Widowhood and Mortality: A Meta-Analysis and Meta-Regression

Abstract

Introduction

The effect of marital status on health and mortality was one of the earliest issues to be systematically studied by sociologists and demographers, with work dating to Durkheim’s classic study on suicide (Durkheim 1897/1951). Over the years, numerous studies have examined this relationship, with many of them focusing on the risk of death among persons who had lost their spouse (e.g., Alter et al. 2007; Clayton 1974; Hart et al. 2007; Helsing et al. 1981; Jones and Goldblatt 1987; Lusyne et al. 2001; Manor and Eisenbach 2003; Schaefer et al. 1995; Stimpson et al. 2007; Young et al. 1963). Indeed, the death of a loved one is widely recognized as one of life’s most potent stressors, due in part to the associated disruption of social support, life routines, and financial status (Stroebe 2001).

Overall, the resulting body of research demonstrates a higher risk of death associated with loss of a spouse (Hughes and Waite 2009), although a few studies report no significant effect and the magnitude of effect varies substantially. This variability is at least partly associated with individual factors that include gender (Mineau et al. 2002; Schaefer et al. 1995; Smith and Zick 1996; Stroebe et al. 2001; Thierry 1999), age (Johnson et al. 2000; Lichtenstein et al. 1998; Manor and Eisenbach 2003; Martikainen and Valkonen 1996a; Mendes De Leon et al. 1993; Schaefer et al. 1995), recency of widowhood (Jagger and Sutton 1991; Kaprio et al. 1987; Mellstrom et al. 1982; Nystedt 2002; Stimpson et al. 2007; Stroebe et al. 2007), and geographical region (Lusyne et al. 2001; Nagata et al. 2003; Rahman et al. 1992; Voges 1996).

The current trend in the literature is toward an increased emphasis on identifying mediating, moderating, and confounding factors in the widowhood/mortality association. Thus, the time is ripe for a meta-analysis that examines known potential moderators and seeks to identify new ones.

Moderating Factors in the Widowhood-Mortality Association

In their recent meta-analysis of the literature on marital status (including widowhood) and mortality among individuals 65 years of age and older, Manzoli et al. (2007) reported that widowed persons had an 11% higher risk of mortality compared with married persons. Moderating factors such as gender and geographic region, however, were nonsignificant. The former finding is surprising given that gender differences in marital status–related mortality have been well established by previous studies (Gove 1973; Hemstrom 1996; Stroebe et al. 2001). As typical examples, it has been found that relative risks (widowhood vs. married) among men were 16% higher (Hemstrom 1996) to 42% higher (Kolip 2005) than relative risks among women. It is interesting to note, however, that gender differences such as these tend to be found in non-elderly samples rather than among the older cohorts examined by Manzoli et al. (2007). This suggests the possibility of a gender-age interaction, an idea that is supported by studies examining both gender and age. Mineau et al. (2002) found that the mortality risk of widowed men relative to married men, compared with the same measure among women, was 46% higher at age 35–44 but only 12% higher at ages 75 and older. Even more dramatically, Smith and Zick (1996) found that the men’s relative risk was 373% higher at ages 25–64 but 22% lower among those 65 and older.

Other potentially important moderators have also received attention. First, studies have considered the possible effect of time period on the magnitude of the widowhood-mortality association. Despite steady improvements in medical treatments, some studies found that widowhood-related relative risks have increased over time. For example, van Poppel and Joung (2001) found increases in relative risks of 43% among men and 87% among women in the Netherlands between the 1850–1859 and the 1960–1969 periods. Unexplained increases, ranging from 6% to 40%, have also been found among Finnish men and women between the 1976–1980 and 1996–2000 periods (Martikainen et al. 2005). However, Mineau et al. (2002) found both increases and decreases in relative risks between an 1860–1874 and an 1895–1904 marriage cohort, depending on age at widowhood. In this study, relative risks increased by between 12% and 19% among persons widowed between the ages of 25 and 44, remained unchanged for ages 45–64, and declined by between 6% and 10% for persons widowed at age 65 or older. In their comparison of Sweden, Belgium, and the Netherlands, Alter et al. (2007) also showed decreased relative risks over time for women who were widowed less than 5 years.

