A Comparison of Risk Willingness Between Same-Sex and Different-Sex Couples: A Quasi-Experimental Approach

Abstract

Household composition in the United States is increasingly diverse; however, research into the diversity of the financial decision maker’s sexual orientation has yet to be explored. This analysis examines whether there are differences in financial risk tolerance between same-sex and different-sex couples using data from the Survey of Consumer Finances. The results from a propensity score matching technique, a Mann–Whitney U test, and interpretations of average treatment effects and average treatment effects of the treated suggest there is a statistical difference in risk tolerance between couples and that, on average, same-sex households are significantly more likely to report higher risk tolerance scores, at the 10% alpha level, when compared to their counterparts. Both treatment effect estimates suggest a high impact of the treatment at the 1% alpha level. This highlights the importance of not assuming homogeneous risk preferences across household types. These findings emphasize the importance of recognizing diversity in household composition. Thus, this study identifies the need for inclusiveness in all segments of financial planning.

1. Introduction

The composition of household decision-makers in the United States has been and is increasingly diverse. The study of that composition through the lens of same-sex household decision-makers has not been as diverse. While the literature on household decision-making is thoroughly abundant, the breakdown and study of the household composition regarding the sexual orientation of those decision-makers is lacking. It is not entirely appropriate to assume that traditional financial planning assumptions in household decision-making, such as risk tolerance measures, are homogeneous across the different types of household compositions. Very few studies have compared the risk tolerance levels of same-sex couples compared to different-sex couples (Hanna and Lindamood 2004; Yin 2020). It may easily be understood why many research areas relating to same-sex couples have been stifled or overlooked. This subject matter has been deemed taboo, criticized, or passed over throughout history. Same-sex couples face differing risks and challenges as compared to different-sex couples. Risks of policy changes and shifting societal norms have historically been a challenge to recognition as a legal entity. As recently as 1996, the United States passed the Defense of Marriage Act, resulting from over a decade of efforts and support from members of Congress. This legislation aimed to maintain the definition of a civil union or marriage to be restricted to “traditional” male-female unions. Not only were marriages and civil unions restricted at the federal level, but benefits at the federal level were also restricted, including Social Security Survivor’s benefits and filing joint tax returns. In 2015, the Supreme Court of the United States struck down all prior legislation restricting the union of same-sex couples, legalized it across the United States, and required the recognition of same-sex marriages across state lines. As a result, the financial planning process for same-sex couples has shifted, as all couples are now given fundamental marital rights that impact the financial planning process.

The impact of the change in this law in the United States has far-reaching implications, particularly in financial planning, where a significant number of decisions are the product of the legality, effort, and desire of both decision-makers in the household. One of the beginning steps in the financial planning process is to determine the risk tolerance(s) of the household decision-makers in an effort to align their investment decisions with their level of comfort. The purpose of this paper is to estimate any potential difference in risk aversion between same-sex couples and different-sex couples. Identifying the differences in risk aversion amongst participants is fundamental to the financial planning process. The theory of decision-making under uncertainty is the underlying theoretical foundation used for this study. This study hypothesizes no difference in the average risk tolerance measure between same-sex and different-sex households.

2. Literature Review

2.1. Conceptual Foundations of Risk Aversion and Financial Risk

Mandrik and Bao (2005) state that the study of risk aversion has implications in many fields, including brand choice, information search, preference for gambles, decision framing, financial portfolio management, and insurance purchases. The study continues to state that because people use different methods to categorize people on different aspects of risk aversion, it is difficult to compare results across studies. Some of the most common methods of measuring risk aversion include choice dilemma studies, self-reporting measures, perceived risk studies, and gambling studies.

Kogan and Wallach (1964) use the Choice Dilemma Questionnaire (CDQ), which relies on measuring attitude toward risk when faced with making a choice in a dilemma. One deficiency of this method is that it does not explicitly address important issues such as construct validation nor issues of reliability such as internal consistency (Mandrik and Bao 2005). Gamble studies often provide unrealistic scenarios and often fail to elicit typical decision behaviors (MacCrimmon and Wehrung 1984). Self-reporting risk measures often fail to provide adequate initial standards by which risk aversion can be measured. These self-reporting measures only ask respondents to assess their overall risk aversion when the study’s focus may be financial or otherwise. It has been empirically proven that people have different risk aversion attitudes towards different scenarios. Perceived risk studies focus on the perceived uncertainty of outcomes and the perceived negative consequences associated with an outcome (Bauer 1960). While the perceived uncertainty of outcomes is directly relative to measuring risk aversion, the perceived negative consequences associated with an outcome may be subject to anchoring bias; therefore, initial conditions for each individual in the study must be accounted for. One can see the problem of heterogeneous variability for each individual.

Risk aversion may be conceptualized as a preference for maintaining a certain level of consumption over uncertain consumption, even if the expected value of the uncertain consumption exceeds that of the level of certain wealth (Finke and Huston 2003). While the risk-averse individual may choose certainty over uncertainty even at a loss, a study by Hanna and Chen (1997) revealed that in all scenarios tested, a portfolio consisting of stocks outperformed a less volatile portfolio over the same 20-year period.

