2.1 Poverty in widowhood

For many elderly women, widowhood is associated with a substantial reduction in household income from all sources and an elevated risk of poverty (Karamcheva and Munnell, Reference Karamcheva and Munnell2007; Gillen and Kim, Reference Gillen and Kim2009). In the years immediately following the death of their husbands, the average income of elderly women declines by over 50% (Zick and Holden, Reference Zick and Holden2000; Sevak et al., Reference Sevak, Weir and Willis2004). The loss of the husband's Social Security retirement benefit accounts for almost half of this reduction (Zick and Holden, Reference Zick and Holden2000). With few sources of retirement income based on their own employment or earnings history (Munnell and Sass, Reference Munnell and Sass2008), many women experience a sharp decline in their income-to-needs and a marked increase in the likelihood of becoming poor in widowhood (Zick and Holden, Reference Zick and Holden2000; Sevak et al., Reference Sevak, Weir and Willis2004). Indeed, among the elderly, the prevalence of poverty is highest for widows, who comprise 20% of all Social Security beneficiaries but account for 43% of all beneficiaries in poverty.Footnote ¹

However, after decades of decline, the poverty rate for elderly widows is now roughly equivalent to the overall poverty rate for the US population. In 2012, 12.6% of elderly widows lived in poverty, down from roughly 40% in the late 1960s (Social Security Administration, 2014b ; Weaver, Reference Weaver2010). The evidence suggests that an increase in survivor benefits starting in 1972 was responsible for a substantial portion of this improvement (Weaver, Reference Weaver2010). The pivotal 1972 Social Security Amendments increased the ceiling on the amount widows were entitled to receive based on their husbands' earnings record up to 100% of their full retirement benefit. This reform reduced the poverty rate of widows by as much as 33% (Weaver, Reference Weaver2010). Despite this considerable progress, elderly widows still have a disproportionate risk of poverty due to the fact that they are much more likely to live just above the poverty line than the rest of the population (Burkhauser et al., Reference Burkhauser, Duncan and Hauser1994; Smeeding and Sandstrom, Reference Smeeding and Sandstrom2005). Almost four in ten elderly widows have an income below 150% of the federal poverty line (FPL) compared with one in five Americans generally.Footnote ²

Social Security is the primary source of income and financial stability for most widows. On average, widows over the age of 65 receive 65% of their total household income from Social Security. Over 20% of widows depend on Social Security alone for all of their income and over 40% depend on Social Security for 80% or more of their income (Social Security Administration, 2014b ). Depending on the widow's own earnings history, some or all of the Social Security income she receives is in the form of a survivor benefit. Of the 4 million survivor benefit recipients, half receive a survivor benefit based entirely on their spouse's retirement benefit.Footnote ³ These widows are not entitled to their own retirement benefit due to citizenship or an insufficient history of earnings in covered employment.Footnote ⁴ On average, these ‘survivor-only’ beneficiaries receive a monthly survivor benefit of $1,200. The other half are ‘dually entitled’ to a retirement benefit based on their own earnings history as well as a survivor benefit. The average combined benefit paid to dually entitled widows totaled $1,400 in 2012, of which approximately half were survivor benefits (Social Security Administration, 2014a ).

Survivor benefits promote an adequate standard of living by establishing an income floor for widows. The husband's actual retirement benefit determines the threshold at which this floor is set. Depending on the timing of his decision, a widow's monthly income from her survivor benefit can vary by as much as 60%. The earlier he chooses to claim before age 70, the more her survivor benefit income is reduced. According to the Social Security Administration (2014a ), widows whose husbands claim early receive almost $4,000 less each year in survivor benefits compared with widows whose husbands did not claim early. Sass et al. (Reference Sass, Sun and Webb2007) estimate that the value of the average widow's annual survivor benefit falls by 25% – from $18,700 to $13,950 – when her husband claims at age 62 as opposed to age 66. They project that this reduction is, in theory, sufficiently large to move 13% of widows into poverty and 23% below 125% of the FPL, assuming that these widows had little or no income from other sources. Their results, however, do not provide evidence of an empirical link between his claim age and his spouse's well-being in widowhood.