Second, the possible effects of widowhood recency have been central to a long line of research. At an early date, Young and colleagues noted that mortality was highest in the first six months following widowhood but declined in subsequent months and years (Young et al. 1963). Subsequent research has largely supported this finding. For example, Nystedt (2002) found that widowhood-related relative risks (RRs) fell from 2.38 in the first six months of widowhood to 1.28 among those for whom widowhood occurred six or more years prior. Likewise, Thierry (1999) found that RRs (widowed vs. married) fell during the first 10 years for men and women of all ages.

The suggestion that the risk of mortality varies depending on the amount of time elapsed from the onset of widowhood has opened the way for physiological investigations of loss and grief from a stress response perspective (Jones and Goldblatt 2006; Martikainen and Valkonen 1996b; Susser 1981). In doing so, these studies have focused on linkage mechanisms, such as immune system disruption (Gerra et al. 2003; Goforth et al. 2009) and cardiovascular effects (Buckley et al. 2010), that may connect the early months of widowhood to a range of chronic diseases and mortality. Others have examined behavioral pathways, such as poor self-care and increases in health-risk behaviors by the surviving spouse, that potentially connect bereavement to near-term negative health consequences (Jin and Christakis 2009; Sharar et al. 2001; Stroebe et al. 2007). A related line of research has focused on the loss of important social (Armenian et al. 1987; Bowling and Charlton 1987; Jylha and Aro 1989; Lusyne et al. 2001; Martikainen and Valkonen 1996b; Mineau et al. 2002) and economic (Nystedt 2002; Rahman 1997; Smith and Waitzman 1994; Zick and Smith 1991b) buffers that may affect health and survival (Subramanian et al. 2008).

The present meta-analysis contributes to this body of knowledge on widowhood and mortality in two important ways. First, we utilize the heterogeneity of research settings found in this literature to assess the impact of multiple potential moderators. Some, such as gender and age, are relatively easy to evaluate within an individual study. Others—such as widowhood recency, time period, and cultural differences—are less frequently addressed by individual studies and are therefore more easily examined by comparing across studies. Meta-analysis is well suited to this task, and our results include tests of gender-age interactions, geographic region, time period, and a number of specific study-design characteristics. Second, the overall magnitude of the association between widowhood and health outcomes has not been examined among non-elderly persons.

Specifically, we test four hypotheses using meta-analysis and meta-regression. First, we assess whether there exists a gender-age interaction such that hazard ratios (HRs) are greater for men than for women, but more so at younger ages. Second, we test the hypothesis that the relative mortality risk associated with widowhood has increased over time. Third, we examine the hypothesis put forth by Young et al. (1963) that more recently experienced widowhood is associated with greater mortality risk. Finally, we test whether there exists an interaction between gender and follow-up duration such that HRs are greater for men than for women, particularly when widowhood is recent. In addition, we examine the possible effects of geographic region, study-level control variables, the composition of the case and control groups, and several other study-design characteristics in the interest of providing findings from which new mediator and moderator hypotheses might be developed.

Methods

In June 2005, we conducted a sensitive search of electronic bibliographic databases to retrieve all publications combining the concepts of psychosocial stress (including widowhood) and all-cause mortality. Overall, 100 search clauses were used for Medline, 97 for EMBASE, 81 for CINAHL, and 20 for Web of Science. (See Online Resource 1, Section 1, for the full search algorithm used for Medline; information on the remaining search algorithms is available from the authors upon request.) This process identified 1,570 unique publications. With these results as a base, the bibliographies of eligible publications, the lists of sources citing an eligible publication, and the sources identified as “similar to” an eligible publication were iteratively hand-searched. The literature was exhausted after eight iterations. (The full description of this iterative search protocol is available from the authors upon request.). The electronic keyword searches in these databases were run again in July 2008, and the search and coding stages were completed in January 2009.

