Recent trends in wealth inequality among older Americans in two surveys

Abstract

We document changes in wealth inequality across American households with a member aged 55 or older, comparing data in the Health and Retirement Study (HRS) with that in the Survey of Consumer Finances (SCF) between 1998 and 2016. We examine net wealth including housing, financial and nonfinancial assets and debt, without the cash value of insurances, DB pensions or Social Security wealth. We find very similar distributions of net wealth in the two surveys between the 25^th and 90^th percentiles, but substantially higher wealth in the SCF at the top of the distribution. Both surveys show an increase in wealth inequality between 1998 and 2016, first mostly due to increased wealth at the top, and, after 2012, due to an increase in the share of households with very little wealth as well. Both surveys agree that wealth inequality by education and race, already substantial in 1998, increased further by 2016.

1. Introduction

Wealth inequality is an important determinant of inequality in wellbeing among older Americans. Households that accumulate more wealth can maintain higher levels of consumption through old age, and they can finance better care and are better positioned to meet unexpected expenses. Wealth inequality among older Americans was found to be substantial in the 1990s, with many households having little, if any, net wealth (1). Since the 1990s, wealth inequality has increased in the total population, according to studies using the Survey of Consumer Finances (2) and data generated from income tax returns (3). These studies have not examined trends in income inequality among older Americans.

The most widely used survey to examine wealth distributions is the cross-sectional Survey of Consumer Finances (SCF), both because of the depth of its wealth questions and its unique sample design to cover high-wealth households. It is designed to represent the U.S. population, and its sample of older households is moderately large. The most widely used survey to analyze the economic circumstances, health, and wellbeing of older Americans is the longitudinal Health and Retirement Study (HRS). It measures total wealth in fewer details, and according to a study using data from 2002, it likely underrepresents the top of the wealth distribution (4). At the same time, the HRS is a longitudinal survey, it is substantially larger than the SCF in terms of the number of households, and it covers a wide range of information on health, behaviors and subjective wellbeing, which are features that allow for more in-depth analysis of the details and consequences of wealth inequality. An important motivation for accumulating wealth is to finance consumption in old age when opportunities to acquire income diminish, which makes the study of wealth inequality at older ages particularly relevant to both the study of wealth and the study of well-being of the elderly. It is therefore important to revisit the question of how well the HRS measures the distribution of wealth among older Americans compared to the benchmark SCF, using data from more recent years, preferably covering many years instead of focusing on a single year.

In this paper, carry out a thorough analysis to compare the wealth distributions in the HRS and the SCF and document trends in wealth inequality using the two surveys between 1998 and 2016. We analyze the time series of various quantiles of the wealth distribution and the share of households with various wealth amounts, and we show the contribution of education, rage, birth cohort and other household characteristics to wealth inequality by estimating linear regressions in various years in comparable ways. In line with the literature ((1), (2), (3), (4), (5)), we examine net wealth including housing, financial and nonfinancial assets and debt, without the cash value of insurances, DB pensions or Social Security. We used Stata version 14.2 for the analysis on a Hewlett-Packard Z640 desktop PC with the Windows 10 (64-bit) operating system.

2. The two surveys

The HRS is a longitudinal survey taken every other year that focuses on the wealth and health of individuals over 50 years old. The sample is selected under a multistage area probability design, and a new cohort of 51- to 56-year-olds is added every six years. The most recent addition was in 2016. The HRS oversamples African Americans and Hispanics to facilitate more in-depth research of racial and ethnic minorities. This oversampling allows for more precise estimates of the contribution of race and ethnicity to wealth inequality, a feature that we’ll see in our regression estimates. At the same time, the oversampling requires using weights to restore representation of the underlying population. The HRS provides appropriate weighting factors that compensate for these unequal selection probabilities and adjust for geographic and demographic differences in response rates (using the American Community Survey, ACS, as a benchmark). See more details at (6). The HRS is representative of households in which at least one member is 51 years old or older. To be more precise, this representativeness is true in 1998, 2004, 2010, and 2016, when new cohorts were added to the survey. In 2000, 2006 and 2012, the HRS represents households with members 53 or older; in 2002, 2008 and 2014, it represents households with members 55 and older. People who fall in the relevant age ranges in each HRS survey are called “HRS age-eligible.”

The SCF is a cross-sectional survey taken once every three years that focuses on household wealth of all ages. It is based on a dual-frame sample design. The first sample, which comprises approximately two-thirds of the total sample, is selected from a standard multistage area-probability design and represents general characteristics of the population (this is the probability sample). Sampling weights for this sample are adjusted to match the original frame population totals and the geographic and demographic distribution benchmarked by the CPS. The second sample is designed to correct for low response rates among wealthy households. It is called the list sample, and it consists of relatively wealthy households selected from records derived from tax data provided by the Statistics of Income Division of the Internal Revenue Service. The SCF creates a ‘wealth index’ to identify the oversample of wealthy individuals. In 1989–1992, this wealth index was simple capitalization of observed income flows using an average rate of return. The wealth index was updated in 1995 and is now a combination of the earlier index and an index estimated from a direct regression of gross assets on income and other tax variables. Households are ranked according to this index, and those with higher predicted wealth are selected with higher probability. The list sample provides approximately one-third of all observations. The list and probability samples are combined such that the probability sample has higher weight in predicting wealth at the lower end of the distribution, and the list sample has higher weight in predicting wealth of the higher end of the distribution. See more details in (7). Both the HRS and the SCF have some groups overrepresented by design. We use weights provided by the surveys to adjust for those throughout the entire analysis.

