Heightened mortality after the death of a spouse: Marriage protection or marriage selection?

Journal of Health Economics 27 (2008) 1326–1342

Contents lists available at ScienceDirect

Journal of Health Economics journal homepage: www.elsevier.com/locate/econbase

Heightened mortality after the death of a spouse: Marriage protection or marriage selection? Javier Espinosa a,∗ , William N. Evans b,c a b c

Department of Economics, Rochester Institute of Technology, Rochester, NY 14623, USA Department of Economics and Econometrics, University of Notre Dame, Notre Dame, IN 46556, USA National Bureau of Economic Research, 1050 Massachusetts Avenue, Cambridge, MA 02138-5398, USA

a r t i c l e

i n f o

Article history: Received 18 July 2007 Received in revised form 27 March 2008 Accepted 3 April 2008 Available online 11 April 2008 JEL classification: I12 J12

a b s t r a c t We test whether the heightened mortality after the death of a spouse represents correlation or causation by examining the heterogeneity in the bereavement effect based on the spouse’s cause of death. Some causes of death are correlated with socioeconomic characteristics while others are not. Equality in the bereavement effect across these two types of death would signal a causal relationship while no bereavement effect for uncorrelated causes of death would indicate an omitted variables bias. Results indicate that the observed effect for women is subject to an omitted variables bias but the estimates for men indicate a causal relationship. © 2008 Elsevier B.V. All rights reserved.

Keywords: Mortality Marriage Widow

1. Introduction In recent years, legislators at the federal and state level have proposed a variety of “healthy marriage” laws designed to discourage divorce and encourage couples with children or pregnant couples to marry. In early January 2005, for example, a welfare reauthorization bill (HR 240) was introduced in the House of Representatives1 that included over $300 million in funds for healthy marriage education and research. Behind these initiatives is the belief that marriage improves the lives of husbands, wives, children and communities. This belief is bolstered by research from a variety of disciplines that documents the social and economic outcomes of the married are better along many dimensions than those who are single (Waite and Gallagher, 2000). Married people, for example, have higher wages (Korenman and Neumark, 1991; Schoeni, 1995), higher incomes (Lerman, 2002), lower rates of poverty (Sawhill and Thomas, 2002), accumulate more wealth (Lupton and Smith, 2003), report higher levels of happiness (Diener et al., 2000; Stack and Eshleman, 1998; Taylor et al., 2006) are less susceptible to psychological disorders (Gove et al., 1983; Barrett, 2000; Quirouette and Gold, 1992), are less prone to suicide (Smith et al., 1988; Kachur et al., 1995; Luoma and Pearson, 2002) and are less likely to be involved in criminal activity (Laub et al., 1998), compared with singles. Researchers have also documented a potential health benefit of marriage—married people have a longer life expectancy than the non-married (Gove, 1973; Kobrin and Hendershot, 1977; Smith and Waitzman, 1994; Hu and Goldman, 1990). Scholars have titled this relationship ‘marriage protection’. Some of the most convincing evidence consistent

∗ Corresponding author. Tel.: +1 585 475 5872; fax: +1 585 475 7120. E-mail addresses: [email protected] (J. Espinosa), [email protected] (W.N. Evans). 1 HR 240. Personal Responsibility, Work, and Family Promotion Act of 2005. January 4, 2005. Section 103.b.2.C. 0167-6296/$ – see front matter © 2008 Elsevier B.V. All rights reserved. doi:10.1016/j.jhealeco.2008.04.001

J. Espinosa, W.N. Evans / Journal of Health Economics 27 (2008) 1326–1342

1327

with the marriage protection hypothesis demonstrates a heightened mortality rate for survivors in the years just after the death of their spouse (Kaprio et al., 1987; Schaefer et al., 1995). One interpretation of the data suggests that as the marriage is ended by death of a spouse, the protective effects of marriage are eroded and the health of the survivor is then compromised. There is, however, an alternative hypothesis called marriage selection (Hu and Goldman, 1990; Goldman, 1993; Waldron et al., 1996; Cheung, 1998; Murray, 2000) that may explain the better health outcomes of married individuals and why we witness a temporal correlation in spousal mortality. In a marriage market where potential husbands and wives find one another, there is a selection process that underlies the matching of two people (Becker, 1973, 1974). Much of the empirical work about marriage markets demonstrates positive assortative mating, which is the occurrence of mating between similar individuals at higher than random frequencies. Given the importance health habits have on life expectancy and mortality (McGinnis and Foege, 1993) and the similarity in health investments by spouses (documented in Section 2 below), we would expect to see life expectancy patterns among the married converge as well. In this article, we estimate the impact of spousal death on the surviving spouse’s mortality, what some researchers have come to call the “bereavement effect”. We do so by trying to extract information about the survivor’s mortality based on the cause of death for their spouse. The starting point for our analysis is the fact that some deaths reveal more information about the surviving spouse than others. Deaths from cirrhosis of the liver, heart disease, and lung cancer are causes that are correlated with various measures of socioeconomic status (SES) and hence, because of assortative mating, characteristics of the surviving spouse. Therefore, any heightened mortality among survivors after the spouse dies from one of these causes is potentially contaminated by an omitted variables bias. In contrast, if some deaths are random in the population, then the fact that a spouse dies from one of these causes should not reveal any information about the survivor. Hence, any bereavement effect that occurs after the death of a spouse from these uninformative causes is more likely to be causal. While we cannot identify which causes of death are ‘random’, we can identify those that are uncorrelated with important observed characteristics. For example, as we document below, some causes of death like adult leukemia and other forms of cancer are uncorrelated with measures of SES like income, education and occupation. Therefore, if the bereavement effect is present after a spouse dies from one of these causes of death, we are more confident that the result is driven by marriage protection rather than a byproduct of marriage selection. To implement this idea, we use a public-use version of the National Longitudinal Mortality Survey (NLMS) to create longitudinal data sets of married couples, aged 50–70. We identify those causes of death that are predicted by SES such as income, education and occupation, and those that are not. We label these two types of deaths as informative and uninformative causes of death, respectively.2 In mortality models that are estimated via Cox proportional hazards, if the bereavement effect is entirely a product of marriage selection, we should find the magnitude of the bereavement effect resulting from uninformative causes of spousal death to be substantially smaller than the magnitude of the effect for informative causes. In contrast, if marital protection explains all of the bereavement effect, then the impacts of informative and uninformative deaths should be similar in magnitude. In Section 2, we summarize the literature on both marriage and mortality and the bereavement effect. We also present some data indicating the robust correlation in health behaviors among older married couples, a result that raises the suspicion that the bereavement effect is driven by marriage selection. In Section 3, we introduce the NLMS, outline the Cox proportional hazard model, and introduce the procedure to isolate what are informative and uninformative deaths. The results from these models are not surprising. Among men, causes of death like pancreatic cancer, genital cancers, and leukemia are considered uninformative while for women, the list includes breast cancer and genitourinary cancer, among others. In Section 4, we demonstrate that for men, the bereavement effect is large whether their wife dies from an informative or uninformative case, suggesting the heightened mortality after the death of a spouse is due to marriage protection. In contrast, for women, we find no bereavement effect for uninformative causes of death, but a large and statistically precise effect for informative causes, suggesting most of the bereavement effect is due to omitted variable bias. We also use an alternative classification of cause of death that is based on professional assessments of what deaths can be prevented and obtain very similar results. Finally, we examine the heterogeneity in the bereavement effect based on the time since the death of the spouse. For females, we find that the bereavement effect dissipates 24 months and more after the death of the spouse. In contrast, the effect for males is as strong 24 or more months after the death of the spouse as it is in the first six months. When we allow these effects to vary by cause of death, we again find evidence that the results for men are causal but the results for women are most likely driven by an omitted variable bias. 2. Marriage, mortality, the bereavement effect, and data on marriage selection 2.1. Literature review The antecedents for the current study flow from two different but related strands of literature. The first group of papers document lower mortality rates among the married. Empirical investigations of the differences in mortality rates between

2 If a spouse’s cause of death can be predicted by SES, it provides the researcher with information regarding the mortality of the surviving spouse; hence, we call this an informative cause of death. The opposing case is a cause of death that is not predicted by SES, thus we call this an uninformative cause of death.

1328

J. Espinosa, W.N. Evans / Journal of Health Economics 27 (2008) 1326–1342

married and unmarried individuals go back over a century to Farr (1858) and Durkheim (1897). The negative association between marriage and mortality has been documented in the US by a number of authors including Gove (1973), Kobrin and Hendershot (1977), Smith and Waitzman (1994), Mergenhagen et al. (1985), Rogers (1995), Kisker and Goldman (1987), Lillard and Waite (1995), and Lillard and Panis (1996). This relationship has been established for other countries including 16 developed countries (Hu and Goldman, 1990), England (Gardner and Oswald, 2004), Canada (Trovato, 1991), Israel (Manor et al., 2000), and Bangladesh (Rahman, 1993). There are a variety of reasons why mortality rates may be lower among the married. Marriage may increase the financial standing of the partners by increasing family income or providing insurance when bad health and/or financial shocks occur (Lillard and Waite, 1995; Waite and Gallagher, 2000). Marriage may also provide important psychological benefits such as reducing stress, improving one’s disposition or integrating a person into a community (Waite, 1995; Gove, 1973). A detailed literature in epidemiology has established that those with greater social ties have better health and lower mortality (Berkman and Syme, 1979; Blazer, 1982; House et al., 1982; Cohen et al., 1997). Finally, marriage may discourage risky behavior such as smoking, heavy alcohol use or illicit drug use (Miller-Tutzauer et al., 1991; Curran et al., 1998; Duncan et al., 2006) or criminal activity (Laub et al., 1998) or encourage healthy behavior such as visiting the doctor (Umberson, 1992; Verbrugge, 1979). In a related line of work, researchers have also documented heightened mortality of the recently widowed, a relationship referred to by some as the “bereavement effect”. This basic statistical relationship between widowhood and mortality has been documented by a number of authors including Cox and Ford (1964), Parkes et al. (1969), Helsing and Szklo (1981), Kaprio et al. (1987), Mendes de Leon et al. (1993), Mineau et al. (2002), Korenman et al. (1997), and Lillard and Waite (1995), just to name a few. Excess mortality among the recently widowed has been demonstrated to exist in a wide variety of age groups, socioeconomic levels, countries and cultures. The impact of a spouse’s death is qualitatively large. After controlling for a variety of factors, Schaefer et al. (1995) and Kaprio et al. (1987) found that mortality rates double for the surviving spouse in the first year after the death of their spouse. Estimates also suggest that the bereavement effect is strongest in the period right after the death of a spouse (Lichtenstein et al., 1998; Manor and Eisenbach, 2003). All of the studies mentioned above demonstrate excess mortality for surviving males, but the results for surviving females are less definitive. Helsing and Szklo (1981) found no excess mortality for widows, and Mineau et al. (2002) found smaller effects for widows compared with widowers, with the results for women varying considerably across birth cohorts. Lillard and Waite (1995) found some excess mortality for women, but the results were sensitive to model specification. Data on couples from Northern California (Schaefer et al., 1995), Finland (Kaprio et al., 1987), and Israel (Manor and Eisenbach, 2003) found similar bereavement effects on mortality for males and females. There are several pathways through which widowhood can become an immediate health risk. Some suggest that excess mortality is generated by the emotional stress of the death of a loved one (Martikainen and Valkonen, 1996; Luoma and Pearson, 2002) or the emotional and physical stress of caring for the dying (Christakis and Iwashyna, 2003). Recent research by Wittstein et al. (2005) suggests that emotional stress can cause the overproduction of particular hormones that can cause a sudden life-threatening heart spasm in otherwise healthy people. Rosenbloom and Whittington (1993) found elderly widowed people suffer from poor nutrition right after the death of their spouse. The loss of a spouse may also reduce contact with established social network, which, given in the literature cited above, may generate poorer health outcomes. Finally, Iwashyna and Christakis (2003) found evidence suggesting that widowhood compromises the quality of medical care sought by the surviving spouse. 2.2. Behavioral correlations The accumulated empirical evidence is convincing that mortality rates of the bereft are higher than their married counterparts. Whether these events can be interpreted as causal relationships remains unclear. The results are potentially explained by marriage selection which has no causal interpretation. There are strong positive correlations between many characteristics of married couples including age, years of education, IQ, height, waist circumference, and even earlobe length (Spuhler, 1968; Vandenburg, 1972; Harrison et al., 1976; Mascie-Taylor and Gibson, 1979; Herbener, 1993; Caspi and Herbener, 1993; Murray, 2000). The similarity in the characteristics of spouses is also matched by a similarity in life events. Married couples that last past middle-age live through the same inter-temporal events, shocks, and income and consumption patterns. Such a convergence of lives can explain why there is evidence that widowed spouses die soon after the loss of their spouse: simply put, they started leading the same lives (Smith and Zick, 1994). Not surprisingly, there are also strong correlations in the investments that husbands and wives make concerning their health. Many married couples share a love of exercise, food, wine, cigarettes, or a sedentary lifestyle. In this section, we document this fact empirically with data for a sample of older married couples from the 1987 to 1990 National Health Interview Survey (NHIS).3 The sample includes information on white, non-Hispanic married people aged 50–70.4