Over time, the share of men claiming at later ages has increased but this trend has been modest. Among recent retirees, almost half claimed at age 62 and only one-quarter waited to claim on or after their full retirement age (Haaga and Johnson, Reference Haaga and Johnson2012) despite the fact that the economic losses associated with early claiming have increased substantially over time as a result of low prevailing interest rates, increased life expectancies, and increases in the delayed retirement credit (Shoven and Slavov, Reference Shoven and Slavov2013). Interestingly, married men appear no less likely than single men to claim early despite the expected effect of this decision on their spouses (Haaga and Johnson, Reference Haaga and Johnson2012; Shoven and Slavov, Reference Shoven and Slavov2014). Of course, job losses, limited financial assets, or poor health may leave many of these men with no choice but to claim their benefits early, though the empirical evidence suggests that even absent these circumstances, men tend to claim early (Li et al., Reference Li, Hurd and Loughran2008; Sun and Webb, Reference Sun and Webb2011). Empirical analyses of the claiming decision reveal that higher educational attainment, sustained labor force activity, and better overall health are consistently associated with delays in claiming among men (Coile et al., Reference Coile, Diamond, Gruber and Jousten2002; Haaga and Johnson, Reference Haaga and Johnson2012; Shoven and Slavov, Reference Shoven and Slavov2014). The key unanswered questions that we seek to address are whether early claiming actually reduces the economic well-being of widows; if so, to what extent; and which types of widows are most affected.

2.2 Benefit calculation

The provisions that determine Social Security survivor benefit eligibility and monthly benefit amounts are complex. To simplify this discussion, we will consider only those eligibility requirements and benefit determinations that apply to non-disabled widows who are not currently caring for a child. In order to receive a survivor benefit, a widow's husband must have received, or been eligible to receive, a retirement benefit based on his earnings history. Second, a widow is not eligible for a survivor benefit until age 60, regardless of whether her husband claimed his retirement benefit before he died. Third, a widow who remarries before age 60 may no longer be eligible to receive a survivor benefit based on her deceased, former, spouse's retirement benefit. Finally, a widow must be the lower-earning spouse to receive a survivor benefit. More specifically, the amount of the deceased husband's retirement benefit must exceed the amount of any retirement benefit his widow is eligible to receive based on her own earnings record.

The monthly survivor benefit amount that a widow is entitled to receive equals 100% of the retirement benefit amount her husband was paid or promised prior to his death. In effect, the husband's retirement benefit acts as a ceiling for the survivor benefit amount. The more substantial his earnings history and the longer he postpones his claiming decision, the higher the ceiling he sets on this amount. Any adjustment in the amount of his monthly retirement benefit applies equally to the survivor benefit. Holding earnings constant, a widow's monthly survivor benefit is smallest when her husband claims at age 62 and largest when he claims at age 70. For each month that he claims early between age 62 and his full retirement age, her survivor benefit, as a share of his full retirement benefit amount, is reduced by 0.56 percentage points, or 6.7 percentage points per year.Footnote ⁵ Between his full retirement age and age 70, the monthly reduction climbs to 0.667 percentage points per month, or 8 percentage points per year.Footnote ⁶ However, the penalty for the husband's decision to claim early on the survivor benefit is subject to a floor of 82.5% of his primary insurance amount (PIA) – the monthly retirement benefit amount the husband receives when he claims at his full retirement age.

In certain circumstances, the widow's age when she claims her survivor benefit can affect the amount she receives. A woman who enters widowhood before her own full retirement age may reduce her survivor benefit by claiming it early, regardless of her husband's claim age. If both the husband and widow claimed early, the widow receives the smaller of two possible survivor benefit amounts: one reduced by her husband's retirement benefit claim age or one reduced by her own survivor benefit claim age. By comparison, the penalty for early claiming by a widow is smaller than that of her spouse. Each month the widow claims before her full retirement age may reduce her survivor benefit by 0.396%, or 4.75% annually. This penalty does not apply to women widowed after their full retirement age – only the husband's claim age can impact the amount these women receive.