The electronic database searches were performed by a research librarian. Two authors trained in meta-analysis coding procedures determined publication eligibility and extracted the data from the identified articles. Data were entered into and publications were tracked throughout the process using spreadsheets. (See Online Resource 1, Section 2, for a full list of variables for which data were sought.) All unpublished work encountered was considered for study inclusion. Although the search was done for articles published in English, we were able to locate and translate the relevant portions of 36 publications in German, Danish, French, Spanish, Dutch, Polish, and Japanese. Figure 1 summarizes the number of publications considered at each step of the search process. Among the 730 publications considered tentatively eligible for study inclusion (based on examinations of title only), 428 were excluded from further consideration upon examination of the abstract. Of the remaining 302 publications that were examined in full, 151 were excluded because of the lack of a valid stress measure (70 publications), unavailable data on the case or control group (37 publications), lack of the all-cause mortality outcome (15 publications), conflation of multiple stressors (13 publications), the lack of a valid comparison group (11 publications), and for other reasons (6 publications). The full database contains 262 publications examining the effects of various stressful events on all-cause mortality. To evaluate coding accuracy, 40 of these publications were randomly selected and recoded (including 446 point estimates). Of the point estimates, 98.6% were error free.

The present analysis uses the subset of articles (n = 123) that reported the effect of widowhood on all-cause mortality. Of these publications, 116 appeared in peer-reviewed journals, four were book chapters, one was an unpublished dissertation, and two were unpublished papers (authors of these papers were contacted for permission to use their results). One publication was translated from Spanish, two from German, one from French, and one from Danish in consultation with fluent speakers of the language; the remaining 118 publications were in English (see Table 1).

In addition to the requirement that a study report one or more point estimates pertaining to all-cause mortality, studies were included in the present analysis if a clear comparison was made between people who lost their spouse and people who were married (or the general population, which consists primarily of married people; see Table 1). In addition to studies with longitudinal designs, cross-sectional studies were included if the sample size was large, a baseline date could be determined accurately, and the manner in which death data was collected approximated a follow-up period. For example, the study by Sheps (1961), which used census data for the denominator and national annual mortality data for the numerator, was coded as having an April 1, 1950, baseline and a one-year follow-up period. In total, the 123 publications provided 1,377 point estimates for analysis.

Statistical methods varied across the included studies, necessitating the conversion of odds ratios, rate ratios, standardized mortality ratios, RRs, and HRs into a common metric. All non-HR point estimates were converted to HRs (the most frequently reported type) using one or both of the following equations (Zhang and Yu 1998):

and

where RR is the relative risk, OR is the odds ratio, HR is the hazard ratio, and r is the death rate for the reference group (i.e., those who are married). For 328 of the 1,377 relative risks, the death rate (i.e., the conversion factor, not the dependent variable) for the reference group was not reported. In these cases, the death rate was estimated using multiple regression. Significant predictors of the death rate were follow-up duration, sample size (log transformed), the proportion of the sample that was male, mean age at enrollment, an indicator for whether the study statistically controlled for gender, an indicator for whether the study statistically controlled for age, and the subjective quality assessment score assigned by the coders (multiple R = .797). Sensitivity analyses were performed to examine the possible effect of including or excluding studies for which we had to estimate the death rate.

As is standard practice, the standard errors reported in the publications were used to calculate the inverse variance weights. When not reported, standard errors were calculated using (1) confidence intervals, (2) t statistics, (3) chi-square statistics, or (4) p values. When upper-limit p values were the only estimate of statistical significance available (e.g., when the reported p value was between .01 and .05), the midpoint of the upper and lower limits was used to estimate the true p value. For 668 of the 1,377 point estimates, no measure of statistical significance was reported and standard errors were estimated using multiple regression. Significant predictors of the standard error were follow-up duration, sample size (log transformed), mean age at enrollment, the magnitude of the HR, and publication date (multiple R = .721). An indicator variable was created so that analyses could be conducted both with and without the estimated standard errors.

Many meta-analysts prefer to use only the most general point estimates reported in a given publication. Although this strategy makes it easier to maintain independence between point estimates and makes the calculations of the inverse variance weights straightforward, it also results in a substantial loss of information. We sought, instead, to maximize the number of point estimates analyzed, capturing variability both between publications and within each publication rather than just the former. For example, when a publication (see hypothetical Study X in Table 2) reported mortality risks by gender subgroups alone, the data require no adjustment. Likewise, when a study reported mortality risks by age group alone (see hypothetical Study Y, Table 2), the data again require no adjustment. When a publication first reported mortality risks by gender and then again by age, however (see hypothetical Study Z, Table 2), this created a violation of independence because each person is represented twice. To correct for this double-counting, each of the variance weights was adjusted to one-half of its original value, thus preserving information on the gender and age variables while counting each subject only once.