2.2. Financial Risk Aversion in Financial Planning

Assessing financial risk aversion is one of the first steps in the financial planning process. Risk aversion, with respect to investing and financial products, represents an investor’s willingness to take or avoid financial risks. In more technical terms, risk aversion is often described as the marginal change in the slope of the utility function relative to the current level of wealth (Finke and Guillemette 2016). If a person’s utility of the expected value of an unknown outcome is greater than their expected utility from the unknown outcome itself, they are said to be risk averse (Pratt et al. 1964). Conversely, if a person’s utility of the expected value of an unknown outcome is less than their expected utility from the unknown outcome itself, they are said to be risk-seeking (Dyer and Sarin 1982). If there is no difference between the expected value of an unknown outcome and the expected utility from the unknown outcome itself, they are considered to be risk neutral.

In the context of investments, volatility is often measured by the standard deviation of the investment portfolio payouts. A safer portfolio will result in less variation in returns, which implies a potentially narrower payout interval in the future that is relative to the amount of real dollars invested throughout the investment timeline (Finke and Guillemette 2016). A riskier portfolio will result in more variation in returns, which implies a potentially larger future payout interval, relative to the amount of real dollars invested throughout the investment timeline (Finke and Guillemette 2016). Previous studies have examined how gender differences impact risk aversion, including longevity expectations, workforce participation, wealth accumulation, access to higher-paying jobs, access to retirement plans, and education (Heo 2016). When taken together, these factors put women at higher risk than men of having financial problems (Fonseca et al. 2012; Heo 2016).

Grable and Lytton (1999) find that higher levels of education and being male are the strongest predictors of willingness to take financial risks. Alternatively, Lee and Hanna (1995) argue that the relationship between income and willingness to take financial risk may be attributed to the importance of wealth when estimating the negative consequences of financial loss. The consequences in terms of utility loss are greatest among those with the least wealth; however, accumulated wealth may not be the most accurate indicator of ability to withstand financial loss (Finke and Huston 2003). Income and risk aversion have also been shown to be indicators of ability to withstand financial loss. While men are generally more willing to invest a larger portion of their investment portfolio in equities, women often feel more comfortable investing in assets that are subject to less volatility (Heo 2016). According to Grable (2000), this difference in willingness to invest in equities accounts for at least 10% of the gap in lifetime wealth accumulation between men and women.

Measuring risk aversion can be challenging. Financial planners and advisors have access to a plethora of risk tolerance questionnaires. For planners to gain valuable information on their clients, a well-designed psychometric assessment should be conducted instead of a simple survey or interview (Fallaw 2021). Kwak and Grable (2024) review the effectiveness of various risk tolerance measurements in order to predict future investment choices. The researchers find that the propensity measurement technique provides the greatest degree of inter-period stability and provides the most accurate predictions of future investment choice.

2.3. Same-Sex Versus Different-Sex Couples and Household Characteristics

Studies that focus on the comparison of same-sex relationships to different-sex relationships are critical because same-sex couples are demographically distinct from different-sex couples. Individuals who report being in same-sex couples are younger, more educated, more likely to be employed, less likely to have children, and slightly more likely to report identifying as female than individuals in different-sex couples (Gates 2013; Umberson et al. 2015).

Money management also varies between same-sex couples and different-sex couples. Typically, in heterosexual couples, the money management system used is a “pooled” joint account with the higher income earner having more say in how the money will be used. However, Burns et al. (2008) find that same-sex couples tend to use “partial pooling” or “independent management” as the preferred money management system. Both of these systems lead to higher independence among partners, even with a large difference in income between partners.

With respect to the labor force participation of women, it is estimated that 76% of women in same-sex couples are in the workforce as opposed to 62% of women in different-sex couples (Badgett et al. 2021). Antecol and Steinberger (2013) suggest the differential gap is explained by a difference in fertility levels and by the fact that women in same-sex couples are more likely to be primary earners compared to women in different-sex couples. Badgett et al. (2021) find that same-sex couples comprising women are penalized from a double gender pay gap. Their study states that even though labor force participation is higher, female same-sex couples had the lowest personal income on average when compared to all other couple types.

Carpenter and Gates (2008) studied partnership rates of heterosexual and gay men and heterosexual women and lesbians in California, and the results from their study found that heterosexual men and women are overwhelmingly more likely to report being or have been married or in a long-term partnership when compared to their counterparts. Using data from the American Community Survey, Badgett et al. (2021) find that heterosexual men and women are more likely to report having children, overwhelmingly biological and that discrimination against same-sex couples by foster care and/or adoption agencies may also contribute to the discrepancy in the number of children between different-sex and same-sex couples.

In a comprehensive study of financial well-being, Carpenter et al. (2024) find that lesbian, gay and bisexual individuals are significantly more financially vulnerable compared to heterosexual individuals. Using data from the Survey of Household Economics and Decision-making (SHED), the study observes that lesbian, gay and bisexual individuals are significantly worse at managing their finances, more likely to have credit card debt, less likely to have an emergency fund, and less likely to have their retirement savings on track (Carpenter et al. 2024).

Very few studies have focused on analyzing differences in risk aversion between same-sex and different-sex couples. Hanna and Lindamood (2004) studied risk aversion and stock ownership among same-sex couples and different-sex couples using the Survey of Consumer Finances. They find that risk aversion levels are similar between same-sex couples and different sex couples with a male respondent. In addition, same-sex male couples are like different-sex couples in stock ownership. However, female same-sex couples are more willing to take risk compared to all other groups; they display more risk aversion by holding less stock for their given risk tolerance level (Hanna and Lindamood 2004). Yin (2020) find that male same-sex couples have lower risk aversion compared to different-sex couples and female same-sex couples, but the actual investment behaviors displayed by the same-sex and different-sex couples were not statistically different.

