Bequests and the Demand for Life Insurance

Abstract

In this study, we investigate how the desire to leave a bequest influences both ownership of life insurance and changes in the demand for life insurance. We find that the desire to leave bequests is related to the ownership of both term and cash value life insurance. We also find that an increase in the self-reported probability of leaving a bequest is positively associated with the purchase of life insurance for small bequests and negatively associated with the likelihood of dropping term coverage. The results provide further empirical support for the use of life insurance to satisfy bequest motives and offer insight into the ways in which bequest motives may impact life insurance demand.

ACKNOWLEDGMENTS

The authors thank the co-editor, Patrick Brockett, and three anonymous reviewers for their valuable comments and suggestions that greatly improved this article.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

1 Life insurance is one method by which individuals can provide an inheritance. However, bequests may be funded using a variety of approaches and resources, including the use of both financial assets (such as stocks and personal savings) and physical assets (such as real estate). Though a full discussion of other approaches and the factors that impact the decision to leave a bequest is beyond the scope of this article, there is extensive literature discussing alternative methods by which individuals can fund bequest motives. Horioka (2009) notes that bequests in Japan may be left in various forms, including “financial assets such as bank and postal deposits, negotiable securities, etc. as well as real assets such as land, housing, etc.” Individuals may also transfer ownership of a family business in order to provide an inheritance (Getz and Petersen 2004). Other studies have found a link between some household financial factors and bequest motives. For example, Dynan, Skinner, and Zeldes (2004) find that high-income households save more, which they argue may be done in an effort to leave a bequest, and Kopczuk and Lupton (2007) and Ventura and Horioka (2020) observe lower spending or decumulation of wealth among the elderly with a bequest motive. Research has also found that investment holdings are related to bequest motives. For instance, E. J. Kim et al. (2012) find a strong relation between stock ownership and bequest motives, where those with a bequest motive were more likely to have recently purchased stock. Similarly, Inkmann and Michaelides (2012) report that stockholders tend to have stronger bequest motives, which they note is consistent with De Nardi, French, and Jones (2010).

2 These papers are discussed in greater detail in Section 2 (Literature Review).

3 Previous research has studied the factors which may impact the decision to leave a bequest (e.g., Hurd and Smith 1999; Wiepking, Scaife, and McDonald 2012; Lee and Tan 2023). The purpose of this study is not to examine factors that may influence the bequest decision; rather, the objective is to investigate whether the desire to leave a bequest impacts life insurance ownership and how changes in the preference to leave a bequest may influence purchase and lapse decisions.

4 As noted by Hurd and Smith (1999), subjective bequest probabilities are an accurate predictor of actual bequests. This is discussed in greater detail in Section 3.2 (Hypotheses and Empirical Methods).

5 There is also a substantial stream of literature that examines whether bequests are intentional or accidental, as well as the reasons why intentional bequests might exist (e.g., Tomes 1981; Bernheim, Shleifer, and Summers 1985; Hurd 1987; Laitner and Juster 1996; Wilhelm 1996; Lockwood 2018). While these studies are clearly important with regard to the influence that bequest motives may have on savings, consumption, and the decumulation of wealth, they do not have a direct impact on this study given that the HRS respondents self-report some intentional desire to leave a bequest. Further, the current study is interested in the relation between bequests and life insurance demand and not whether the decision to leave a bequest is consistent with life-cycle models of consumption, which has commonly been the focus of these studies.

