NBER WORKING PAPER SERIES
WEALTH INEQUALITY IN THE UNITED STATES SINCE 1913:
EVIDENCE FROM CAPITALIZED INCOME TAX DATA
Emmanuel Saez
Gabriel Zucman
Working Paper 20625
http://www.nber.org/papers/w20625
NATIONAL BUREAU OF ECONOMIC RESEARCH
1050 Massachusetts Avenue
Cambridge, MA 02138
October 2014
We thank Tony Atkinson, Mariacristina DeNardi, Matthieu Gomez, Barry W. Johnson, Maximilian
Kasy, Lawrence Katz, Arthur Kennickell, Wojciech Kopczuk, Moritz Kuhn, Thomas Piketty, Jean-Laurent
Rosenthal, John Sabelhaus, Amir Sufi, Edward Wolff, and numerous seminar and conference participants
for helpful discussions and comments. Juliana Londono-Velez provided outstanding research assistance.
We acknowledge financial support from the Center for Equitable Growth at UC Berkeley, and the
MacArthur foundation. A complete set of Appendix tables and figures supplementing this article is
available online at http://eml.berkeley.edu/~saez and http://gabriel-zucman.eu/uswealth The views
expressed herein are those of the authors and do not necessarily reflect the views of the National Bureau
of Economic Research.
NBER working papers are circulated for discussion and comment purposes. They have not been peer-
reviewed or been subject to the review by the NBER Board of Directors that accompanies official
NBER publications.
© 2014 by Emmanuel Saez and Gabriel Zucman. All rights reserved. Short sections of text, not to
exceed two paragraphs, may be quoted without explicit permission provided that full credit, including
© notice, is given to the source.
Wealth Inequality in the United States since 1913: Evidence from Capitalized Income Tax
Data
Emmanuel Saez and Gabriel Zucman
NBER Working Paper No. 20625
October 2014
JEL No. H2,N32
ABSTRACT
This paper combines income tax returns with Flow of Funds data to estimate the distribution of household
wealth in the United States since 1913. We estimate wealth by capitalizing the incomes reported by
individual taxpayers, accounting for assets that do not generate taxable income. We successfully test
our capitalization method in three micro datasets where we can observe both income and wealth: the
Survey of Consumer Finance, linked estate and income tax returns, and foundations' tax records. Wealth
concentration has followed a U-shaped evolution over the last 100 years: It was high in the beginning
of the twentieth century, fell from 1929 to 1978, and has continuously increased since then. The rise
of wealth inequality is almost entirely due to the rise of the top 0.1% wealth share, from 7% in 1979
to 22% in 2012—a level almost as high as in 1929. The bottom 90% wealth share first increased up
to the mid-1980s and then steadily declined. The increase in wealth concentration is due to the surge
of top incomes combined with an increase in saving rate inequality. Top wealth-holders are younger
today than in the 1960s and earn a higher fraction of total labor income in the economy. We explain
how our findings can be reconciled with Survey of Consumer Finances and estate tax data.
Emmanuel Saez
Department of Economics
University of California, Berkeley
530 Evans Hall #3880
Berkeley, CA 94720
and NBER
saez@econ.berkeley.edu
Gabriel Zucman
Department of Economics
London School of Economics and Political Science
Houghton Street
London WC2A
g.zucman@lse.ac.uk
1 Introduction
Income inequality has sharply increased in the United States since the late 1970s, yet currently
available evidence suggests that wealth concentration has not grown nearly as much. One
possible explanation is that rising inequality is purely a labor income phenomenon: despite an
upsurge in top wage and entrepreneurial incomes (Piketty and Saez, 2003), the working rich
might not have had enough time yet to accumulate a lot of wealth—perhaps because they have
low saving rates, face high tax rates, or have low returns on assets. Should this be true, the
implications for analyzing the US economy and for policy-making would be far-reaching.
Our paper, however, challenges this view. On the basis of new, annual, long-run series, we
find that wealth inequality has considerably increased at the top over the last three decades.
By our estimates, almost all of this increase is due to the rise of the share of wealth owned by
the 0.1% richest families, from 7% in 1978 to 22% in 2012, a level comparable to that of the
early twentieth century (Figure 1).