Two measures of study quality were adopted. First, a three-level subjective rating was assigned to each publication. Publications were rated as “low quality” if they contained obvious reporting errors or applied statistical methods incorrectly. Publications were rated as “high quality” if models were well specified (i.e., the correct model was used relative to the state of the art at the time of publication) and if discussions and reporting of study results were detailed. Next, principal components analysis was used to construct a 10-point scale using the following: (1) the five-year impact factor (ISI Web of Knowledge 2009) of the journal in which the article was published (an impact factor of 1 was assigned when the impact factor was not available); (2) the number of citations received per year since publication according to ISI Web of Knowledge; and (3) the number of authors, since studies with a larger author body may have a more diverse pool of scholarly expertise, thereby decreasing the likelihood of methodological or theoretical errors.

Both Q tests (which assess the probability that the observed variability among effect estimates, across studies and/or subgroup, within a meta-analysis is due solely to chance) and I² tests (which use the results of the Q test to calculate the degree of heterogeneity present) were used to assess heterogeneity in the data (Huedo-Medina et al. 2006). Q test results from preliminary analyses revealed substantial heterogeneity. In light of this, all meta-analyses and meta-regression analyses were calculated by maximum likelihood using a random-effects model, the random effects being applied at the level of the HR data. Analysis was performed with PASW Statistics 18.0 using matrix macros provided by Lipsey and Wilson (2001). The possibility of selection and publication bias was examined using a funnel plot of the log HRs against sample size. Because of heterogeneity in the data, funnel plot asymmetry was tested using Eggers’ test (Egger and Davey-Smith 1998) and weighted least squares regressions of the log HRs on the inverse of the sample size (Moreno et al. 2009; Peters et al. 2006).

Analyses performed include meta-analyses of subgroups and multivariate meta-regression analyses. The covariates used in the analyses were dictated by data availability. Variables such as race or ethnicity, which were used as grouping variables or included in interaction terms in only a small number of studies, could not be used in the analyses. The following covariates were used: (1) whether the death rate had to be estimated in order to derive the HR (yes or no); (2) whether the standard error was estimated (yes or no); (3) the proportion of respondents who were male; (4) the mean age of the sample/subgroup at enrollment, divided by 10; (5) the age range of the sample/subgroup at enrollment, divided by 10; (6) age of the study (the years elapsed since the beginning of the enrollment period), divided by 10; (7) age of the publication (the years elapsed since publication), divided by 10; (8) the duration of the enrollment period, in years; (9) the time elapsed between the end of enrollment and the beginning of follow-up, in years; (10) the follow-up duration, in years; (11) whether the general population was used as the comparison group (yes or no), as opposed to only married persons; (12) whether the study sample consisted of persons with preexisting health problems and/or unusually high levels of stress (yes or no); (13) geographic region (China/Japan, eastern Europe,¹ western continental Europe²/Israel, the United Kingdom, Scandinavia,³ the United States, or Bangladesh/Lebanon); (14) sample size, log transformed; (15) subjective scale of study quality (1–3 range); (16) a continuous composite measure of study quality (0–10 range); (17) a series of variables indicating whether sex, age, socioeconomic status, and health were statistically controlled; and (18) interaction terms between gender, mean age, and follow-up duration.

Results

Table 3 provides descriptive statistics on the 1,377 mortality risk estimates included in the current meta-analysis. Data were obtained from 123 studies published between 1955 and 2007, covering 22 countries, and representing more than 500 million people. Both men and women are well represented in the data set, and 87.5% of the study samples had a mean age of 40 years or older. The median of the maximum follow-up was five years. Of the HRs analyzed, 91.9% were reported in studies assigned a subjective quality rating of average or high; the mean five-year impact factor was 4.2; and the mean number of citations received per year since publication was 2.4.