6 Although the focus of this study is to investigate the relation between the intention to leave a bequest (as proxied by the self-reported probability of leaving a bequest) and life insurance decision making, significant research exists that studies why households choose to leave a bequest, as well as why individuals expect to receive an inheritance. K. Kim et al. (2012) contends that there are four general factors which impact the decision, including (1)(1) $${\mathit{\text{Any Whole}}\mathrm{ }\left(\mathit{\text{Any Term}}\right)}_{i,t}=\alpha +{\beta }_1{\mathit{\text{Prob}}.\mathit{\text{Bequest}}\mathrm{ }10k\mathit{\text{plus}}(\mathit{\text{Prob}}.\mathit{\text{Bequest}}\mathrm{ }100k\mathit{\text{plus}})}_{i,t}+{\mathit{\text{Controls}}}_{i,t}+{\eta }_t+{\epsilon }_{i,t}$$(1) economic resources, (2)(2) $${\mathit{\text{New Whole}}\left(\mathit{\text{Term}}\right)}_{i,t}=\alpha +{\beta }_1{\mathrm{\Delta }\mathit{\text{Bequest}}\mathrm{ }10k\mathit{\text{plus}}\left(\mathrm{\Delta }\mathit{\text{Bequest}}\mathrm{ }100k\mathit{\text{plus}}\right)}_{i,t}+{\mathit{\text{Controls}}}_{i,t}+{\mathit{\text{Life Change Controls}}}_{i,t}+{\eta }_t+{\epsilon }_{i,t}$$(2) family characteristics, (3)(3) $${\mathit{\text{Lapsed Whole}}\left(\mathit{\text{Term}}\right)}_{i,t}=\alpha +{\beta }_1{\mathrm{\Delta }\mathit{\text{Bequest}} 10k\mathit{\text{plus}}\left(\mathrm{\Delta }\mathit{\text{Bequest}} 100k\mathit{\text{plus}}\right)}_{i,t}+{\mathit{\text{Controls}}}_{i,t}+{\mathit{\text{Life Change Controls}}}_{i,t}+{\eta }_t+{\epsilon }_{i,t}$$(3) beliefs about family obligations, and (4)(4) $$\mathit{\text{RMSE}}=\frac{1}{\sqrt{n}}\sqrt{\sum _{k=1}^n{(y_k-\widehat{y_k)}}^2}$$(4) current support exchanges, whereby family expectations regarding bequests may be a function of current support provided by/received within a family.

7 For example, if a household does not have life insurance, that could suggest that either (1)(1) $${\mathit{\text{Any Whole}}\mathrm{ }\left(\mathit{\text{Any Term}}\right)}_{i,t}=\alpha +{\beta }_1{\mathit{\text{Prob}}.\mathit{\text{Bequest}}\mathrm{ }10k\mathit{\text{plus}}(\mathit{\text{Prob}}.\mathit{\text{Bequest}}\mathrm{ }100k\mathit{\text{plus}})}_{i,t}+{\mathit{\text{Controls}}}_{i,t}+{\eta }_t+{\epsilon }_{i,t}$$(1) the household has no bequest motive or (2)(2) $${\mathit{\text{New Whole}}\left(\mathit{\text{Term}}\right)}_{i,t}=\alpha +{\beta }_1{\mathrm{\Delta }\mathit{\text{Bequest}}\mathrm{ }10k\mathit{\text{plus}}\left(\mathrm{\Delta }\mathit{\text{Bequest}}\mathrm{ }100k\mathit{\text{plus}}\right)}_{i,t}+{\mathit{\text{Controls}}}_{i,t}+{\mathit{\text{Life Change Controls}}}_{i,t}+{\eta }_t+{\epsilon }_{i,t}$$(2) the household is not risk averse. The opposite of this argument can also be made for instances where a household owns life insurance.

8 The higher degree of welfare is based on the inclusion of opt-out provisions in the long-term policy. When bequest type (high vs. low) is nonverifiable, low bequest types sell back their policies and purchase new ones at an actuarially fair price, and high bequest types prefer to keep their policies rather than selling them back.

9 While the authors initially argue that either option could be used by the policyholders, for purposes of their theoretical development, they assume that “‘in equilibrium’ the policyholder who retains his bequest motive will always keep the policy rather than surrender and repurchase.”

10 In a similar vein, Li et al. (2022) report that married households are more likely to have life insurance relative to “never-married” households, which provides support for the use of life insurance to address the desire to leave a bequest.

11 While Inkmann and Michaelides (2012) do empirically examine the relation between bequest motives and life insurance demand, there are some specific aspects of the study that make further investigation warranted. First, the authors focus on term life insurance and endowment insurance, where endowment insurance is effectively a form of term insurance with a savings component. As noted by the authors, endowment life insurance is largely used in the United Kingdom because of its tax benefits. Further, these policies can expire, at which point the savings (which consist of premiums paid and reinvested returns) are paid to the policyowner. This differs from whole (cash value) life insurance sold in the United States, which provides insurance coverage for the entirety of one’s life and not simply for a specified period of time (i.e., a term). Given that it is not uncommon for whole life insurance to be used when the insured intends to leave a bequest (Mulholland et al. 2015), the fact that whole life insurance is not addressed in the study is of significant importance. Second, while the study does capture the intent to leave a bequest, it does not consider the probability of leaving a bequest. The authors employ a binary variable equal to 1 for households that have a nonzero probability of leaving a bequest as their measure of intent. The use of a binary variable does allow the authors to capture a desire to leave a bequest, but it does not allow them to consider how likely a household is to leave a bequest. Finally, though the authors investigate how the desire to leave a bequest influences the ownership of term and endowment life insurance, they do not explicitly focus on how bequests influence new life insurance purchases or the lapsation of existing policies.