The unit of observation in the HRS is the age-eligible individual if they do not have a spouse or partner and the couple if they do. In what follows, we shall sometimes refer to this unit as “the couple,” which also refers to the individual if they have no spouse or partner. The HRS collects information on each individual in the couple. Wealth in the HRS refers to assets and debts held by members of the couple. Thus, the HRS measures wealth at the level of the couple. The wealth questions are asked of the “financial respondent,” or the person most knowledgeable about the household’s finances. The respondent and spouse are asked to determine jointly who they believe is most knowledgeable about their finances.

The unit of analysis in the SCF is the primary economic unit (PEU), which is the “economically dominant” single individual or couple in a household and all other individuals in the household that are financially interdependent with that individual. Thus, members who live together but are economically independent are not in the PEU. The head of household is the male in a mixed-sex couple or the older individual in a same-sex couple. The SCF collects background information on the household head and their spouse (except for race and ethnicity, which is collected on the survey respondent only, who may or may not be the household head), plus the age of all other members of the household, including those outside the PEU if there are any. The SCF measures wealth at the level of the entire PEU. The core wealth questions in the SCF refer to members of the PEU or are directed at the head of household; additional questions are asked about the assets and debts of non-PEU household members.

To make the samples comparable both across surveys and across time, we require that at least one member of the couple is 55 years old or older at the time of their interview (the lowest common age-eligibility threshold across years in the HRS). In the SCF, this requirement means the household heads and their spouses or partners. We impose this age-eligibility restriction on the HRS as well (this restriction excluded age-eligible couples with the oldest member 51 to 54 years old and non-age-eligible households that were formed by splitting from a previously age-eligible couple).

Table 1 shows the number of observations in each survey used in this analysis. The first two columns show the total number of households in each survey (couples in the HRS, PEUs in the SCF); the last two columns show the number of households in our analytical samples. The HRS sample is substantially larger. Focusing on our analytical samples, the HRS sample was eight times as large in 1998; by 2016, it was slightly less than five times as large. The SCF sample increased by a factor of two over the years, while the size of the HRS sample remained approximately the same.

Table 1.

Number of observations in the two surveys

Survey	All households		Analytical sample
year	HRS	SCF	HRS	SCF
1998	14,257	4,305	13,109	1,618
2000	13,117	-	12,423	-
2001	-	4,442	-	1,672
2002	12,319	-	12,134	-
2004	13,585	4,519	11,912	1,840
2006	12,558	-	11,692	-
2007	-	4,417	-	1,940
2008	11,842	-	11,616	-
2010	15,130	6,482	12,896	2,706
2012	14,191	-	12,980	-
2013	-	6,015	-	2,691
2014	13,098	-	12,794	-
2016	14,766	6,248	12,861	3,066

Notes. Analytical sample: households with a member aged 55 or older.

The demographic composition of the (properly weighted) HRS and SCF samples are similar, and both are similar to the demographic composition of the benchmark CPS (table A1 in the Appendix shows the results). One difference is the slightly lower share of couples versus singles with other household members in the HRS, which is due to differences in how couples are identified in the HRS versus the SCF and the CPS. Another difference is that the SCF respondents are slightly more educated than those in the HRS, and the CPS respondents are in-between.

3. Wealth measurement in the two surveys

The HRS measures wealth by asking questions about twelve asset groups and three kinds of debt. The value of the primary residence and the second home, and debt owed on them (mortgages and home equity lines of credit), are asked in one part of the questionnaire (housing). Defined contribution (DC, including 401(k)) pension balances are asked of each respondent in the employment section of the questionnaire (together with defined benefit plan participation). The remaining nine assets are elicited in the assets part of the questionnaire, asked of the financial respondent. For each asset category, the net value is asked (e.g., “if you sold all that and then paid off any debts on it, about how much would you get?”). In this analysis, we use the asset and debt items available in the RAND HRS file, except for a minor modification regarding DC pension balances (see below).

Wealth is measured in substantially more detail in the SCF, reflecting the fact that this survey focuses on household finances. Questions about the primary residence and the money owed on it, and questions about pension accounts are similar to those in the HRS. Most other assets are asked in more disaggregated categories, and within category, they are asked of separate members of the PEU, listing each such asset or account and assessing the value of each. Typically, asset values and debts are asked in separate questions. In this analysis, we use aggregated asset and debt items available in the public use Summary Extract Public Data Files.

