I. INTRODUCTION

Economic models often hypothesize that the markets for life insurance and health insurance are markets that contain asymmetric information problems leading to adverse selection (Nicholson and Snyder 2009, 550; Perloff 2004, 659). For example, consumers of life insurance may know facts about their probability of death, which are difficult, if not impossible, for a life insurer to detect. These factors may include whether individuals buckle their seatbelts, have genetic predispositions to breast cancer, or go rock climbing on weekends (Rothschild and Stiglitz 1976).

The adverse selection model predicts that those who die within a given period of time are more likely to hold life insurance after controlling for age, earnings, wealth, and a variety of other socio-economic and demographic factors. Further, the adverse selection model predicts that those with higher self-perceived risks of mortality are more likely to purchase life insurance and also have larger life insurance policies.

A competing model to the adverse selection model is advantageous selection, originally proposed by Hemenway (1990). This model posits that risk-averse behavior, associated with a concern for one's children, spouses, or other potential life insurance policy beneficiaries, leads both to increased demand for life insurance and lower mortality risk. This model predicts that those with higher perceived risk of mortality are less likely to purchase life insurance and will have smaller life insurance policies.

In this article, we look for the existence of adverse selection in the market for life insurance. Life insurance policies include term, whole, universal, limited, and other policies. (1) We focus our analysis on the simplest and most common form of life insurance, which is term life insurance. This type of life insurance policy requires regular and constant premium payments in exchange for a payout in the event of the death of the insured. The payout is the policy's face value. The cost of a policy's premium is a function of the duration of the contract, the face value of the policy, and the actuarial likelihood of dying within that duration.

Previous research analyzing the empirical validity of either model has been hamstrung by a lack of micro-data that provide precise information about the purchase of life insurance, the risk of death, and other motives for purchasing life insurance. Moreover, these previous studies did not have the data to clearly identify whether a person died, and instead used survey attrition as a proxy for mortality (Cawley and Philipson 1999, 839).

Using both survey data and administrative records from the Social Security Administration (SSA), we tested the competing predictions of the adverse selection and advantageous selection by examining how an individual's health status and probability of death predict whether the person purchases term life insurance, and the size of their policy if they buy one. Our data have at least two advantages over previous studies. First, we eliminate the measurement error of previous studies because we know the date of death from administrative records. This is an advantage because using attrition from longitudinal surveys as a measure of death is potentially problematic. That is because attrition can be caused by multiple factors: the respondent moved to an unknown location, (2) the respondent refused to continue participating in the survey, or the respondent died without the surveyor's knowledge. Attrition rates of longitudinal surveys can reach 20% per year; thus, measuring death by attrition is associated with measurement error (Alderman 2001; Deeg et al. 2002).

Second, our administrative data provide us with a very long survey panel where individuals are observed until the present day or their death. For example, while the survey may indicate that a person has life insurance, the administrative data may reveal that the person died several years later, when they were no longer part of the survey.

Our initial findings show that both self-perceived risk and actual risk of death predict a lower likelihood of owning life insurance. Furthermore, we find that conditional to owning a life insurance policy, these policies have a smaller value when individuals perceive their risk of death as higher than average. These findings are consistent with the advantageous selection hypothesis, and with the hypothesis that insurers, rather than consumers, are better at estimating a consumer's risk of death.

To further study the relative importance of information asymmetry, we analyze the market for employer-provided life insurance. In this market, insurers have less ability to discriminate among individuals. After including controls for socio-economic status, our estimation results for ownership are negative for Fair and Poor health, but positive for Very Good and Good. The relationship between ownership and mortality risk is insignificant after adding all controls. The relationship between the amount of life insurance coverage owned (face value) and health is mostly statistically insignificant. While the relationship between coverage and mortality risk is similar to our initial findings, they are roughly half the magnitude. These findings suggest that while advantageous selection is still present (i.e., there is not a large positive relationship between mortality risk), it is price discrimination that is the primary force overcoming adverse selection in this market.

II. RELATED LITERATURE

Markets contain asymmetric information if either sellers or buyers have more information than their counterparties, giving rise to adverse selection. The concept of adverse selection was explained by the Nobel Prize committee in the following way:

[a]t any given price, a seller of high-quality units is less willing to sell than is the seller of low-quality units. Rational buyers anticipate this, suspecting that the item they face is of low quality. This rational suspicion depresses prices, which further discourages sellers of high-quality units, who continue to leave the market until only low-quality items remain for sale. Such a downward quality bias is called adverse selection. Adverse selection may thus hinder mutually beneficial transactions. (Weibull 2001)

Economic theory predicts that individuals with private information about the likelihood of an adverse event in their lives are more likely to buy health insurance or a life insurance policy. Asymmetric information where buyers have more information than sellers of life insurance results in more claims than predicted by the insurer. Adverse selection models predict that if asymmetric information is sufficiently pronounced, insurers cannot price policies to be actuarially fair, let alone profitable.

The idea of adverse selection was first proposed by Akerlof (1970) and was later formalized by Rothschild and Stiglitz (1976). In a survey of empirical work on asymmetric information in several types of insurance markets, Cohen (2010) finds mixed support for the adverse selection hypothesis. For instance, the market for long-term care insurance shows no evidence of adverse selection (see Finkelstein and McGarry 2006), whereas the market for annuities does (see Finkelstein and Poterba 2004).

Even though the Nobel Prize committee has stated that a prime example of adverse selection can be found in insurance (Weibull 2001), the empirical literature on this subject tends to conclude that evidence for adverse selection exists in some insurance markets but not in others (Einav and Finkelstein 2001). This inconsistency suggests that there are countervailing forces in markets for insurance that mitigate or remove the selection problems.

One possible explanation for the mixed support of the adverse selection hypothesis comes from the theory of advantageous selection (Hemenway 1990). Advantageous selection posits that an individual's risk aversion leads both to the uptake of insurance and to behavior that makes unlikely the occurrence of the adverse event for which they have purchased insurance. As a result, an adverse event has a lower occurrence probability for individuals with insurance, relative to others who have not purchased insurance. Thus, the very act of buying insurance suggests that the insured are less likely to incur adverse outcomes that result in making claims on the policy. More recently, empirical work by Cutler, Finkelstein, and McGarry found that individuals who engage in riskier behavior are less likely to have each type of insurance (Cutler, Finkelstein, and McGarry 2008).

Adverse selection and advantageous selection are two forces that push in opposite directions. Adverse selection causes a given insurance price pool to appeal to the qualifying sicker portion of the population pool, suggesting a negative correlation between health and insurance ownership. Advantageous selection causes those who are risk averse to seek insurance at a higher rate than others within a risk pool. Given that risk-averse individuals are less likely to engage in unhealthy behaviors such as smoking, drinking, and being obese (Anderson and Mellor 2008), advantage selection suggests a positive correlation between health and insurance ownership. In observed data, any correlation between adverse events and insurance coverage is likely generated by both adverse selection and advantageous selection. Both effects predict opposite signs of the correlation between the uptake of insurance and health status.