12 The HRS is sponsored by the National Institute on Aging (grant number NIA u01ag009740) and is conducted by the University of Michigan.

13 An older population has been used in other research that focuses on life insurance. For example, the English Longitudinal Study of Ageing (ELSA) used by Inkmann and Michaelides (2012) consists of respondents who are 50 years and older and their spouses.

14 The RAND HRS files consolidate the biannual core HRS files into a single file, which contains variables that are constructed in a consistent manner across the sample period.

15 Furthermore, as evidenced by the years which are covered by the SCF panel datasets, the use of the SCF datasets would require us to rely on either (1)(1) $${\mathit{\text{Any Whole}}\mathrm{ }\left(\mathit{\text{Any Term}}\right)}_{i,t}=\alpha +{\beta }_1{\mathit{\text{Prob}}.\mathit{\text{Bequest}}\mathrm{ }10k\mathit{\text{plus}}(\mathit{\text{Prob}}.\mathit{\text{Bequest}}\mathrm{ }100k\mathit{\text{plus}})}_{i,t}+{\mathit{\text{Controls}}}_{i,t}+{\eta }_t+{\epsilon }_{i,t}$$(1) an older sample that may yield results that are not necessarily generalizable to the current nature of the industry and policyholder demographics, or (2)(2) $${\mathit{\text{New Whole}}\left(\mathit{\text{Term}}\right)}_{i,t}=\alpha +{\beta }_1{\mathrm{\Delta }\mathit{\text{Bequest}}\mathrm{ }10k\mathit{\text{plus}}\left(\mathrm{\Delta }\mathit{\text{Bequest}}\mathrm{ }100k\mathit{\text{plus}}\right)}_{i,t}+{\mathit{\text{Controls}}}_{i,t}+{\mathit{\text{Life Change Controls}}}_{i,t}+{\eta }_t+{\epsilon }_{i,t}$$(2) data that largely encompass a significant economic recession.

16 SCF respondents are asked “Some people think it is important to leave an estate or inheritance to their surviving heirs, while others don’t. Which is closer to your (and your {husband/wife/partner/spouse}’s feelings? Would you say it is very important, important, somewhat important, or not important?” (Survey of Consumer Finances 2019).

17 Respondents are able to report having one, two, three, four, or “five or more” policies. Because an answer of “five or more” eliminates our ability to accurately capture the total number of policies owned by the respondent (information that is used to assist in determining whether a respondent purchased or lapsed a policy, as well as whether a household owns term life insurance), we omit respondents who report having “five or more” life insurance policies. This accounts for 1.57% of the sample.

18 We remove these respondents as the contemporaneous purchase and lapsing of life insurance prevents us from accurately capturing the type of life insurance that has been purchased or lapsed. These situations are very limited and account for less than 0.60% of the sample.

19 While Fang and Kung (2020) do examine situations in which individuals lose the bequest motive or the bequest motive diminishes over time, their study is largely theoretical in nature.

20 Further, Campbell (1980) contends that an increase in bequest intensity should result in an increase in the “proportion of human capital that will be insured,” which can be achieved through the purchase of another life insurance policy.

21 For instance, Daily, Hendel, and Lizzeri (2008) assume in their models that the need for life insurance ceases once the bequest motive is eliminated. They argue that when the bequest motive no longer exists “the insured prefers to avoid moving resources from the life to the death state.”

22 Term life insurance is a form of insurance that remains in force for a prespecified period of time and that does not include a cash value or savings element. Alternatively, whole life insurance is a form of insurance that remains in force for the entirety of the insured’s life as long as premium payments are made. Whole life policies can also include a variety of features depending on the type of whole life owned (such as ordinary whole life, universal life, variable life, etc.), including the accumulation of cash value, the ability to borrow the cash value, and the potential payment of dividends. While differences exist across the different forms of whole life insurance, the HRS data do not allow us to distinguish across the various forms.

23 The question also includes a clarifying definition, which states, “‘Policies that build up a cash value’ are sometimes called ‘whole life’ or ‘straight life policies.’”