Table 1 displays the distribution of monthly survivor benefit amounts associated with the hypothetical claim ages of a husband and those of his widow for her survivor benefit. In this example, we have made the following assumptions: both spouses have a full retirement age of 66, the husband's PIA is equal to $1,300, and the widow is not entitled to receive her own retirement benefit. The amounts in the shaded area of the table indicate the combination of claim ages in which the husband's early claim age is entirely responsible for determining his widow's survivor benefit amount. These shaded cells also represent the point at which widows should no longer postpone their decision to claim their benefit. The amounts in the non-shaded area demonstrate that a widow can significantly reduce her survivor benefit amount by claiming early, even if her husband waited until age 70 to claim.

Table 1.

The effect of widow and spouse claim ages on the value of the survivor benefit

Notes: FRA, full retirement age. The amounts in the table assume the husband's primary insurance equals $1,300 and that both spouses are born between 1943 and 1954 and whose FRA is, therefore, age 66. This table is a modified version of the one presented by Shoven and Slavov (Reference Shoven and Slavov2013).

When the husband and the widow each have their own, earned Social Security retirement benefit, the relative value of each benefit is also an important determinant of a widow's survivor benefit amount. Widows not entitled to their own retirement benefit due to a limited earnings history are entitled to receive a survivor benefit equivalent to 100% of their deceased spouses' retirement benefit. Widows entitled to their own retirement benefit on account of their own earnings history receive a smaller share of their husbands' retirement benefit amount. These ‘dually entitled’ widows receive a survivor benefit equivalent to the difference between their own retirement benefit and that of their deceased spouse. Thus, the value of a widow's survivor benefit falls as a widow's own retirement benefit approaches parity with that of her deceased husband.Footnote ⁷

3.1 Discrete-time hazard model

We use a discrete-time hazard model to assess the relationship between the timing of the husband's Social Security claiming decision and the likelihood that his spouse enters poverty in widowhood. This analytical framework is useful for describing whether, and when, a subject experiences some event (e.g., poverty) when time is measured in discrete intervals (Singer and Willett, Reference Singer and Willett2003). For this purpose, we constructed a panel data set comprised of women who we observe entering widowhood. We re-organized the panel data into a person-period data set in which widow i contributed j rows, where j represents the number of periods in which widow i was observed. The first row of data for each widow represents her initial period of widowhood and the point at which she becomes ‘at risk’ of entering poverty for the first time as a widow. Each widow remains in the estimation sample as long as she is at risk. A widow is no longer at risk (1) after her first transition into poverty or (2) through censoring (discussed below). The result from this framework is a data set that is unbalanced due to the fact that the women have spells of widowhood of differing lengths.

Discrete-time methods are used to examine the hazard rate which is defined here as the conditional probability that the widow enters poverty in period j given that she did not enter poverty in a prior period. Within this framework, the hazard is typically modeled as a logit function, which has the following form:

(1) $${\rm ln}\left( {\displaystyle{{\lambda ({t_{ij}})} \over {1 - \lambda ({t_{ij}})}}} \right) = \beta ClaimAge\_Husband_{i} \,+ \delta X_{ij} + \theta Z_i + \alpha D_j $$

The claim age of the husband of widow i is identified by ClaimAge_Husband and is measured in years and months. The time-varying and time-invariant covariates measured for widow i in time period j are indicated by X and Z, respectively. The time structure of the hazard (log-odds that the event occurs) is measured by D, a vector of period-specific intercepts that indicate each of the j time periods. We are most interested in the parameter estimate β on the husband's claim age. This coefficient indicates the estimated change in the log-odds that a widow enters poverty associated with a 1-year increase in her husband's claim age, adjusting for the other covariates. We report all coefficients from equation (1) as odds-ratios by computing their exponentiated values. We also report the average marginal effect (AME) of ClaimAge_Husband on the probability that a widow enters poverty to provide a more meaningful metric of the magnitude of the implied changes. This conversion allows for a more substantive interpretation of the results from equation (1), while also accounting for the nonlinearity of this relationship, which is inherent in the estimation of a logistic regression model. The relationship implied in the regression results is constrained to be constant across each time period. Alternate specifications that relax this constraint failed to detect a non-constant effect and resulted in more poorly fitting models.