The set of asset and debt items that are covered in the HRS and the SCF overlap to a large extent, except for the cash value of life insurance (not in the HRS). Moreover, the balances of pension accounts are measured and imputed somewhat differently in the HRS. Table A2 in the Appendix lists the wealth measures in the two surveys. Directly comparable items are the value of the primary residence, the net value of primary residences (value less debt), the net value of all other real estate (second homes plus other real estate in the HRS, other residential real estate and nonresidential real estate in the SCF), the value of DC pension accounts from the current employer, and the value of IRA and Keogh accounts. We cannot create directly comparable values or net values of the other items or aggregates of the other items because HRS asks for net values for most assets (“if sold and paid off anything owed on them”), whereas the SCF summary measures include asset values, and it has variables for debts owed on those assets and other debts combined. Note that, in principle, we could reconstruct the net values for some items using the raw SCF data, but we use the summary extract public data files that do not allow for that.

As we mentioned, the way DC pension account balances are collected is somewhat different in the two datasets. Before 2012, HRS collected account balances on DC pension plans from the current employer (for both members of the couple); starting with 2012, it added account balances on all other DC pension plans. To maintain comparability in time, the RAND release of the HRS data that we use contains DC pension account balances from the current employer of the respondents. The SCF has collected balances on all DC pension accounts throughout the entire time period. To create comparable summary wealth measures in both surveys, we include DC pension account balances from the current employer, but we do not include account balances from previous employers (unless they are rolled over to IRAs, in which case they are included).

It is important to note that our measure of total wealth doesn’t include the capitalized value of expected Social Security benefits or expected benefits from company-sponsored defined benefit (DB) pension plans. These benefits are the most important sources of consumption for a large part of the population, especially in the bottom half of the wealth distribution (8). An even broader concept of wealth may include a measure of human capital to capture the future earnings potential of people in the labor force. We restrict our analysis to the narrower concept of wealth primarily because of measurement problems that affect not only our analysis but also most other analyses in the literature. To estimate the capitalized value of Social Security and DB pension benefits for future retirees we would need good data on lifetime incomes and, for DB pensions, details of the pension plans, and we would have to use the complex rules of Social Security and DB pension plans together with several assumptions. Data on lifetime earnings and DB pension plans is not available in the SFC. It is available in the HRS for some waves and for a large subset of the respondents, and corresponding estimates of the present value of Social Security benefits and DB pension wealth exist for selected years (see, e.g., (9) and (10)) but not for the entire time period and for all respondents. Note our measure of total net wealth is conceptually the same (excluding Social Security and DB pension benefit present values) as, and empirically very close to, the wealth measures used in most of the literature (including (1), (2), (3), (4), and (5)).

Both surveys attempt to reduce item nonresponse and use imputation for items with missing values. In the HRS, respondents are presented “unfolding brackets” when they are unwilling or unable to give a number answer to the value of the wealth item in question. Unfolding brackets are a sequence of questions asking whether the amount is larger, smaller, or about the same as a certain value. In the sequence, each subsequent value depends on the answer to the previous question. In the end, the value is narrowed down to be in a range called a bracket. The RAND release of the HRS data contains imputed values of all asset and income types using a consistent method for all waves. In this analysis, we use these RAND imputed values; see (11) for details. The RAND HRS file contains the current-job DC pension account balances, but it does not contain imputed values for the missing values. For the account balance values that are identified in the RAND HRS data, we imputed the values in a way that is similar to the other RAND imputations, except for the use of information from other survey waves. For total wealth, we use the RAND HRS summary net wealth measure (“ATOTB”), and we add the DC pension accounts from current jobs. We converted all values to 2016 prices.

Similar to the HRS, SCF respondents are also invited to give range responses when they are unwilling or unable to give a number answer for the value of the wealth item in question. The implementation is different: respondents are first invited to state a range value themselves, and then, if they are unable or unwilling to do so, they are presented with a series of ranges simultaneously (a “range card”). These range values are then used for imputing point values. The publicly released SCF data contain imputed values for all asset categories and summary wealth measures. The result of the imputation is five implicates for each observation; each implicate is made by drawing repeatedly from an estimate of the conditional distribution of the data. When estimating confidence intervals for statistics or regression coefficients, one must account for the fact that one observation occurs five times in the dataset. The SCF releases “Summary Extract Public Data Files” that include the imputed total values of all wealth categories listed above as well as summary wealth measures. The values in the data are converted to dollars in 2016 prices. For total wealth, we use the broadest summary wealth measure (“networth”), and we subtract “current” and “future” DC pension account balances (from previous employers not rolled over to an IRA) and the cash value of life insurance.

4. Wealth quantiles

We start by estimating quantiles of the wealth distribution in the two surveys. The wealth distribution is very skewed, with little wealth below the 25^th percentile. In this section, we analyze the distribution above the 25^th percentile; we will turn to the bottom part of the wealth distribution in the next section.