3 “The NHIS is the principal source of information on the health of the civilian non-institutionalized population of the United States and is one of the major data collection programs of the National Center for Health Statistics (NCHS).” The survey contains data on roughly 60,000 households. More information about the survey is available at http://www.cdc.gov/nchs/about/major/nhis/hisdesc.htm (accessed March 25, 2008). 4 This sample roughly overlaps the time period and ages for the primary data set we use in later analyses.

J. Espinosa, W.N. Evans / Journal of Health Economics 27 (2008) 1326–1342

1329

Table 1 Conditional means of health behaviors, 1987–1990 NHIS and 1992/1993 CPS tobacco use supplement, white married males and females, 50–70 years of age Variable

Obese? Fair or poor health at survey? Bed days in past 12 months? Short-term hospital stay past 12 months? Current smoker?

Data set

NHIS NHIS NHIS NHIS CPS tobacco use supplements

Married women Pr(Xw = 1|Xh = 1)

Pr(Xw = 1|Xh = 0)

Difference

0.233 (0.008) [2870] 0.450 (0.008) [3998] 0.500 (0.006) [6724] 0.118 (0.006) [2618] 0.401 (0.008) [3509]

0.124 (0.002) [17,426] 0.097 (0.002) [16,298] 0.275 (0.004) [13,572] 0.091 (0.002) [17,678] 0.123 (0.003) [15,316]

0.109 (0.007) 0.353 (0.006) 0.225 (0.007) 0.027 (0.006) 0.277 (0.007)

Pr(Xh = 1|Xw = 1)

Pr(Xh = 1|Xw = 0)

Difference

0.258 (0.008) [3077] 0.507 (0.009) [3281] 0.482 (0.006) [8093] 0.143 (0.008) [1975] 0.459 (0.008) [4085]

0.133 (0.003) [19,094] 0.117 (0.002) [18,890] 0.255 (0.004) [14,078] 0.114 (0.002) [20,196] 0.147 (0.003) [17,433]

0.125 (0.007) 0.390 (0.007) 0.227 (0.006) 0.029 (0.008) 0.312 (0.007)

Married men

Obese? Fair or poor health at survey? Bed days in past 12 months? Short-term hospital stay past 12 months? Current smoker?

NHIS NHIS NHIS NHIS CPS tobacco use supplements

Table entries are fraction (standard error) [observations].

First, we consider whether knowing information about a husband’s health habits conveys any information about the wife’s health behavior. We look at four discrete outcomes: whether the person is obese (has body mass index > 30), whether they self-report fair or poor health, whether they had any bed rest days in the past 12 months, and whether they had a short-term hospital stay in the past 12 months.5 We calculate the conditional probability that a wife (husband) answers yes to each of these questions (Xw = 1) given that their husband (wife) answers yes or no (Xh = 1 or 0). The results from this exercise are reported in Table 1. Equality of the conditional probabilities means that the behaviors and outcomes for wives and husbands are statistically independent. In the married women sample, for each of the four variables, the differences in the conditional probabilities are large, and we easily reject the null hypothesis that the conditional probabilities are equal. Wives with obese husbands are twice as likely to report they are obese. Wives with husbands in fair or poor health are 4.5 times as likely to report they are in poor health compared with when their husband are not in fair or poor health. In the lower half of the table, we repeat the exercise for males. In each of these cases, we can reject the null of equality in the conditional means, and the relative differences and absolute differences in probabilities are similar to the sample of married women. The NHIS survey that we use in this analysis does not ask respondents whether they smoked. Supplemental surveys to the NHIS do ask about smoking habits, but these surveys are typically only administered to one person in the household. Therefore, we cannot conduct this same analysis with this key health behavior. In the last row of each panel in Table 1, we generate probability estimates for whether a married respondent currently smoked based on whether their spouse smokes from the September 1992, January 1993 and May 1993 Tobacco Use Supplements from the Current Population Survey (CPS).6 Both married women and married men are about three times as likely to report they smoke when their spouse smokes, compared with when their spouse does not smoke. 3. Empirical methodology Although there is convincing evidence that mortality is higher after the death of the spouse, the data in Table 1 does raise the possibility that this result is simply an omitted variables bias. In this section, we outline our empirical methodology designed to help distinguish whether the bereavement effect is marriage protection or marriage selection. We begin with a description of the data set used for this project. 3.1. Analysis sample and descriptive statistics The primary data for this analysis are public-use versions of the National Longitudinal Mortality Study (NLMS). The data are the product of a merging of person-level responses from the Census long-form data and the Current Population Survey (CPS)

5 Body mass index = 703 × weight (lb)/height2 (in2 ) and in the survey, height and weight are self-reported. On the questionnaire, self-reported health status is ascertained by asking “Would you say [your] health in general is excellent, very good, good, fair, or poor?”; bed rest days is determined from “During the past 12 months, about how many days did illness or injury keep [you] in bed more than half the day? (including days while an overnight patient)”; and short-term hospitals stays is generated from a series of questions including “How many nights [were you] in the hospital?”. 6 The CPS is a monthly household survey of roughly 60,000 households designed to measure labor market conditions for the non-institutionalized population. Respondents are in the CPS for 4 months, out for 8, then back in the survey for 4 months. Therefore, the three smoking supplements represent three unique samples and are designed to be pooled together.

1330

J. Espinosa, W.N. Evans / Journal of Health Economics 27 (2008) 1326–1342

to death certificate information from the National Death Index (NDI).7 We use a public-use version of the NLMS that contains data from five monthly CPS samples from 1979 through 1981. The CPS data provides information on household income, labor force status, individual education and other demographic variables; however, they do not provide any information on health status or health behaviors at the time of the survey. The endpoint to the mortality follow-up is fixed to nine years from entry into the CPS and the data indicates the days from the initial survey until death or the end of the follow-up, which ever comes first. Data from the NDI include information on the cause of death, which is recorded using the ICD-9 coding system.8 The NLMS contains CPS and NDI information for 637,162 individuals. The longitudinal component to this otherwise cross-sectional data set enables us to investigate how marital status impacts mortality. We focus our attention on married, white, non-Hispanic people between the ages of 50 and 70.9 Using household identification numbers and variables in the NLMS that indicate a respondent’s relationship to the head of household, we match the records of husbands and wives. Although spouses of 50–70 year olds are primarily within this age range, some are not, so there are different sized samples for “married men” and “married women”. Our two NLMS samples consist of 37,777 married men and 34,465 married women. Table 2 provides the means for several of the key variables of interest from our analysis sample. Not surprisingly, a smaller proportion of women die within nine years of being surveyed (9 percent) than do married men (18 percent). Likewise, 8 percent of married men become widowers after their initial interview, and 22 percent of married women lose their spouse. As a different exercise, we calculate mortality rates over a shorter period of time after the initial interview, and in Table 2, we also report the proportion of married men and married women that died within five years of entering the survey and the proportion that were widowed over the same period. The 5-year mortality and widowhood rates are about one-half the 9-year rates. Table 2 also provides means and standard deviations for the income, education, occupation10 and age variables for the NLMS data. In public-use version of the NLMS, family income is reported in $5000 intervals through $25,000, then a group from $25,000 to 50,000 and a final group with incomes in excess of $50,000.11 From more detailed education variables, we construct five dummy variables measuring education (no high school, some high school, high school graduate, some college education and a college graduate). We also construct 12 dummies for each 1-digit occupation code. The average married woman is about 3 years older than the average married male. The median male and female belongs to a family that has between $20,000 and 25,000 in income. About 36 percent of males have less than a high school education and less than 17 percent are college graduates. The corresponding numbers for females are 32 and 9 percent, respectively. Among males that identify an occupation, craftsman is the mode, but for women, the majority report not working or occupation is missing. Among working women, the most frequently reported occupation is clerical.

3.2. Baseline Cox proportional hazard models In our sample, an observation consists of a married couple that enters the NLMS between 1979 and 1981. The observation period begins with this entry point and extends nine years, or until death, which ever comes first. The data are therefore considered right-censored, because we do not observe events (i.e. widowhood and death) after the end of the mortality follow-up period. While there are several modeling techniques for the analysis of survival data, we follow previous work by employing a Cox proportional hazard model, which is a partial maximum likelihood estimation method (Cox, 1972). The Cox model begins with the assumption that the hazard for individual i can be written as hi (t) = 0 (t) exp(Xi ˇ),

(1)

where for simplicity, we assume Xi is a vector of time invariant characteristics. The model assumes that the hazard at time t for individual i is a function of the baseline hazard 0 (t), plus a function of observed characteristics. The baseline hazard is left unspecified but it is assumed to be constant across people, meaning that the proportional hazard for person i relative to person j for fixed time t is only a function of observed characteristics hi (t)/hj (t) = exp (Xi ˇ)/exp (Xj ˇ). To circumvent making an assumption about the baseline hazard, the Cox model specifies a partial likelihood, which can be described as follows. If we order the data from the shortest to longest spell t1 , t2 , . . ., tn , the conditional probability that person 1 dies at time period

7 For more information about the NLMS, please see Sorlie et al. (1995) and Rogot et al. (1992), and the NLMS web page http://www.census.gov/nlms/ (accessed March 25, 2008). 8 ICD-9 is the International Statistical Classification of Diseases and Related Health Problems, 9th revision. 9 We restrict our attention to whites because the statistical procedure outlined below to identify uninformative deaths did not identify any deaths categories as uninformative for minorities. This is due to two features of our procedure. First, we need to guard against Type II errors which required that we re-aggregate many death categories into large groups. Coupled with the small samples for black and other minorities, we could not isolate small enough death categories that could be used as uninformative deaths. 10 In the CPS, for persons who were employed at the time of the survey, the occupation variable relates to the job worked during the preceding week. For unemployed persons and those not currently in the labor force were to give their most recent occupation. 11 During the time period we are considering, the family income variable from the basic monthly CPS was a 14-level categorical variable with incomes greater than $75,000 being the top group. The public-use version of the data presents a less detailed version of family income to maintain confidentiality.