As mentioned previously, most empirical studies that include marriage or any type of living arrangement have only taken the traditional definition, male–female, into consideration. As laws in the United States and other countries have changed to provide equal rights and benefits to all with respect to their sexual orientation, more households are reporting different definitions of marriage or living arrangements compared to what was considered “traditional”. To assume, or simply overlook, that the gender composition of household decision-makers is homogeneous or not important is incorrect. A test of differences in risk aversion between same-sex and different-sex household decision-makers is extremely relevant. Several theories exist that may explain why there is a difference between same-sex and different-sex couples with respect to financial risk willingness. Agency-communion theory suggests that men are more driven by goals that further self-interest, and women are more driven by goals that further coexistence. Within comparative households of same-sex vs. different-sex, these suggested gender-driven differences may influence the household’s willingness to take on financial risk. To further the discussion on how a household’s gender composition may impact financial risk willingness, household bargaining theory suggests decisions made for the household are a result of negotiations among those members, and each member has their own interests and differing levels of power, often determined by income, human capital, societal norms, and legal constraints. Labor supply theory discusses differences in gender participation in the workforce. Due to discrimination in wages earned between men and women for similar positions, men have traditionally exhibited a higher participation in the workforce, primarily due to earning a higher wage. Households whose both participants are women may be negatively impacted by a double gender wage penalty, further impacting their financial risk willingness.

2.4. Methodological Considerations

A common dilemma when isolating the comparison of interest between two groups is ensuring the sampling procedure or assignment is random, and the distribution of the group characteristics, and covariates, are balanced. If not balanced, the study will produce estimates that are not representative of the population due to selection bias. Selection bias may occur when observed or unobserved covariates are not accounted for in a statistical model or controlled for in the design, which results in spurious estimates of causal effects (Rosenbaum 2010). In observational studies for causal effects, treatments are assigned to experimental units without the benefit of randomization (Rosenbaum and Rubin 1984). As a result, treatment groups may differ systematically with respect to relevant characteristics and, therefore, may not be directly comparable, resulting in biased estimates (Rosenbaum and Rubin 1984). Observational studies must consistently deal with data that is often collected without random assignment, and many research designs fail to adequately control for selection bias at the analysis level through the balancing of covariates between the groups of interest. Fortunately, there exists a scalar function of the covariates, namely the propensity score, that summarizes the information required to balance the distribution of the covariates (Rosenbaum and Rubin 1984).

With respect to random assignments, many times survey participants cannot be assigned to a group as it is too difficult, unethical, or simply impossible. For example, Higgins et al. (2011) examined how racial profiling impacted the propensity to be searched after a traffic stop between two groups, Blacks and Whites and Hispanics and Whites. In this scenario, it is evident that random assignment to an ethnicity or race is impossible. When random assignment is not feasible or controlled for, the result is a poor research design that has limited capability to make inferential or causal statements and conclusions.

Ideally, an experimental design approach that allows for random assignment can better examine the current vs. the counterfactual, control vs. treatment. This type of design would produce estimates that have much stronger inferential and/or causal interpretations. According to the counterfactual framework for modeling causal effects, the true treatment effect for the group of interest is the difference between the treated outcome and the counterfactual (Holland 1986; Rubin 1974). As it is impossible to observe both the current and the counterfactual simultaneously, a reasonable alternative is to estimate an average treatment effect (ATE) for the population (Rubin 1974; Winship and Morgan 1999). In order to correctly apply propensity score matching to estimate an average treatment effect and average treatment effect of the treated (ATT), several assumptions must be met: the conditional independence assumption (CIA), the stable unit treatment value assumption (SUTVA), and the common support assumption (CSA).

The conditional independence assumption states that assignment to the treatment group is independent of the treatment effect conditional on a set of observed covariates, propensity score (Rubin 1980). If the distributions of the propensity scores are balanced between the treatment conditions, the distributions of the covariates used to obtain the propensity score are also equal between the treatment conditions (Rubin 1980). Formally, treatment assignment and the observed covariates are conditionally independent given the propensity score, that is

x ⊥ z|e(x)

where x = observed covariates; z = assignment condition (treatment or control); e(x) = propensity score (Rosenbaum and Rubin 1983).

The stable unit treatment value assumption (SUTVA) states that the outcome does not depend on the assignment procedure, whether it is randomized or self-selected, and the treatment is the same for all participants in the treatment group (Holmes 2014; Rosenbaum and Rubin 1983). Cox (1958) states that the observation of one unit should be unaffected by the particular assignment of treatment to the other units. This states that when a treatment subject is matched with a control subject, both have the same likelihood of being assigned to either group, and all participants within the treatment group receive the same treatment. This assumption ensures that the estimates for the average treatment effect and the average treatment effect of the treated are consistent across the treated group.

The common support assumption requires enough overlap or common support of the distributions between the treatment and control groups concerning their propensity scores. This assumption postulates that any participants with the same propensity score have the same likelihood of being assigned to either the treatment or control groups based on the covariates used to assign the score. Smith and Todd (2005) explain this by identifying the range of propensity scores with a positive density within both distributions, control, and treatment. This analysis can be estimated in several ways but are not limited to using comparable histograms to visually inspect overlap, comparing minima and maxima values of propensity scores, and using inferential statistics to test significant difference between the distributions of the two groups. This analysis uses inferential statistics to test common support.