For each quantile, we estimated 95% confidence intervals by bootstrapping. The bootstrap draws were clustered at the household level. In the HRS, that took care of potential serial correlation between survey waves; in the SCF, that took care of the fact that each household had five imputed values of total wealth. Figure 1 shows the 25^th, 50^th, 75^th, and 90^th percentile estimates and their confidence intervals between 1998 and 2016. The left panel shows the results for the HRS; the right panel shows the results for the SCF. Figure 2 shows the analogous results for the 90^th, 95^th, and 98^th percentiles. Table A3 in the Appendix contains the quantiles themselves in the two surveys; Figure A1 in the appendix shows the same quantile estimates in a different way to allow for a more direct survey-by-survey comparison, plotting the estimates and confidence intervals of each quantile together from the two surveys.

Figure 1.

Quantiles of total wealth: 25^th, 50^th, 75^th, and 90^th percentiles

Notes. Net total wealth (including current job DC pensions; not including the cash value of life insurance and past job DC pensions unless converted to IRAs). Percentile estimates with bootstrap 95% confidence intervals. Households with a member of age 55 or older.

Figure 2.

Quantiles of total wealth: 90^th, 95^th, and 98^th percentiles

Notes. Net total wealth (including current job DC pensions; not including the cash value of life insurance and past job DC pensions unless converted to IRAs). Percentile estimates with bootstrap 95% confidence intervals. Households with a member aged 55 or older.

According to Figure 1, the wealth distributions are very similar in the two surveys up to the 90^th percentile. The SCF shows a higher 90^th percentile value in 2010 and 2016 but a lower value in-between, but its confidence intervals are wide. Above the 90^th quantile, the two surveys differ more systematically. In particular, the 95^th and 98^th wealth percentile values are higher in the SCF except in 1998. The differences are larger at the 98^th percentile, and they have increased over time. By 2016, the SCF 98^th percentile was twice the HRS 98^th percentile.

These estimates suggest that the HRS misses the upper part of the wealth distribution that the SCF covers, likely due to differences in their sampling procedures (recall that the SCF explicitly targets high-wealth individuals on top of selecting a population-representative sample). In fact, when we remove the upper 2 percent of the wealth distribution from the SCF, the estimated quantiles in the two surveys also become very close at the top. Table A4 and Figure A2 in the appendix show the corresponding results. In fact, dropping the upper 2 percent of the SCF sample may be too much for earlier years. The similarity of the wealth distributions below the 98^th percentile of the SCF is supported by additional evidence: the share of wealth items by the wealth distribution. Up to the 98^th percentile of the SCF wealth distribution, the patterns are very similar in the two surveys, except that the relative share of retirement accounts and nonretirement-related other wealth (mostly financial wealth) is larger in the HRS. At the same time, magnitudes and patterns for the share of housing wealth, other real estate wealth and business wealth are very similar except for the very top of the SCF distribution.

Turning to the substantive question of our paper, we can see very similar trends in wealth inequality in the two surveys. Wealth inequality above the 25^th percentile increased substantially between 1998 and 2008, decreased somewhat between 2008 and 2012, and then increased again. The increase in inequality before 2008 went in parallel with increased levels of wealth at all percentiles above the 25^th. However, after 2012, median wealth stayed roughly the same, whereas wealth at the very top of the distribution increased substantially, especially the top 2 percent as measured by the SFC.

5. Absolute wealth

After analyzing the wealth quantiles, we turn to the share of U.S. households with various amounts of wealth. Recall that wealth is total wealth measured in 2016 prices. Thus, when we compare households with various amounts of wealth across time, we keep consumer prices constant. First, we compare the share of households with less than threshold wealth amounts, and then we turn to the share of households with more than selected threshold wealth amounts. Figures 3 and 4 show the results in the two surveys, with appropriate bootstrap confidence intervals.

Figure 3.

Share of households below threshold values of wealth; lower half of the distribution

Notes. Net total wealth (including current job DC pensions; not including the cash value of life insurance and past job DC pensions unless converted to IRAs). Share estimates with bootstrap 95% confidence intervals. Households with a member of age 55 or older.

Figure 4.

Share of households above threshold values of wealth; upper half of the distribution

Notes. Net total wealth (including current job DC pensions; not including the cash value of life insurance and past job DC pensions unless converted to IRAs). Percentile estimates with bootstrap 95% confidence intervals. Households with a member of age 55 or older.

Figure 3 shows the time series of the share of households with zero or negative wealth, with wealth less than 5 thousand dollars, less than 50 thousand dollars, and less than 200 thousand dollars. The latter two thresholds correspond to the 25^th and 50^th wealth percentiles in 1998. The left panel shows the HRS estimates; the right panel shows the SCF estimates. Using the HRS, we estimated a larger value for households with very low levels of wealth. In 1998, the estimated share of households with zero or negative wealth was approximately 7 percent, which increased to 10 percent by 2016. The corresponding SCF estimates are 5 percent in both 1998 and 2016. Similarly, the estimated share of households with less than $2,000 of wealth was approximately 13 percent in 1998 in the HRS, which increased to 16 percent by 2016; the corresponding SCF estimates are 10 percent in 1998 and 12 percent in 2016. There are similar absolute differences between the two surveys for the share of households with less than $50,000, but the share of households with less than $200,000 is more similar across the HRS and the SCF.