J. Espinosa, W.N. Evans / Journal of Health Economics 27 (2008) 1326–1342

1331

Table 2 Sample characteristics, white married males and females, 50–70 years of age, public-use NLMS Variable

Married males

Married females

Died Within 9 years of survey Within 5 years of survey

0.178 0.091

0.093 0.044

Become widow/widower Within 9 years of survey Within 5 years of survey

0.066 0.032

0.206 0.112

Family income <$5K ≥$5K, <$10K ≥$10K, <$15K ≥$15K, <$20K ≥$20K, <$25K ≥$25K, <$50K ≥$50K

0.041 0.132 0.169 0.142 0.154 0.285 0.076

0.052 0.166 0.186 0.138 0.143 0.248 0.067

Education
0.206 0.161 0.343 0.124 0.165

0.158 0.160 0.463 0.125 0.094

Age 50–59 60–70

0.544 0.456

0.572 0.428

<0.001 0.047 0.026 0.062 0.036 0.167 0.046 0.047 0.148 0.106 0.052 0.263

0.006 0.061 0.004 0.044 0.002 0.008 0.133 0.035 0.039 0.055 0.014 0.598

Occupation Home maker Service Laborer Operaters Operate equipment Craftsmen Clerical Sales Managers Professionals Farming Missing/never worked Individuals

37,777

34,465

Table entries are the sample fraction within each category.

t1 , given that anyone could have died at that time, is  (t1 ) exp(X1 ˇ)

0

 (t) exp(Xi ˇ) i 0

=

exp(X1 ˇ)



i

exp(Xi ˇ)

.

(2)

By definition, the baseline hazard drops out and the term on the right in Eq. (2) represents the partial likelihood for person 1. The model is partial likelihood because the estimation does not exploit all the information in the data, namely, the baseline hazard 0 (t). As a result, the estimates for ˇ are consistent but not efficient. However, the benefit of the procedure is that the researcher does not have to specify the form for the baseline hazard 0 (t). The Cox model outlined in (1) and (2) is easily adapted to include incomplete spells and time-varying covariates. In the study, we are primarily interested in a set of time-variant variables that indicate the period after respondent’s spouse has died. Above, we raise the suspicion that unobserved or omitted variables are potentially biasing single-equation estimates by illustrating the strong correlation in behavior between spouses using data not available in the NLMS. In an additional effort to support these concerns, we run a series of Cox proportional hazard models that begin with only the widowhood variable and progressively add more covariates. As we add more variables, if the estimated bereavement effect declines, this signals that widowhood is correlated with observed characteristics. Since we suspect, to some degree, that observed and unobserved characteristics are positively correlated, this would also indicate a concern that the bereavement effects are capturing unobserved factors as well in the specifications without many covariates. The final results in this table will also indicate a baseline specification to which we will compare our eventual models that allow for heterogeneity in the bereavement effect based on cause of death. The results from these models are reported in Table 3. In specification 1, the only covariate is the time-varying indicator for when the spouse dies (widow) and we progressively add to the model four sets of time-invariant variables measured at baselines: 4 indicators for level of education (specification 2), 6 indicators for family income (specification 3), indicators for

1332

J. Espinosa, W.N. Evans / Journal of Health Economics 27 (2008) 1326–1342

Table 3 Partial maximum likelihood estimates, Cox proportional hazard models, death during follow-up, white married males and females, 50–70 years of age, public-use NLMS Model (1) Married males, aged 50–70 (n = 37,777) Widow 1.787 (0.096) Include Education Income Occupation Age Married females, aged 50–70 (n = 34,465) Widow 1.617 (0.078) Include Education Income Occupation Age

Model (2)

Model (3)

Model (4)

Model (5)

1.702 (0.091)

1.542 (0.083)

1.439 (0.077)

1.316 (0.071)

Yes

Yes Yes

Yes Yes Yes

Yes Yes Yes Yes

1.521 (0.074)

1.361 (0.067)

1.346 (0.066)

1.196 (0.059)

Yes

Yes Yes

Yes Yes Yes

Yes Yes Yes Yes

Table entries are hazard ratios (standard errors) from Cox proportional hazard models. There are up to 6 income, 4 education, 10 or 11 occupation, and 21 age dummy variable included, depending on the specification.

occupation (specification 4), and 20 indicators for age (specification 5). There are two panels to Table 3—in the top panel are the Cox regression results for married men and in the bottom panel are the results for married women. Since all of the covariates are dummy variables, we only report the hazard ratio. Of course, we expect the bereavement effect estimate to begin to fall as more variables are introduced to explain the excess mortality of widows over people still married to their spouses. The results in Table 3 show for men and women that the magnitude of the bereavement effect drops by approximately 50 percent as we move from the parsimonious to the fully specified model. For men, the hazard ratio begins at 1.787 and falls to 1.316, and the fall is from 1.617 to 1.196 for women. The findings provide suggestive evidence that widowhood is strongly correlated with observed characteristics, and since we suspect correlation between observed and unobserved factors, these results provide suggestive evidence that marriage selection explains part of the temporal correlation in spousal mortality. The results in the last column of Table 3 are comparable to other studies and provide the launching board for the next section. The point to take away is that after a wife dies the husband has a 32 percent greater risk of dying (hazard ratio (h.r.) = 1.32), and after a husband dies, a wife has a 20 percent greater hazard (h.r. = 1.20). Both of these results are statistically distinguishable from the null hypothesis that the h.r. = 1 at conventional levels. 3.3. Informative and uninformative causes of death In an effort to isolate whether the bereavement effect estimated in Table 3 represents marriage selection or protection, we exploit the information contained in the heterogeneity in the bereavement across certain causes of death. If a cause of death is democratic in its affliction, then the realization of the disease and the subsequent death of a spouse reveals less information regarding the health of the surviving spouse; therefore, in these special cases, we are less concerned that the bereavement effect is driven by omitted variables that measure individual and couple behavior. In what follows, we provide evidence that certain causes of death (COD) are uncorrelated with our measures of SES: income, education and occupation at the time of survey. This evidence leads to our claim that these COD are not explained well by the theory of assortative mating. Our measures of SES are shared by the bulk of research that has examined the SES/mortality gradient for all-cause and cause-specific mortality (Berkman and Kawachi, 2000). The genesis for much of the work in social sciences is the research of Kitagawa and Hauser (1973) who matched survey data from the 1960 Census long-form, conducted in April of 1960, to death records from the May–October 1960 period. The stylized facts from their work are that mortality rates decline with income and education but at a decreasing rate. This relationship is present for all age groups, but Kitagawa and Hauser find less variation in mortality across socioeconomic groups for the elderly, a result that has been replicated in more recent surveys (Hurd et al., 1999; Snyder and Evans, 2006). With the NLMS data, we use detailed 3-digit ICD-9 cause of death information to construct a series of dummy variables to indicate whether a person died in the 9-year follow-up from a particular cause. For example, a dichotomous variable for the cause of death “ischemic heart disease” takes on the value of 1 for an individual who dies of ischemic heart disease. For both married men and women, we then run a series of logistic regressions that identify whether the death from a particular cause is predicted by the income, occupation and education of the respondent—our SES measures, controlling for a cubic term in age. For each regression, we conduct four −2 log likelihood tests: whether the income dummies are jointly zero, whether the education dummies are jointly zero, whether the occupation dummies are jointly zero, and whether the income, education and occupation dummies are jointly zero. If we can reject any of the four null hypotheses, then we consider the cause to be

J. Espinosa, W.N. Evans / Journal of Health Economics 27 (2008) 1326–1342

1333

Table 4 Uninformative causes of death (UCOD) for white married males and females, 50–70 years of age, public-use NLMS Cause of death (first 3 digits of ICD-9) Men Pancreatic cancer (157) Other cancers (170–239, various)a Genital cancer (179–187) Lymphoma, leukemia (200–208) Cardiomyopathy (425) Pneumonia (480–487) Non-mv accidents, murder/suicide (800–807; 826–999) Total # of deaths % of all deaths % of sample Women Other cancers (140–239)b Breast cancer (174) Genitourinary cancer (179–189) General circulatory disease (390–459)c Accidents (800–999) Total # of deaths % of all deaths % of sample

Number 107 278 212 203 65 149 174

Proportion of UCOD (%) 9 23 18 17 6 13 14

1188 15.7 3.4 301 216 149 241 79

30 22 15 24 8

986 34.1 2.6

a These diseases include: neoplasms of bone, connective tissue and skin (ICD-9 codes 170–175); neoplasms of unspecified sites (190–199); benign neoplasms and carcinoma in situ (210–239). b ICD 9 codes included are: 140–152;154–156; 157–161;163–173;175–178; 190–199. c These diseases include all general circulatory diseases except the following: acute myocardial infarction (410), other forms of chronic ischemic heart disease (414), other forms of heart disease (420–429), intracerebral hemorrhage (431), other and unspecified intracranial hemorrhage (432), and occlusion of cerebral arteries (434).

an informative-COD group (ICOD). Likewise, if we cannot reject all four hypothesis tests, we consider the cause of death to be uninformative (UCOD). We are careful to not disaggregate death causes into too small of cells so that we have a high Type II error rate for the −2 log likelihood tests. To determine the minimum mortality rate that a death category must have, we first run logit models for all 3-digit death categories. We then find the death category with the smallest mean where we can reject one of the four tests outlined above.12 If a 3-digit death category falls below this minimum threshold, we aggregate it with other deaths from the same 2-digit category so the pooled deaths have a mean above the minimum threshold.13 In Table 4, we report the causes of death that fall into the uninformative categories for both men and women. The list includes the number of deaths attributed to each cause, as well as each cause’s proportion of total UCOD deaths. We find that 15 percent of deaths for men married to women aged 50–70 and 34 percent of deaths for women married to men aged 50–70 fall into causes that are uncorrelated with observed characteristics. We test the validity of our identification strategy using an analysis sample of women married to men aged 50–70 and men married to women aged 50–70 and extracted from the NLMS. For each sex, we run two logit models. In one specification, the dependent variable is the probability the individual dies of an UCOD within the sample time frame (the entire 9-year follow-up), and for the other specification, the dependent variable is the probability of dying of an ICOD under the same period. The other explanatory variables consist of a full set of dummies for income, education, occupation and age groups. The marginal effects from these logit models are listed in Table 5. We find an inverse, monotonic relationship between mortality and the SES variables for income and education when we model the probability of dying from an ICOD. Among the occupation dummy variables, 5 of 10 are statistically significant at p < 0.05 in the male ICOD model and 2 of 11 are statistically significant for women in the ICOD model. This is not the case for the group of UCOD. Here, none of the 21 coefficients in the two UCOD models are statistically significant. The marginal effects we find in the UCOD models are all very small, and in every case but one (income less than $5000 in the male UCOD model), we cannot reject the null hypothesis that the individual coefficient estimates are equal to zero. The p-values for the various hypothesis tests are also listed in Table 5. For both men and women in the NLMS sample, we easily reject the null that the ICODs are not correlated with income, education and occupation. All p-values for this outcome are <0.0001. In contrast to the results for ICODs, the results for UCODs differ between men and women. For the wives of

12 For men married to women aged 50–70, malignant neoplasms of rectum, rectosigmoid junction, and anus is the 3-digit death category with the minimum mortality rate (37 deaths) and a p-value of 0.04 for the hypothesis test that all the education variables are equal to zero. For women married to men aged 50–70, diabetes mellitus is the 3-digit death category with the minimum mortality rate (54 deaths) and a p-value of 0.05 for the hypothesis test that all the income variables are equal to zero. 13 When we re-group into large categories, we use the ICD-9 72-category grouping as a guide.