3. Data

The data used in this study derives from the Survey of Consumer Finances’ (SCF) 2016 and 2019 waves, which are pooled together. These waves were selected as they both contained a better measurable variation in the dependent variable, risk willingness, that was unavailable in the previous waves. Each wave of the SCF is a cross-sectional, triennial survey supported by the Federal Reserve Board in Partnership with the Department of Treasury. The National Opinion Research Center (NORC) at the University of Chicago has collected data since 1992. Each wave of the SCF randomly selects individuals from different economic strata to participate in the voluntary study that collectively produces nationally representative data sets. Each wave of the SCF oversamples wealthy households in the United States.

The unit of measurement is the primary economic unit (PEU). The PEU is the self-reported, economically dominant individual or couple within the household who reports aggregate household information on demographics, income sources, housing characteristics, and a number of attitudinal and expectation questions (Hanna et al. 2018). Any individual demographic information is representative of the PEU respondent. To control missing data and conceal each respondent’s identity, the SCF introduces five inclusive implicates into each wave for every PEU respondent (Hanna et al. 2018). This study uses all implicates from each wave and applies the appropriate weights to ensure proper variance estimates. This analysis’ total number of initial observations is 10,646 (53,230) PEU responses. The final analysis includes 102 (510) total responses. This reduced number of tested observations may potentially have an impact on the generalizability of the study.

4. Dependent Variable

This analysis’ dependent variable, or outcome variable, is a self-reported measure of financial risk willingness scaled from 0 to 10. The question asks, “On a scale from zero to ten, where zero is not at all willing to take risks and ten is very willing to take risks, what number would you (and your {husband/wife/partner}) be on the scale?” (Survey of Consumer Finances 2016 and 2019). This response is measured as such on a scale from 1 to 10 continuously. Categories 0 and 1 are combined due to the minimal number of respondents who reported in the 0 category. While the survey allows for respondents to report a risk willingness score of 0, inherent economic conditions, such as inflation and its impact on purchasing power, expose all to some level of risk. An average of self-reported risk willingness from the two samples, control and treatment, is used for the Mann–Whitney U Test of comparison in the second stage of the analysis.

In the first stage of the analysis, a control and treatment group must be identified, different-sex and same-sex, respectively. This study’s primary variable of interest is identifying the decision maker(s) based on whether they are a same-sex or different-sex household. Each respondent’s gender is identified, and the gender of their spouse/partner is also identified to create a dichotomous variable representing the household status as a same-sex or different-sex household that the PEU determines. The variable SSH (Same Sex Household) is coded as 0 if they are a different-sex household, control, and SSH is coded as 1 if they are a same-sex household, treatment. There are 310 same-sex households identified in the 2016 wave and 270 same-sex households identified in the 2019 wave. There is a total of 510 same-sex households in the pooled analysis sample. In total, same-sex households comprise 1.8% of the initial analysis sample.

5. Independent Variables

The log of income is measured continuously in this analysis. Income is the total inflow of monies into the household as reported by the PEU, and income does not represent losses due to investments or other balance sheet inventories. Therefore, any value less than zero was omitted from this analysis, and it is interpreted that the PEU did not answer the question. The average income in this analysis is $160,045; the SCF oversamples wealthy households. The use of propensity score matching to establish comparative samples alleviates a measure of inherent bias as a result of this oversampling method.

Gender is a dichotomous variable. Gender in this analysis is coded as 1 if the PEU is male and 0 if the PEU is female. The gender variable is used to help better match groups in the first stage of the matching procedure. Approximately 80% of the PEUs who respond in the sample are male. In the SCF, the question for marital status asks, “What is your current legal marital status? Are you married, separated, divorced, widowed, or have you never been married?” The five categories are divided and combined into two categories: married and not married. The respondent is considered unmarried if the legal status does not equate to marriage. A dichotomous variable is created for marital status, 1 for married and 0; thus, approximately 60% of the sample reports to be married.

The SCF asks, “What is the highest level of school completed or degree you have received?” There are 16 different categories available for response. Categories 1–8 are combined and create the “did not graduate” category, representing a level that has not achieved a high school diploma. Responses 9–16 represent the categories that correspond to the highest degree achieved, which included: high school diploma, some college but no degree, two categories for associate degree, bachelor’s degree, master’s degree, professional school degree, and doctorate. The two categories for an associate degree are combined, and the professional school and doctorate categories are combined. The variable for education is categorical, with seven total categories. A dichotomous variable is created for each category, and the variable is coded as one if the PEU responds to that category and zero otherwise.

This analysis measures age continuously. Each PEU’s age is determined by subtracting their date of birth from the year of the wave in which they participated. The average age of the responding PEU in this analysis is 54 years old. Race in this analysis is measured in four different categories, and a dichotomous variable is created for each category that is coded as one if the PEU responds to that category and 0 otherwise. The four variables in this analysis are White, Black, Hispanic, and Other.

The presence of children in this analysis is a dichotomous variable, and the variable is coded as one if the respondent(s) have children and 0 otherwise. The inclusion of children within a household has been shown to reduce available financial resources, which may constrain a household’s ability to take on additional financial risk. Considerations for a bequest are also controlled for in this analysis. The PEU was asked if they plan to leave an inheritance or estate for their heirs, and if they answered yes, the variable is coded as 1 and 0 otherwise. In many studies, the correlation between children and leaving a bequest is high, but children are not the only recipients of bequests. Many organizations often benefit from bequests also. Based on this sample, a correlation matrix analysis comparing bequest motivations and children reveals a correlation coefficient of 0.0143. This indicates an acceptable level of correlation between the covariates for analysis purposes.

