laws, controlling for many state-specific effects. Thus, on the one hand we wish to examine consumer choice over an extended period of time. On the other hand, information about health insurance choice is limited. The CPS only began asking data about health insurance purchases continuously since 1982 (it also appeared on the 1980 questionnaire). Further, questions change over time, making comparisons over years difficult. We began by considering CPS over years 1982 to 1995, the most recent available data.

            Information on three types of health insurance coverage is provided by the CPS; these are group health, private health insurance other than group health, which we call “individual,” and public health insurance. Public health insurance includes coverage from a military program, such as CHAMPUS, Medicare and Medicaid. Our analysis excludes the elderly (age greater than 64) so that Medicare is a relatively small percentage for our population. Our response variable is defined as:

Insurance families may have any combination of group, individual and public insurance. Thus, in principle, with three types of health insurance coverage, there are eight (2³) possible outcomes for the response variable y. Unfortunately, with the CPS data base, it was not possible to identify each of the eight types consistently over 1982-1995. For example, families that had both group and individual insurance are coded differently before and after 1988. By combining categories, we are able to provide a consistent classification of insurance families over the years.

            Our classification allows us to generate interesting testable hypotheses. We are primarily interested in the effect that underwriting laws have in causing families with individual health insurance to drop this coverage. For example, we would like to test whether a family coded as y=2 that has individual but not group insurance remains in the following year at y=2 or moves to either government insurance, y =3, or no insurance, y =4. We assess this effect through the logit parameters. It is possible that a family has group and individual health insurance is forced to drop their individual coverage due to the introduction of underwriting laws. Unfortunately, due to the coding of CPS data, we not able to test this hypothesis. We remark that our data does not allow us to follow individual families directly; a more complete panel data set would provide more powerful inferences than the CPS data.

APPENDIX C

Variable Definitions

Variable	Definition
Individual Characteristics
Age	is the age of the individual.
Faminc	is the family’s income, in thousands of dollars.  Dollars are in constant 1991 terms.
Sex	is an indicator variable.  It is coded 1 if the individual is male, and zero otherwise.
Unemploy	is an indicator variable. It is coded 1 if the individual was unemployed in the prior year and zero otherwise.
Parttime	is an indicator variable.  It is coded 1 if the individual worked 20 or fewer hours on average per week during the preceding year, and 0 otherwise.
Mar	is an indicator variable. It is coded 1 if the individual is married, and 0 otherwise.
Edu	is the number of years of education the individual has acquired.
Dislim	is an indicator variable. It is coded 1 if the individual is physically impaired and zero otherwise.
Race_o	is an indicator variable. It is coded 1 if the individual’s race is other than black or white, and zero otherwise.
Race_b	is an indicator variable. It is coded 1 if the individual’s race is black, and zero otherwise.
Race_w	is an indicator variable. It is coded 1 if the individual’s race is white, and zero otherwise.
Legal Characteristics
Lawdisa	is an indicator variable. It is coded 1 if the individual lives in a state during a year where an underwriting law restricting discrimination based on physical impaired is in effect and zero otherwise.
Lawgend	is an indicator variable. It is coded 1 if the individual lives in a state during a year where an underwriting law restricting discrimination based on gender is in effect and zero otherwise.
Lawage	is an indicator variable. It is coded 1 if the individual lives in a state during a year where an underwriting law restricting discrimination based on age is in effect and zero otherwise.
Interaction Variables
L_dis	is an indicator variable. It equals Lawdisa times Dislim.
L_ndis	is an indicator variable. It equals Lawdisa times ( one minus Dislim).
Dependent Variables
Group	is an indicator variable. It is coded 1 if the individual purchases group insurance and zero otherwise.
Individual	is an indicator variable. It is coded 1 if the individual purchases private, nongroup insurance and zero otherwise.
Government	is an indicator variable. It is coded 1 if the individual purchases government insurance and zero otherwise.
Uninsure	is an indicator variable. It is coded 1 if the individual does not purchase health  insurance and zero otherwise.