It is not straightforward to explain the differences between the two surveys. The HRS oversamples disadvantages minorities, but it does so using the same sampling frame and method as for the rest of the population. Nevertheless, while having very similar estimates for households below $200,000 (the approximate median), the HRS estimates show a larger share of the poorest households and a larger increase in their share. Thus, it appears that the population the HRS represents has more of the poorest households relative to middle-class households than the population the SCF represents.

Turning to the upper part of the distribution, Figure 4 shows the share of households above $200,000, above $500,000, above $1,000,000 and above $5,000,000. These threshold values roughly correspond to the median, 75^th percentile, 90^th percentile and 98^th percentiles of total wealth in 1998. The estimates are very similar in the two surveys up to the highest wealth category: the share of households with $5,000,000 or more increased from 1.5 percent to 2.5 percent according to the HRS; it increased from 2 percent to close to 5 percent by the SCF. These figures reflect the same differences that we have seen when we examined the wealth quantile estimates: the HRS appears to miss the very top of the wealth distribution that is represented by the SCF.

The differences at the two tails of the distribution notwithstanding, there are several similar trends in the two surveys. The share of households with no or very little positive wealth increased throughout the entire time period, especially since 2007. Similarly, the share of households with great wealth (more than $1,000,000) increased during the same time period. This increase took place in two windows, first between 1998 and 2007 and then after 2012. Reflecting these trends, the share of the broadly defined middle class, with net wealth of $50,000 to $1,000,000 (roughly between the 25^th and 90^th percentiles), decreased between 1998 and 2016.

6. Regression analysis

Our third analysis examines how several household and personal characteristics are associated with wealth in the two surveys and how those associations changed over time. The most important household characteristic variables we included are household composition (whether single or couple), birth cohort (by the year of birth of the financial respondent in the HRS and the household head in the SCF), race and ethnicity (of the financial respondent in the HRS and the survey respondent in the SCF), and years of schooling (average if a couple). We estimated two sets of regressions, each for one of the four years in which both surveys were fielded: 1998, 2004, 2010, and 2016. We dropped members of birth cohorts from survey waves where their share was less than 2 percent.

The first set of regressions has log total net wealth as the dependent variable, restricted to households with positive wealth; the second set of regressions are linear probability models for whether the household had positive net wealth. The standard errors are clustered at the household level in both surveys. In the HRS, this means clustering at the level of original households, some of which have split since the beginning of the survey. Allowing for this kind of clustering acknowledges the potential correlation of wealth between households that used to be related. This acknowledgement turns out to be a minor issue. However, clustering in the SCF means acknowledging the fact that each household is represented five times in the SCF data due to implicates of the wealth imputations. Table 2 shows the results of the log wealth regressions; Table A5 in the Appendix shows the results of the linear probability models.

Table 2.

Yearly cross-sectional regressions of log total wealth on household characteristics

Dep. var.	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
ln wealth	HRS 1998	HRS 2004	HRS 2010	HRS 2016	SCF 1998	SCF 2004	SCF 2010	SCF 2016
Single female	−1.10**	−1.15**	−0.98**	−1.11**	−1.09**	−1.45**	−1.30**	−1.32**
	(0.039)	(0.046)	(0.047)	(0.055)	(0.117)	(0.135)	(0.098)	(0.111)
Single male	−0.83**	−0.99**	−0.79**	−0.88**	−0.65**	−0.60**	−0.90**	−0.99**
	(0.060)	(0.076)	(0.065)	(0.075)	(0.161)	(0.147)	(0.125)	(0.140)
Couple	Ref.	Ref.	Ref.	Ref.	Ref.	Ref.	Ref.	Ref.
Other HH member	−0.10*	−0.08	−0.07	−0.10	−0.20	−0.12	0.03	0.14
	(0.040)	(0.046)	(0.048)	(0.054)	(0.167)	(0.151)	(0.104)	(0.110)
Birth cohort
1892–1923	0.07	0.14*	0.05		0.22	0.65**	0.36
	(0.042)	(0.055)	(0.076)		(0.127)	(0.192)	(0.209)
1924–1930	−0.02	0.10	0.14*	−0.03	0.05	0.52**	0.25	0.43
	(0.046)	(0.053)	(0.063)	(0.097)	(0.138)	(0.148)	(0.159)	(0.249)
1931–1941	Ref.	Ref.	Ref.	Ref.	Ref.	Ref.	Ref.	Ref.
1942–1947	−0.02	0.01	−0.02	−0.16*	−0.34*	0.22	0.02	−0.48**
	(0.070)	(0.056)	(0.055)	(0.066)	(0.167)	(0.139)	(0.117)	(0.152)
1948–1953		−0.24**	−0.22**	−0.33**		0.11	−0.30**	−0.52**
		(0.078)	(0.056)	(0.065)		(0.187)	(0.110)	(0.137)
1954–1959			−0.38**	−0.38**			−0.34*	−0.46**
			(0.084)	(0.063)			(0.135)	(0.129)
African American	−0.92**	−0.93**	−1.00**	−1.14**	−0.86**	−1.03**	−1.04**	−1.25**
	(0.066)	(0.069)	(0.070)	(0.084)	(0.179)	(0.200)	(0.152)	(0.158)
Hispanic	−0.61**	−0.36**	−0.40**	−0.31**	0.03	−0.53	−0.53**	−0.82**
	(0.098)	(0.107)	(0.101)	(0.098)	(0.309)	(0.327)	(0.198)	(0.249)
Non-Black non-Hispanic	Ref.	Ref.	Ref.	Ref.	Ref.	Ref.	Ref.	Ref.
Education (years)	0.23**	0.25**	0.27**	0.28**	0.24**	0.26**	0.29**	0.39**
	(0.007)	(0.008)	(0.009)	(0.010)	(0.019)	(0.023)	(0.018)	(0.022)
Constant	5.70**	5.76**	5.65**	5.80**	5.45**	5.50**	5.57**	5.64**
	(0.030)	(0.035)	(0.037)	(0.046)	(0.089)	(0.100)	(0.089)	(0.113)
Observations	11,603	10,569	10,798	9,270	7,754	8,864	12,628	12,935
Clusters (households)	11502	10386	10565	9029	1554	1779	2537	2593
R-squared	0.267	0.260	0.247	0.254	0.248	0.306	0.283	0.338

