The Many Measures of Wealth: Income versus Net Worth
Having described patterns of fertility, we then turned our attention to how we would measure wealth in our analyses. Wealth has many meanings, and different measures of wealth could influence fertility outcomes in different ways. This, then, is another reason why we selected the NLSY79 cohort for our analyses: it represents a rich source of information on economic resources (Zagorsky 1999). Each wave includes detailed questions on various sources of income, as well as information on assets and debt. More specifically, assets minus debt is calculated to construct a variable that represents an individual’s “net worth.” This “net worth” is what economists have in mind when they talk about “wealth” (e.g., Emmons and Noeth 2015). In what follows, we refer to this as “net worth” because the term “wealth” has a broader meaning in human behavioral ecology (e.g., Borgerhoff Mulder and Beheim 2011), and we do not wish to confuse the two. Although income feeds into net worth (Semyonov and Lewin-Epstein 2013), they are not necessarily strongly related: for a given income level, there can be stark variations in net worth (Braveman et al. 2013; Jez 2014; Keister 2000, 2003; ESM Figure 3). Indeed, inequality in net worth is much more pronounced than inequality in earnings (Semyonov and Lewin-Epstein 2013).
A high net worth is considered more beneficial to individuals than receiving a certain level of income because, among other things, it increases financial stability by providing a buffer for emergencies, it can be passed on to subsequent generations, and it can be invested in ways that generate further wealth (Jez 2014; Keister 2003). Importantly for our purposes here, income and net worth (sometimes also referred to as the flow and stock of resources, respectively) can have different effects on particular kinds of outcome measures, such as health and education (Braveman et al. 2013; Jez 2014). It therefore seems reasonable to test whether the same is true of reproductive outcomes, especially as net worth is not often measured or included in studies of health, education, or fertility (Braveman et al. 2013; Jez 2014; Stulp and Barrett 2016a). This is also true of most studies in the human evolutionary sciences. Here, we examine the effects of both income and net worth on fertility decisions using both cross-sectional and longitudinal analyses.
Variable Construction
For our analyses, we constructed a value of respondent income on the basis of income generated from the labor market—in other words, income from wages, salaries, and tips, from business and farm work, and from military income (sometimes referred to as “earnings”; see Breen and Chung 2015 for a similar strategy). Note that this excludes income from other sources, such as alimony, child support, unemployment compensation, or food stamps, because (a) they were not consistently measured across years and (b) the amount of money received from some of these sources depends on household characteristics, including the number of dependent children. Our measure therefore contrasts with a previous study that used income from wages only (Hopcroft 2014). In addition, we used income data from 1983 onward only, as this was the year in which questionnaires were standardized and identical across all respondents.
For the years 1985–1990, 1992–2000, 2004, 2008, and 2012, respondents were asked questions about their assets as well as income. In addition to providing information on “raw assets,” these data are represented in the database as a constructed variable of “net worth.” Here, for each survey year, the total value of assets a respondent possessed (including the worth of their home, cash savings, stock and bond portfolios, estate, business, and automobile assets, along with retirement and other saving plans) has been subtracted from their debt (including mortgage debt, property, business and automobile debt). Net worth can therefore take a negative value for those with more debts than assets. This newly constructed variable was checked for inconsistencies and missing data were imputed (details in the ESM). Net worth thus provides us with a measure of economic “wealth” and illustrates how the rich data characteristic of large-scale databases can be exploited to generate measures that are meaningful and directly relevant to the target question.
Top-Coding, Response Bias, Retention Bias, and Reliability of Measures
For reasons of confidentiality, all income variables in the NLSY79 are top-coded (i.e., data points above an upper bound are censored), and in varying ways (Zagorsky 1999; e.g., the highest 2% of incomes are given the average value of that 2%). This artificial truncation of the wealth continuum obviously presents a limitation on any analysis performed using these data. Since 2% is a very small fraction, however, and since we are not specifically interested in the very wealthy, but in the entire wealth distribution, this limitation is unlikely to distort our findings, and we decided to include the top-coded values in our analyses. Nevertheless, top-coding is one of the reasons why income variables need to be transformed in order to allow comparability across survey years and respondent age (see below). Other studies using the same dataset have variously discarded the top-coded individuals (Breen and Chung 2015; Zagorsky 1999) or retained the top-coded values (Hopcroft 2014) during analyses.