We are also interested in the coefficients on two separate terms (not shown) that identify the interaction of the husband's claim age with relevant measures of the Social Security benefits paid to the widow. The coefficients on these terms assess the differential effect of a husband's claim age on his widow's likelihood of entering poverty according to the share of the total household income she derives from Social Security in widowhood and according to the percentage of her Social Security income she receives in the form of a survivor benefit (based on her husband's earnings history) as opposed to her own retirement benefit. The control variables include demographic, socio-economic, and health status measures to account for the characteristics and circumstances of the widow as well as those for the husband, prior to his death.

4.1 HRS

To estimate our hazard models, we use data from the HRS, a longitudinal panel study that follows a representative sample of elderly Americans as they make the transition from work into retirement. Every 2 years the HRS collects socio-economic, health, and demographic information from more than 26,000 Americans over the age of 50. Using the RAND data file of the HRS waves from 1992 to 2012, we are able to piece together data on Social Security claiming decisions, the onset of widowhood, and the financial circumstances of widowed women who receive survivor benefits. Earnings records from the Social Security Administration (SSA) are linked to data in the HRS for a subsample of respondents. We used data from this linked sample to identify husbands entitled to a retirement benefit and widows entitled to a survivor benefit.

4.2 Sample restrictions

The HRS contains data on 7,726 women who were widowed at some point in their lives. Yet, our estimation sample comprises only a small fraction of these women due primarily to data limitations and, to a far lesser extent, sample restrictions that we impose to isolate the appropriate sample for examining our research question and to address concerns related to endogeneity.

Most widows were excluded from our sample due to data limitations. First, we limit our sample to widows whose husbands died between 1994 and 2012 such that we start with a sample equal to 3,202 widows. This restriction is necessary to observe the onset of widowhood for each widow in the sample and thus avoid significant and intractable problems associated with left-censored data (Singer and Willett, Reference Singer and Willett2003). This restriction also has the benefit of allowing us to incorporate data on the deceased husbands of the widows in our sample into our hazard models as we can observe them in at least one wave of the HRS prior to their deaths. Second, we must observe the husband's claim age for each widow in this analysis and missingness in this variable results in the loss of 1,668 widows, which reduces the usable sample to 1,534.Footnote ⁸ We also must observe the earnings history of both spouses in order to establish whether the widow is the secondary earner and, therefore, entitled to receive a survivor benefit. However, these data are only available for those HRS respondents who consented to having their HRS survey responses matched to their administrative records from the SSA. This restriction further reduces the sample to 802 widows who remain ‘reasonably representative’ of the full HRS sample (Haider and Solon, Reference Haider and Solon2000).

The sample was then restricted to include only those widows entitled to receive a survivor benefit. To identify these widows, we compared the expected present value of the Social Security retirement benefit promised to each spouse based on their own earnings history data from the SSA.Footnote ⁹ Only those women with a relatively less substantial earnings history compared with their husband were included in the sample, which consequently equals 625.

Three additional restrictions are required due to data limitations in the HRS. The first restriction excludes women widowed before their full retirement age, which reduces the sample to 348. This restriction is necessary because the HRS data do not distinguish between the claim ages for retirement, spousal, or survivor benefits for secondary earners entitled to each. Without this information, we cannot establish whether women widowed before their full retirement age receive a survivor benefit amount determined by their husband's claim age or their own survivor benefit claim age.Footnote ¹⁰

The second restriction removes widows entitled to one of several survivor benefits bequeathed to her from previous marriages. As explained above, a widow who remarries before age 60 forfeits her entitlement to a survivor benefit based on the earnings record of her previous husband. However, women with multiple spells of widowhood may be entitled to receive one of multiple survivor benefits when at least one of her marriages began after her 60th birthday. In these instances, she must choose between the survivor benefits bequeathed to her from previous marriages. However, it is not possible for us to determine which deceased spouse is the source of the survivor benefits paid to these women. To ensure that we have the claim age of the correct husband, we remove these widows from the sample, which then equals 332. In other words, our sample is limited to women entering the spell of widowhood in which they are first eligible to receive one survivor benefit.