Note. Only households (HH) with positive wealth were included. Clusters are original households in the HRS (potential splits by the time of the interview) and households in the SCF (with up to 5 observations for imputation implicates).

^**p<0.01;

^*p<0.05.

Most of the regression coefficients are very similar in the HRS and the corresponding SCF samples. Exceptions include some of the family composition variables (some of which are measured differently in the two surveys), some of the cohort indicators, and the Hispanic indicator. Most of these differences are statistically not significant because their standard errors are large, especially in the much smaller SCF sample. A larger and statistically significant difference between the two surveys is the coefficient estimates on education in 2016, with an unusually high value in the SCF sample.

Focusing on results that are similar across the two surveys, singles tend to have substantially lower wealth, and this association remained relatively stable across the years (it increased in the SCF, but the estimates are imprecise). Cohort differences are strong and stable across time, with older cohorts holding more wealth. This trend is especially true with respect to the youngest cohorts in the data, early baby boomers (born 1948–1953) and mid baby boomers (born 1954–1959), whose wealth remained 20 to 50 percent less than the cohorts born between 1931 and 1941. The wealth disadvantage of younger cohorts may have increased with time, but the estimates are not very precise.

Racial and ethnic disparities in wealth are known to be large, in large part due to substantial differences in education and lifetime earnings (see, e.g., (12) and (13)). Racial and ethnic differences decreased in terms of education but not in earnings during the twenty years covered in our analysis (14). It is therefore interesting to know whether, and to what extent, racial and ethnic disparities grew during these twenty years. Our regression results can shed light on this because the results from the two surveys tell a similar story. African Americans tend to have substantially less wealth than households with non-Black non-Hispanic respondents of the same birth cohort, family structure and years of schooling. Their disadvantage is not only very large, but it increased with time, with an especially large increase after 2010. The time trend of their disadvantage is only marginally statistically significant, but it is consistent across the two surveys. The wealth disadvantage of households with Hispanic respondents is also large, although smaller than that of African Americans, and it may have actually decreased with time, although the estimates are not precise enough to conclude this decrease with a definitive time trend.

Another strong result is the substantial gradient by years of schooling and the steady increase of that gradient through time. In 1998, our estimates suggest that households with one more year of average education (averaged across members of the couple) had approximately 25 percent more wealth (conditional on cohort, race, ethnicity and family structure). This difference grew to be more than 30 percent by 2016 (we convert log point differences to percent differences). In the SCF, the increase was very similar up to 2010, followed by a very large jump in 2016. Whether that jump is a chance event or the beginning of a new trend that the HRS has not captured, the substantial and increasing education gradient of wealth is a strong and qualitatively similar result across the two surveys.

Turning to associations with whether the household holds positive wealth, we find strong coefficient estimates of the same sign for all right-hand-side variables that we found in the log wealth regressions. In particular, singles are close to 10 percent less likely to have positive wealth, African Americans are over 15 percent less likely, Hispanics are more than 10 percent more likely, and the gradient with respect to years of education is strong, at approximately 2 percentage points. Here, the HRS coefficient estimates tend to be larger, but the qualitative patterns are the same across the two surveys. In contrast with the results for log wealth, we do not find strong time trends in the coefficients. This outcome suggests that observable family characteristics became more important in shaping wealth inequality not at the bottom but the middle and/or the top of the distribution. This finding also means that the time trends of coefficients in the log wealth regressions were not affected by changing the selection into the samples with positive wealth.