The log of net worth is measured continuously in this analysis. Net worth represents the sum of total assets minus total liabilities. Included in the calculation of total assets are financial and non-financial assets. Calculating total liabilities includes all lines of credit and loans, mortgage(s), etc. A positive net worth results from total assets exceeding total liabilities.

A wave variable is created to identify each wave category used in the analysis. There is a total of two waves, which are coded chronologically. This identifies the point in time that the survey was conducted.

6. Model

This analysis uses a 1-for-1 nearest neighbor without replacement propensity score matching technique, t-tests, and the Mann–Whitney U test (MWU test) to estimate the potential difference in risk tolerance between households that identify as same-sex decision makers and those that do not. Propensity score matching is used to help control selection bias that may arise from the differences in distributions of the independent variables examined when comparing different groups. This allows for correction and minimization of selection bias at the design level instead of covariate adjustment at the analysis level (Table 1).

In addition to propensity score matching and the MWU-test of significance (Table 2), this analysis reports an average treatment effect (ATE) and an average treatment effect of the treated (ATT) (Table 3), which are better treatment effects than what may be obtained from using other methods that do not control adequately for selection bias (Heckman et al. 1997; Higgins et al. 2011).

Propensity score matching is a two-stage process. In the first stage, Stata 15’s psmatch2 command utilizes a probit regression model, estimated via maximum likelihood. However, a logistic regression model is also appropriate to determine the propensity of all respondents to experience a treatment of interest (Higgins et al. 2011). The treatment of interest for this study is same-sex household status, and the control in this study is different-sex household status. In each SCF wave, same-sex households comprised approximately 0.5% and 1.5% of the sample. Propensity score matching is used to help balance the matching covariates between the treatment and control groups that result from non-random sampling. It is possible that the balance of the covariates is not improved, or even worsened, after the matching procedure; therefore, it is necessary to test the balance of the treatment and control groups before and after the matching process by utilizing a measure of standardized bias. The standardized bias is calculated by taking the difference in the proportions of a cofounder variable of interest in each control and treatment group, then dividing by the pooled standard deviation of the two groups and multiplying by 100 (Mayne et al. 2015).

While a t-test can measure the difference between two group means, it does not measure any potential difference in the distributions of the covariates. Measuring the potential imbalance of the distributions of covariates prior to matching identifies the presence of selection bias. Distributions of covariates are likely balanced if there is no relation between the treatment conditions and the covariates or no relation between propensity scores and the covariates (Rosenbaum and Rubin 1984). In most cases, the distributions are unequal when the sample selection is non-random.

The propensity score is determined as such (Rosenbaum and Rubin 1983):

p(T) = Pr{T = 1|S} = E{T|S},

(1)

where p(T) is the propensity to being identified between the treatment group or the control group in this study, T indicates that a household identifies as either a same-sex household or different-sex household, and S is a vector that contains the covariates upon which the two groups will be matched (Higgins et al. 2011). The propensity score is determined via a probit regression that assesses the likelihood of identifying with either the treatment or control groups. The vector of covariates used to match the two groups includes a variety of socioeconomic factors which are based on the responses from the Primary Economic Unit (PEU): income, age, education, net worth, marital status, ethnicity, gender, presence of children, bequest motivation, and a wave variable to control for each wave.

In the second stage of the process, the treatment and control groups are matched to the counterparts, same-sex vs. different-sex status, based upon the propensity scores derived from the first stage to establish two comparative samples. This study uses a 1-for-1 nearest neighbor, without replacement matching technique. After matching, the sample size is reduced to 510 responses, 255 for the control group and 255 for the treatment group. A standardized bias estimate is conducted to estimate the balance of the distributions of the covariates after the matching procedure, reported in Table 3, which measures the mean difference relative to the variability of the values in the covariate distribution (Rosenbaum and Rubin 1985). For continuous covariates, the STANDARDIZED BIAS estimate is measured by dividing the difference in means between the two groups by the pooled standard deviation of the groups multiplied by 100 (Clark 2015). For binary categorical variables, the STANDARDIZED BIAS estimate is the difference between the proportions of the characteristics in each of the two groups divided by the pooled standard deviation multiplied by 100 (Austin 2009).

Once it has been determined that the two groups have been adequately balanced and selection bias has been minimized, this study uses a Mann–Whitney U test as the hypothesis test to compare the average difference in risk tolerances between the matched same-sex households and different-sex households. The Mann–Whitney U test is used instead of a standard t-test as it is the t-test’s non-parametric counterpart and accounts for independent variances for the control and treatment groups.

Additionally, an average treatment effect (ATE) and an average treatment effect of the treated (ATT) are also reported in Table 3. These two statistics provide counterfactual estimates for the treatment and control groups. The average treatment effect (ATE) is defined as the difference in averages between two groups that are estimated when compared to their estimated counterfactual counterparts: what would be the difference in average outcome if all participants were under the treatment assumption and what would be the difference in average outcome if all participants were under the control assumption. These averages are estimated as such: the treatment group is weighted by the inverse of the propensity score derived from the first stage of the analysis to produce a counterfactual estimate for all participants in the treatment group; the control group is weighted by one minus the propensity score derived from the first stage of the analysis to produce a counterfactual estimate for all participants in the control group.