24 Hurd and Smith (1999) find that self-reported probabilities with regard to the desire to leave a bequest in the HRS dataset are accurate predictors of the actual probability of leaving a bequest. A similar assumption is also made by Hurd and Smith (2002) and Inkmann and Michaelides (2012). Furthermore, prior research that gives consideration to a bequest motive has commonly relied on self-reported probabilities of leaving a bequest, including McGarry (2001), Hurd (2009), Huffman, Maurer, and Mitchell (2019), and Gottlieb and Mitchell (2020). While such a measure is commonly used in prior literature, we acknowledge that the overall accuracy of the measure cannot truly be validated until the death of the respondent. However, we contend that this measure is a significant improvement over other binary measures and indirect proxies that have been relied upon in other studies.

25 “Controls” represents a vector of household demographic and financial characteristics that may impact the ownership of whole life and/or term life insurance. They include age, marital status, presence of children, race, ethnicity, health, education, work status, and financial condition. The variables are specifically identified and described in Section 3.3.

26 Respondent-level fixed effects are inappropriate within the context of our models, given the existence of time-invariant control variables such as gender, race, and ethnicity. Wooldridge (2009, 482) provides a detailed discussion of the problems which arise when including fixed effects in the presence of time-invariant variables.

27 We focus specifically on voluntary lapsations. A respondent is considered to have voluntarily lapsed a policy based on the response to a question in the HRS survey that asks, “Was this lapse or cancellation something you chose to do, or was it done by the provider, your employer, or someone else?” A respondent answering that the lapse was the decision of an employer/provider/someone else is not considered to have voluntarily lapsed a policy.

28 Although the HRS specifically asks about policy lapses, we cannot rule out the possibility that the respondent surrendered the policy instead. While similar, a policy surrender occurs when an individual cancels a life insurance policy in exchange for the policy’s cash surrender value, while a lapse simply represents the cancelation of a life insurance policy. The HRS data do not allow us to distinguish between lapse and surrender activity. As such, we follow the approach adopted by prior literature and use the term “lapse” to encapsulate both lapse and surrender activity (e.g., Kuo, Tsai, and Chen 2003; Kiesenbauer 2012; Eling and Kiesenbauer 2013; Fier and Liebenberg 2013; Nolte and Schneider 2017; Cole and Fier 2021).

29 We identify respondents who canceled a policy based on a question in the HRS survey that asks “have you allowed any life insurance policies to lapse or have any been canceled” since the prior survey. A similar question is asked with regard to the purchase of a new policy.

30 As shown by Hwang, Chan, and Tsai (2022), the factors that influence the decision to surrender a policy can differ from the factors associated with the decision to allow a life insurance policy to lapse; however, the HRS data do not allow us to differentiate between these two possibilities.

31 We do not use the percentage change in order to account for the potential that respondents may have a self-reported probability of leaving a bequest equal to zero in the prior survey but a nonzero probability in the most recent survey.

32 “Controls” represent the same vector of household demographic and financial characteristics used in Equation (1)(1) $${\mathit{\text{Any Whole}}\mathrm{ }\left(\mathit{\text{Any Term}}\right)}_{i,t}=\alpha +{\beta }_1{\mathit{\text{Prob}}.\mathit{\text{Bequest}}\mathrm{ }10k\mathit{\text{plus}}(\mathit{\text{Prob}}.\mathit{\text{Bequest}}\mathrm{ }100k\mathit{\text{plus}})}_{i,t}+{\mathit{\text{Controls}}}_{i,t}+{\eta }_t+{\epsilon }_{i,t}$$(1) . “Life Change Controls” include a vector of variables to capture specific household changes during the sample period. This includes changes to financial measures, employment status, marital status, and the presence of children.

33 Equations (1)(1) $${\mathit{\text{Any Whole}}\mathrm{ }\left(\mathit{\text{Any Term}}\right)}_{i,t}=\alpha +{\beta }_1{\mathit{\text{Prob}}.\mathit{\text{Bequest}}\mathrm{ }10k\mathit{\text{plus}}(\mathit{\text{Prob}}.\mathit{\text{Bequest}}\mathrm{ }100k\mathit{\text{plus}})}_{i,t}+{\mathit{\text{Controls}}}_{i,t}+{\eta }_t+{\epsilon }_{i,t}$$(1) through (3)(3) $${\mathit{\text{Lapsed Whole}}\left(\mathit{\text{Term}}\right)}_{i,t}=\alpha +{\beta }_1{\mathrm{\Delta }\mathit{\text{Bequest}} 10k\mathit{\text{plus}}\left(\mathrm{\Delta }\mathit{\text{Bequest}} 100k\mathit{\text{plus}}\right)}_{i,t}+{\mathit{\text{Controls}}}_{i,t}+{\mathit{\text{Life Change Controls}}}_{i,t}+{\eta }_t+{\epsilon }_{i,t}$$(3) were also reestimated using a standard probit model without the inclusion of random effects and the (unreported) results were qualitatively similar to those reported in this study.