The results of a number of meta-analyses are presented in Table 4 (see Table 5 for sample size information), stratified by the level of statistical adjustment of the risk estimate. They reveal that widowed individuals were more likely to die than their married, nonwidowed counterparts. The mean HR was 1.73 among statistically unadjusted point estimates (95% CI, 1.68–1.79; n = 693 HRs); 1.20 among age-adjusted point estimates (95% CI, 1.15–1.26; n = 284); and 1.20 among point estimates adjusted for age and additional covariates (95% CI, 1.16–1.25; n = 400). Exclusion of HRs based on estimated death rates, and of HRs in which the standard error was estimated, does not substantively alter the mean HRs (see Table 4). The mean HR among studies with a low subjective quality rating did not differ significantly from 1.00 (HR, 1.44; 95% CI, 0.90–2.32; n = 2), but this may be due solely to the small sample size. The mean HR was elevated among studies with an average quality rating (HR, 1.17; 95% CI, 1.08–1.26; n = 104) and the highest quality rating (HR, 1.22; 95% CI, 1.16–1.28; n = 294). Thus, after controlling for multiple covariates including age and including only high-quality studies, widowhood was associated with a 22% higher risk of mortality.

Subgroup Meta-Analyses and Meta-Regression Analyses

From this point forward, the discussion will focus on the more conservative findings of HRs adjusted for age and additional covariates (see Table 4). Results of these analyses reveal that widowhood had a deleterious effect for both genders, but the magnitude of the effect was greater for men (HR, 1.27; 95% CI, 1.19–1.35; n = 166) than for women (HR, 1.15; 95% CI, 1.08–1.22; n = 177). Furthermore, the results of meta-regression analyses, modeling all main effects (Model 1), main effects plus three interaction terms (Model 2), and a final parsimonious model (Model 3), confirm that the increase in risk of death for men who lost their spouse was substantially higher than the increase in risk for women who lost their spouse (see Table 6).

An interesting result comes from comparing groups by average age at study enrollment. As shown in Table 4, widowhood has a harmful effect on mortality in almost all age groups, but the magnitude of the effect decreases with age. The mean HR associated with widowhood was high yet nonsignificant for people aged 30–39 (HR, 1.94; 95% CI, 0.98–3.87; n = 3). The lack of significance, however, is probably due to the limited number of studies that included individuals in this age range, the small number of widows, and the low mortality rate in this age range. It became significant in the 40–49 age group, in which widows had a 15% higher risk of death than married persons (HR, 1.15; 95% CI, 1.02–1.29; n = 33), and remained so for all other age groups. Although the risk was highest for those aged 50–59 (HR, 1.38; 95% CI, 1.15–1.67; n = 27), it then decreased for those aged 60–69 (HR, 1.24; 95% CI, 1.16–1.34; n = 103), 70–79 (HR, 1.19; 95% CI, 1.07–1.32; n = 52), and 80 years or older (HR, 1.18; 95% CI, 1.11–1.24; n = 182). The results of the initial meta-regression analysis (Model 1 of Table 6) reflect this downward trend among the latter four age groups (suggesting a 10% decrease for each additional 10 years; p < .001).

The effects of gender and age on the magnitude of the HR are more complex than the meta-analyses with only main effects reveals. Specifically, both Model 2 (full model) and Model 3 (parsimonious model) show a significant interaction effect between these two variables (see Table 6). In Model 3, the exponentiated regression coefficients are 1.75 (95% CI, 1.54–2.00) for gender, 0.93 (95% CI, 0.91–0.94) for mean age, and 0.94 (95% CI, 0.92–0.96) for the interaction between gender and mean age. Taken together, these results indicate that in middle age, the excess mortality risk associated with widowhood is substantially greater for men than for women, but that this excess risk also declines more rapidly with age for men than it does for women. By age 90, the difference in excess mortality risk between men and women is negligible; in fact, the HR for widowhood is approximately 1.0 (no excess risk) for both men and women. Figure 2 shows the predicted hazard rate by age, separately for men and women, based on the estimates from Model 3 of Table 6 (see Online Resource 1, Section 3, for details).

The results presented in Table 4 also show that the effects of widowhood on mortality remained quite stable throughout the 120 years represented by the studies that were sampled for the current analysis. Widowhood had a significant harmful effect on mortality in studies with a baseline before 1940 (HR, 1.14; 95% CI, 1.05–1.24; n = 68), and also in studies with a baseline after 1960. The harmful effect, however, was lower in the 1960s and 1970s than in more recent decades. The mean HR was 1.18 (95% CI, 1.03–1.36; n = 37) in studies with a baseline between 1960 and 1969, and was 1.17 (95% CI, 1.08–1.26; n = 81) in studies with a baseline between 1970 and 1979. It increased to 1.24 (95% CI, 1.16–1.33; n = 146) in studies with a baseline between 1980 and 1989, and to 1.27 (95% CI, 1.16–1.24; n = 68) in studies with a baseline between 1990 and 1999. The meta-regression results (Table 6) confirmed that the effect of widowhood on mortality was lower in previous decades. HRs were 2% lower for each additional 10 years that had elapsed since the baseline data were collected (p < .001).