The average treatment effect of the treated (ATT) only considers a counterfactual estimate of the treatment group, not the control group. The ATT is defined as the difference in averages between two groups that are estimated when the treatment group is compared to its estimated counterfactual counterpart: what would be the average outcome if only the treatment group remained under the treatment assumption and what would be the average outcome if the treatment group were under the control assumption. In essence, a measure of the difference between the treatment group’s average risk tolerance and its estimated counterfactual’s average risk tolerance. The counterfactual estimates are calculated as such: the treatment group is weighted by the inverse of the propensity score derived from the first stage of the analysis to produce a counterfactual estimate for all participants in the treatment group.

7. Results

Prior to the propensity score matching process, the mean standard bias estimate was 5.4%; after the matching process, the mean standard bias estimate was reduced to 4.6%. Before the propensity score matching process, the median standard bias was estimated to be 10.1%, and after the matching process, the median standardized bias estimate was reduced to 4.6%. Harder et al. (2010) recommend a less than 20% standardized bias estimate to be considered a suitable balance between comparison groups. Caliendo and Kopeinig (2008) recommend a standardized bias estimate of less than 5% to be considered a suitable balance. While other studies have recommended additional varying acceptable measures of standardized bias, this analysis supports Harder et al. (2010)’s recommendation. A list of covariates with before and after measurements of standardized bias can be located in Table 1. Figure 1 illustrates a comparison of standardized bias for all covariates before and after the matching process.

Table 1. Descriptive statistics (averages) and before/after matching standardized bias measurement.

	Before Propensity Score Matching				After Propensity Score Matching
Covariates	Same-Sex	Non-SS	SB (%)	t-Test	Same-Sex	Non-SS	SB (%)	t-Test
Dependent
Risk Tolerance	5.69	5.39	8.0	1.85	5.87	5.54	13.3	2.18 *
Demographics
Income	11.84	11.76	5.0	1.02	11.84	11.82	1.0	0.15
Gender (Male)	0.50	0.82	−71.7	−18.64 *	0.50	0.51	−1.8	−0.25
Marital Status	0.53	0.62	−18.6	−4.24 *	0.53	0.49	6.4	1.00
Age	50.67	54.23	−25.3	−5.62 *	50.67	50.20	3.3	0.05
Net Worth	12.62	12.87	−8.9	−1.94	12.62	12.83	−2.1	−1.08
White	0.80	0.75	12.4	2.67 *	0.80	0.79	0.5	0.09
Black	0.08	0.11	−9.1	−1.93	0.08	0.06	6.7	1.54
Hispanic	0.07	0.09	−6.9	−1.48	0.07	0.07	−0.7	−0.12
Other (Race)	0.04	0.06	−5.8	−1.31	0.05	0.05	−0.8	−0.14
Children	0.64	0.78	−33.6	−8.91 *	0.64	0.63	−0.4	−0.07
Bequest	0.62	0.69	−16.2	−3.73 *	0.62	0.60	4.6	0.71
Education
Did Not Grad.	0.04	0.07	−17.5	−3.40 *	0.04	0.05	−5.9	−1.07
Diploma	0.15	0.19	−10.1	−2.61 *	0.15	0.13	5.2	0.88
Some College	0.08	0.14	−14.2	−2.97 *	0.08	0.06	6.9	1.29
Assoc.	0.10	0.10	0.3	0.06	0.10	0.13	−9.9	1.49
Bachelor’s	0.28	0.25	5.5	1.25	0.28	0.29	−1.3	−0.21
Master’s	0.21	0.14	19.0	4.65 *	0.21	0.25	−7.7	−1.12
Professional	0.12	0.09	8.5	2.01 *	0.12	0.08	12.2	2.00 *

Data: Survey of Consumer Finance; Number of Obs.: 510; * Indicates significance @ 5%.

Table 1. Descriptive statistics (averages) and before/after matching standardized bias measurement.

	Before Propensity Score Matching				After Propensity Score Matching
Covariates	Same-Sex	Non-SS	SB (%)	t-Test	Same-Sex	Non-SS	SB (%)	t-Test
Dependent
Risk Tolerance	5.69	5.39	8.0	1.85	5.87	5.54	13.3	2.18 *
Demographics
Income	11.84	11.76	5.0	1.02	11.84	11.82	1.0	0.15
Gender (Male)	0.50	0.82	−71.7	−18.64 *	0.50	0.51	−1.8	−0.25
Marital Status	0.53	0.62	−18.6	−4.24 *	0.53	0.49	6.4	1.00
Age	50.67	54.23	−25.3	−5.62 *	50.67	50.20	3.3	0.05
Net Worth	12.62	12.87	−8.9	−1.94	12.62	12.83	−2.1	−1.08
White	0.80	0.75	12.4	2.67 *	0.80	0.79	0.5	0.09
Black	0.08	0.11	−9.1	−1.93	0.08	0.06	6.7	1.54
Hispanic	0.07	0.09	−6.9	−1.48	0.07	0.07	−0.7	−0.12
Other (Race)	0.04	0.06	−5.8	−1.31	0.05	0.05	−0.8	−0.14
Children	0.64	0.78	−33.6	−8.91 *	0.64	0.63	−0.4	−0.07
Bequest	0.62	0.69	−16.2	−3.73 *	0.62	0.60	4.6	0.71
Education
Did Not Grad.	0.04	0.07	−17.5	−3.40 *	0.04	0.05	−5.9	−1.07
Diploma	0.15	0.19	−10.1	−2.61 *	0.15	0.13	5.2	0.88
Some College	0.08	0.14	−14.2	−2.97 *	0.08	0.06	6.9	1.29
Assoc.	0.10	0.10	0.3	0.06	0.10	0.13	−9.9	1.49
Bachelor’s	0.28	0.25	5.5	1.25	0.28	0.29	−1.3	−0.21
Master’s	0.21	0.14	19.0	4.65 *	0.21	0.25	−7.7	−1.12
Professional	0.12	0.09	8.5	2.01 *	0.12	0.08	12.2	2.00 *