34 The RAND HRS dataset captures the race of the respondents. The race variables consist of White/Caucasian, Black/African American, and “Other.” The “Other” category includes American Indian, Alaska Native, Asian, Native Hawaiian, Pacific Islander, or “something else.” Thus, the Black and Other Race variables are relative to those individuals that identify as White/Caucasian.

35 Because of the overlap between race and ethnicity, models were reestimated with the omission of the Hispanic control variable. Results obtained from these alternative specifications are similar to those reported in this study and are available from the authors upon request.

36 HRS respondents can self-report their health status as “Excellent,” “Very Good,” “Good,” “Fair,” and “Poor.” The binary variable representing a self-reported health status of “Poor” is relative to the four remaining potential health statuses.

37 A respondent is considered to have a college education if he/she has attained a(n) associate or bachelor of arts degree, master’s degree, law degree, medical degree, or PhD.

38 Total debt includes all mortgages/land contracts (primary residence), the value of other home loans (primary residence), the value of other debt, and the value of all mortgages/land contracts (secondary residence).

39 Total respondent earnings are equal to “the sum of Respondent’s wage/salary income, bonuses/overtime pay/commissions/tips, 2nd job or military reserve earnings, and professional practice or trade income.”

40 Wealth components of the net worth variable include (1)(1) $${\mathit{\text{Any Whole}}\mathrm{ }\left(\mathit{\text{Any Term}}\right)}_{i,t}=\alpha +{\beta }_1{\mathit{\text{Prob}}.\mathit{\text{Bequest}}\mathrm{ }10k\mathit{\text{plus}}(\mathit{\text{Prob}}.\mathit{\text{Bequest}}\mathrm{ }100k\mathit{\text{plus}})}_{i,t}+{\mathit{\text{Controls}}}_{i,t}+{\eta }_t+{\epsilon }_{i,t}$$(1) the value of the primary residence, (2)(2) $${\mathit{\text{New Whole}}\left(\mathit{\text{Term}}\right)}_{i,t}=\alpha +{\beta }_1{\mathrm{\Delta }\mathit{\text{Bequest}}\mathrm{ }10k\mathit{\text{plus}}\left(\mathrm{\Delta }\mathit{\text{Bequest}}\mathrm{ }100k\mathit{\text{plus}}\right)}_{i,t}+{\mathit{\text{Controls}}}_{i,t}+{\mathit{\text{Life Change Controls}}}_{i,t}+{\eta }_t+{\epsilon }_{i,t}$$(2) the value of any secondary residence, (3)(3) $${\mathit{\text{Lapsed Whole}}\left(\mathit{\text{Term}}\right)}_{i,t}=\alpha +{\beta }_1{\mathrm{\Delta }\mathit{\text{Bequest}} 10k\mathit{\text{plus}}\left(\mathrm{\Delta }\mathit{\text{Bequest}} 100k\mathit{\text{plus}}\right)}_{i,t}+{\mathit{\text{Controls}}}_{i,t}+{\mathit{\text{Life Change Controls}}}_{i,t}+{\eta }_t+{\epsilon }_{i,t}$$(3) the net value of real estate (not the primary residence), (4)(4) $$\mathit{\text{RMSE}}=\frac{1}{\sqrt{n}}\sqrt{\sum _{k=1}^n{(y_k-\widehat{y_k)}}^2}$$(4) the net value of vehicles, (5)(5) $$\mathit{\text{MAE}}=\frac{1}{n}\sum _{k=1}^n\left|y_k-\widehat{y_k}\right|$$(5) the net value of businesses, (6) the net value of IRAs and Keogh accounts, (7) the net value of stocks, mutual funds, and investment trusts, (8) the value of checking, savings, or money market accounts, (9) the value of CDs, government savings bonds, and T-bills, (10) the net value of bonds and bond funds, and (11) the net value of all other savings. Debt components of the net worth variable include (1)(1) $${\mathit{\text{Any Whole}}\mathrm{ }\left(\mathit{\text{Any Term}}\right)}_{i,t}=\alpha +{\beta }_1{\mathit{\text{Prob}}.\mathit{\text{Bequest}}\mathrm{ }10k\mathit{\text{plus}}(\mathit{\text{Prob}}.\mathit{\text{Bequest}}\mathrm{ }100k\mathit{\text{plus}})}_{i,t}+{\mathit{\text{Controls}}}_{i,t}+{\eta }_t+{\epsilon }_{i,t}$$(1) the value of all mortgages/land contracts (primary residence), (2)(2) $${\mathit{\text{New Whole}}\left(\mathit{\text{Term}}\right)}_{i,t}=\alpha +{\beta }_1{\mathrm{\Delta }\mathit{\text{Bequest}}\mathrm{ }10k\mathit{\text{plus}}\left(\mathrm{\Delta }\mathit{\text{Bequest}}\mathrm{ }100k\mathit{\text{plus}}\right)}_{i,t}+{\mathit{\text{Controls}}}_{i,t}+{\mathit{\text{Life Change Controls}}}_{i,t}+{\eta }_t+{\epsilon }_{i,t}$$(2) the value of other home loans (primary residence), (3)(3) $${\mathit{\text{Lapsed Whole}}\left(\mathit{\text{Term}}\right)}_{i,t}=\alpha +{\beta }_1{\mathrm{\Delta }\mathit{\text{Bequest}} 10k\mathit{\text{plus}}\left(\mathrm{\Delta }\mathit{\text{Bequest}} 100k\mathit{\text{plus}}\right)}_{i,t}+{\mathit{\text{Controls}}}_{i,t}+{\mathit{\text{Life Change Controls}}}_{i,t}+{\eta }_t+{\epsilon }_{i,t}$$(3) the value of other debt, and (4)(4) $$\mathit{\text{RMSE}}=\frac{1}{\sqrt{n}}\sqrt{\sum _{k=1}^n{(y_k-\widehat{y_k)}}^2}$$(4) the value of all mortgages/land contracts (secondary residence). Similar to Liebenberg et al. (2012), respondents reporting a negative household net worth are coded with a value of zero.