Another factor that reduces the impact of adverse selection is the hypothesis that the providers of insurance are better able to estimate mortality risk than consumers. When this is the case, insurers are able to use prices to discourage high-risk individuals from purchasing insurance and avoid adverse selection. For example, an analysis for automobile insurance in France suggested that price discrimination could result in a negative relationship between coverage and risk under certain assumptions (Chiappori and Salanie 2000, 74).

The presence of adverse selection in the market for life insurance has been examined by Cawley and Philipson (1999), who analyze data from the Health and Retirement Study (HRS). They measure the probability of death in two ways: as perceived risk and as actual risk. The perceived risk measure comes from survey responses to the question of the likelihood of reaching a particular birthday. The authors' measure of actual risk is derived by using observed deaths over two waves of the HRS in order to calibrate a logistic model that is then applied to all members of the sample. Some of the deaths are reported by surviving family members, but others are implied by sample attrition. Because sample attrition implies that the death status is unknown, the authors report two models: one where all attrited respondents are assumed dead and one where they are all assumed alive (Cawley and Philipson 1999, 837, 839-40). That study finds no evidence of adverse selection using either self-perceived or actual risk. They hypothesize that this is because "sellers may know their costs of production better than consumers" (Cawley and Philipson 1999, 827). The authors acknowledge their measure of risk contains measurement error. Furthermore, the HRS only surveys people over the age of 50, so their findings may not hold for younger individuals.

He (2009) uses the same HRS data as Cawley and Philipson but makes an important sample restriction that reverses their findings. He asserts that those who have a high risk of death are more likely to die and are thus under-represented in cross-sectional samples (He 2009, 1091). In an effort to correct for this, she restricts the sample under consideration to those who do not have life insurance during the first wave the question was asked. He then compares the mortality rates of those who bought life insurance in the second wave the question was asked. She finds that those who reported buying life insurance in the second wave were more likely to die than those who did not.

This is an important insight and likely an important issue with the HRS data used in He's analysis of 51-61 year olds. However, this method of correction potentially introduces as many problems as it solves. Unfortunately, survey responses do not correspond perfectly with the truth. Every respondent answers questions with some measurement error. Misunderstanding the question, proxies answering for others, and uncertainty all factor into survey responses (see Abowd and Stinson 2013). Some of those who changed their answers from No to Yes truly did acquire life insurance during the time between waves, whereas some are merely misreporting in one wave or the other. Respondents who misreport their life insurance status will be over-represented in He's sample of those who buy life insurance. If misreporting error is correlated with mortality, these results are harder to interpret. Unfortunately, life insurance ownership was only asked in two waves of the HRS, so it is difficult to get a measure of how many respondents provided noisy answers to this question.

Hendel and Lizzeri (2003) use a different method to assess the presence of adverse selection in life insurance markets. They extract premium data from CompuLife, a life insurance quotation software. Hendel and Lizzeri then model the nature of contracting in the market for term life insurance and test the relation between front-loading and premiums. They find a negative correlation between front-loading and total premiums, and conclude that this result is not consistent with asymmetric information.

Working with aggregate mortality tables from the United States, United Kingdom, and Japan, McCarthy and Mitchell (2010) compare mortality rates among the total population relative to the mortality rates of those with life insurance. They suggest that their findings indicate insurance companies are better than individuals in assessing mortality risks of these individuals (McCarthy and Mitchell 2010). However, they also suggest that their results are confounded by their inability to control for income and wealth, which are correlated with both longevity and insurance ownership.

III. DATA

Our data come from version 6 of the Survey of Income and Program Participation (SIPP) Gold-Standard File (GSF) from the U.S. Census Bureau. This data set matches respondents from the SIPP with administrative records from the SSA and the Internal Revenue Service (IRS). SIPP is a series of publicly available national panels, with sample sizes ranging from 14,000 to 36,700 interviewed households. Each panel of households is unique and distinct from other panels. To analyze the most comparable set of questions, we use the four most recent panels in the GSF: 1996, 2001, 2004, and 2008. (3) The duration of each panel ranges from 2 1/2 years to 4 years (U.S. Census Bureau 2011). Each survey wave is administered every 4 months. While many questions are asked in every wave, the majority of the variables in our study come from topical module questions asked once a year. The SIPP Gold Standard pairs these survey responses with a variety of administrative data, including earnings histories and mortality. (4)

In our panels, the life insurance questions were directed to every survey participant over the age of 15. We restrict our sample to those who were of prime working age (21-62), because term life insurance is largely a hedge against lost wages. Because we focus on term life insurance, we also exclude consumers with only whole life policies.

The administrative data provide information whether and when the individual died. The U.S. Census, the agency developing the GSF, obtains this mortality measure by using a hierarchy of administrative sources: (i) SSA's [Master Beneficiary Record] file, (ii) the Census Personal Characteristics File with death information coming from the SSA Numident and Master Death Files, and (iii) SSA's [Supplemental Security Income Record] file (U.S. Census Bureau 2010, 16).

As we know whether a person actually died, we can overcome one of the biggest limitations of using survey data to analyze issues relating to death rates: we do not have to infer death from attrition. Another advantage of these data is that they allow us to use an extensive set of socio-economic control variables, such as education, wealth, income, marital status, age, race, and gender.

Table 1 reports summary statistics for the variables in our models. We report summary statistics for the pooled data set. (5) In our data set, 49% of the individuals own life insurance and 30% have employer-provided life insurance. The mean face value of life insurance is $133,829, and the mean face value of employer-provided life insurance is $88,567 in 2010 real dollar terms. While the questions on health status, life insurance ownership, and type are unchanged during this period, there was a change in the question about the size of the life insurance policy. However, this latter change has little bearing on the data because even with this altered question, individuals report similar sizes in life insurance policies when asked about the face value of their life insurance policy. (6)

IV. EMPIRICAL MODEL

We employ two measures of an individual's probability of death. One is the consumer's self-reported health status, (7) which serves as a proxy for their self-perceived risk of death. Our other measure is the consumer's recorded death from administrative records.

We test for adverse selection by estimating

(1) [LifeIns.sub.c] = [[beta].sub.1] [VeryGoodHealth.sub.c] + [[beta].sub.2][GoodHealth.sub.c] + [[beta].sub.3][FairHealth.sub.c] + [[beta].sub.4][PoorHealth.sub.c] + [beta]X + [[mu].sub.c]

where [LifeIns.sub.c] is an indicator variable equaling one if respondent c held term life insurance in the first wave she was asked. The variables [VeryGoodHealth.sub.c], [GoodHealth.sub.c], [FairHealth.sub.c], and [PoorHealth.sub.c] are indicator variables measuring whether the individual reported their health status as either Very Good, Good, Fair, or Poor. In these regressions, coefficients have to be interpreted relative to the omitted category, which is ExcellentHealth.