 

 

 

 

 

APPENDIX D

Underwriting Regulations By State

Year of Enactment

 

 

STATE	RACE	DISA	GEND	AGE
AZ
AR
CA
CO			94-
DE	88-
IL
KS
KY
ME	93-	93-	93-	93-
MD	94-	94-	94-	94-
MA	92-	92-	92-	92-
MI
MN	95-	95-	93-	95-
MT
NE
NV
NJ	92-	92-	92-	92-
NM	85-
NY		93-	93-	93-
ND	95-	95-	95-	95-
OH
OR	92-	92-	92-	92-
SC	95-	95-	95-	95-
SD
TN
TX	95-
UT	80-
VT	92-	92-	92-	92-
WA	95-	95-	95-	95-
WI	80-

 

 

 

 

 

 

Table 1a.  Averages of Indicator Control Variables
Data: 91-95, Single Person Households (RELHD=2), Employer Group Size<10
Year	Number	Sex	Unemploy	Part-time	Marital Status	Dislim	Race_o	Race_b	Race_w
91	1,806	0.6645	0.0537	0.0615	0.0266	0.0759	0.0377	0.0764	0.8859
92	1,746	0.6741	0.0527	0.0556	0.0326	0.0825	0.0372	0.0647	0.8981
93	1,795	0.6396	0.0563	0.0669	0.0273	0.0925	0.0384	0.0719	0.8897
94	1,819	0.6520	0.0605	0.0699	0.0258	0.0962	0.0561	0.0753	0.8686
95	1,614	0.6512	0.0613	0.0558	0.0335	0.0849	0.0353	0.0713	0.8934

 

 

 

Table 1b.  Summary Statistics for Other Control Variables
Data: 91-95, Single Person Households (RELHD=2), Employer Group Size<10
Variable	Year	Number	Mean	Median	Minimum	Maximum	Standard Deviation
Age	91	1,806	39.30	38	15	64	12.08
	92	1,746	39.60	38	16	64	11.99
	93	1,795	40.55	40	18	64	12.12
	94	1,819	40.37	39	17	64	12.18
	95	1,614	40.39	40	17	64	12.13
Faminc	91	1,806	25.51	18.52	-11.52	172.90	24.86
	92	1,746	24.20	17.89	-14.43	229.24	24.35
	93	1,795	26.38	19.54	-10.86	295.42	27.14
	94	1,819	24.44	18.06	-17.78	186.86	23.98
	95	1,614	27.70	19.18	-10.25	394.99	35.15
Education	91	1,806	13.33	14	0	18	2.85
	92	1,746	13.25	14	0	18	2.75
	93	1,795	13.48	14	0	18	2.68
	94	1,819	13.49	14	0	18	2.66
	95	1,614	13.41	14	0	18	2.73

 

 

Table 1c.  Averages of Indicator Dependent Variables
Data: 91-95, Single Person Households (RELHD=2), Employer Group Size<10,  Multinomial Logit
Year	Number	Uninsure	Individual	Government	Group
91	1,806	0.3898	0.2807	0.0377	0.2918
92	1,746	0.4210	0.2761	0.0401	0.2629
93	1,795	0.4045	0.2240	0.0507	0.3209
94	1,819	0.4046	0.2336	0.0451	0.3167
95	1,614	0.4126	0.2280	0.0471	0.3123

 

 

 

 

 

 

 

Table 1d.  Means by Dependent Variable for Non-law Variables
Data: 91-95, Single Person Households (RELHD=2), Employer Group Size<10, Multinomial Logit
	Number	Sex	Unemploy	Part-time	Marital Status	Dislim	Race_o	Race_b	Race_w	Age	Fam inc	Edu
Uninsure	3,567	0.721	0.051	0.065	0.035	0.083	0.050	0.095	0.855	38.5	18.1	12.7
Individual	2,184	0.621	0.060	0.071	0.033	0.071	0.026	0.050	0.924	41.6	28.1	13.9
Government	387	0.618	0.238	0.217	0.021	0.463	0.088	0.111	0.801	42.9	14.3	12.3
Group	2,642	0.604	0.035	0.028	0.019	0.048	0.034	0.053	0.913	40.4	35.4	14.1
Total	8,780	0.656	0.057	0.062	0.029	0.086	0.041	0.072	0.887	40.0	25.6	13.4