Another problem pertinent to our analyses (and to secondary databases more generally) is that of response bias. This can be a particular problem when dealing with sensitive issues, including measures of net worth and income (Ross and Reynolds 1996). Zagorsky’s (1999) detailed investigation of the income and assets data in the NLSY79, however, revealed that only a very small proportion of individuals either refused to answer or did not know the specific value of their assets. Moreover, such refusal was less likely for assets such as mortgages, vehicles, and possessions, and more likely for items such as cash savings, stocks, and bonds. There were also associations, albeit small, between net worth and the likelihood of refusing to answer or not knowing the answer to a question. In contrast to this slight reluctance to report on their assets, those who possessed more assets were more likely to participate every year than those with fewer. Thus, although there is clear evidence for some response (and retention) biases, high retention rates combined with high response rates meant that little imputation of values was needed (Zagorsky 1999).
It is also important to mention those measures of income we did not include in our analyses. Most notably, the NLSY79 incorporates a constructed variable of “household income,” which combines the income of all respondents related by blood or marriage residing in the household (i.e., income from siblings, parents, and/or children living in the household, as well as spousal income). It does not, however, include partner income in the case of unmarried couples, even when the partner resides in the same household: the criterion for inclusion in the household measure is that people be related by either “blood or marriage.” For our purposes, this is problematic because partner income is likely to be more relevant to understanding childbearing decisions than the income of other relatives in the household. Although there are separate variables for partner and spousal incomes, they are not measured consistently, particularly for partners. Moreover, respondents were less certain about their spouse’s income than about their own income, and even much less certain about their non-spousal partner’s income. Given this, we focused only on the income and net worth of the main respondent and did not include information on partners or spouses. This is, of course, a clear limitation, because childbearing decisions are likely to be dependent on the income and net worth of both partners in a relationship.
Ethnic Differences in Income and Net Worth
Examining the patterns of income and net worth across the life span, we find stark differences across ethnicities (Fig. 3). These replicate earlier studies from both the NLSY79 and other US samples (Emmons and Noeth 2015; Keister 2000). White respondents of both sexes acquire much greater net worth than black and Hispanic respondents, although there is substantial variation within whites. These ethnic differences in net worth are argued to result from differences in inheritance patterns (Keister 2000), education (Emmons and Noeth 2015), financial decision-making (Emmons and Noeth 2015; Keister 2000), discrimination, cumulative disadvantage, and early learning experiences (Emmons and Noeth 2015). White men also have much higher income compared with the other ethnic groups. For women, differences in income are much less pronounced across ethnicities, and women in general earn much less than their male counterparts.
These findings are important to consider in light of our research question. First, increases in income and net worth differ by sex and ethnicity in complex ways. For example, we might expect net worth to play a larger role than income in white women’s reproductive decision-making because (a) there is large variation in net worth and (b) average levels of net worth are much higher in white women than in women of other ethnicities, whereas this is much less pronounced in the case of income. Second, although we should be cautious of committing the ecological fallacy (Pollet et al. 2015), when we examine these aggregate patterns, it is clear that the groups with the highest income and net worth (white men and women) also report the lowest number of children (see Figs. 1–3). This suggests that other factors exert an influence on fertility besides wealth, or at least “material wealth,” which again implies that different ethnic groups may follow different reproductive strategies.
The Cross-Sectional Analysis of the Association between Wealth and Lifetime Reproductive Success
Following this descriptive assessment of the available data, we now turn to our cross-sectional analysis of wealth in relation to LRS: a question that can only be dealt with effectively using cross-sectional data, given the nature of the dependent variable. Another reason for performing such an analysis is because it permits a direct comparison with another recent analysis of the wealth-fertility relationship in the NLSY79. Thus, we can provide a neat illustration of how differences in sample selection, along with variable selection and construction, can affect the outcomes of analyses.