Data: Survey of Consumer Finance; Number of Obs.: 510; * Indicates significance @ 5%.

Prior to the matching process, 11 of the 18 matching covariates had a statistically significant difference in means between the control and treatment groups. Additionally, STANDARDIZED BIAS estimates in 3 of the 18 matching covariates were over the 20% acceptable threshold. After the matching process, only 1 of the 18 matching covariates had a statistically significant difference in the means between the control and treatment groups. More importantly, 18 of the 18 matching covariates were below the 20% STANDARDIZED BIAS acceptable threshold. Figure 1 illustrates a comparison of the STANDARDIZED BIAS percentages across covariates before and after the matching process.

After the matching procedure, the Mann–Whitney U test is used to estimate any statistically significant difference in the average risk tolerance between the control and treatment groups after the matching process. The null hypothesis for this analysis is that the average risk tolerance for both the treated and control groups is equal. The average reported risk tolerance for the sample prior to matching is 5.39/10; the average reported risk tolerance for the control group after matching is 5.54/10; the average risk tolerance for the treatment group after matching is 5.87/10. The results of the MWU test are listed in Table 2 and are as follows where Risk Tolerance is the scaled dependent variable representing risk tolerance reported on a scale from 1 to 10. SS is the identifying variable between the control (0) and the treatment (1):

Ho: Risk Tolerance (sss = 0) = Risk Tolerance (sss = 1)

z = −1.867

Prob > |z| = 0.0618

The results from this sample suggest no statistically significant difference in the average reported risk tolerances between the control and treatment groups after matching at the 5% significance level; therefore, the null hypothesis cannot be rejected at the 5% alpha level. Alternatively, the null hypothesis can be rejected at the 10% significance level, and it can be assumed that there is a statistically significant difference in the average reported risk tolerance between the control and treatment groups after matching at the 10% alpha level. The average risk tolerance score for the control group, DS-couples, is 5.54/10, and the average risk tolerance score for the treatment group, SS-couples, is 5.87/10. The difference in reported risk tolerance is 0.33 higher for the treatment group, with a standard error of approximately 0.15.

Table 2. Mann–Whitney hypothesis test results.

Null	Ho: Risk Tolerance (sss = 0) = Risk Tolerance (sss = 1)
Alternative	Ha: Risk Tolerance (sss = 0) /= Risk Tolerance (sss = 1)
	z = −1.867 *
	Prob > |z| = 0.0618

* Indicates significance @ 10%.

Table 2. Mann–Whitney hypothesis test results.

Null	Ho: Risk Tolerance (sss = 0) = Risk Tolerance (sss = 1)
Alternative	Ha: Risk Tolerance (sss = 0) /= Risk Tolerance (sss = 1)
	z = −1.867 *
	Prob > |z| = 0.0618

* Indicates significance @ 10%.

As mentioned previously, the average treatment effect (ATE) is defined as the difference between the average outcome if all participants were under the treatment assumption and the average outcome if all participants were under the control assumption. In this analysis, the ATE is 0.258 with a standard error of 0.063. Based on this sample, it can be inferred that same-sex households have a higher average reported risk tolerance level when compared to their counterparts, significant at the 1% alpha level. Using the ATE and pooled standard deviation of our dependent variable from our samples, a Cohen’s D statistic can be used to estimate the magnitude of the difference in risk tolerance between the two groups. In this analysis, an estimated Cohen’s D statistic of 0.902 indicates that the effect size is large.

Conversely, the average treatment effect on the treated (ATT) does not consider the difference in the potential counterfactuals of all participants. Instead, the ATT only considers the difference in the potential counterfactuals of the participants who report to be in the treatment group. Based on this sample, the ATT is 0.487 with a standard error of 0.134. This estimate is significant at the 1% alpha level. The results of the ATE and ATT are listed in Table 3.

Table 3. Average treatment effect (ATE) and average treatment effect of the treated (ATT) of Reported risk tolerance difference between same-sex and different-sex couples.

Average Treatment Effect (ATE)		Average Treatment Effect of the Treated (ATT)
Estimate	Standard Error	Estimate	Standard Error
0.258 *	0.063	0.487 *	0.134

* Indicates significance at 1% alpha.

Table 3. Average treatment effect (ATE) and average treatment effect of the treated (ATT) of Reported risk tolerance difference between same-sex and different-sex couples.

Average Treatment Effect (ATE)		Average Treatment Effect of the Treated (ATT)
Estimate	Standard Error	Estimate	Standard Error
0.258 *	0.063	0.487 *	0.134

* Indicates significance at 1% alpha.