 

 

Table 1e.  Frequency Table of Dependent Variable by Law Variables
Data: 91-95, Single Person Households (RELHD=2), Employer Group Size<10, Multinomial Logit
	Number	Uninsure	Individual	Government	Group	Total
Number		3,567	2,184	387	2,642	8,780
Percentage		40.63	24.87	4.41	30.09	100
Lawdisa=0	7,612	0.4091	0.2499	0.0448	0.2962	1.0000
=1	1,168	0.3878	0.2414	0.0394	0.3313	1.0000
Lawgend=0	7490	0.4110	0.2493	0.0443	0.2955	1.0000
=1	1290	0.3791	0.2457	0.0426	0.3326	1.0000
Lawage=0	7612	0.4091	0.2499	0.0448	0.2962	1.0000
=1	1168	0.3878	0.2414	0.0394	0.3313	1.0000

 

 

 

 

 

 

 

 

 

Table 2.  Logistic Regressions Using Group as the Dependent Variable with only Control Variables.
Data: 91-95, Single Person Households (RELHD=2), Employer Group Size<10. Year and state effects are included, although coefficients are not reported.
	Parameter Estimates	p-values
Inter	-1.7644	0.0001
Age	0.00166	0.4373
Faminc	0.1609	0.0001
Sex	-0.4831	0.0001
Unemploy	-0.2694	0.0340
Parttime	-0.8994	0.0001
Mar	-0.4511	0.0082
Edu	0.0802	0.0001
Dislim	-0.4932	0.0001
Race_O	-0.2812	0.0458
Race_B	-0.3717	0.0006
-2 Log Likelihood	10,740.299

 

 

Table 3.  Multinomial Logistic Regression with only Control Variables
Data: 91-95, Single Person Households (RELHD=2), Employer Group Size<10 Year and state effects are included, although coefficients are not reported.
	None versus Group		Private versus Group		Government versus Group
	Parameter Estimates	p-values	Parameter Estimates	p-values	Parameter Estimates	p- values
Inter	2.219	0.0000	-0.791	0.0155	-0.852	0.2042
Age	-0.010	0.0000	0.009	0.0003	-0.000	0.9695
Faminc	-0.299	0.0000	-0.070	0.0000	-0.332	0.0000
Sex	0.706	0.0000	0.204	0.0013	0.528	0.0000
Unemploy	-0.007	0.9585	0.340	0.0204	0.954	0.0000
Parttime	0.761	0.0000	0.867	0.0000	1.321	0.0000
Mar	0.385	0.0399	0.514	0.0081	-0.077	0.8521
Edu	-0.140	0.0000	0.019	0.1520	-0.123	0.0000
Dislim	0.289	0.0147	0.180	0.1640	2.268	0.0000
Race_O	0.391	0.0113	-0.146	0.4256	1.138	0.0000
Race_B	0.551	0.0000	0.018	0.8980	0.549	0.0125
-2 Log Likelihood	18,780.4

 

 

Table 4.  Comparison of Multinomial Logistic Regressions
Data: 91-95, Single Person Households (RELHD=2), Employer Group Size<10. The base model is described in Table 3.
Model	-2 Log Likelihood	Difference in -2 Log Likelihood from the base model	Difference in the number of parameters from the base model	p-value
Base	18,780.4
Omit Year and State Effects	19,330.8	550.4	162	0.0000
Omit Year Effects	18,826.4	46	12	0.0000
Omit State Effects	19,285.9	505.5	150	0.0000
Omit State Effects but include PCEMPLOY and PCINC			142	NA
Include PCEMPLOY and PCINC	18,776.9	3.5	6	0.7440
Include LAWDISA and LAWDISA*dislim	18,767.1	13.3	6	0.0385
Include LAWGEND and LAWGEND*sex	18,775.7	4.7	6	0.5828
Include LAWAGE and LAWAGE*age	18,775.6	4.8	6	0.5697
Include LAWDISA and LAWDISA*age	18,775.6	4.8	6	0.5697
Include LAWDISA and LAWDISA*faminc	18,771.6	8.8	6	0.1851