Specifically, Hopcroft (2014) presented a cross-sectional analysis of NLSY79 data in which she showed that income in 2010 (when respondents were between 45 and 53 years old) was positively associated with the number of children ever born in men, but negatively associated in women. Here, we extend these analyses and offer our own cross-sectional analyses of the same data, in which we consider income across the life span. This is because, although the number of surviving children is appropriately measured at the end of the reproductive life span, the use of a single income measure taken at the same point in time is less appropriate. That is, income in 2010 was used as a predictor of the number of children born many years earlier, when it can have no causal influence on the decision to bear children. The use of such a measure also makes the implicit assumption that income in early life is strongly correlated to income in later life, which is not the case in these data (see ESM Figures 1–2). Given that income and net worth are measured in many more rounds prior to 2010, the use of only a single measure to represent these variables means a large amount of available information goes to waste. Finally, selecting one particular year allows the researcher unwarranted “degrees of freedom” in the choice of the year selected. As a result, we decided to investigate the association between wealth across the life span and lifetime reproductive success in order to provide a more comprehensive analysis. That is, rather than selecting a single year’s income and net worth as representative of an individual’s material wealth, we constructed income and net worth measures for each wave of data collection. Thus, our analysis represents a series of cross-sectional “snapshots” taken across the life span, as we explain below. Moreover, we simultaneously assess the effect of income and net worth by including both variables in all statistical models. This, in turn, required us to address two additional analytical issues: (1) the use of income and net worth in our statistical analysis and (2) selection of, and controlling for, confounding variables.
Constructing the Income Variable
The top-coding of income (see above) along with the typical skew of income distributions (see Fig. 3), economic inflation over time, and the increase of income with age means that the income variable required transformation in order to generate consistency across years and ages before it could be used as a predictor. To account for these effects, and to ensure we could include the top-coded 2% in our sample, we converted this variable into quintiles within ages. A value of 1 for income, for instance, means that this individual was in the lowest 20% of his or her income group relative to individuals of the same age, whereas a value 5 indicated that the individual was in the highest 20% income group. We performed this standardization separately for the sexes because of the large difference in income between men and women. We performed a similar standardization for net worth. Quintile measures were included in our statistical models as continuous variables. Quintiles were used because a larger number of categories (e.g., deciles) was not feasible because too few cases were available or there was too little variation in income or net worth to enable them to be grouped into more categories; see Grundy and Read (2015) for similar analytical strategies. Hart (2015) also used quintiles but included them as a categorical variable in his analyses. In our case, sample sizes were too low to follow this strategy.
Choice and Justification of Confounding Factors
In order to assess the effect of income and net worth on LRS, it is necessary to control for other variables that affect either the independent or dependent variable(s). Here, we included the following factors: country of birth (US or other), religion, whether the respondent lived in a rural or urban environment at age 14, region of the United States where the interview was held (Northeast; North Central; South; West), the number of siblings in 1979, maternal education, and respondent education. For reasons discussed above, we analyzed different sexes and ethnicities separately.
Religious differences are well-known to influence fertility (e.g., McQuillan 2004), as are geographical differences (reflected in the rural-urban designation, and location in the US). Country of birth has also been shown to be particularly important in explaining variation in fertility within Hispanics (Sweeney and Raley 2014). We included maternal education as a proxy for respondents’ socioeconomic background during childhood, which is thought to be predictive of fertility. Paternal education arguably would be a better proxy, but 14.7% of values on paternal education were missing, whereas only 5.5% were missing for maternal education. Thus, using paternal education would not only reduce sample size, but also bias the sample toward “nuclear” families. One’s own education is also a strong determinant of fertility (Skirbekk 2008), particularly in women, as well as being associated with income and net worth (e.g., Boshara et al. 2015).