Given that we also know whether an individual died between the year the individual responded to the SIPP survey and 2010, we can model actual risk of death by using actual deaths with the following specification:

(2) [LifeIns.sub.c] = [[gamma].sub.1][Died.sub.c] + [beta]X + [[eta].sub.c]

This model is identical to Equation (1) but replaces the self-reported health status categories with a binary variable, indicating whether the respondent died between SIPP participation (8) and May 30, 2010. (9)

In Equations (1) and (2), the vector X captures individual-specific characteristics. These include factors that may affect a respondent's demand for life insurance. Following Cawley and Philipson (1999), we control for the bequest motive, income, wealth, and demographic information. (10) Bequest motives, leading to potentially more demand for life insurance, are measured by the number of children and by whether the individual is married in the year they indicated owning life insurance. Beneficiaries of those who purchase life insurance are likely to experience a larger income shock with the death of a high-income buyer of life insurance relative to the beneficiaries of low-income buyers. This reasoning suggests that individuals with higher incomes are more likely to own term life insurance. This is one reason why we include income in our regression specifications. Moreover, if term life insurance is a normal good, those with higher incomes are more likely to purchase this insurance. Our measure of income is the log of self-reported income in SIPP in the year the life insurance questions were asked. (11) Our regressions include wealth, measured as self-reported net worth from the GSF. (12) We include wealth because it is a means by which survivors can self-insure against the death of their provider, and predict that those with more wealth are less likely to own term life insurance.

Consistent with Cawley and Philipson (1999), our regressions also include education, race, gender, age, and age squared. (13) In some specifications, our covariates include the spouse's income and life insurance ownership. This is because having a spouse with high earnings might reduce the demand to hedge against the loss of your own earnings with the purchase of life insurance. For the regression specifications that include the spouse's income and life insurance ownership information, we restrict our sample to married individuals.

V. RESULTS

A. Baseline Specifications

We first test whether individuals have private information about the probability of their death. If not, then asymmetric information on health cannot play an important role in the decision whether to purchase life insurance, because individuals do not have sufficiently accurate information to self-select into insurance plans that are based on future risk.

Table 2 shows the results of regressing an indicator on health status for whether a person died based on health status. The indicator equals one if the person died between the date of the survey and 2010, and zero otherwise. We estimate logistic regressions and report marginal effects.

In these regressions, the incidence of the death variable comes from Social Security records, whereas the variables on health status come from the SIPP. Health status is measured via indicators reflecting whether the individual responded that they are in Excellent health, Very Good health, Good health, Fair health, or Poor health. In all specifications, the reference group is those who report themselves to be in Excellent health. (14)

Table 2 shows that self-perceived health status is a statistically significant predictor of death, even after controlling for demographic variables, that is, the quadratic of age, gender and race, and socio-economic status variables, that is, log of income, and total net worth. (15) For example, Table 2, column 3, shows that those who report they are in Very Good health are 22% more likely to die in the observed period than those in Excellent health. The size of the coefficients increases monotonically for each health level. Those who report themselves to be in Poor health are 225% more likely to die than those in Excellent health.

These findings show that individuals' self-reported health status is a predictor of their likelihood of death, suggesting that individuals have some information about their true risk of dying. With respect to the adverse selection hypothesis, this finding implies that individuals may have private information about their health, and this information gives those who are in poorer health an incentive to purchase life insurance.

Table 3 shows the estimation results from a regression of an indicator for owning life insurance on a self-reported health status. As in the previous table, we estimate logistic regressions and report marginal effects. In all specifications, we cluster the standard errors at the household level. Excellent health status is the omitted category in each of the specifications. Table 3, columns 1-4, includes different sets of control variables. The coefficients on the health status indicators are either statistically indistinguishable from zero or negative and statistically significant, indicating that relative to individuals who report themselves in Excellent health, those with a worse self-reported health status are less likely to own life insurance.

The negative coefficients on the health status indicators in Table 3 are not consistent with adverse selection. Nor can these results be explained by omitted variables such as total net worth, income, or age, because we include these variables among others as control variables. Instead, these results are consistent with the advantageous selection hypothesis and the hypothesis that insurance companies can distinguish between potential buyers of life insurance through price discrimination.

Table 4 shows the marginal effects from logistic regressions where the dependent variable is whether an individual owns term life insurance and our key explanatory variable is an indicator for whether the individual died. Thus, we test whether the likelihood of death can explain the purchase of life insurance, as is predicted by the adverse selection model. Table 4 shows that in all specifications, those who died are much less likely to own life insurance. The magnitude of this point estimate indicates that those who die between being interviewed in SIPP and 2010 are 39 percentage points less likely to report owning life insurance (Table 4, column 3). (16) This suggests that either insurers are effective in distinguishing between consumers with respect to the actual risk consumers face, or that the advantageous selection effect dominates the adverse selection effect.

Next, we estimate similar specifications as those in Tables 3 and 4, but now the dependent variable is the log of face value of the policies held by those owning life insurance, using only the sample of individuals who own life insurance. Here, we assume that the errors of the specifications in the previous tables are uncorrelated with the specifications that explain the amount of insurance bought.

Table 5 shows the results from testing whether those with worse self-perceived health have life insurance policies with larger face values. We regress the log of the face value of the life insurance policy on self-reported health status and the same set of controls used in Tables 3 and 4. (17) We estimate these regressions using ordinary least squares (OLS) and cluster standard errors at the household level. The results in Table 5 show that in each specification individuals in worse health have smaller policies relative to those in excellent health. For example, column 3 shows that the face value of the life insurance policy for those in Poor health is less than 50% of the face value of policies held by those in Excellent health. This once again suggests that advantageous selection effects are dominating the adverse selection effects, or that insurers have better information about individuals' health status than the individuals themselves. (18)

Next, we test whether the occurrence of death--between being interviewed in SIPP and 2010--can explain the face value of life insurance purchased (Table 6). Here, our sample consists of individuals owning life insurance. We specify a regression with the log of the face value of the life insurance policy as the dependent variable. Our explanatory variables are an indicator of whether the individual died in the aforementioned specified time interval, along with the same set of controls used in previous specifications. In these specifications, the point estimates on death are negative and statistically significant. For example, the point estimate from Table 6, column 3 indicates that those who die between being interviewed and 2010 have policies whose value is 28% smaller than those who do not die in this time interval.

To test the robustness of these findings, we also restricted the sample to just one SIPP panel, to one gender (testing males and females separately), to just those employed, to those with children younger than 18, and to just those earning more than $30,000 a year. We also modified our specifications to include interaction terms (e.g., between age categories and health status). Our results remained consistent across these specifications.

B. Results--Employer-Provided Life Insurance

Next, we seek to understand the cause of these negative coefficients. To understand if insurer price discrimination has any explanatory value for the results in Tables 3-6, we examine the market for group life insurance.

When the life insurance policy is provided as part of a group contract, an insurer has fewer opportunities to use price discrimination among consumers in offering life insurance policies. For example, federal employees are offered term life insurance at a flat rate of $0,325 per month per $1,000 of coverage (U.S. Office of Personnel Management 2012). Many employer life insurance policies have a fixed rate, but require some form of examination prior to receiving coverage. This provides a blunt threshold restriction allowing insurers to avoid the very ill, but allows for less precise pricing than the market for privately purchased policies.

First, we restrict the sample to working individuals who were asked about their life insurance status and determine whether they have purchased employer-provided life insurance. If insurers are less able to price discriminate with the latter type of life insurance, then we expect that those in worse health are more likely to purchase employer-provided life insurance than privately provided insurance, but only if consumers have private information about their health status that insurers cannot include in the employer-provided policy price. We then predict there to be more evidence of adverse selection. In particular, we expect the coefficients on Health Status and Died to be less negative than they were in Tables 3 through 6.