The third restriction removes the widows of husbands who received disability benefits. The claim age in the HRS can refer to either disability or retirement claiming. Workers who receive disability have no official retirement benefit claim age because the SSA automatically converts their disability benefit to their retirement benefit when recipients reach their full retirement age. To avoid conflating disability claiming with retirement benefit claiming, we remove widows whose husbands received disability benefits at any point in time before dying such that the sample equals 272. We also removed widows with spouses who reported receiving Social Security benefits before age 61 and 9 months, the age when workers are first eligible to file for their benefit, reducing the sample to 251. These 21 spouses may have received Social Security benefits before age 62 due to disability or their reported claim ages may be subject to measurement error.

We also imposed one last sample restriction to reduce, but not rule out, the potential for omitted variable bias in our estimation. We remove those women who were already living below the poverty line in the period prior to the one in which they entered widowhood such that the sample is reduced to 236. These widows may have continued to live in poverty regardless of when their husbands claimed Social Security, in which case, our hazard models would mistake the continuation of their pre-widowhood poverty trend for an initial poverty transition in widowhood. Also, if early claiming is more prevalent among already poor households, then the inclusion of previously poor widows may bias our results towards finding a higher (or lower) correlation between a husband's claim age and his widow's poverty risk.

Finally, missingness in the variables included in the specification of equation (1) results in a final sample for analysis comprises 197 widows. The effect of each restriction on the sample size is summarized in Table 2A of the online Appendix.

4.3 Estimation sample

The data for these widows are pooled into a person-year data set containing 482 non-unique observations. We follow these widows through their first four periods of their widowhood. We truncated our observation window because we observe very few widows for an extended period of time. As additional waves of the HRS become available, a more comprehensive analysis will become possible as the number of women with longer widowhood durations increases.

Table 2 summarizes the distribution of poverty entry rates by period for widows in the estimation sample. The rates in the table describe a widow's unadjusted likelihood of first entering poverty in a given period conditional on being at risk of entering poverty at the beginning of that period, often referred to as the hazard rate. Examining the first row of the table, we see that of the 197 widows in the sample, 20 entered poverty in the first interval of their widowhood. Looking at the results in the second row, we see that the number of widows at risk of entering poverty has fallen to 138. This decline represents removal of the 20 widows who entered poverty in the first interval from the estimation sample as well as the 39 widows considered censored at the time of the first follow-up. The censored observations in this analysis are those widows we do not observe entering poverty because it has not occurred by the end of the observation period or due to some other intervening event such as death, remarriage, or attrition from the HRS survey sample.

Table 2.

Hazard of falling into poverty among female widows who receive a survivor benefit based on their spouse's earnings history, 1994–2012

The third column of Table 2 presents the proportion of widows known to have not become poor in previous intervals, who then became poor. Looking again at the first two periods, we see that 10.2% of widows entered poverty in their first 2 years of widowhood. In the second period, 4.3% of the 138 non-censored widows at risk of becoming poor for the first time as a widow entered poverty in the second period. The distribution of these hazard rates over time describes the baseline hazard function defined by the coefficients on the period identifiers in equation (1). The hazard of poverty entry appears to be the highest for those widows immediately following their entry into widowhood. The hazard rate then declines and then rises again as the duration of widowhood increases. This implies that widows are at the highest risk of entering poverty immediately following the death of their husband and over extended durations of widowhood. This pattern is consistent with those from previous studies tracking the change in poverty rates among women at the onset of widowhood and over time (Zick and Holden, Reference Zick and Holden2000; Sevak et al., Reference Sevak, Weir and Willis2004).

4.4 Dependent and explanatory variables

The dependent variable in our analysis, poverty, indicates the occurrence of the event of interest, whether or not a widow's income falls below the poverty threshold in a given wave. The RAND Data file identifies the poverty status of the households for the 5 biennial waves between 2002 and 2012. Following the steps outlined by RAND, we determined the poverty status of each household in the HRS from 1992 to 2000. The dichotomous variable poverty is coded 1 if the widow's household income fell below the applicable poverty threshold set by the US Census in the previous year and 0 otherwise.