Follow-up duration was also a significant predictor in the meta-regression analyses (see Table 6), and the meta-analyses suggest that the effects of widowhood on mortality are substantively higher during the first two years of follow-up (see Table 4). The excess risk associated with widowhood was 58% in studies with only six months of follow-up (HR, 1.58; 95% CI, 1.32–1.88; n = 33), 33% in studies with one year of follow-up (HR, 1.33; 95% CI, 1.11–1.61; n = 30), and 51% in studies with two years of follow-up (HR, 1.51; 95% CI, 1.27–1.79; n = 15). In studies that followed individuals for 16–20 years, the excess risk decreases to 22% (HR, 1.22; 95% CI, 1.02–1.47; n = 11), 27% for 21–25 years of follow-up (HR, 1.27; 95% CI, 1.09–1.49; n = 14), and 11% for 25 years or more of follow-up (HR, 1.11; 95% CI, 1.02–1.20; n = 54). The final regression (Model 3 of Table 6) indicates that the mean HR decreases by 2% (p < .001) for every additional 10 years of follow-up. This pattern of results suggests that the excess risk associated with widowhood is greatest during the first few years after the death of a spouse, but persists at reduced levels for 20 years or more. The hypothesized interaction between follow-up and gender, however, is not supported (p = .244; Model 2 of Table 6).

Finally, the results presented in Table 4 show that the effect of widowhood on mortality is relatively homogenous in different regions of the world. The mean HR was 1.22 for Scandinavia (95% CI, 1.13–1.32; n = 92); 1.19 for the United States (95% CI, 1.12–1.26; n = 146); 1.16 for the United Kingdom (95% CI, 1.01–1.33; n = 24); 1.01 for eastern Europe (95% CI, 0.57–1.80; n = 2); 1.25 for western continental Europe (95% CI, 1.14–1.36; n = 101); 1.17 for China and Japan (95% CI, 1.00–1.37; n = 24); and 1.22 for Bangladesh/Lebanon (95% CI, 0.97–1.54; n = 11). Model 3 in Table 6 suggests that in the United States, eastern Europe, western Europe, and Bangladesh/Lebanon, widowed people have a somewhat higher risk for mortality than in the United Kingdom (the reference group), in Scandinavia (p = .844), and in China and Japan (p = .716). This model shows that the magnitude of the effect is 14% higher in the United States (p < .001), 18% higher in eastern Europe (p = .001), and 13% higher in western Europe (p < .001). Although Model 3 shows that the mean HR is 57% higher in Bangladesh/Lebanon (p < .001), this is likely due solely to the fact that there are more than three times as many unadjusted HRs (n = 36) as there are HRs adjusted for age and additional covariates (n = 11) for this region.

The results presented in Table 6 show that other significant predictors of differences among reported HRs include the time elapsed between a study’s end of participant enrollment and beginning of follow-up (a 6% increase in risk for each additional year; p < .001), and whether the risk estimate was adjusted for age (a 14% decrease when age was controlled; p < .001), SES (a 11% decrease when controlled; p < .001), social ties (a 9% increase when controlled; p = .013), and previous stress (a 9% decrease when controlled; p = .034). The results presented in Table 6 also show that HRs in studies that estimated the underlying death rate were significantly lower than in studies that did not (a 12% decrease; p < .001). Finally, contrary to the common conception that the average effect size decreases as the study quality improves, we found that the mean magnitude of the effect actually increased in studies that were evaluated as having a higher quality (a 14% increase in the hazard for each 1-point increase in the 3-point subjective study-quality measure; p < .001).