In Table 7, we report marginal effects from repeating the logistic analysis shown in Table 3, but here we use a dependent variable that equals one if the respondent owns employer-provided life insurance and is zero otherwise. Our variables of interest are health status and, as in our previous regressions, column 1 includes panel and demographic characteristics. Column 2 adds socio-economic variables, column 3 adds bequest motive variables, and column 4 restricts our sample to married individuals and includes the previously described characteristics of the spouse.

Columns 2, 3, and 4 of Table 7 show that the point estimates on Very Good and Good health are positive and statistically significant, indicating that these individuals are more likely to own life insurance than those in Excellent health. This is consistent with adverse selection. However, Table 7 also shows that the point estimates on those with Fair and Poor health are negative and statistically significant, indicating that these groups are less likely to own life insurance than those in Excellent health. One explanation for this finding is that although insurers have a reduced ability to price discriminate for employer-based life insurance policies, insurers can only deny coverage to those with Fair or Poor health status. This suggests that insurer pricing is the most important contributor to our findings from the baseline specification in subsection A of Section V.

Table 8 shows results from testing whether death between interview (19) and 2010 can predict employer-provided life insurance ownership. We estimate a logistic regression where the dependent variable equals one when the individual owns employer-provided life insurance, and zero otherwise. The point estimates on the indicator whether the person died within our sample period is negative and statistically significant in the first column of Table 8, but when we add socio-economic controls (columns 2-4), we are unable to reject the null hypothesis. Again, this supports the hypothesis that price discrimination is the most important driver of our findings from Table 4.

In Tables 9 and 10, we estimate similar regressions as in Tables 7 and 8, but the dependent variable is the face value of employer-provided life insurance policies. Table 9 shows that many of the point estimates on health status are not statistically significant. This contrasts with our findings in Table 5, which shows that policy size is monotonically decreasing with the deterioration of reported health status. While that latter table shows evidence of advantageous selection dominating adverse selection, most results in Table 9 no longer support the advantageous selection hypothesis. Given that the sample has already been restricted to only include working people, this is not because of changes in the probability of being employed, but rather reflects the importance of price discrimination in avoiding adverse selection.

In Table 10, the point estimates on whether the person died in the sample period are negative and statistically significant at the 5% level in three of the four specifications and are negative at the 10% level in all four. This suggests that a higher probability of death is a predictor of owning larger life insurance policies. These point estimates on death for employer-provided life insurance are roughly half the size of corresponding point estimates on death in Table 6, which are for all term life insurance policies, group, and individual.

These four analyses suggest that adverse selection is present in the market for life insurance, but that it is overcome through price discrimination and advantageous selection. The relative lack of positive coefficients suggests that advantageous selection is still present to temper the impact of adverse selection, but it is price discrimination that is the primary force removing it.

VI. LIMITATIONS

One limitation of our study is that our proxy of self-reported risk of death, that is, self-reported health status, is subject to measurement error. For example, an individual who knows she has a genetic predisposition to breast cancer might enjoy excellent health now, but know that her probability of death is higher than average. Moreover, the meanings of the health prompts (Excellent, Very Good, etc.) are highly subjective and are subject to measurement error. Formal assessments of the accuracy of self-reported health data on mortality have been mixed. A Swedish study found that "self-rated health is as good as or even better than that of most of the more specific questions" (Lundberg and Manderbacka 1996, 218). A large analysis of 27 data sets concludes, "in the great majority of cases, self-ratings add something more to the prediction of mortality" relative to specific, objective measures like blood pressure (Idler and Benyamini 1997, 34). However, a more recent U.S. study found that "there is a substantial amount of error in individuals' self-assessment of health" (Zajacova and Dowd 2011, 977).

Our data set also lacks pricing data for life insurance policies owned or offered, making it impossible to create a consumer-level model of how price discrimination occurs. It also lacks important variables in life insurance pricing, like smoking status, making it impractical to estimate insurance prices that would have been available to respondents.

We also lack the data to identify individuals who have jobs where they are eligible for group policy life insurance. Unfortunately, this prevents us from separating the full data set to those not in the market for employer life insurance, which could provide more helpful insights into the relative explanatory power of price discrimination and adverse selection. This limitation also opens up the possibility that the correlations observed in Tables 7-9 are because of the difference in distribution of mortality risk across firms that do or do not offer group life insurance policies.

VII. CONCLUSION

Our analysis makes use of administrative records and survey records to assess the presence of adverse selection in the market for life insurance. We fail to find evidence of adverse selection, finding the opposite predicted signs on both self-reported health, and true mortality risk. This suggests that either insurers are able to use price discrimination to exclude those with high probability of death from purchasing life insurance, or that consumers who are risk adverse are both more likely to buy life insurance and less likely to die. To gain some evidence which of these hypotheses have more explanatory power, we examine the market for employer-provided policies. In this market, the insurer has less ability to price discriminate but the behavioral response suggested by advantageous selection should still be in effect. For these types of policies, we find weak evidence for adverse selection. This suggests that most of the lack of support for the adverse selection hypothesis in the individual insurance markets is attributable to price discrimination techniques used by insurers. However, the fact that we do not find stronger evidence of adverse selection, for example, very few point estimates are significant and positive, in this area of employer-provided policies where pricing is so blunt suggests that there is some explanatory power in the advantageous selection hypothesis.

The applicability of our findings to other insurance markets, for example health insurance and homeowners' insurance, is left to future work. However, our findings suggest that granular pricing may help to avoid some of the adverse selection issues that are postulated in these markets. Policies that limit price discrimination in insurance markets will likely result in increased evidence of adverse selection and its attendant problems.

doi: 10.1111/ecin.12212

ABBREVIATIONS

GSF: Gold-Standard File

HRS: Health and Retirement Study

IRS: Internal Revenue Service

OLS: Ordinary Least Squares

SIPP: Survey of Income and Program Participation

SSA: Social Security Administration

REFERENCES

Abowd, J. M., and M. H. Stinson. "Estimating Measurement Error in Annual Job Earnings: A Comparison of Survey and Administrative Data." Review of Economics and Statistics, 95(5), 2013, 1451-67.

Akerlof, G. A. "The Market for 'Lemons': Quality Uncertainty and the Market Mechanism." Quarterly Journal of Economics, 84(3), 1970, 488-500.

Alderman, H. A. "Attrition in Longitudinal Household Survey Data." Demographic Research, 5(4), 2001, 79-124.

Anderson, L. R., and J. M. Mellor. "Predicting Health Behaviors with an Experimental Measure of Risk Preference." Journal of Health Economics, 27(5), 2008, 1260-74.

Cawley, J., and T. Philipson. "An Empirical Examination of Information Barriers to Trade in Insurance." American Economic Review, 89(4), 1999, 827-46.

Chiappori, P. A., and B. Salanie. "Testing for Asymmetric Information in Insurance Markets." Journal of Political Economy, 108(1), 2000, 56-78.

Cohen, A. "Testing for Adverse Selection in Insurance Markets." Journal of Risk and Insurance, 77(1), 2010, 39-84.