Our explanatory variable of interest, ClaimAge_Husband, measures the age of the widow's husband when he claimed Social Security before his death. To construct this continuous measure we divided the husband's claim age, measured in months, by 12.Footnote ¹¹ We calculate a similar measure of secondary interest for each widow in the sample, ClaimAge_Widow, using her reported claim age. This measure can refer to the widow's claim age for her own retirement benefit, her spousal benefit, or her survivor benefit; the HRS data does not distinguish between them. On account of our sample restrictions, the widow's claim age cannot affect the survivor benefit amount she receives. However, it can affect the value of any retirement benefit she receives. We include ClaimAge_Widow to disentangle changes in Social Security income associated with changes in ClaimAge_Husband on her survivor benefit amount from those changes associated with the impact of her claim age on her own retirement benefit, should she receive one.

4.5 Control variables

Of course, many of the same factors that affect a husband's claim age also influence the likelihood that his widow will become poor in widowhood. For example, married couples with a history of low-earnings will have lower Social Security benefits and may also be more inclined to claim early due to a further reduction in their earnings potential in old age and inadequate retirement wealth. In this case, it is the couples' history of low-earnings or the factors that contributed to low earnings, such as low education, poor health, and/or unemployment that explains why widows in these households enter poverty in widowhood as opposed to the timing of the husbands' claiming decision. With this understanding, we control for the relevant characteristics of the widow's husband and their household, measured prior to his death, in addition to those measured for the widow. The variables for the husband are measured once at the most recent follow-up interview prior to his death. The variables measured for the widow are a combination of lagged and contemporaneous time-varying and time-invariant covariates. Despite our effort to control for as many of these confounding factors as possible, our research design is not sufficiently rigorous to warrant a causal interpretation of our results. Table 1A in the Appendix provides a detailed description of all covariates included in this analysis.

When specifying our model, we are careful to avoid overcontrolling for those factors that may mediate the relationship between ClaimAge_Husband and poverty. Most notably, we do not control for Social Security income because we assume that this factor is the only plausible mechanism through which ClaimAge_Husband affects poverty. More precisely, a husband's claim age affects the value of his widow's survivor benefit and therefore helps determine the amount of Social Security income she receives after he dies. The change in her Social Security income attributable to the effect of his claim age on her survivor benefit may, in turn, affect her poverty status in widowhood. If we were to adjust our estimates for her Social Security income by controlling for it in our model, we would remove the only variation relevant for evaluating our relationship of interest, the indirect link between his claim age and her poverty risk. Instead, we control for the widow's income from all sources other than Social Security.

We also control for the relative importance of Social Security to the widow's total household income as well as the share of her Social Security income that she receives in the form of a survivor benefit. To estimate these measures, we calculate the ratio of her Social Security income to her total household income (SocSecInc/TotInc) and the proportion of her Social Security income that she receives in the form of a survivor benefit (SurvivorBen/SocSecInc).Footnote ¹² As these variables approach parity, the widow's dependence upon Social Security and the survivor benefit increases, respectively, and the husband's claim age becomes a more important determinant of her total household income and her risk of entering poverty in widowhood. To determine whether the relationship between the husband's claim age and his widow's poverty risk is a function of SocSecInc/TotInc and SurvivorBen/SocSecInc, each of the measures is interacted with ClaimAge_Husband.