41 We acknowledge that a given respondent’s financial literacy could potentially influence life insurance ownership (Lin, Hsiao, and Yeh 2017); however, the HRS only includes questions regarding financial literacy in the 2004 wave. As such, we are unable to directly control for the potential effect of financial literacy on life insurance ownership.

42 The Δ Ln Earnings and Δ Ln Net Worth variables are winsorized at the 1st and 99th percentiles in order to reduce the effect of potential outliers.

43 Pairwise correlations as well as variance inflation factors (VIFs) were examined for each set of estimated models. VIFs never exceed a value of 3, suggesting an absence of “harmful collinearity” in our models (Kennedy 2008). However, given the relatively high correlations between the employment (Working and Retired) and income variables in the Equation (1)(1) $${\mathit{\text{Any Whole}}\mathrm{ }\left(\mathit{\text{Any Term}}\right)}_{i,t}=\alpha +{\beta }_1{\mathit{\text{Prob}}.\mathit{\text{Bequest}}\mathrm{ }10k\mathit{\text{plus}}(\mathit{\text{Prob}}.\mathit{\text{Bequest}}\mathrm{ }100k\mathit{\text{plus}})}_{i,t}+{\mathit{\text{Controls}}}_{i,t}+{\eta }_t+{\epsilon }_{i,t}$$(1) models, we reestimated the models excluding the highly correlated variables, and the results for the primary independent variables of interest remained similar to those reported in this study. Results for these alternative model specifications are available from the authors upon request.