Cutler, D. M., A. Finkelstein, and K. McGarry. "Preference Heterogeneity and Insurance Markets: Explaining a Puzzle of Insurance." American Economic Review, 98(2), 2008, 157-62.

Deeg, D. J. H., T. van Tilburg, J. H. Smit, and E. D. de Leeuw. "Attrition in the Longitudinal Aging Study Amsterdam: The Effect of Differential Inclusion in Side Studies." Journal of Clinical Epidemiology, 55(4), 2002, 319-28.

Einav, L. A., and A. Finkelstein. "Selection in Insurance Markets: Theory and Empirics in Pictures." Journal of Economic Perspectives, 25(1), 2001, 115-38.

Finkelstein, A., and K. McGarry. "Multiple Dimensions of Private Information: Evidence from the Long-Term Care Insurance Market." American Economic Review, 96, 2006, 938-58.

Finkelstein, A., and J. Poterba. "Adverse Selection in Insurance Markets: Policyholder Evidence from the U.K. Annuity Market." Journal of Political Economy, 112, 2004, 183-208.

Gottschalck, A., and J. Moore. Evaluation of Questionnaire Design Changes on Life Insurance Policy Data. Washington, DC: U.S. Census Bureau, 2006.

He, D. "The Life Insurance Market: Asymmetric Information Revisited." Journal of Public Economics, 93(9), 2009, 1090-97.

Hemenway, D. "Propitious Selection." Quarterly Journal of Economics, 105(4), 1990, 1063-69.

Hendel, I., and A. Lizzeri. "The Role of Commitment in Dynamic Contracts: Evidence from Life Insurance." Quarterly Journal of Economics, 118(1), 2003, 299-327.

Idler, E. L., and Y. Benyamini. "Self-rated Health and Mortality: A Review of Twenty-seven Community Studies." Journal of Health and Social Behavior, 38(1), 1997, 21-37.

Lundberg, O., and K. Manderbacka. "Assessing Reliability of a Measure of Self-rated Health." Scandinavian Journal of Public Health, 24(3), 1996, 218-24.

McCarthy, D., and O. S. Mitchell. "International Adverse Selection in Life Insurance and Annuities, " in Ageing in Advanced Industrial States, edited by S. Tuljapurkar, A. H. Gauthier, and N. Ogawa. New York: Springer-Verlag, 2010, 119-35.

Nicholson, W., and C. Snyder. Intermediate Microeconomics: and Its Applications. Mason, OH: Cengage Learning, 2009.

Perloff, J. M. Microeconomics. Boston: Pearson Addison-Wesley, 2004.

Rothschild, M., and J. Stiglitz. "Equilibrium in Competitive Insurance Markets: An Essay on the Economics of Imperfect Information." Quarterly Journal of Economics, 90(4), 1976, 629-49.

U.S. Census Bureau. "Codebook for SIPP Gold Standard and Synthetic Beta Files." 2010. Accessed August 1, 2011. http://www.census.gov/sipp/SSB_Codebook.pdf.

--. "Overview of the Survey of Income and Program Participation." 2011. Accessed October 20, 2011. http:// www.census.gov/sipp/overview.html.

--. "Calculating Migration Expectancy Using ACS Data." 2012. Accessed August 31, 2012. http://www .census.gov/hhes/migration/about/cal-mig-exp.html.

U.S. Office of Personnel Management. "Federal Employees' Group Life Insurance." 2012. Accessed August 31, 2012. http://www.opm.gov/insure/life/rates/em_rates .asp.

Weibull, J. W. "Award Ceremony Speech." December 10, 2001. Accessed November 1, 2011. http://www .nobelprize.org/nobel_prizes/economics/laureates/ 2001/presentation-speech.html.

Wisconsin Office of the Commissioner of Insurance. "Cash Value Life Insurance." N.d. Accessed March 1, 2012. http://oci.wi.gov/consumer/lifeann-cashval.htm.

Zajacova, A., and J. B. Dowd. "Reliability of Self-rated Health in US Adults." American Journal of Epidemiology, 174(8), 2011, 977-83.

(1.) Of the individuals in our sample who own life insurance, over 50% report holding only term policies, about 30% hold only whole policies, and the remainder have both types of policies.

(2.) This can account for a great deal of attrition, as the average American moves 11.7 times in their lifetime (U.S. Census Bureau 2012).

(3.) The 1984, 1990, 1991, 1992, and 1993 panels are also available in the SIPP GSF. We do not use these waves because the 1984 panel does not contain all of our variables of interest and the 1990-1993 panels did not ask about all of the outcomes we study in this paper at the same time. We used these five additional panels as a robustness check, creating comparable data sets, and reached the same conclusions as those in this paper. We do not report these results here, and these tables are available upon request.

(4.) For additional information on the SIPP Gold Standard File please see Section III of http://www.census.gov/sipp/ FinalReporttoSocialSecurityAdministration.pdf.

(5.) In all regressions, we control for changes across the panels with an indicator variable. We have also replicated the analysis on each panel individually and find the same conclusions as those reported in this paper. These regressions are available upon request.

(6.) In the 2004 panel, the question to determine the value of the insurance policy held was changed from asking about the face value--the amount a policy pays out in the event of the policy holder's death--to the cash value--i.e., the amount of money the policy could be sold for or borrowed against (Wisconsin Office of the Commissioner of Insurance). However, by definition, term life policies have no cash value. Even though the text of the question changed, and changed to something to which there is no clear answer, consumers by and large continued to report the face value of their policies. Gottschalck and Moore (2006) found that "the median value for term policies (excluding respondents who reported a zero dollar amount) remained unchanged" from the 2001 to 2004 panels. As our analysis excluded those who describe their policies as whole life, we believe there is no significant cause for concern in the accuracy of the 2004 and 2008 panels.

(7.) The question text is, "Would you say your health in general is excellent, very good, good, fair, or poor?" (8.) All post-1996 SIPP panels are included in these regressions, so the interview took place in 1996, 2001, 2004, or 2008.

(9.) The last date available in the administrative records at the time this analysis was performed.

(10.) We have also estimated our regressions using earnings instead of income and reach the same conclusions. These results are available upon request. Income and wealth variables are both in 2010 constant dollars.

(11.) We also used the log of administrative earnings from the GSF and our result did not change. To minimize the number of tables to send through disclosure review, we do not report those results here.

(12.) We do not log net worth in order to use data from the many consumers with negative net worth, which might increase demand for life insurance.

(13.) Unfortunately, the SIPP does not have information on smoking status nor for other measures of health, as for example, blood pressure.

(14.) We estimated all regressions using both SIPP weights and no weights, and the results from these two specifications are very similar. We do not report the weighted regression results. A drawback of the weights is that the weights from the SIPP are not nationally representative in the GSF, because some respondents are missing the information to link them to administrative records. The weights available in the GSF do not account for this issue.

(15.) Following Cawley and Philipson (1999), we include income rather than earnings in our controls. We have also estimated all regressions with earnings and our conclusions are unchanged.

(16.) For simplicity, we have omitted any discussion of moral hazard here, but the presence of moral hazard would tend to push these coefficients toward positive numbers. The fact that they are so strongly negative suggests any moral hazard effect is also overwhelmed by advantageous selection.