Table A1.

Household composition, age and education in the HRS, the SCF, and the CPS (pooled 1998–2016, respondent or spouse 55 years old or older, weighted with survey weights)

	HRS	SCF	CPS
Household composition
Single female household	0.23	0.24	0.24
Single male household	0.11	0.11	0.12
Couple household	0.34	0.41	0.39
Single and other member(s)	0.17	0.10	0.11
Couple and other member(s)	0.15	0.14	0.14
Average household size	2.1	1.9	2.0
Age a
25th percentile	59	59	59
Median	66	65	65
75th percentile	75	74	74
Average age	68	67	67
Education
Years of educationa,b	12.7	13.0	12.9
Less than high school (both if couple)	0.19	0.16	0.14
College degree (any member if couple)	0.30	0.33	0.30
College degree (both members if couple)	0.18	0.21	0.19

^aAverage value for couples.

^bYears of education top-coded at 17.

^cAge is top-coded at 85 in CPS; average may be biased downward.

Table A2.

Wealth measures in the HRS and the SCF

Health and Retirement Study (HRS)*	Survey of Consumer Finances (SCF)**
Housing
Primary residence (value)	Principal residence (value)
Mortgages and land contracts owed on primary residence (value)	Mortgages, heloc (value); home equity loans on principal residence (value)
Other debt owed on primary residence (value)
Secondary residence (value)	Other residential real estate (value)
Mortgages and land contracts owed on secondary residence (value)	Debt on other residential real estate (value)
DC pensions and IRAs
Pension accounts from current job (value) (other pension accounts)	Pension accounts from current job (“thrift accounts,” value) Future pensions (accounts from previous employment; value) Current pensions (accounts from which currently receiving pension; value)
IRA, Keogh accounts (value)	IRA, Keogh accounts (value)
Other assets
Real estate (not primary or secondary residence; net value)	Nonresidential real estate (net value)
Vehicles (net value)	Vehicles (value, imputed)
Businesses (net value)	Businesses (net value)
Stocks, mutual funds (net value)	Mutual funds (value), company stocks (value),
Checking, savings, money market accounts (value)	Checking accounts (value), saving accounts (value), money market accounts (money market deposit accounts and money market mutual funds; value) Call accounts at brokerages (value)
CDs, govt. savings bonds, Treasury bills (value)	CDs (value)
Bonds (corporate, municipal, govt, foreign, bond funds; net value)	Bonds (corporate, govt, mortgage-backed; value), saving bonds (value), Cash value of life insurance
Other savings or assets (incl. rights in trust or estate; money owed by others, jewelry, etc.; net value)	Other managed assets (incl. trusts; value), other non-financial assets (value), other financial assets (value)
Debt
All debts not previously mentioned (value)	Debt on other lines of credit (value) Debt on installment loans (vehicles, education, etc; value), Credit card debt (value). Student debt (value). Other debt (margin loans, pension loans, etc.)

^*RAND files and imputations;

^**Summary Extract Public Data Files and imputations

Table A3.

Total net household wealth: quantiles and means, USD ‘000, 2016 prices

	HRS									SCF
year	p5	p10	p25	p50	p75	p90	p95	p98	mean	p5	p10	p25	p50	p75	p90	p95	p98	mean
1998	0	1	54	192	502	1,135	1,860	3,434	535	0	4	66	186	450	1,045	1,819	4,771	621
2000	0	2	58	210	554	1,259	2,042	3,887	586
2001										0	7	68	210	580	1,432	2,775	7,072	816
2002	0	3	65	224	607	1,353	2,181	3,907	609
2004	0	3	61	238	664	1,481	2,337	4,198	659	0	5	66	234	720	1,488	2,544	6,733	852
2006	0	1	60	259	724	1,592	2,646	4,848	719
2007										0	5	72	249	648	1,415	3,436	8,312	942
2008	0	1	56	251	726	1,625	2,665	4,712	709
2010	−2	0	44	222	655	1,489	2,464	4,352	640	0	2	50	204	582	1,718	3,092	6,373	840
2012	−3	0	37	196	609	1,469	2,308	4,222	616
2013										0	2	42	180	506	1,395	2,885	7,110	789
2014	−1	0	37	207	671	1,581	2,527	4,372	691
2016	0	0	39	211	670	1,621	2,603	4,543	673	0	3	45	207	650	1,905	3,750	8,813	1,063

Table A4.