1334

J. Espinosa, W.N. Evans / Journal of Health Economics 27 (2008) 1326–1342

Table 5 Maximum likelihood estimates of logistic regression, death from informative or uninformative causes in 9-years after initial interview white married males and females, 50–70 years of age, public-use NLMS Covariates

Males

Females

Observations: die within 9 years of the survey from an

Observations: die within 9 years of the survey from an

Informative death

Uninformative death

Informative death

Uninformative death

Income <$5K $5K ≤ Income < $10K $10K ≤ Income < $15K $15K ≤ Income < $20K $20K ≤ Income < $25K $25K ≤ Income < $50K No high school Some high school High school graduate Some college Occupation: never work, missing value, etc. Farming Household Service Laborers Operatives, not transport Operatives, equipment operative Craftsmen Clerical Sales Managers

0.031 (0.006) 0.025 (0.005) 0.017 (0.004) 0.011 (0.004) 0.011 (0.004) 0.006 (0.003) 0.016 (0.003) 0.015 (0.003) 0.008 (0.002) 0.007 (0.003) 0.025 (0.005) −0.008 (0.003) NA 0.008 (0.004) 0.002 (0.004) 0.004 (0.004) −0.002 (0.003) 0.002 (0.003) 0.011 (0.004) 0.009 (0.004) 0.005 (0.003)

0.006 (0.003) 0.002 (0.002) 0.002 (0.002) 0.002 (0.002) 0.002 (0.002) 0.001 (0.002) −0.001 (0.001) −0.001 (0.001) −0.001 (0.001) 0.001 (0.001) 0.003 (0.002) −0.001 (0.002) NA 0.004 (0.003) 0.004 (0.003) 0.001 (0.003) −0.001 (0.002) −0.001 (0.002) 0.001 (0.002) −0.001 (0.002) −0.001 (0.002)

0.006 (0.002) 0.004 (0.002) 0.003 (0.001) 0.002 (0.001) 0.001 (0.001) 0.001 (0.001) 0.004 (0.002) 0.003 (0.001) 0.002 (0.001) 0.001 (0.001) 0.009 (0.002) −0.005 (0.002) 0.001 (0.004) 0.002 (0.002) −0.002 (0.004) −0.002 (0.005) 0.003 (0.002) −0.001 (0.003) 0.002 (0.002) 0.001 (0.002) 0.002 (0.002)

−0.003 (0.002) 0.002 (0.002) 0.002 (0.002) −0.001 (0.002) 0.001 (0.002) −0.001 (0.002) 0.002 (0.002) 0.002 (0.002) <0.001 (0.002) <0.001 (0.002) 0.001 (0.002) −0.003 (0.004) −0.001 (0.005) < 0.001 (0.003) −0.001 (0.006) −0.002 (0.008) −0.002 (0.003) −0.007 (0.004) −0.001 (0.002) −0.001 (0.003) 0.001 (0.003)

p-Values on −2 log likelihood test statistics Education coefficients are all 0 Income coefficients are all 0 Occupation coefficients are 0 Education, income and occupation coefficients are all 0

<0.0001 <0.0001 <0.0001 <0.0001

0.542 0.168 0.036 0.014

0.005 <0.0001 <0.0001 <0.0001

0.502 0.071 0.690 0.203

Mean of outcome −2 log likelihood

0.185 30,071.31

0.034 59,960.58

0.051 13,987.50

0.026 8,871.83

Table entries are marginal effects (standard errors) from a logistic regression. The reference person is someone married to a 50–70-year-old white nonHispanic woman or man, and who is 50-year-old or younger, has a college degree, $50K or more in family income, and is/was employed in a professional occupation. Other covariates (not reported) include dummy variables for each age 51 through 70, plus a dummy for people 71 and greater in age.

men aged 50–70, we are able to support our assertion that income, occupation and education are uninformative because the p-value for the hypothesis test is 0.203. For the husbands of women aged 50–70, we are not able to reject the hypothesis test on the education parameters, p-value = 0.542, but we have weaker statistics for the other tests. While this suggests some correlation between SES and the probability of dying from an UCOD, the point estimates and marginal effects of the income variables are very small and close to negligible in comparison with the magnitudes of the point estimates and marginal effects under the ICOD specification. Another point of interest is whether a husband’s (wife’s) death from an UCOD or ICOD is correlated with the surviving wife’s (husband’s) type of death. If the probability of dying from an UCOD is significantly greater if the spouse died of an UCOD rather than of an ICOD, then we would argue that our classification system is incorrectly producing diseases that do reveal information regarding the mortality of the surviving spouse. In Table 6 we investigate this issue by comparing the probability of a widow dying from an UCOD (ICOD) given the spouse died of an UCOD or an ICOD. Like Table 1, we compare the conditional probabilities by taking the difference of these two probabilities and check whether the difference is statistically different than zero. As the table shows, given that a spouse dies Table 6 Conditional means of death by type of cause, white married males and females, aged 50–70, public-use NLMS Unconditional

Conditional on Spouse dies (1)

Spouse died of UCOD (2)

Spouse died of ICOD (3)

|Difference| |(3) – (2)|

Males Pr[Die of UCOD by end of follow-up] Pr[Die of ICOD by end of follow-up]

0.028 (<0.001) 0.149 (0.002)

0.028 (0.003) 0.124 (0.007)

0.030 (0.006) 0.115 (0.011)

0.027 (0.004) 0.129 (0.008)

0.003 (0.007) 0.014 (0.014)

Females Pr[Die of UCOD by end of follow-up] Pr[Die of ICOD by end of follow-up]

0.031 (<0.001) 0.062 (0.001)

0.023 (0.002) 0.054 (0.053)

0.018 (0.004) 0.044 (0.006)

0.022 (0.002) 0.057 (0.003)

0.004 (0.005) 0.014 (0.008)

Table entries are mean (standard error).

J. Espinosa, W.N. Evans / Journal of Health Economics 27 (2008) 1326–1342

1335

of an UCOD or ICOD, we are unable to reject the hypothesis that the probability the surviving spouse dies from an UCOD is different from the probability the surviving spouse dies of an ICOD. 4. Results 4.1. Bereavement effects for informative and uninformative deaths In the last column of Table 3 we report the bereavement effect results from a Cox hazard model that includes as covariates a complete set of fixed effects for age, income, education and occupation. Now we estimate two additional models where we replace the time-variant widowhood variable with two time-variant variables that distinguish between types of widows: widow by UCOD or ICOD, and widow by unpreventable or preventable cause of death. The UCOD (ICOD) variable has a value of 1 for the day, and each subsequent day, an individual becomes a widow by an UCOD (ICOD). In each of these models we include mutually exclusive set of age dummy variables plus we control for SES in these specifications by including timeinvariant variables for income, education, age and occupation. We also test the null hypothesis that the bereavement effect from an ICOD (preventable COD) is equal to the effect from an UCOD (unpreventable COD). The results for various Cox proportional hazard models are reported in Table 7. In the first two models, we present hazard ratio estimates for men, and in the final two models, we report the estimates for women. We also report the standard error for the various covariates that measure widow/widower status. Model (1) estimates are repeated from Model (5) of Table 3 and these numbers show that widowers have a 1.32 hazard ratio or widowers have 32 percent higher hazard of exiting life excess mortality compared to married men, and the bereavement effects for widows are smaller at 20 percent. For both of these models, we can reject the null that the hazards equal one at conventional levels. The second column in each group lists the Cox proportional hazard model results when we split the widowhood variable into UCOD and ICOD. For men, the widowhood effect is nearly the same magnitude regardless of whether the spouse dies of an informative or uninformative cause of death. For both groups of death causes, we can reject the null that the hazard ratios equal one and the p-value on the −2 log likelihood test that the coefficients are identical is very high (p-value = 0.79), so we cannot reject the null hypothesis that the two estimated bereavement effects are equivalent. The similarity in results suggests that the omitted variable bias is not large, and marital selection is not a major force in producing the observed bereavement effect for men. Thus, the loss of marital protection may indeed be the culprit in a widower’s demise. For women in the NLMS, however, we find a different story. After breaking up the time-variant variable by cause of death, we find that widows of UCOD spouses have a bereavement effect (as measured by the hazard ratio) that is not statistically different from one. On the other hand, widows of ICOD spouses have a hazard ratio that is statistically distinguishable from 1 (p-value < 0.001). One must be careful in interpreting these results. The substantial difference in the hazard ratios for ICOD and UCOD deaths raises some suspicion of an omitted variable bias in the bereavement effect for women. However, the large standard errors on the hazard for the UCOD variable make it such that we cannot reject the null hypothesis that the ICOD/UCOD hazard ratios are the same (p-value = 0.22). These results are suggestive of an omitted variables bias but not definitive. The ability to differentiate between causation and correlation based on the results in Table 7 is a function of our ability to isolate those causes that are essentially random in the population. While we have followed a procedure to isolate causes of death that are uncorrelated with SES characteristics, there is still some chance that the list of diseases in the UCOD category are somehow reflective of the surviving spouse. This is a much bigger concern for the married males sample because the likely outcome of including informative causes of death in the UCOD group would be to drive that coefficient closer to the ICOD value. The lack of any noticeable bereavement effect for UCOD in the married females sample is reassuring. In contrast, some causes of death grouped into the UCOD category are potentially ICODs in the sense that anecdotal evidence and research point to individual behavior as an important component in explaining the likelihood of death. For example, “sun-tanning” is a behavior which raises the chances of developing skin cancer (Holly et al., 1995; Kricker et al., 1991). Researchers have found that cigarette smoking is a risk factor for pneumonia (LaCroix et al., 1989), acute myocardial infarction (Croft and Hannaford, Table 7 Partial maximum likelihood estimates, Cox proportional hazard models, death during follow-up, white married males and females, 50–70 years of age, public-use NLMS Covariate

Married males Model (1)

Widower Widower due to uninformative-COD Widower due to informative-COD

1.316 (0.071)

−2 log likelihood p-Value, −2 log likelihood test

129,943.05

Married females Model (2)

Model (1)

Model (2)

1.196 (0.059) 1.360 (0.121) 1.293 (0.085) 129,942.84 0.647

1.041 (0.129) 1.216 (0.063) 63,773.35

63,771.79 0.212

Table entries are hazard ratios (standard errors) from Cox proportional hazard models. The −2 log likelihood tests test for the equality of the coefficients in the ICOD and UCOD models. Other covariates (not reported) include a complete set of dummy variables for income, education, occupation and age.