Analysis of Data Heterogeneity

The between-groups Cochrane’s Q for the meta-analysis of all 1,377 HRs was statistically significant (p < .001), and the I² statistic was quite high (I², 99.2; 95% CI, 98.8–99.5), indicating that important moderating variables exist and supporting the decisions to use random-effects models and conduct subgroup meta-analyses. Because the discussion of the meta-analysis focused on HRs adjusted for age and additional covariates, the corresponding heterogeneity test results were carefully examined. As shown in Table 5, the Q tests for these subgroup meta-analyses were statistically significant for only two cases: the 40–49 age group (p = .018) and the two-year follow-up group (p < .001). I² tests for these subgroups indicate heterogeneity was moderate for the 40–49 age group (I², 37.2; 95% CI, 4.2–58.8) and high for the two-year follow-up group (I², 67.0; 95% CI, 43.4–80.8). The results from these two subgroup meta-analyses should therefore be treated conservatively. In all remaining subgroup analyses, however, Q tests and I² tests were nonsignificant, indicating that heterogeneity was adequately accounted for by the use of a random effects model.

Meta-regressions were also used to examine possible sources of heterogeneity in the data. The model fit statistics for Model 3 of Table 6 (R², .590; p < .001 for the Cochrane’s Q of the model) indicate that this model captured a very substantial portion of the heterogeneity in the data. Nevertheless, the unexplained heterogeneity variance component (which measures the nonrandom variance remaining in the model residuals after the effects of all independent variables have been taken into account) for the models shown in Table 6 remained highly significant (each p < .001), confirming the need to use a random effects model for all analyses.

Discussion

The results of the present meta-analyses and meta-regression analyses show that, overall, the relative risk of death for those who lost their spouse was 22% higher than the risk among married persons, among high-quality studies that adjusted for age and additional covariates. The adverse effects of widowhood on mortality, however, were not uniform across all subgroups. As hypothesized, the effects were greater for men (an average increased risk of 27%) than for women (an average increased risk of only 15%), with these risks and the difference between them being more pronounced at younger ages and less pronounced at older ages. By age 90, no difference was found between widowed and married persons among men or women. This aspect of the findings is consistent with Manzoli et al. (2007), who also found no difference in relative risk between men and women at older ages. They are also consistent with Mineau et al. (2002) and Smith and Zick (1996), who found that the relative mortality risk was higher for widowed men than for widowed women and that the relative risk was higher at younger ages than at older ages. Because these previous studies examined only the independent effects of gender and age, however, the documentation of an interaction between gender and age in the present study is a major finding.

A comparison between findings from earlier and more recent studies revealed that the excess risk of mortality among widowed persons has been slowly increasing over time. This both supports the hypothesis put forth earlier in this article and suggests that future meta-analyses should strive to include the results from both early and recent studies in order to evaluate the impact of societal trends. The role of marriage in networks of interpersonal ties has shifted over time (Henrard 1996; Manzoli et al. 2007), and multiple facets of this shift may be reflected in the time trend of increasing HRs. In previous decades, widowed men almost always remarried. Because widowed women have always outnumbered widowed men, the long-term widowed group was predominately female. Declining rates of remarriage in Western nations (Bramlett and Mosher 2002; Bumpass and Sweet 1991) have increased the relative number of men who are in the long-term widowed group in more recent years. Because widowed men have a higher relative risk than do widowed women, the growing proportion of male widowers would cause the overall HR to rise over time. In addition, rates of cohabitation have increased. Research has shown that, controlling for age, those who choose cohabitation tend to have lower SES and therefore are likely to be less healthy than those who marry (Manning and Smock 2002). Presuming that those who cohabitate would have married in previous decades, the growth of the less-healthy cohabitating group increased the average health level of the denominator (married) group over time. This also would cause the overall HR to rise. Factors that help buffer the stress of widowhood have also become less available. Societal decentralization and the geographic dispersal of the family have altered the quantity and quality of interpersonal social support available to widows (Lopata 1978; Popenoe 1993). The erosion of pensions and other similar supplemental sources of income since the 1970s has brought new challenges for widows in maintaining their pre-widowhood material quality of life (Marin and Zolyomi 2010). The loss of buffers like this would also help explain rising HRs over time. Finally, the married population has benefited most from certain health care advances, such as the prevention of childbearing-related deaths. Likewise, compared with widowed persons, married persons have benefitted much more from family-oriented primary care strategies (McDaniel et al. 2005). Positive health changes such as these, which bring about a reduction in the mortality rate for the denominator population, can also explain the increase in the HR over time.