Number of siblings is often positively associated with fertility, which can occur through a number of pathways: (a) having multiple brother and sisters may be indicative of high fecundity, and there is evidence for heritable variation in fertility that would support this inference (Kohler et al. 1999; Tropf et al. 2015a, b); (b) having multiple siblings may shape fertility intentions (whether positively or negatively). Our selection of control variables is thus substantially different from that of Hopcroft (2014), who included only sex, education, and “intelligence.” It is important to emphasize here that neither set of decisions is inherently correct, and that analyzing the data in a variety of ways furthers our understanding of the relationships that exist, and the extent to which they are robust. At the same time, it is also true that failing to account for highly influential factors such as sex and ethnicity may result in biased estimates.
The Association between Income and Net Worth and the Number of Surviving Children
To examine how measures of wealth at each age were associated with the number of surviving children in later life, we used data from the respondents across multiple waves of data collection. We thus examined how wealth from all individuals who reported their income at a given age was associated with the number of surviving children produced over their entire reproductive life span (see ESM for further details). Figure 4 shows the effects of income and net worth across the life span on lifetime reproductive success (for respondents who are aged 45 or older) from our Poisson regression analysis. So, for example, the effect seen at age 20 indicates the strength and direction of the effect of income earned when 20 years old on the number of surviving children at older ages. In women, across all ethnicities, a clear picture emerges: income is negatively associated with LRS across the life span, although the effects tend to become slightly less negative at older ages. For net worth, there is a consistent positive affect on LRS for white women only, between the ages of around 25 to 35.
In white and Hispanic men, the effect of income on LRS is almost always positive. For black men, in contrast, there is no consistent pattern. When it comes to net worth, we observe almost the opposite effect: for white and Hispanic men, net worth does not seem to be associated with LRS, whereas for black men, there is an apparent positive association between net worth and LRS, particularly between the ages of 25 and 35.
In addition to providing these snapshots of wealth at different ages, we also tried to create measures that captured overall income and net worth generated across an individual’s life. We calculated a median income across the life span for those respondents aged 45 and over who had reported on their incomes at least five times. We rounded this variable and again generated five wealth categories. We then did the same for net worth.
We found very similar results when using these measures (see Fig. 4 for model estimates). Across all women, median income across the life span was consistently negatively related to LRS. Furthermore, in white women, the median value of net worth was positively associated with LRS. Median income was positively associated with LRS in both white and Hispanic men, and there was a negative association for black men. In contrast, in black men, a positive association between net worth and LRS was observed (see ESM Figure 3 for correlations between income and net worth, that were positive but not particularly high). These results are very similar when analyses are conducted without controlling for confounding factors—if anything, the associations between income and LRS are slightly larger in magnitude (see ESM Figure 4).
Cross-Sectional Analyses in Context
Overall, these results are in line with other studies which also show that, generally speaking, income is positively associated with LRS in men, and negatively in women (e.g., Barthold et al. 2012; Fieder and Huber 2007, 2012; Hopcroft 2006, 2014; Nettle and Pollet 2008; Weeden et al. 2006). However, our analyses extend this earlier work in several ways. First, we show that the effect of income differs across ethnicities, suggesting that studies that fail to include or control for ethnicity may generate a slightly misleading picture. Moreover, we find that net worth is positively associated with LRS among women and black men, suggesting that, for some groups, the stock of resources exerts a greater influence on reproductive decisions than the flow. We also observe that the effects of income and net worth are typically most pronounced during peak childbearing ages (i.e., between 25 and 35), suggesting that resources are particularly important at those ages. This indicates that using measures of income earned late in life could provide a biased account of the association between resources and LRS.
Although Fig. 4 plots the associations between our measures of wealth and LRS across the life span of our respondents, it consists nevertheless of multiple cross-sectional analyses. To gain a better understanding of the process of reproductive decision-making, and specifically how wealth might factor into the probability of producing offspring, longitudinal analyses are needed. We turn to these below.