 

 

 

Table 5

Multinomial Logit Estimation

Group Size 10 and Under

 

 

Sample size
Total	Group	Private	Gov’t	None
8780	2642	2184	387	3567

 

 

-2LogLik= 18767.132

 

 

              

 

Table 6.  Logistic Regressions Using Group as the Dependent Variable with only Control Variables.
Data: 88-95, Single Person Households (RELHD=2), Employer Group Size<25. Year and state effects are included, although coefficients are not reported.
	Parameter Estimates	p-values
Inter	-1.794	0.0001
Age	0.000	0.9789
Faminc	0.190	0.0001
Sex	-0.427	0.0001
Unemploy	-0.598	0.0001
Parttime	-1.069	0.0001
Mar	-0.427	0.0001
Edu	0.085	0.0001
Dislim	-0.472	0.0001
Race_O	-0.252	0.0045
Race_B	-0.326	0.0001
-2 Log Likelihood	25090.169

 

 

 

 

 

 

 

 

 

 

 

Table 7.  Multinomial Logistic Regression with only Control Variables
Data: 88-95, Single Person Households (RELHD=2),  Employer Group Size<25 Year and state effects are included, although coefficients are not reported.
	None versus Group		Private versus Group		Government versus Group
	Parameter Estimates	p-values	Parameter Estimates	p-values	Parameter Estimates	p-values
Inter	2.550	0.0000	-1.184	0.0000	-0.977	0.0298
Age	-0.009	0.0000	0.011	0.0000	0.011	0.0012
Faminc	-0.362	0.0000	-0.077	0.0000	-0.339	0.0000
Sex	0.614	0.0000	0.183	0.0000	0.472	0.0000
Unemploy	0.284	0.0038	0.781	0.0000	1.012	0.0000
Parttime	0.735	0.0000	1.187	0.0000	1.570	0.0000
Mar	0.354	0.0033	0.464	0.0003	0.004	0.9882
Edu	-0.149	0.0000	0.026	0.0020	-0.128	0.0000
Dislim	0.288	0.0003	0.164	0.0617	2.134	0.0000
Race_O	0.300	0.0023	-0.041	0.7324	1.108	0.0000
Race_B	0.432	0.0000	-0.010	0.9120	0.802	0.0000
-2 Log Likelihood	44036.86

 

 

 

 

 

 

Table 8.  Comparison of Multinomial Logistic Regressions
Data: 88-95, Single Person Households (RELHD=2), Employer Group Size<25. The base model is described in Table 3.
Model	-2 Log Likelihood	Difference in -2 Log Likelihood from the base model	Difference in the number of parameters from the base model	p-value
Base	44036.86
Omit Year and State Effects	44949.51	912.65	162	0.0000
Omit Year Effects	44099.67	62.81	12	0.0000
Omit State Effects	44887.60	850.74	150	0.0000
Omit State Effects but include PCEMPLOY and PCINC	44793.01		142	NA
Include PCEMPLOY and PCINC	44028.20	8.66	6	0.1934
Include LAWDISA and LAWDISA*dislim	44024.51	12.35	6	0.0545
Include LAWGEND and LAWGEND*sex	44034.52	2.34	6	0.8859
Include LAWAGE and LAWAGE*age	44034.24	2.62	6	0.8548
Include LAWDISA and LAWDISA*age	44034.24	2.62	6	0.8548
Include LAWDISA and LAWDISA*faminc			6
Include LAWMS1 and LAWMS1*mar			6

 

 

 

 

 

Table 9

Multinomial Logit Estimation

Group Size 25 and Under

 

 

 

Sample size
Total	Group	Private	Gov’t	None
20849	7838	4495	812	7704

 

 

-2LogLik= 44024.514