1336

J. Espinosa, W.N. Evans / Journal of Health Economics 27 (2008) 1326–1342

1989; Wilhelmsen et al., 1984), and cancer of several types including pancreas (Tai et al., 2007), lip and oral cavity (Macfarlane et al., 1995), nasal cavity (Brinton et al., 1984), pharynx and larynx (Franceschi et al., 1990), bladder (Taylor et al., 1998), and stomach (Brown and Devesa, 2002). Alcohol is a risk factor for all types of accidents (Thun et al., 1997), cardiomyopathy (Klatsky et al., 1990; Afonso et al., 2007), acute myocardial infarction, and cancers of the stomach, esophagus, small intestine, and duodenum (Brown and Devesa, 2002; Hansson et al., 1994). Obesity has been linked to acute myorcardial infarction (Yusuf et al., 2005), uterine cancer (Austin et al., 1991), kidney cancer (Key et al., 2002), and breast cancer (Morimoto et al., 2002; van den Brandt et al., 2000), as well as several other diseases already listed. Some of the causes of death grouped into the UCOD category, though, are not rooted in individual behavior. In these cases, genetics, radiation, or other individual morbidity are more often cited as the risk factors (Fentiman et al., 2006; Velicer et al., 2006; Smith et al., 2001, 2003; Adami et al., 1998). In theory, we could define UCOD as deaths that are uncorrelated with SES characteristics and behaviors. However, the public-use NLMS does not have any control variables that measure health status or health behavior. Therefore, we cannot produce a more refined group of UCOD diseases. We can however, test the sensitivity of the hazard ratios to changes in the set of UCOD. Using similar duration models for both men and women and given the full set of UCOD as before, we take one cause of death category out of the UCOD grouping and place it in the ICOD grouping. We then re-estimate the basic model from Model (2) of Table 7. Following this, we return the cause of death to the UCOD grouping, select another cause of death to transfer from the UCOD to ICOD group and follow the same procedure as above. We do this for each cause of death category in the UCOD groupings for men and women. The results of this exercise provide a distribution of hazard ratios for men and women. For widowers, the estimated hazard ratios range from 1.30 to 1.43 depending on which cause of death is removed from the set of UCOD. For each specification, we cannot reject the null hypothesis that the UCOD and ICOD hazard ratios are equal which supports our finding of a bereavement effect among men. The distribution of hazard ratios is narrow and centered about the original UCOD estimate (1.36). For widows, the story is slightly different as the UCOD hazard ratios are more dispersed, ranging from 0.87 to 1.16. We cannot reject the null hypothesis that the hazard ratios for ICOD (1.22) and UCOD (1.04) are equal under the original specification (p-value = 0.218). However, when we remove the “other accidents” category from the UCOD grouping, we find stronger evidence for a significant difference in the estimates (p-value = 0.104) and more evidence to suggest that women do not experience a bereavement effect. We also find for widows that when pneumonia is removed from the UCOD grouping, the hazard ratio jumps to 1.16 which is very close to the ICOD estimate (h.r. = 1.20). Lastly, ignoring the two outliers (h.r. = 0.87 and 1.16), the distribution narrows to hazard ratios ranging from 0.94 to 1.06. 4.2. An alternative classification: preventable and unpreventable causes of death The method we use to identify UCOD is a statistical exercise and does not rely on epidemiological research; therefore, we borrow from this line of literature to develop an additional layer of analysis that parallels the models above. The results of this exercise provide us a method to verify the results of the UCOD/ICOD specifications. Generally, causes of death vary for many reasons, but one important dimension by which they differ is in their preventability. For example, liver cirrhosis is often the result of alcohol abuse, and because there are means to alter this behavior, the disease can be considered preventable. On the other hand, the causes of multiple sclerosis are to date largely unknown; thus, this disease can be considered unpreventable. Phelan et al. (2004) developed a preventability scoring index for causes of death that ranges from 1.0 for a cause that is almost entirely unpreventable (e.g., gall bladder cancer) to 5.0 for a cause that is almost entirely preventable (e.g., accidental poisoning). Two of the authors, both of whom are physicians and epidemiologists, independently rated the causes of death “in terms of the degree to which death was amenable to prevention or delay during the 1980s in the United States” (p. 272). Prevention can be due to either intervention by the person (abstaining from heavy drinking) or by the medical community via timely diagnosis and treatment. Using the preventability scoring, we develop a set of causes of death that cannot be explained by individual tastes for preventative behavior. For our purposes, the set of unpreventable causes are those diseases with a score of 2.0 or lower according to the preventability scoring index.14,15 In Table 8, we list the unpreventable diseases from Phelan et al., the ICD-9 code, and the number of deaths over the 9-year follow-up in both the white married male and female subsamples. The causes of death listed in the table that are in bold-faced type are those that appear as uninformative deaths in either the male or female samples listed above. There are some notable overlaps between the two lists but in general, most of the death categories that Phelan et al. list as preventable are infrequent events, and based on our grouping procedure to prevent Type II errors outlined above, we would not have run individual logit models to determine informative/uninformative status for these particular causes. Note also that the

14 The preventability index in Phelan et al. is based on deaths in the entire population. In comparison, our sample of interest is married whites aged 50–70. The list in Phelan et al. may be larger or smaller if one considers this more restrictive group. 15 We conduct supplementary analysis and allow the constraint on the preventability score to vary. We only consider expansions of the group of preventable causes: diseases with (1) scores at or below 2.5 and (2) scores at or below 3.0. The hazard ratio for UCOD widowers remains stable as the preventability constraint is relaxed, however, the hazard ratio for UCOD widows increases from 1.05 (preventability score (p.s.) ≤ 2.0) to 1.20 and 1.26 (p.s. ≤ 2.5 and 3.0, respectively), which is what we would expect if we start to add more suspect causes into the UCOD group.

J. Espinosa, W.N. Evans / Journal of Health Economics 27 (2008) 1326–1342

1337

Table 8 Preventable and unpreventable causes of death from Phelan et al. (2004), and frequency of events in white married males and females, 50–70 years of age, public-use NLMS Disease (ICD-9 code)

# of deaths in sample (married whites, aged 50–70) Males

Malignant neoplasm of gallbladder/extrhepatic bile ducts (156) Multiple sclerosis (340) Anterior horn cell disease (335) Cardiomyopathy (425) Disorder of lipid plasma protein metabolism (272) Leukemia of unspecified cell type (208) Lymphosarcoma and reticulosarcoma, (200) Malignant neoplasm of brain (191) Malignant neoplasm of ovary and other uterine adnexa (183) Malignant neoplasm of if pancreas (157) Multiple myeloma and immunoproliferative neoplasms (203) Myeloid leukemia (205) Myoneural disorders (358) Muscular dystrophies and other myopathies (359) Polyartheritis nodosa and allied conditions (446)

13 3 17 65 2 20 17 51 0 107 33 39 2 3 1

Total number of deaths % of all deaths % of sample

373 4.9 1.1

Females 8 10 7 20 4 12 8 33 67 60 27 21 0 2 1 280 9.7 0.74

unpreventable causes of death are decidedly less frequent than uninformative cases of death, with only 1 percent of males and 0.74 percent of females dying from these causes. Preventability is close in spirit to the idea of an informative cause of death in that a preventable cause of death likely provides information regarding the mortality of the surviving spouse. Alternatively, an unpreventable cause of death would not reveal information regarding the mortality of the surviving spouse because of its nature. As a check to this supposition, we estimate models similar to those in Table 5. Specifically, for males and females, we estimate two logit models. In the first, the outcome of interest is an indicator that equals one if the respondent died during the nine years of follow-up from a preventable cause of death. In the second, the outcome variable equals one if the respondent died from an unpreventable cause. In these models, we include full sets of dummy variables for age, education, income and occupation. The marginal effects for income and education are reported in Table 9 and to conserve space, we suppress the reporting for the age and occupation variables. As in Table 5, the marginal effects in both the male and female preventable death logits show a pronounced income and education gradient. The p-values for the tests that the income, education, occupation are all zero and incredibly small. In contrast, none of the income, education or occupation marginal effects are statistically significant in the logits for death Table 9 Maximum likelihood estimates of logistic regression, death from preventable or unpreventable causes in 9-years after initial interview white married males and females, aged 50–70, public-use NLMS

Income < $5K $5K ≤ Income < $10K $10K ≤ Income < $15K $15K ≤ Income < $20K $20K ≤ Income < $25K $25K ≤ Income < $50K No high school Some high school High school graduate Some college p-Values on −2 log likelihood test statistics Education coefficients are 0 Income coefficients are all 0 Occupation coefficients are all 0 Education, income and occupation coefficients are all 0 Mean of outcome

Males Died by the end of the follow-up from an

Females Die by the end of the follow-up from an

Preventable death

Unpreventable death

Preventable death

Unpreventable death

0.0021 (0.0017) 0.0009 (0.0010) 0.0009 (0.001) 0.0012 (0.0011) 0.0007 (0.0008) −0.0004 (0.0005) −0.0004 (0.0005) −0.0004 (0.0005) −0.0003 (0.0005) −0.0005 (0.0006)

0.0051 (0.0027) 0.0066 (0.0023) 0.0045 (0.0021) 0.0016 (0.0019) 0.0010 (0.0019) −0.0006 (0.0017) 0.0064 (0.0023) 0.0044 (0.0021) 0.0027 (0.0017) 0.0018 (0.0018)

0.0011 (0.0013) 0.0005 (0.0009) 0.0009 (0.0010) 0.0002 (0.0009) 0.0016 (0.0011) 0.0012 (0.0009) 0.0010 (0.0010) 0.0007 (0.0009) 0.0000 (0.0006) −0.0001 (0.0007)

0.0361 (0.0066) 0.0275 (0.0051) 0.0188 (0.0043) 0.0122 (0.0039) 0.0126 (0.0039) 0.0058 (0.0033) 0.0164 (0.0036) 0.0156 (0.0036) 0.0080 (0.0026) 0.0079 (0.0030) <0.0001 <0.0001 <0.0001 <0.0001

0.616 0.362 0.636 0.709

0.005 <0.0001 0.0059 <0.0001

0.465 0.279 0.799 0.650

0.185

0.034

0.051

0.026

Table entries are marginal effects (standard errors) from a logistic regression. Other covariates include dummy variables for each age 51 through 70 and a dummy for ages 71 and over. The reference person is someone married to a 50–70-year-old white non-Hispanic woman or man, and who is 50 years old or younger, has a college degree, $50K or more in family income, and is/was employed in a professional occupation.