(17.) We compute the face value of the life insurance policy as the sum of term and whole life policies. This is because the face values of each type of policy are not separated out in SIPP.

(18.) We also estimate this table using all respondents and coded those without life insurance as having a policy with a face value of zero and added one to all face values to allow logging. The trend of our results is unchanged. Those who report worse health are less likely to own life insurance. For example, those in Good, Fair and Poor health in column 3 have point estimates of -0.51, -1.40, and -2.18, respectively. In column 4 the point estimates are -0.28, -0.83, and -1.45 for Good, Fair, and Poor, respectively. All of these estimates are significant at the 0.001 level.

(19.) All post 1996 SIPP panels are included in these regressions, so that the interview took place either in 1996, 2001, 2004, or in 2008.

DAVID HEDENGREN and THOMAS STRATMANN *

* The authors are grateful to Rebecca Chenevert, Graton Gathright, Martha Stinson, Marina Vomovitsky, and Chris Wignall of the U.S. Census Bureau, and two anonymous reviewers for their helpful comments and feedback on early versions of this article.

Hedengren: Senior Marketing Manager, VP of Analytic Marketing, JPMorgan Chase, Columbus, OH 43015. Phone 703-634-9479, Fax 703-993-1133, E-mail [email protected]

Stratmann: University Professor of Economics and Law, Department of Economics, George Mason University, Fairfax, VA 22030. Phone 703-993-4920, Fax 703-9931133, E-mail [email protected]

TABLE 1
Descriptive Statistics

                                                             Standard
                                           N        Mean     Deviation

Owns life insurance                     124.977     .49         .50
Owns employer provided life insurance   124.977     .30         .46
Face value of life insurance            50.426    $133,829    172,451
Face value of employer provided         29.632    $88,567     104,662
Died                                    124.958     .029        .17
Male                                    124.977     .47         .50
Married                                 124.977     .59         .49
Income                                  124,977   $36,410     45,043
Any savings?                            124.977     .31         .46
Age                                     124,977      41         11
Total net worth                         124.977   $210,521   1,392,179
Total kids in family                    124,977     .92         1.2
Less than high school                   124,977     .10         .30
High school graduate                    124.977     .26         .44
Some college                            124,977     .33         .47
Bachelor's degree                       124,977     .17         .38
Graduate degree                         124,977     .088        .28
White                                   124,977     .82         .38
Black                                   124,977     .11         .31
Other                                   124,977     .07         .25

Notes: Observations are by individual and are from the 1996, 2001,
2004, and 2008 panels of the SIPP Gold Standard File V. 6. Each
individual in this data set appears in only one of the panels.
Dollar values are in real 2010 dollars.

http://www.census.gov/sipp/
FinalReporttoSocialSecurityAdministration.pdf.

TABLE 2
Regression of Health Status on the Likelihood of Death

                                             (1)

Very Good                             0.437 *** (0.0647)
Good                                  1.161 *** (0.0612)
Fair                                  2.150 *** (0.0634)
Poor                                  3.119 *** (0.0674)
N                                          124,963
Pseudo [R.sup.2]                            0.160
Panel indicators included?                   Yes
Demographic variables included?               No
Socio-economic variables included?            No

                                             (2)

Very Good                             0.315 *** (0.0650)
Good                                  0.878 *** (0.0622)
Fair                                  1.708 *** (0.0658)
Poor                                  2.540 *** (0.0713)
N                                          124,963
Pseudo [R.sup.2]                            0.227
Panel indicators included?                   Yes
Demographic variables included?              Yes
Socio-economic variables included?            No

                                             (3)

Very Good                             0.222 *** (0.0665)
Good                                  0.701 *** (0.0650)
Fair                                  1.476 *** (0.0715)
Poor                                  2.245 *** (0.0784)
N                                          118,748
Pseudo [R.sup.2]                            0.232
Panel indicators included?                   Yes
Demographic variables included?              Yes
Socio-economic variables included?           Yes

Notes: Results are from a logistic regression. The dependent variable
is equal to one if an individual died by 2010, and zero otherwise.
The variables Very Good, Good, Fair, and Poor are indicator variables
reflecting the self-reported health condition of an individual.
Excellent is the reference category. Observations are by individual
and are from the 1996,2001,2004, and 2008 panels of the SIPP Gold
Standard File V. 6. Each individual in this data set appears in only
one of the panels. Demographic variables are an individual's gender,
race, age, and age squared. Socio-economic variables are education,
log of personal income, and total net wealth of the individual's
household. Bequest variables are the number of children and an
indicator for being married. Spouse variables are the log of the
spouse's income, and an indicator for whether the spouse owns life
insurance. We report marginal effects and standard errors in
parentheses. Standard errors are clustered by household.

* p < .05, ** p < .01, *** p < .001.

http://www.census.gov/sipp/
FinalReporttoSocialSecurityAdministration.pdf.

TABLE 3
Regression of Health Status on the Likelihood of Owning Life
Insurance

                                              (1)

Very Good                             -0.109 *** (0.0161)
Good                                  -0.510 *** (0.0176)
Fair                                  -1.195 *** (0.0255)
Poor                                  -1.796 *** (0.0416)
N                                           124,977
Pseudo [R.sup.2]                            0.0735
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            No
Bequest variables included?                   No
Spouse variables included?                    No

                                              (2)

Very Good                              -0.00509 (0.0174)
Good                                  -0.223 *** (0.0192)
Fair                                  -0.689 *** (0.0277)
Poor                                  -1.103 *** (0.0454)
N                                           118,767
Pseudo [R.sup.2]                             0.148
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            Yes
Bequest variables included?                   No
Spouse variables included?                    No

                                              (3)

Very Good                              0.00979 (0.0175)
Good                                  -0.182 *** (0.0193)
Fair                                  -0.582 *** (0.0279)
Poor                                  -0.966 *** (0.0456)
N                                           118,767
Pseudo [R.sup.2]                             0.165
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            Yes
Bequest variables included?                   Yes
Spouse variables included?                    No

                                              (4)

Very Good                               0.0329 (0.0230)
Good                                  -0.130 *** (0.0256)
Fair                                  -0.430 *** (0.0412)
Poor                                  -0.731 *** (0.0696)
N                                           65,744
Pseudo [R.sup.2]                             0.273
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            Yes
Bequest variables included?                   Yes
Spouse variables included?                    Yes

Notes: Results are from a logistic regression. The dependent variable
is equal to one if an individual holds life insurance, and zero
otherwise. The variables Very Good, Good, Fair, and Poor are
indicator variables reflecting the self-reported health condition of
an individual. Excellent is the reference category. Observations are
by individual and are from the 1996, 2001, 2004, and 2008 panels of
the SIPP Gold Standard File V. 6. Each individual in this data set
appears in only one of the panels. Demographic variables are an
individual's gender, race, age, and age squared. Socio-economic
variables are education, log of personal income, and total net wealth
of the individual's household. Bequest variables are the number of
children and an indicator for being married. Spouse variables are the
log of the spouse's income, and an indicator for whether the spouse
owns life insurance. We report marginal effects and standard errors
in parentheses. Standard errors are clustered by household.

* p < .05, ** p < .01, *** p < .001.