An interesting finding emerges from the current analyses concerning the difference in the effect of widowhood on mortality by the duration of follow-up, which exhibits the pattern hypothesized by Young et al. (1963). Our findings add to the literature (e.g., Nystedt 2002; Thierry 1999) on this topic by providing consensus estimates for the length of the period immediately following widowhood when the surviving spouse is at his/her greatest risk. As seen in Table 4, the risk of mortality is especially high in studies that followed individuals for two years or less. The excess risk decreases substantially among studies with longer follow-up durations, although it remains elevated among studies with up to 15 years of follow-up. Although the onset of widowhood coincided with the enrollment period for only a small fraction of the studies, these findings suggest that the immediate stress caused by widowhood is indeed an important factor in increasing the risk of mortality. Further analyses should investigate the specific physiological and/or behavioral mechanisms that lead to the increase of risk during the first two years after losing a spouse. The results suggest the possibility that different mechanisms dominate the early and later stages of widowhood. Comorbidity effects (Cheung 2000; Elwert and Christakis 2008; Lillard and Panis 1996; Smith and Zick 1996) may combine with stress effects in the early years of widowhood but decline as the influence of the lost spouse diminishes. Practitioners and counselors should focus their attention on the first years of widowhood without losing sight of the continuing risk. Identification of the underlying pathophysiology and determining the changing contributions of physiological and behavioral mechanisms over time will contribute to better targeting of supportive interventions.

Finally, the analysis by region of the world suggests that the risk of death following widowhood is approximately equal in most regions. The magnitude of the effect in the less-developed countries—eastern European countries as well as Bangladesh/Lebanon—is of particular interest. Economic support may have increased importance in poorer countries, where the decrease in income associated with the loss of a spouse may substantially reduce the quality of nutrition and healthcare. We cannot, however, make firm conclusions about the mean effect in developing nations because of the small number of studies conducted in them. Although the mean HR is suggestively high in eastern Europe among HRs adjusted for age alone, there are not enough studies to evaluate whether this pattern would hold among HRs adjusted for age and additional covariates. The results for Bangladesh/Lebanon should be treated with caution as well, considering the small number of studies conducted.

Limitations

A major limitation of the reported analyses, shared by many meta-analyses, is the “file drawer effect,” or more specifically, the nonreporting in the literature of nonsignificant findings (Berman and Parker 2002; Egger and Davey-Smith 1998). This tendency may lead to an overestimation of the mean HRs. Therefore, one should be especially careful in interpreting mean HRs that are relatively close to 1, even when these are significant (as is the case with some of the results in the current meta-analysis). A funnel plot of the log HRs against sample size appears somewhat asymmetric around the mean HR, suggesting the possibility of publication bias (Fig. 3). The results of formal tests for publication bias differ, with Eggers’ test (Egger and Davey-Smith 1998) indicating significant bias (p < .001) and that of Peters et al. (Moreno et al. 2009; Peters et al. 2006) indicating no significant bias (p = .178). The results of the more conservative Eggers test suggest that the HRs that are missing from the analysis are small studies with large HRs. The nature of the bias is such that our results would tend to underestimate the mean HR rather than overestimate it.

A second limitation stems from the nature of the data. Most of the research on widowhood and mortality was conducted in the developed world. Very few studies were conducted in eastern European countries, the Middle East, or South Asia, and there were none from Africa or South America. The sample sizes in the studies from the developing world are small, and conclusions cannot be drawn about potential differences between the developed and the developing world. Also, because most of the results come from the developed countries, the findings from the different analyses presented here should not be extrapolated to populations in developing countries. In addition, important moderators—such as race, ethnicity, and occupational class—were not examined because of data unavailability. Future studies, stratified by these factors or including appropriate interaction terms, are needed.

Conclusion

In conclusion, the analyses reported here show that widowhood substantially increases the risk of death among broad segments of the population. Future research should focus on understanding the health, socioeconomic, physiological, and behavioral factors through which this effect is manifested, especially for younger men and during the first two years following the loss of a spouse. In addition, results from the few studies that were conducted in the developing world suggest that widowed people in these countries may be at greater risk. Further research in developing countries may help explain not only the cultural differences in the experience of widowhood but also the differential mechanisms that mediate the risk of death following widowhood.

Acknowledgments

The authors are grateful for the support provided by Grant HL-76857 from the National Institutes of Health. The funding source had no involvement in the collection, analysis, and interpretation of the data; in the writing of the report; and in the decision to submit the article for publication.

Notes

References