1338

J. Espinosa, W.N. Evans / Journal of Health Economics 27 (2008) 1326–1342

Table 10 Partial maximum likelihood estimates, Cox proportional hazard models, death during follow-up, white married males and females, aged 50–70, public-use NLMS Covariate

Married males Model (1)

Widower Widower due to unpreventable COD Widower due to preventable COD Widow due to either uninformative or unpreventable COD Widow due to other cause of death

1.316 (0.071)

−2 log likelihood p-Value, −2 log likelihood test

129,943.1

Married females Model (2)

Model (3)

Model (1)

Model (2)

Model (3)

1.196 (0.059) 1.377 (0.237) 1.310 (0.074)

1.058 (0.232) 1.202 (0.060) 1.311 (0.114) 1.318 (0.088)

129,943.0 0.777

129,943.1 0.956

1.064 (0.128) 1.219 (0.063) 63,773.4

63,773.0 0.560

63,772.1 0.281

Table entries are hazard ratios (standard errors) from Cox proportional hazard models. The −2 log likelihood tests test for the equality of the preventable and unpreventable CODs in Model (2), or the equality of the coefficients on uninformative or unpreventable and other causes in Model (3). Other covariates include a complete set of dummy variables for income, education, occupation and age.

by an unpreventable cause of death. At the bottom of the table, the p-values for the joint tests that the income, education, occupation and all three sets are jointly zero are all very high. Much like uninformative causes of death, unpreventable causes of death should provide little information about the surviving spouse. As in Table 7, we estimate Cox proportional hazard models for both sexes and allow the bereavement effect to vary between preventable and unpreventable causes of death. For the purpose of verification, the results from these specifications should produce similar inferences to the earlier results, even though the set of causes of death differ. In Table 10, we report estimates of the bereavement effect where the coefficients are allowed to vary by preventability of the spouse’s death (Model 2). The results for this analysis tell a very similar story to the estimates for informative/uninformative causes. For men, the hazard ratio for the widowers of spouses who die of unpreventable causes of death (h.r. = 1.377) is not statistically different than the ratio for the widowers of spouses who die of more preventable causes (h.r. = 1.377) and both hazards are statistically distinguishable from 1. For women, we find suggestive evidence of a selection bias in that the hazard ratios for spouse’s deaths by a preventable cause are large and statistically different from 1, but the hazard for a spouse’s unpreventable death is small. Like the previous column, however, we cannot reject the null hypothesis that the two hazard ratios are equal. Not surprisingly, given the smaller number of unpreventable deaths compared to uninformative deaths, the standard error on the hazard for the spouse’s death from an unpreventable cause has doubled compared to the similar estimate from column (2). In an attempt to reduce the standard error on the cause of death that provides little information about the surviving spouse, in Model (3) of Table 10, we allow the bereavement effect to vary based on whether the deaths are either uninformative or unpreventable, versus all other causes of death. The results from this exercise are identical to the results from the previous models: the bereavement effect for males is the same regardless of the cause of death but for women, the uninformative causes of death show little bereavement effect but large standard errors. 4.3. Does the bereavement effect decline over time? A number of authors have examined whether the bereavement effect varies with the length of time since the death of a spouse and most of the evidence suggests that mortality is greatest right after the death of a spouse (Lichtenstein et al., 1998; Manor and Eisenbach, 2003). After controlling for a variety of factors, Schaefer et al. (1995) and Kaprio et al. (1987), for example, found mortality rates double for the surviving spouse in the first year after the death of their spouse. The evidence does however show some heterogeneity across samples. Johnson et al. (1976) found higher mortality rates for widowed males and females, but the impact drops off sharply as the time since widowhood increases for females. In contrast, the authors show a smaller dropout in the impact of bereavement on mortality for widowed males. In this section, we examine this same question with our analysis samples allowing for two alterations to the basic model. First, we generate a baseline specification where we reproduce some of the models from the previous literature and allow the bereavement effect to vary based on the time since the death of the spouse (Model 1). In this case, we consider bereavement effects that are within 6 months of the death of the spouse, 6–24 months, and greater than 24 months. Next, we allow these three estimates to vary based on whether the death was an uninformative or informative cause of death (Model 2). The hazard ratios from these models are reported in Table 11. In Model (1), we reproduce some of the results from the literature. For males, the bereavement effect does not dissipate as we lengthen the time from the death of the spouse. In all cases for males, the hazard ratio is roughly 1.30 and for the coefficients that are between 6 and 24 months and greater than 24 months, we can easily reject the null at conventional levels that the hazard ratio is 1. The estimated bereavement effect for males are so close that the hazard ratio for males is within one one-hundredth of a point for deaths within 6 months and deaths greater than 24 months. The p-value that the coefficients are the same is 0.92. The estimates for Model (1) for women tell a different story. The bereavement effect declines monotonically as we lengthen window of observations after the death of the spouse. The hazard ratios for the three time periods are 1.40, 1.31, and 1.13, respectively. The first two are statistically distinguishable from 1 at the 0.05 p-value and the last is at the 0.10 p-value. But,

J. Espinosa, W.N. Evans / Journal of Health Economics 27 (2008) 1326–1342

1339

Table 11 Partial maximum likelihood estimates of Cox proportional hazard model, death during follow-up, white married males and females, aged 50–70, public-use NLMS Covariate number

Covariate

Males Model (1)

(1) (2) (3)

Spouse died <6 months ago Uninformative death Informative death

1.327 (0.197)

(4) (5) (6)

Spouse died 6–24 months ago Uninformative death Informative death

1.275 (0.121)

(7) (8) (9)

Spouse died >24 months ago Uninformative death Informative death

1.335 (0.093)

Females Model (2)

Model (1) 1.404 (0.177)

1.399 (0.350) 1.291 (0.236)

1.092 (0.387) 1.463 (0.196) 1.265 (0.104)

1.454 (0.223) 1.185 (0.143)

1.029 (0.231) 1.308 (0.114) 1.127 (0.070)

1.301 (0.156) 1.353 (0.114)

−2 log likelihood

129,942.89

p-Value on −2 log likelihood tests (1) = (4) = (7) (2) = (3) (5) = (6) (8) = (9) (2) = (3) and (5) = (6) and (8) = (9)

0.922 0.796 0.291 0.786 0.740

Model (2)

129,941.65

1.033 (0.166) 1.141 (0.075) 63,770.16

63,768.10

0.195 0.438 0.315 0.558 0.582

Table entries are hazard ratios (standard errors) from Cox proportional hazard models. Other covariates include dummy variables for each age 51 through 70, plus a complete set of dummies for income, education and occupation.

given the standard errors on estimates, the p-value on the test that the coefficients are the same is almost 0.20. So in general, the bereavement effect for women dissipates considerably for women but not for men. When we allow the bereavement effect to vary both by time since the death of the spouse and cause of death, we reproduce the results from our previous analyses. In Model (2) in Table 11, we see for widowed males, that there is no qualitative difference in the bereavement effect based on the cause of death of their spouse either within 6 months or after 24 months of the death of their spouse. There is a noticeable difference in the hazard ratios for informative and uninformative deaths within 6–24 months but as with all cases, these differences are not statistically significant. For males, the bereavement effect hazard ratio for uninformative deaths within 6–24 months is statistically different from 1 at the p = 0.05 level and the same coefficient for 24 and more months is statistically significant at the p = 0.10 level. In no case can we reject the null that the coefficients for uninformative and informative deaths are the same. For women, Model (2) shows qualitatively small hazard ratios for all uninformative deaths. The standard errors on these estimates are however so large that we cannot reject the null that the hazard ratios for uninformative and informative deaths within time intervals are equal. As with the previous models, the results for males are convincing, the results for women are suggestive. 5. Conclusion Although a number of papers have established a heightened mortality for the survivor after the death of the spouse, the similarities in lifestyles, age, experience, and behaviors between a married couple call into question whether this result is indicative of the protective effects of marriage or whether the result is simply marriage selection. For males, the death of their spouse from informative causes of death reveals a lot of information about their SES so we learn little about the causal effect of bereavement on mortality by looking at mortality patterns after a spousal death from one of those causes. In contrast, the fact that the bereavement effect is as large for surviving husbands when their wives die from uninformative causes, that is, causes of death that reveal little about the survivor, suggests the bereavement effect is causal. Our results for women are less clear cut. The bereavement effect for surviving wives when their husband dies of an uninformative cause is small but with a large standard error, making it statistically indistinguishable from the effect for informative causes. There are a number of caveats to our work. The analysis only considers one measure of the health benefits of marriage (lower mortality) and the paper addresses an indirect measure of the health benefits of marriage (the rapid demise in health for some after the death of a spouse). That said, we have made progress analyzing this particular and well-studied aspect of the marriage/health literature. Second, death is our only longitudinal variable, so we do not observe some transitions that may contaminate the work. For example, a couple that divorces or separates after the survey date would be classified as married in our analysis. A more detailed longitudinal data set like the Panel Study of Income Dynamics would provide a longer list of time-varying covariates but the sample size drop would make the exercise useless. Given the sample sizes for current longitudinal data sets, adding more covariates does not seem like a viable strategy. Larger sample sizes in a mortality sample would however allow us to use finer breakdowns in causes of death when identifying uninformative causes of death. This could potentially generate greater agreement between the uninformative and unpreventable causes of death. Finally, it