Total net household wealth: quantiles and means, USD ‘000, 2016 prices. SCF: top 2% of the wealth distribution removed in each year

	HRS									SCF without top 2%
year	p5	p10	p25	p50	p75	p90	p95	p98	mean	p5	p10	p25	p50	p75	p90	p95	p98	mean
1998	0	1	54	192	502	1,135	1,860	3,434	535	0	4	65	180	430	886	1,452	2,342	370
2000	0	2	58	210	554	1,259	2,042	3,887	586
2001										0	6	65	204	547	1,246	1,974	3,730	500
2002	0	3	65	224	607	1,353	2,181	3,907	609
2004	0	3	61	238	664	1,481	2,337	4,198	659	0	4	63	228	676	1,333	2,041	3,339	528
2006	0	1	60	259	724	1,592	2,646	4,848	719
2007										0	5	69	242	611	1,215	2,386	4,428	567
2008	0	1	56	251	726	1,625	2,665	4,712	709
2010	−2	0	44	222	655	1,489	2,464	4,352	640	0	2	47	194	542	1,400	2,216	3,896	520
2012	−3	0	37	196	609	1,469	2,308	4,222	616
2013										−1	2	40	172	467	1,219	2,088	3,719	471
2014	−1	0	37	207	671	1,581	2,527	4,372	691
2016	0	0	39	211	670	1,621	2,603	4,543	673	0	2	42	200	596	1,564	2,782	4,884	608

Table A5.

Yearly cross-sectional regressions of positive total wealth on household characteristics (linear probability models)

Dep. var.	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
In wealth	HRS 1998	HRS 2004	HRS 2010	HRS 2016	SCF 1998	SCF 2004	SCF 2010	SCF 2016
Single female	−0.096**	−0.100**	−0.100**	−0.108**	−0.085**	−0.087**	−0.069**	−0.092**
	(0.006)	(0.007)	(0.008)	(0.008)	(0.018)	(0.018)	(0.014)	(0.015)
Single male	−0.072**	−0.080**	−0.082**	−0.088**	0.012	−0.005	−0.072**	−0.060**
	(0.008)	(0.010)	(0.011)	(0.011)	(0.021)	(0.017)	(0.020)	(0.019)
Couple	Ref.	Ref.	Ref.	Ref.	Ref.	Ref.	Ref.	Ref.
Other hh member	−0.019**	−0.002	−0.023**	−0.018*	−0.035	0.032	0.015	−0.010
	(0.006)	(0.007)	(0.008)	(0.008)	(0.027)	(0.021)	(0.017)	(0.018)
Birth cohort
1892–1923	0.030**	0.024**	0.031**		0.009	0.043*	0.034
	(0.006)	(0.007)	(0.011)		(0.019)	(0.022)	(0.027)
1924–1930	0.009	0.018*	0.011	0.015	−0.009	0.044*	0.016	0.024
	(0.007)	(0.008)	(0.009)	(0.013)	(0.021)	(0.019)	(0.021)	(0.032)
1931–1941	Ref.	Ref.	Ref.	Ref.	Ref.	Ref.	Ref.	Ref.
1942–1947	0.014	0.001	−0.004	−0.001	−0.019	−0.036	−0.016	−0.014
	(0.010)	(0.008)	(0.009)	(0.009)	(0.025)	(0.019)	(0.017)	(0.019)
1948–1953		−0.041**	−0.054**	−0.032**		−0.033	−0.053**	−0.017
		(0.013)	(0.009)	(0.009)		(0.025)	(0.016)	(0.018)
1954–1959			−0.063**	−0.060**			−0.074**	−0.029
			(0.013)	(0.010)			(0.022)	(0.017)
African American	−0.152**	−0.151**	−0.167**	−0.161**	−0.118**	−0.150**	−0.128**	−0.093**
	(0.012)	(0.013)	(0.013)	(0.014)	(0.038)	(0.035)	(0.026)	(0.024)
Hispanic	−0.116**	−0.105**	−0.135**	−0.097**	−0.061	−0.092	−0.128**	−0.106**
	(0.016)	(0.016)	(0.018)	(0.018)	(0.055)	(0.047)	(0.035)	(0.033)
Non-Black non-Hispanic	Ref.	Ref.	Ref.	Ref.	Ref.	Ref.	Ref.	Ref.
Education (years)	0.020**	0.017**	0.018**	0.018**	0.020**	0.011**	0.013**	0.016**
	(0.001)	(0.001)	(0.001)	(0.002)	(0.003)	(0.003)	(0.003)	(0.003)
Constant	0.974**	0.976**	0.971**	0.978**	0.968**	0.978**	0.975**	0.972**
	(0.004)	(0.005)	(0.006)	(0.007)	(0.012)	(0.012)	(0.012)	(0.015)
Observations	12,604	11,396	12,328	10,645	8,037	9,120	13,400	13,345
Clusters (households)	12484	11157	12006	10311	1609	1827	2680	2669
R-squared	0.150	0.127	0.130	0.119	0.121	0.123	0.083	0.092

Figure A1.

Time series of the wealth percentiles in the HRS and the SCF.

(Total net household wealth, less DC pension plans from past jobs and the cash value of life insurance, $ ‘000 in 2016 prices.)

Figure A2.

Time series of the wealth percentiles in the HRS and the SCF; SCF: top 1.5% of the wealth distribution removed in each year

Figure A3.

The share of broad categories of wealth by the distribution of total wealth in the HRS and the SCF