44 The five categories or “bins” that respondents are assigned to in Figure 1 are (1)(1) $${\mathit{\text{Any Whole}}\mathrm{ }\left(\mathit{\text{Any Term}}\right)}_{i,t}=\alpha +{\beta }_1{\mathit{\text{Prob}}.\mathit{\text{Bequest}}\mathrm{ }10k\mathit{\text{plus}}(\mathit{\text{Prob}}.\mathit{\text{Bequest}}\mathrm{ }100k\mathit{\text{plus}})}_{i,t}+{\mathit{\text{Controls}}}_{i,t}+{\eta }_t+{\epsilon }_{i,t}$$(1) 0%, (2)(2) $${\mathit{\text{New Whole}}\left(\mathit{\text{Term}}\right)}_{i,t}=\alpha +{\beta }_1{\mathrm{\Delta }\mathit{\text{Bequest}}\mathrm{ }10k\mathit{\text{plus}}\left(\mathrm{\Delta }\mathit{\text{Bequest}}\mathrm{ }100k\mathit{\text{plus}}\right)}_{i,t}+{\mathit{\text{Controls}}}_{i,t}+{\mathit{\text{Life Change Controls}}}_{i,t}+{\eta }_t+{\epsilon }_{i,t}$$(2) 1% to 25%, (3)(3) $${\mathit{\text{Lapsed Whole}}\left(\mathit{\text{Term}}\right)}_{i,t}=\alpha +{\beta }_1{\mathrm{\Delta }\mathit{\text{Bequest}} 10k\mathit{\text{plus}}\left(\mathrm{\Delta }\mathit{\text{Bequest}} 100k\mathit{\text{plus}}\right)}_{i,t}+{\mathit{\text{Controls}}}_{i,t}+{\mathit{\text{Life Change Controls}}}_{i,t}+{\eta }_t+{\epsilon }_{i,t}$$(3) 26% to 50%, (4)(4) $$\mathit{\text{RMSE}}=\frac{1}{\sqrt{n}}\sqrt{\sum _{k=1}^n{(y_k-\widehat{y_k)}}^2}$$(4) 51% to 75%, and (5)(5) $$\mathit{\text{MAE}}=\frac{1}{n}\sum _{k=1}^n\left|y_k-\widehat{y_k}\right|$$(5) 76% to 100%.

45 The Wald χ² statistic presented in Tables 3 through 5 is comparable to the F-statistic frequently reported with linear regression models. Similar to the F-statistic, the Wald χ² indicates whether the overall regression is significant, where failure to reject the null hypothesis indicates that none of the independent variables help explain the dependent variable under consideration.

46 We use the margins command in Stata to compute the probability of life insurance ownership for values of the self-reported probability of leaving a bequest in increments of five, starting from zero (no self-reported probability of leaving a bequest) to a maximum value of 100 (i.e., a 100% probability of leaving a bequest), holding all other variables at their means.

47 Recent statistics provided by the U.S. Department of Labor indicate that 60 to 83% of workers (depending on the type of employer) have access to life insurance through their employer and the take-up rates are high – 97 to 98%. For more information, see https://www.bls.gov/news.release/pdf/ebs2.pd. Additionally, Beam and McFadden (2012) note that “an increasing number of group life insurance plans have been designed to provide postretirement as well as preretirement life insurance coverage.”

48 While Hau (2000) does provide some support for this notion, ultimately Hau’s (2000) findings with regard to the relation between household debt and term life insurance holdings are inconclusive.

49 Note that it is possible for the self-reported probability of leaving a bequest to decrease over time so the range for the calculations starts at –100 (i.e., a respondent reporting zero probability of leaving a bequest in the current period and a 100% probability of leaving a bequest in the prior period) and increases in increments of five to a maximum of 100 (i.e., a respondent reports a 100% probability of leaving a bequest in the current period and a 0% probability of leaving a bequest in the prior period).

50 The number of observations is reduced from 71,118 in Tables 3 and 4 to 20,728 for whole life models and 41,609 for term life models in Table 5 because only those respondents owning a given type of life insurance in the previous survey are included in the lapse models.

51 Some of the methods employed to check model accuracy, validation, and robustness do not permit the inclusion of random effects and instead require the estimation of a more standard probit or logistic regression. Receiver operating characteristics, areas under the curve (AUCs), and k-fold cross-validation are conducted using a standard probit model without the use of random effects. Firth’s penalized maximum likelihood approach is estimated using a logistic regression.

52 Results for unreported models are available from the authors upon request.

53 An important and related issue with respect to model complexity is the inclusion of many binary control variables, which leads to a large number of potential combinations, resulting in a small cell count and ultimately impacting the maximum likelihood estimation. While we acknowledge that such an issue could impact the results and conclusions of this study, as mentioned previously we find similar results when eliminating all control variables with the exception of the bequest variables, which suggests the results are not influenced in a meaningful way by the potential of a small cell count. We thank an anonymous reviewer for bringing this potential problem to our attention and for encouraging additional consideration and analysis.