8. Conclusions

This analysis uses data from the 2016 and 2019 waves of the Survey of Consumer Finances to test the potential difference in risk tolerance between couples who identify as same-sex household financial decision-makers and couples who identify as different-sex household financial decision-makers. The ending analysis sample size is 102 (510) primary economic unit responses.

This analysis utilizes a two-stage propensity score matching and testing process to minimize selection bias resulting from a non-random sample. First, a propensity score balances the covariates used to match the two groups of interest. The propensity scores matching method is a 1-for-1, nearest neighbor matching without replacement technique. Second, a Mann–Whitney U test tests the difference in means between the control and treatment groups. To ensure the matching process is successful in its purpose, the measure of standardized bias is reported twice, once before and after matching, to estimate the reduction in selection bias at the design level.

The results suggest a statistically significant difference in average reported risk tolerance between same-sex and different-sex couples at the 10% alpha level, with same-sex couples having the higher average reported risk tolerance. The average risk tolerance score for the control group, DS-couples, is 5.54/10, and the average risk tolerance score for the treatment group, SS-couples, is 5.87/10. The average treatment effect, a measure of the counterfactual of all respondents, indicates same-sex couples have a 0.258 higher average reported risk tolerance compared to different-sex couples. The average treatment effect of the treated, a measure of the counterfactual only for those in the treatment group, indicates same-sex couples report having a 0.487 higher average risk tolerance in general.

Same-sex couples are demographically distinct from different-sex couples; therefore, it is important not to assume that risk aversion will be the same between the two groups. Measuring risk aversion is an initial step in the financial planning process, and everything subsequent in the process is a function of the client’s risk aversion. Financial planners must adapt and refine their training and approach when working with same-sex couples as ignoring diversity in households may lead to implicit bias and ineffective planning strategies. Planners should begin by assessing each partner’s risk tolerance with a robust psychometric measure and then interpreting the results in the context of the couple’s shared goals. Since same-sex couples have a higher risk tolerance in general, the planner should design investment portfolios that take advantage of growth opportunities while still aligning with time horizons, cash-flow needs, and legal/benefit considerations specific to same-sex households. Financial planners should address potential gaps between stated risk tolerance and actual investment behavior by providing education about volatility, providing clear expectations during economic downturns, and scheduling periodic check-ins. These steps can ensure that both partners remain engaged with the plan and accomplish their goals.

A survey from the Nations Endowment for Financial Education found that 30% of LGTBQ+ individuals have faced bias, discrimination, and/or exclusion in the financial services sector. An additional survey, LGBTQ+ Money Study, found that these individuals are less likely to use important financial tools, do not feel ready to make important financial decisions, more likely to carry financial stress, less likely to be prepared for a financial emergency, more likely to have significant student loan debt and credit card debt, and more likely to face discrimination when compared to their non-LGBTQ+ counterparts. The results from this study suggest that financial planners will benefit by being aware and understanding the nature of diversity amongst their clients. While prior studies have ignored or overlooked such measures of diversity, these results emphasize the importance of propensity score matching in quasi-experimental procedures as a proper way to estimate factors that may contribute to differences in risk aversion measures and other relatable fields of study. The use of propensity score matching to help balance the distribution of comparison variables has significant implications for personal financial planning, as comparison groups of interest are often of different sizes and characteristics. Studying diversity is important to counter the underrepresentation of certain socio-demographic groups. When control measures are not in place to account for such underrepresentation, biased estimates may lead to spurious conclusions and, in fact, may do more damage than good. In all scientific fields of study, it is important to dedicate research efforts towards all groups that make up society.

Obscured cofounders that may not be readily available in the survey data, and therefore cannot be controlled for, have the ability to skew estimates and bias resulting interpretations. These may include identifying social norms, discrimination in attitudes of financial planners, cost of accessing financial advice, sources of financial advice, and estimates of household bargaining power within the household, to name a few. Additionally, the low percentage of survey respondents who were able to be identified as same-sex poses additional limitations to this study. A limitation of this research is that it is difficult to accurately measure a household’s risk aversion when potentially only one person is responding for all household members. Even if two or more risk aversion measures are reported, can it be assumed that a simple average of the two represents the actual risk aversion of the household members? Including the bargaining power relative to financial risk willingness among all household members would help provide more accurate average estimates for both the control and treatment groups. Another limitation of this study is the limited sample size of the control and treatment groups. Those in the survey identified as same-sex households only comprised 1.8% of the initial sample. Pooling more waves together in this analysis may have bolstered the number of same-sex respondents, yet this was impossible as previous waves of the SCF utilized different measures of risk willingness than were used in this study. While both groups were matched on gender as one of the confounding variables, identifying the sex of the same-sex groups and measuring and comparing risk willingness accordingly may give more insight into the differences in reported risk aversion when compared to different-sex couples. Unfortunately, reducing the sample size of the control and treatment groups may have provided less reliable estimates.

Future research should explore differences in risk aversion within the same-sex financial decision-makers’ household, men–men vs. women–women. As mentioned previously, several theories including agency-communion theory, household bargaining theory, labor supply theory, and discrimination theory may offer an explanation as to why there are reported differences in a household’s financial risk willingness when gender is the subject of focus. Our study extends the results of what prior studies grounded in the aforementioned theories continue to suggest. By testing the financial risk willingness between same-sex vs. different-sex couples first, this study sets the pathway for additional research to explore gender preferences in more detail. Additional insight into contributing factors of differences in risk aversion can better assist financial planners in meeting the specific needs of their clients.