1340

J. Espinosa, W.N. Evans / Journal of Health Economics 27 (2008) 1326–1342

would be useful to have list of unpreventable diseases for those 50–70 instead of the current metric that focuses on deaths across all age groups. As with most papers, our results raise more questions than they answer. The most compelling being why for uninformative causes of death is there a large bereavement effect for surviving males and a lack of an effect for women? Some possible hypotheses cannot explain the disparity in results. We do not believe the results for women are simply a Type II error. Fewer men outlive their wives and we were able to find a bereavement impact on men so the lack of results for women is most likely not due to a lack of power. More importantly, there are fewer uninformative deaths among women making the results for men even more dramatic. Some papers in this literature suggest that among the elderly, living alone may increase mortality but our calculations from the 1980 Census indicate there is no difference in the fraction living alone among widows or widowers, aged 50–70 (roughly 59 percent for both groups). Likewise, there is some evidence that the bereavement effect is larger for people of younger ages but we find little difference in our sample for the average age when husbands and wives lose their spouse (65.6 years for widows, 66.4 for widowers). Previous literature points to a number of possible explanations for the disparity in results between men and women. As we outlined above, many older men rely heavily on their spouses for activities that are key for health such as scheduling doctor visits, reminding them about taking their medicine, cooking, etc. In this situation, males would suffer more from the loss of spouse than married women. Likewise, women may have deeper social networks than men to help shield them from the psychological costs of bereavement. There is some evidence that the impact of widowhood on the incidence of depression is greater for men than women (Lee et al., 2001) although the results do not always show a gender difference (Feinson, 1986). Most women also may expect to outlive their husbands so the psychological shock to them of the death of as spouse may not be as dramatic. Working against these explanations is however the fact widowers appear to make some adjustments that shield them from the stress of widowhood. For example, Helsing et al. (1981) estimate that remarriage reduces the chance of death for widowers but not widows, and older widowers are more likely to remarry than elderly widows (Smith et al., 1991). Although both the marriage protection and marriage selection hypotheses explain the observed patterns in the data, only the former purports a causal relationship between marriage and mortality. Deciphering which story is correct can have interesting policy implications. Standard cost-effectiveness studies produce an estimate of the cost of treatment per quality adjusted life year saved. The denominator in this value is exclusively the patient being treated. However, if widows die at much higher rates in the first year after the death of the husband and this event can be attributed to marriage protection, standard cost-effectiveness estimates may understate the benefits of certain treatments (Christakis, 2004). Acknowledgements The authors wish to thank Seth Sanders, John Wallis, Suzanne Bianchi, Robert Kaestner, Jon Skinner and Nicholas Christakis for a number of helpful comments. This research was supported in part by a grant from the Russell Sage Foundation. References ´ O., Bergstrom, ¨ Adami, J., Nyren, R., Ekbom, A., Engholm, G., Englund, A., Glimelius, B., 1998. Smoking and the risk of leukemia, lymphoma, and multiple myeloma (Sweden). Journal of Cancer Causes and Control 9 (1), 49–56. Afonso, L., Mohammad, T., Thatai, D., 2007. Crack whips the heart: a review of the cardiovascular toxicity of cocaine. American Journal of Cardiology 100 (6), 1040–1043. Austin, H., Austin, J.M., Partridge, E.E., Hatch, K.D., Shingleton, H.M., 1991. Endometrial cancer, obesity, and body fat distribution. Cancer Research 51 (2), 568–572. Barrett, A.E., 2000. Marital trajectories and mental health. Journal of Health and Social Behavior 41 (4), 451–464. Becker, G.S., 1973. A theory of marriage: part I. Journal of Political Economy 81 (4), 813–846. Becker, G.S., 1974. A theory of marriage: part II. Journal of Political Economy 82 (2.2), S11–S26. Berkman, L.F., Kawachi, I., 2000. Social Epidemiology. Oxford University Press, New York. Berkman, L.F., Syme, S.L., 1979. Social networks, host resistance, and mortality: a nine-year follow-up study of Alameda County residents. American Journal of Epidemiology 109 (2), 186–204. Blazer, D.G., 1982. Social support and mortality in an elderly community population. American Journal of Epidemiology 115 (5), 684–694. Brinton, L.A., Blot, W.J., Becker, J.A., Winn, D.M., Browder, J.P., Famer, J.C., Fraumeni, J.F., 1984. A case–control study of cancers of the nasal cavity and paranasal sinuses. American Journal of Epidemiology 119 (6), 896–906. Brown, L.M., Devesa, S.S., 2002. Epidemiologic trends in esophageal and gastric cancer in the United States. Surgical Oncology Clinics of North America 11 (2), 235–256. Caspi, A., Herbener, E.S., 1993. Marital assortment and phenotypic convergence—longitudinal evidence. Social Biology 40 (1–2), 48–60. Cheung, Y.B., 1998. Can marital selection explain the difference in health between married and divorced people? From a Longitudinal Study of a British Birth Cohort. Public Health 112 (2), 113–117. Christakis, N.A., 2004. Social networks and collateral health effects have been ignored in medical care and clinical trials, but need to be studied. British Medical Journal 329, 184–185. Christakis, N.A., Iwashyna, T.J., 2003. The health impact of health care on families: a matched cohort study of hospice use by decedents and mortality outcomes in surviving, widowed spouses. Social Science and Medicine 57 (3), 465–475. Cohen, S., Doyle, W.J., Skoner, D.P., Rabin, B.S., Gwaltney Jr., J.W., 1997. Social ties and susceptibility to the common cold. Journal of the American Medical Association 277 (24), 1940–1944. Cox, P.R., Ford, J.R., 1964. The mortality of widows shortly after widowhood. Lancet 1 (7325), 163–164. Cox, R., 1972. Value of objective examinations. Nature 237 (5357), 489–492. Croft, P., Hannaford, P.C., 1989. Risk factors for acute myocardial infarction in women: evidence from the Royal College of General Practitioners’ Oral Contraception Study. British Medical Journal 298 (6667), 165–168.

J. Espinosa, W.N. Evans / Journal of Health Economics 27 (2008) 1326–1342

1341

Curran, P.J., Muthen, B.O., Harford, T.C., 1998. The influence of changes in marital status on developmental trajectories of alcohol use in young adults. Journal of Studies on Alcohol 59 (6), 647–658. Diener, E., Gohm, C.L., Suh, E., Oishi, S., 2000. Similarity of relations between marital status and subjective well-being across cultures. Journal of Cross-Cultural Psychology 31 (4), 419–436. Duncan, G.J., Wilkerson, B., England, P., 2006. Cleaning up their act: the impact of marriage and cohabitation on licit and illicit drug use. Demography 43 (4), 691–710. Durkheim, E., 1897. Suicide: A Study in Cociology (J.A. Spaulding, G. Simpson, Trans.). The Free Press of Glenco, New York. Farr, W., 1858. The influence of marriage on the mortality of the French people. In: Hastings, G.W. (Ed.), Transactions of the National Association for the Promotion of Social Sciences. Parker and Son, London, pp. 504–513. Feinson, M.C., 1986. Aging widows and widowers: are there mental health differences? International Journal of Aging and Human Development 23 (3), 241–255. Fentiman, I.S., Fourquet, A., Hortobagyi, G.N., 2006. Male breast cancer. Lancet 367 (9510), 595–604. Franceschi, S., Talaminin, R., Barra, S., Baron, A., Negri, E., Bidoli, E., Serraino, D., La Vecchia, C., 1990. Smoking and drinking in relation to cancers of the oral cavity, pharynx, larynx, and esophagus in Northern Italy. Cancer Research 50 (20), 6502–6507. Gardner, J., Oswald, A., 2004. How is mortality affected by money, marriage and stress? Journal of Health Economics 23 (6), 1181–1207. Goldman, N., 1993. Marriage selection and mortality patterns: inferences and fallacies. Demography 30 (2), 189–208. Gove, W.R., 1973. Sex, marital status and mortality. American Journal of Sociology 79, 45–67. Gove, W.R., Hughes, M.M., Style, C.B., 1983. Does marriage have positive effects on the psychological well-being of the individual? Journal of Health and Social Behavior 24, 122–131. ´ O., Bergstrom, ¨ Hansson, L.E., Baron, J., Nyren, R., Wolk, A., Adami, H.O., 1994. Tobacco, alcohol and the risk of gastric cancer. A population-based case–control study in Sweden. International Journal of Cancer 57 (1), 26–31. Harrison, G.A., Gibson, J.B., Hiorns, R.W., 1976. Assortative marriage for psychometric, personality and anthropometric variation in a group of Oxfordshire villages. Journal of Biosocial Science 8 (2), 145–153. Herbener, C.A., 1993. Marital assortment and phenotypic convergence—longitudinal evidence. Social Biology 40 (1–2), 48–60. Helsing, K., Szklo, M., 1981. Mortality after bereavement. American Journal of Epidemiology 114 (1), 41–52. Helsing, K., Szklo, M., Comstock, G., 1981. Factors associated with mortality after widowhood. American Journal of Public Health 71 (8), 802–809. Holly, E.A., Aston, D.A., Cress, R.A., Ahn, D.K., Kristiansen, J.J., 1995. Cutaneous melanoma in women I. Exposure to sunlight, ability to tan, and other risk factors related to ultraviolet light. American Journal of Epidemiology 141 (10), 923–933. House, J.S., Robbins, C., Metzner, H.L., 1982. The association of social relationships and activities with mortality: prospective evidence from the Tecumseh Community Health Study. American Journal of Epidemiology 116 (1), 123–140. Hu, Y., Goldman, N., 1990. Mortality differentials by marital status: an international comparison. Demography 27 (2), 233–250. Hurd, M., McFadden, D., Merrill, A., 1999. Predictors of mortality among the elderly. National Bureau of Economics Working Paper No. 7440. Cambridge, MA. Iwashyna, T.J., Christakis, N.A., 2003. Marriage, widowhood, and health-care use. Social Science and Medicine 57 (11), 2137–2147. Johnson, R.C., DeFries, J.C., Wilson, J.R., McClearn, G.E., Vandenberg, S.G., Ashton, G.C., Mi, M.P., Rashad, M.N., 1976. Assortative marriage for specific cognitive abilities in two ethnic groups. Human Biology 48 (2), 343–352. Kachur, A.J., Potter, L.B., James, S.B., Powell, K.E., 1995. Suicide in the United States, 1980–1992. Centers for Disease Control and Prevention, National Center for Injury Prevention and Control, Atlanta. Kaprio, J., Koskenvuo, M., Rita, H., 1987. Mortality after bereavement: a prospective study of 95,647 widowed persons. American Journal of Public Health 77 (3), 283–287. Key, T., Allen, N., Spencer, E., Travis, R., 2002. The effect of diet on risk of cancer. Lancet 360 (9336), 861–868. Kisker, E.E., Goldman, N., 1987. Perils of single life and benefits of marriage. Social Biology 34 (3–4), 135–152. Kitagawa, E.M., Hauser, P.M., 1973. Differential Mortality in the United States: A Study in Socioeconomic Epidemiology. Harvard University Press, Cambridge. Klatsky, A.L., Armstrong, M.A., Friedman, G.D., 1990. Risk of cardiovascular mortality in alcohol drinkers, ex-drinkers and nondrinkers. American Journal of Cardiology 66 (17), 1237–1242. Kobrin, F.E., Hendershot, G., 1977. Do family ties reduce mortality? Evidence from the United States, 1966–1968. Journal of Marriage and the Family 39 (4), 737–745. Korenman, S., Goldman, N., Fu, H., 1997. Misclassification bias in estimates of bereavement effects. American Journal of Epidemiology 145 (11), 995– 1002. Korenman, S., Neumark, D., 1991. Does marriage really make men more productive? Journal of Human Resources 26 (2), 282–307. Kricker, A., Armstrong, B.K., English, D.R., Heenan, P.J., 1991. Pigmentary and cutaneous risk factors for non-melanocytic skin cancer—a case–control study. International Journal of Cancer 48 (5), 650–662. LaCroix, A.Z., Lipson, S., Miles, T.P., White, L., 1989. Prospective study of pneumonia hospitalizations and mortality of U.S. older people: the role of chronic conditions, health behaviors, and nutritional status. Public Health Report 104 (4), 350–360. Laub, J.H., Nagin, D.S., Sampson, R.J., 1998. Trajectories of change in criminal offending: good marriages and the desistance process. American Sociological Review 63 (2), 225–238. Lee, G.R., DeMaris, A., Bavin, S., Sullivan, R., 2001. Gender differences in the depressive effect of widowhood in later life. The Journals of Gerontology: Series B, Psychological Sciences and Social Science 56 (1), S56–61. Lerman, R.I., 2002. Marriage as a Protective Force Against Economic Hardship. The Urban Institute, Washington, DC. Lichtenstein, P., Gatz, M., Berg, S., 1998. A twin study of mortality after spousal bereavement. Psychological Medicine 28 (3), 635–643. Lillard, L.A., Panis, C.W.A., 1996. Marital status and mortality: the role of health. Demography 33 (3), 313–327. Lillard, L., Waite, L., 1995. Til’ death do us part: marital disruption and mortality. American Journal of Sociology 100 (5), 1131–1156. Luoma, J.B., Pearson, J.L., 2002. Suicide and marital status in the United States, 1991–1996: is widowhood a risk factor? American Journal of Public Health 92 (9), 1518–1522. Lupton, J.P., Smith, J.P., 2003. Marriage, assets, and savings. In: Marriage and the Economy: Theory and Evidence from Advanced Industrial Societies. Cambridge University Press, New York, pp. 129–152. Macfarlane, G.J., Zheng, T., Marshall, J.R., Boffetta, P., Niu, S., Brasure, J., Merletti, F., Boyle, P., 1995. Alcohol, tobacco, diet and the risk of oral cancer: a pooled analysis of three case–control studies. European Journal of Cancer Part B: Oral Oncology 31B (3), 181–187. Manor, O., Eisenbach, Z., 2003. Mortality after spousal loss: are there socio-demographic differences? Social Science and Medicine 56 (2), 405–413. Manor, O., Eisenbach, Z., Israeli, A., Friedlander, Y., 2000. Mortality differentials among women: the Israel longitudinal mortality study. Social Science and Medicine 51 (8), 1175–1188. Martikainen, P., Valkonen, T., 1996. Mortality after the death of a spouse: rates and causes of death in a large Finnish cohort. American Journal of Public Health 86 (8), 1087–1093. Mascie-Taylor, C.G., Gibson, J.B., 1979. A biological survey of a Cambridge suburb: assortative marriage for IQ and personality traits. Annals of Human Biology 6 (1), 1–16. McGinnis, J.M., Foege, W.H., 1993. Actual causes of death in the United States. JAMA 270 (18), 2207–2212. Mendes de Leon, C.F., Kasl, S.V., Jacobs, S., 1993. Widowhood and mortality risk in a community sample of the elderly—a prospective-study. Journal of Clinical Epidemiology 46 (6), 519–527. Mergenhagen, P.M., Lee, B.A., Gove, W.R., 1985. Till death do us part: recent changes in the relationship between marital status and mortality. Sociology and Social Research 70 (1), 53–56.