4.6 Descriptive statistics

Table 3 displays the sample means of the explanatory variables included in this analysis. The means in the second column of the table are estimated using the widow's data from the first period of her widowhood. The last column includes sample means for a more representative sample of widows in the HRS. Looking at the means, a few notable characteristics of our sample emerge. First, neither spouse postpones their claiming decision long after turning age 62 which is consistent with the distribution of claim ages in the population. Second, our sample is overwhelmingly white. One reason for this is that African American women are much less likely to be eligible for survivor benefits on account of their marital history (Iams and Tamborini, Reference Iams and Tamborini2012). In addition, African American women were disproportionately restricted from the estimation sample because they were more likely to enter widowhood prior to their full retirement age than their white counterparts due to the black-white life expectancy differentials. While necessary to avoid potential confounds, this restriction also increased the average age of the widows in the sample. Another notable feature of the sample is that widows tend to receive a majority of their income from Social Security and that this income also tends to be in the form of a survivor benefit. On average, widows in the sample receive 68% of their income from Social Security, more than half (55%) of which they received in the form of a survivor benefit. These averages are consistent with those reported by Social Security Administration (2014a ) for the population of widows who received this benefit in 2012. The dependence of widows on these benefits reflects, in part, the lower labor force participation of the widows relative to their deceased spouses. For the most part, the estimation sample is similar to the larger, more representative sample of widows in the HRS but the exclusion restrictions appear to result in a more advantaged sample of widows, a matter to which we return later.

Table 3.

Characteristics of widows in their first year of widowhood

¹ This sample includes widows who receive a survivor benefit and were widowed after their full retirement age.

² The means for the husbands were derived from the last wave in which they were observed prior to their deaths.

³ This sample includes all widows observed entering widowhood in the HRS with data linked to the Social Security Administration and with non-missing values for the variables in the table. To maximize the sample size of this group we do not report the means for the variables with high levels of missingness such as the claim ages of each spouse.

6.1 Endogeneity and sample size issues

In this paper we have provided suggestive evidence that a spouse claiming his Social Security benefits early is associated with an increased probability that their widow will enter poverty. We are limited in our ability to make a more causal statement by a relatively small sample and concerns regarding endogeneity of the claiming decision. While the sample size is small, there is no larger data set that includes the variables requisite to consider this question – panel data including data on retirement age, lifetime earnings, etc. for elderly individuals. Additionally, small sample sizes are common among analyses of recently widowed women (Holden et al., Reference Holden, Burkhauser and Feaster1988; Zick and Smith, Reference Zick and Smith1991; Zick and Holden, Reference Zick and Holden2000; Sevak et al., Reference Sevak, Weir and Willis2004; Sass et al., Reference Sass, Sun and Webb2007; Gillen and Kim, Reference Gillen and Kim2009). As noted, the small sample size is more the result of data limitations in the HRS than the inclusion criteria we imposed to reduce bias. However, we did prioritize internal over external validity when the analysis required a trade-off between the two because this analysis probes whether a husband's claim age is a determinant of his widow's poverty status in widowhood. The combination of data limitations and our inclusion criteria resulted in a less representative estimation sample that comprised widows who were older, healthier, and had a higher socio-economic status compared with the full sample of widows in the HRS. These restrictions result in a relatively affluent, less generalizable estimation sample but it is not clear whether the relationships established here would be strengthened, attenuated, or absent in a more representative sample.Footnote ¹⁴ On the one hand, it is possible that the financial well-being of the widows in our sample is particularly sensitive to the timing of their husband's claiming decisions. Wives tend to reduce their labor supply as their husband's earnings increase (Blau and Kahn, Reference Blau and Kahn2007). This means the husband's earnings advantage, while evident across the income distribution, is largest for women with higher-earning husbands (Winkler, Reference Winkler1998). As a result, the more affluent widows in our sample may receive a larger share of their Social Security income in the form of a survivor benefit than their less affluent counterparts. On the other hand, the women in this sample tend to be less dependent on Social Security for their income in widowhood (see Table 3). Taken together, a less affluent, more representative sample of elderly widows may receive less of their Social Security income in the form of a survivor benefit but, at the same time, they may be more dependent upon the amount they receive to maintain an income considered minimally sufficient to meet their needs. Therefore, it is not clear a priori how the results would change using a more representative sample.

More worrisome than sample size is the potential endogeneity in claiming decision-making that is not due to the included control variables. For example, it is possible that the observed relationship in our analysis is confounded, not by the circumstances of the husband prior to his death, but by micro- and macro-level factors prior to age 62 such as health and market wages. Unfortunately, too few waves of data are available in the HRS to control for these potentially confounding factors as well. As a result, we err on the side of caution in interpreting our estimates as suggestive rather than causal and assume that our point estimates represent the upper-bound of any true effect.