TABLE 4
Death as a Predictor of Owning Life Insurance

                                              (1)

Died                                  -0.790 *** (0.0364)
N                                           124,958
Pseudo [R.sup.2]                            0.0481
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            No
Bequest variables included?                   No
Spouse variables included?                    No

                                              (2)

Died                                  -0.474 *** (0.0408)
N                                           118,748
Pseudo [R.sup.2]                             0.141
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            Yes
Bequest variables included?                   No
Spouse variables included?                    No

                                              (3)

Died                                  -0.386 *** (0.0407)
N                                           118,748
Pseudo [R.sup.2]                             0.159
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            Yes
Bequest variables included?                   Yes
Spouse variables included?                    No

                                              (4)

Died                                  -0.292 *** (0.0658)
N                                           65,737
Pseudo [R.sup.2]                             0.270
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            Yes
Bequest variables included?                   Yes
Spouse variables included?                    Yes

Notes: Results are from a logistic regression. The dependent variable
is equal to one if an individual holds life insurance, and zero
otherwise. The variable Died is equal to one if an individual died by
2010, and zero otherwise. Observations are by individual and are from
the 1996, 2001, 2004, and 2008 panels of the SIPP Gold Standard File
V. 6. Each individual in this data set appears in only one of the
panels. Demographic variables are an individual's gender, race, age,
and age squared. Socio-economic variables are education, log of
personal income, and total net wealth of the individual's household.
Bequest variables are the number of children and an indicator for
being married. Spouse variables are the log of the spouse's income,
and an indicator for whether the spouse owns life insurance. We
report marginal effects and standard errors in parentheses. Standard
errors are clustered by household.

* p < .05, ** p <.01, *** p <.001.

http://www.census.gov/sipp/
FinalReporttoSocialSecurityAdministration.pdf.

TABLE 5
Regression of Health Status on the Face Value of Life Insurance

                                              (1)

Very Good                             -0.122 *** (0.0193)
Good                                  -0.317 *** (0.0218)
Fair                                  -0.606 *** (0.0350)
Poor                                  -0.987 *** (0.0647)
N                                           50.426
[R.sup.2]                                   0.0600
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            No
Bequest variables included?                   No
Spouse variables included?                    No

                                              (2)

Very Good                             -0.0501 * (0.0196)
Good                                  -0.158 *** (0.0228)
Fair                                  -0.348 *** (0.0362)
Poor                                  -0.609 *** (0.0662)
N                                           48,639
[R.sup.2]                                   0.0960
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            Yes
Bequest variables included?                   No
Spouse variables included?                    No

                                              (3)

Very Good                             -0.0399 * (0.0194)
Good                                  -0.133 *** (0.0223)
Fair                                  -0.299 *** (0.0356)
Poor                                  -0.554 *** (0.0654)
N                                           48.639
[R.sup.2]                                    0.109
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            Yes
Bequest variables included?                   Yes
Spouse variables included?                    No

                                              (4)

Very Good                             -0.0641 ** (0.0239)
Good                                  -0.159 *** (0.0279)
Fair                                  -0.276 *** (0.0478)
Poor                                  -0.600 *** (0.0921)
N                                           31.160
[R.sup.2]                                   0.0951
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            Yes
Bequest variables included?                   Yes
Spouse variables included?                    Yes

Notes: The dependent variable is the log face value of the purchased
life insurance. The variables Very Good, Good, Fair, and Poor are
indicator variables reflecting the self-reported health condition of
an individual. Excellent is the reference category. Observations are
by individual and are from the 1996, 2001, 2004, and 2008 panels of
the SIPP Gold Standard File V. 6. Each individual in this data set
appears in only one of the panels. Demographic variables are an
individual's gender, race, age, and age squared. Socio-economic
variables are education, log of personal income, and total net wealth
of the individual's household. Bequest variables are the number of
children and an indicator for being married. Spouse variables are the
log of the spouse's income, and an indicator for whether the spouse
owns life insurance. We report marginal effects and standard errors
in parentheses. Standard errors are clustered by household.

* p < .05, ** p < .01, *** p < .001.

http://www.census.gov/sipp/
FinalReporttoSocialSecurityAdministration.pdf.

TABLE 6
Death as a Predictor of the Face Value of Life Insurance

                                              (1)

Died                                  -0.500 *** (0.0459)
N                                           50,525
[R.sup.2]                                   0.0509
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            No
Bequest variables included?                   No
Spouse variables included?                    No

                                              (2)

Died                                  -0.310 *** (0.0454)
N                                           48,732
[R.sup.2]                                   0.0933
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            Yes
Bequest variables included?                   No
Spouse variables included?                    No

                                              (3)

Died                                  -0.275 *** (0.0449)
N                                           48,732
[R.sup.2]                                    0.107
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            Yes
Bequest variables included?                   Yes
Spouse variables included?                    No

                                              (4)

Died                                  -0.236 *** (0.0615)
N                                           31,213
[R.sup.2]                                   0.0931
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            Yes
Bequest variables included?                   Yes
Spouse variables included?                    Yes

Notes: The dependent variable is the log face value of the purchased
life insurance. The variable Died is equal to one if an individual
died by 2010, and zero otherwise. Observations are by individual and
are from the 1996, 2001, 2004, and 2008 panels of the SIPP Gold
Standard File V. 6. Each individual in this data set appears in only
one of the panels. Demographic variables are an individual's gender,
race, age, and age squared. Socio-economic variables are education,
log of personal income, and total net wealth of the individual's
household. Bequest variables are the number of children and an
indicator for being married. Spouse variables are the log of the
spouse's income, and an indicator for whether the spouse owns life
insurance. We report marginal effects and standard errors in
parentheses. Standard errors are clustered by household.

* p < .05, ** p < .01, *** p < .001.

http://www.census.gov/sipp/
FinalReporttoSocialSecurityAdministration.pdf.

TABLE 7
Regression of Health Status on the Likelihood of Owning
Employer-Provided Life Insurance

                                              (1)

Very Good                               -0.004 (0.0172)
Good                                  -0.288 *** (0.0192)
Fair                                  -0.772 *** (0.0328)
Poor                                  -1.300 *** (0.0766)
N                                           97,950
Pseudo [R.sup.2]                            0.0420
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            No
Bequest variables included?                   No
Spouse variables included?                    No

                                              (2)

Very Good                             0.139 *** (0.0193)
Good                                  0.0618 ** (0.0219)
Fair                                  -0.154 *** (0.0367)
Poor                                  -0.353 *** (0.0813)
N                                           93,122
Pseudo [R.sup.2]                             0.161
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            Yes
Bequest variables included?                   No
Spouse variables included?                    No

                                              (3)

Very Good                             0.141 *** (0.0193)
Good                                  0.0694 ** (0.0219)
Fair                                  -0.138 *** (0.0367)
Poor                                  -0.332 *** (0.0812)
N                                           93,122
Pseudo [R.sup.2]                             0.162
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            Yes
Bequest variables included?                   Yes
Spouse variables included?                    No

                                              (4)

Very Good                             0.169 *** (0.0245)
Good                                  0.109 *** (0.0286)
Fair                                   -0.109 * (0.0512)
Poor                                    -0.135 (0.116)
N                                           52,715
Pseudo [R.sup.2]                             0.169
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            Yes
Bequest variables included?                   Yes
Spouse variables included?                    Yes

Notes: Results are from a logistic regression. The dependent variable
equals one if the consumer owns employer-provided life insurance, and
zero otherwise. The variables Very Good, Good, Fair, and Poor are
indicator variables reflecting the self-reported health condition of
an individual. Excellent is the reference category. Observations are
by individual and are from the 1996, 2001, 2004, and 2008 panels of
the SIPP Gold Standard File V. 6. Each individual in this data set
appears in only one of the panels. Demographic variables are an
individual's gender, race, age, and age squared. Socio-economic
variables are education, log of personal income, and total net wealth
of the individual's household. Bequest variables are the number of
children and an indicator for being married. Spouse variables are the
log of the spouse's income, and an indicator for whether the spouse
owns life insurance. We report marginal effects and standard errors
in parentheses. Standard errors are clustered by household.