1342

J. Espinosa, W.N. Evans / Journal of Health Economics 27 (2008) 1326–1342

Miller-Tutzauer, C., Leonard, K., Windle, M., 1991. Marriage and alcohol use: a longitudinal study of “Maturing Out”. Journal of Alcohol Studies 52 (5), 434–440. Mineau, G.P., Smith, K.R., Bean, L.L., 2002. Historical trends of survival among widows and widowers. Social Science and Medicine 54, 245–254. Morimoto, L.M., White, E., Chen, Z., Chlebowski, R.T., Hays, J., Kuller, L., Lopez, A.M., Manson, J., Margolis, K.L., Muti, P.C., Stefanick, M.L., McTiernan, A., 2002. Obesity, body size, and risk of postmenopausal breast cancer: the Women’s Health Initiative (United States). Journal of Cancer Causes and Control 13 (8), 741–751. Murray, J.E., 2000. Marital protection and marital selection: evidence from a historical-prospective sample of American men. Demography 37 (4), 511–521. Parkes, C.M., Benjamin, B., Fitzgerald, R.G., 1969. Broken heart: a statistical study of increased mortality among widowers. British Medical Journal 1 (646), 740–743. Phelan, J.C., Link, B.G., Diez-Roux, A., Kawachi, I., Levin, B., 2004. “Fundamental causes” of social inequalities in mortality: a test of the theory. Journal of Health and Social Behavior 45 (3), 265–285. Quirouette, C., Gold, D.P., 1992. Spousal characteristics as predictors of well-being in older couples. International Journal of Aging and Human Development 34 (4), 257–269. Rahman, O., 1993. Excess mortality for the unmarried in rural Bangladesh. International Journal Epidemiology 22 (3), 445–456. Rogers, R.G., 1995. Marriage, sex and mortality. Journal of Marriage and the Family 57 (2), 515–526. Rogot, E., Sorlie, P.D., Johnson, N.J., Schmitt, C., 1992. A Mortality Study of 1.3 million Persons by Demographic, Social, and Economic Factors: 1979–1985 Follow-up. National Institutes of Health, Bethesda MD. Rosenbloom, C.A., Whittington, F.J., 1993. The effects of bereavement on eating behaviors and nutrient intakes in elderly widowed persons. Journal of Gerontology: Social Sciences 48 (4), S223–S229. Sawhill, I., Thomas, A., 2002. For richer or for poorer: marriage as an antipoverty strategy. Journal of Policy Analysis and Management 21 (4), 587–599. Schaefer, C., Quesenberry Jr., C.P., Wi, S., 1995. Mortality following conjugal bereavement and the effects of a shared environment. American Journal of Epidemiology 141 (12), 1142–1152. Schoeni, R.F., 1995. Marital status and earnings in developed countries. Journal of Population Economics 8 (4), 351–359. Smith, J.C., Mercy, J.A., Conn, J.M., 1988. Marital status and the risk of suicide. American Journal of Public Health 78 (1), 78–80. Smith, K.R., Waitzman, N.J., 1994. Double jeopardy: interaction effects of marital and poverty status on the risk of mortality. Demography 31 (3), 487–507. Smith, K.R., Zick, C.D., 1994. Linked lives, dependent demise? Survival analysis of husbands and wives. Demography 31 (1), 81–89. Smith, K.R., Zick, C.D., Duncan, G.J., 1991. Remarriage patterns among recent widows and widowers. Demography 28 (3), 361–374. Smith, M., Eastham, E., Gleason, D., Shasha, D., Tchekmedyian, S., Zinner, N., 2003. Randomized controlled trial of zoledronic acid to prevent bone loss in men receiving androgen deprivation therapy for nonmetastatic prostate cancer. The Journal of Urology 169 (6), 2008–2012. Smith, M.R., McGovern, F.J., Zietman, A.L., Fallon, M.A., Hayden, D.L., Schoenfeld, D.A., Kantoff, P.W., Finkelstein, J.S., 2001. Pamidronate to prevent bone loss during androgen-deprivation therapy for prostate cancer. New England Journal of Medicine 345 (13), 948–955. Snyder, S.E., Evans, W.N., 2006. The effect of income on mortality: evidence from the Social Security Notch. Review of Economics and Statistics 88 (3), 482–495. Sorlie, P.D., Backland, E., Keller, J.B., 1995. U.S. Mortality by economic, demographic and social characteristics: The National Longitudinal Mortality Study. American Journal of Public Health 85, 949–956. Spuhler, J.N., 1968. Assortative mating with respect to physical characteristics. Eugenics Quarterly 15 (2), 128–140. Stack, S., Eshleman, J.R., 1998. Marital status and happiness: a 17-nation study. Journal of Marriage and the Family 60 (2), 527–536. Tai, M.H., Upham, B.L., Olson, L.K., Tsao, M.S., Reed Jr., D.N., Trosko, J.E., 2007. Cigarette smoke components inhibited intercellular communication and differentiation in human pancreatic ductal epithelial cells. International Journal of Cancer 120 (9), 1855–1862. Taylor, J.A., Umbach, D.M., Stephens, E., Castranio, T., Paulson, D., Robertson, C., Mohler, J.L., Bell, D.A., 1998. The role of N-acetylation polymorphisms in smoking-associated bladder cancer: evidence of a gene–gene-exposure three-way interaction. Cancer Research 58 (16), 3603–3610. Taylor, P., Funk, C., Craighill, P., 2006. Are we happy yet? Pew Research Center Social Trends Report, February 13, 2006. http://pewresearch.org/assets/social/pdf/AreWeHappyYet.pdf. Thun, M.J., Peto, R., Lopez, A.D., Monaco, J.H., Henley, S.J., Heath Jr., C.W., Doll, R., 1997. Alcohol consumption and mortality among middle-aged and elderly U.S. adults. New England Journal of Medicine 337 (24), 1705–1714. Trovato, F., 1991. Sex, marital status, and suicide in Canada: 1951–1981. Sociological Perspectives 34 (4), 427–445. Umberson, D., 1992. Gender, marital status, and the social control of behavior. Social Science and Medicine 34 (8), 907–917. van den Brandt, P.A., Spiegelman, D., Yaun, S.S., Adami, H.O., Beeson, L., Folsom, A.R., Fraser, G., Goldbohm, R.A., Graham, S., Kushi, L., Marshall, J.R., Miller, A.B., Rohan, T., Smith-Warner, S.A., Speizer, F.E., Willett, W.C., Wolk, A., Hunter, D.J., 2000. Pooled analysis of prospective cohort studies on height, weight, and breast cancer risk. American Journal of Epidemiology 152 (6), 514–527. Vandenburg, S.G., 1972. Assortive mating, or who marries whom? Behavior Genetics 2 (2–3), 127–157. Velicer, C.M., Kristal, A., White, E., 2006. Alcohol use and the risk of prostate cancer: results from the VITAL cohort study. Nutrition and Cancer 56 (1), 50–56. Verbrugge, L.M., 1979. Marital status and health. Journal of Marriage and the Family 41 (2), 267–285. Waldron, H., Hughes, M.E., Brooks, T.L., 1996. Marriage protection and marriage selection—prospective evidence for reciprocal effects of marital status and health. Social Science and Medicine 43 (1), 113–123. Waite, L.J., 1995. Does marriage matter? Demography 32 (4), 483–507. Waite, L.J., Gallagher, M., 2000. The Case for Marriage: Why Married People are Happier, Healthier, and Better Off Financially. Broadway Books, New York. ¨ Wilhelmsen, L., Svardsudd, K., Korsan-Bengtsen, K., Larsson, B., Welin, L., Tibblin, G., 1984. Fibrinogen as a risk factor for stroke and myocardial infarction. New England Journal of Medicine 311 (8), 501–505. Wittstein, I.S., Thiemann, D.R., Lima, J.A., Baughman, K.L., Schulman, S.P., Gerstenblith, G., Wu, K.C., Rade, J.J., Bivalacqua, T.J., Champion, H.C., 2005. Neorohumoral features of myocardial studding due to sudden emotional stress. New England Journal of Medicine 352 (6), 539–548. ˆ Yusuf, S., Hawken, S., Ounpuu, S., Bautista, L., Franzosi, M., Commerford, P., Lang, C., Rumboldt, Z., Onen, C., Lisheng, L., 2005. Obesity and the risk of myocardial infarction in 27,000 participants from 52 countries: a case–control study. Lancet 366 (9497), 1640–1649.