* p < .05, ** p < .01, *** p < .001.

http://www.census.gov/sipp/
FinalReporttoSocialSecurityAdministration.pdf.

TABLE 8
Death as a Predictor of Owning Employer-Provided Life Insurance

                                              (1)

Died                                  -0.415 *** (0.0499)
N                                           97,939
Pseudo [R.sup.2]                            0.0335
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            No
Bequest variables included?                   No
Spouse variables included?                    No

                                              (2)

Died                                   -0.0779 (0.0559)
N                                           93,111
Pseudo [R.sup.2]                             0.160
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            Yes
Bequest variables included?                   No
Spouse variables included?                    No

                                              (3)

Died                                   -0.0712 (0.0559)
N                                           93,111
Pseudo [R.sup.2]                             0.161
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            Yes
Bequest variables included?                   Yes
Spouse variables included?                    No

                                              (4)

Died                                    0.0152 (0.0783)
N                                           52,711
Pseudo [R.sup.2]                             0.168
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            Yes
Bequest variables included?                   Yes
Spouse variables included?                    Yes

Notes: Results are from a logistic regression. The dependent variable
equals one if the consumer owns employer-provided life insurance, and
zero otherwise. The variable Died is equal to one if an individual
died by 2010, and zero otherwise. Observations are by individual and
are from the 1996, 2001, 2004, and 2008 panels of the SIPP Gold
Standard File V. 6. Each individual in this data set appears in only
one of the panels. Demographic variables are an individual's gender,
race, age, and age squared. Socio-economic variables are education,
log of personal income, and total net wealth of the individual's
household. Bequest variables are the number of children and an
indicator for being married. Spouse variables are the log of the
spouse's income, and an indicator for whether the spouse owns life
insurance. We report marginal effects and standard errors in
parentheses. Standard errors are clustered by household.

* p < .05, ** p < .01, *** p < .001.

http://www.census.gov/sipp/
FinalReporttoSocialSecurityAdministration.pdf.

TABLE 9
Regression of Health Status on the Face Value of Employer-Provided
Life Insurance

                                              (1)

Very Good                             -0.0781 *** (0.0228)
Good                                  -0.216 *** (0.0265)
Fair                                  -0.361 *** (0.0466)
Poor                                        -0.269 *
                                            (0.110)
N                                            27,844
[R.sup.2]                                    0.0347
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?             No
Bequest variables included?                    No
Spouse variables included?                     No

                                              (2)

Very Good                               0.0119 (0.0225)
Good                                    -0.0337 (0.0263)
Fair                                   -0.0935 * (0.0456)
Poor                                         0.0134
                                            (0.111)
N                                            26,935
[R.sup.2]                                    0.0854
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            Yes
Bequest variables included?                    No
Spouse variables included?                     No

                                              (3)

Very Good                               0.0144 (0.0225)
Good                                    -0.0265 (0.0262)
Fair                                    -0.0743 (0.0455)
Poor                                         0.0286
                                            (0.111)
N                                            26,935
[R.sup.2]                                    0.0884
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            Yes
Bequest variables included?                   Yes
Spouse variables included?                     No

                                              (4)

Very Good                               0.0119 (0.0276)
Good                                   -0.00376 (0.0334)
Fair                                    -0.0440 (0.0585)
Poor                                         0.0629
                                            (0.127)
N                                            16,742
[R.sup.2]                                    0.0888
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            Yes
Bequest variables included?                   Yes
Spouse variables included?                    Yes

Notes: The dependent variable is the log face value of employer-
provided life insurance. The variables Very Good, Good, Fair, and
Poor are indicator variables reflecting the self-reported health
condition of an individual. Excellent is the reference category.
Observations are by individual and are from the 1996, 2001, 2004, and
2008 panels of the SIPP Gold Standard File V. 6. Each individual in
this data set appears in only one of the panels. Demographic
variables are an individual's gender, race, age, and age squared.
Socio-economic variables are education, log of personal income, and
total net wealth of the individual's household. Bequest variables are
the number of children and an indicator for being married. Spouse
variables are the log of the spouse's income, and an indicator for
whether the spouse owns life insurance. We report marginal effects
and standard errors in parentheses. Standard errors are clustered by
household.

* p < .05, ** p < .01, *** p < .001.

http://www.census.gov/sipp/
FinalReporttoSocialSecurityAdministration.pdf.

TABLE 10
Death as a Predictor of the Face Value of Owning Employer-Provided
Life Insurance

                                              (1)

Died                                  -0.289 *** (0.0559)
N                                           27,843
[R.sup.2]                                   0.0316
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            No
Bequest variables included?                   No
Spouse variables included?                    No

                                              (2)

Died                                   -0.130 * (0.0536)
N                                           26,934
[R.sup.2]                                   0.0852
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            Yes
Bequest variables included?                   No
Spouse variables included?                    No

                                              (3)

Died                                   -0.122 * (0.0534)
N                                           26,934
[R.sup.2]                                   0.0883
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            Yes
Bequest variables included?                   Yes
Spouse variables included?                    No

                                              (4)

Died                                    -0.125 (0.0707)
N                                           16,741
[R.sup.2]                                   0.0888
Panel indicators included?                    Yes
Demographic variables included?               Yes
Socio-economic variables included?            Yes
Bequest variables included?                   Yes
Spouse variables included?                    Yes

Notes: The dependent variable is the log face value of employer-
provided life insurance. The dependent variable equals one if the
consumer owns employer provided the life insurance, and zero
otherwise. The variable Died is equal to one if an individual died by
2010, and zero otherwise. Observations are by individual and are from
the 1996, 2001, 2004, and 2008 panels of the SIPP Gold Standard File
V. 6. Each individual in this data set appears in only one of the
panels. Demographic variables are an individual's gender, race, age,
and age squared. Socio-economic variables are education, log of
personal income, and total net wealth of the individual's household.
Bequest variables are the number of children and an indicator for
being married. Spouse variables are the log of the spouse's income,
and an indicator for whether the spouse owns life insurance. We
report marginal effects and standard errors in parentheses. Standard
errors are clustered by household.

* p < 0.05, ** p < 0.01, *** p < 0.001.

http://www.census.gov/sipp/
FinalReporttoSocialSecurityAdministration.pdf.