Although the top 0.1% is a small group—it includes about 160,000 families with net assets
above $20 million in 2012—carefully measuring its wealth is important for two reasons. First, the
public cares about the distribution of economic resources. Since wealth is highly concentrated
(much more than labor income, due to the dynamic processes that govern wealth accumulation),
producing reliable estimates requires to pay careful attention to the very top. This is difficult
to achieve with survey data and motivates our attempt at using tax records covering all the
richest families. The top 0.1% also matters from a macroeconomic perspective: it owns a sizable
share of aggregate wealth and accounts for a large fraction of its growth. Over the 1986-2012
period, the average real growth rate of wealth per family has been 1.9%, but this average masks
considerable heterogeneity: for the bottom 90%, wealth has not grown at all, while it has risen
5.3% per year for the top 0.1%, so that almost half of aggregate wealth accumulation has been
due to the top 0.1% alone.
To construct our series on the distribution of wealth, we capitalize income tax data. Starting
with the capital income reported by individuals to the Internal Revenue Service—which is broken
down into many categories: dividends, interest, rents, profits, mortgage payments, etc.—for each
asset class we compute a capitalization factor that maps the total flow of tax income to the
total amount of wealth recorded in the Flow of Funds. We then combine individual incomes and
aggregate capitalization factors by assuming that within a given asset class the capitalization
factor is the same for everybody. For example, if the ratio of Flow of Funds fixed income claims
to tax reported interest income is 50, then $50,000 in fixed income claims is attributed to an
1
individual reporting $1,000 in interest. By construction, the wealth distribution we estimate
is consistent with the Flow of Funds totals. Our paper can thus be seen as a first attempt at
creating distributional Flow of Funds statistics that decompose aggregate wealth and saving by
fractiles. This allows us to jointly analyze growth and distribution in a consistent framework.
A number of authors have used the capitalization in the past, notably King (1927), Stewart
(1939), Greenwood (1983) in the United States, and Atkinson and Harrison (1978) in the United
Kingdom. But these studies typically provide estimates for just a few years in isolation, do not
use micro-data, or have a limited breakdown of capital income by asset class. Compared to
earlier attempts, our main advantage is that we have more data.1
The capitalization method faces a number of potential obstacles. We carefully deal with each
of them and provide checks showing that the method works well in practice. First, not all assets
generate taxable investment income—owner-occupied houses and pensions, in particular, do not.
These assets are well covered by a number of sources and we account for them by combining the
available information—surveys, property taxes paid, pension distributions, wages reported on
tax returns, etc.—in a systematic manner. Second, within a given asset class, richer households
might have different rates of returns than the rest of the population, in particular because of
tax avoidance. We have conducted a large-scale reconciliation exercise between income tax and
national accounts data to track unreported income and we impute missing wealth (e.g., held
through trusts) when necessary. We then investigate all the situations where both wealth and
capital income can be observed at the micro level—in the Survey of Consumer Finances (SCF),
matched estate and individual income tax data, and publicly available tax returns of foundations.
In each case, we find that within asset-class realized returns are similar across groups, and that
top wealth shares obtained by capitalizing income are very close to the directly observed top
shares in both level and trend. At the individual level, the relationship between capital income
and wealth is noisy, but the capitalization method works nonetheless because the noise cancels
out when considering groups of thousands of families, which is what matters for our purposes.2
1King (1927) and Stewart (1939) had to rely on tax tabulations by income size (instead of micro-data).
Atkinson and Harrison (1978) lack sufficiently detailed income data (they had access to tabulations by size of
capital income but with no composition detail). Greenwood (1983) comes closest to our methodology. She
uses one year (1973) of micro tax return data and various capital income categories but does not use the
Flow of Funds to estimate returns by asset class so that her estimates are not consistent with the Flow of
Funds aggregates. She relies instead on market price indexes to infer wealth from income. Asset price indexes,
however, have shortcomings (such as survivor bias for equities) that can cause biases when analyzing long-time
periods. Recently, Mian et al. (2014) also use the capitalization method and zip code level income tax statistics
to measure wealth by zip code.
2A number of studies have documented the noisy relationship at the individual level between income and
wealth, see, e.g., Kennickell (2001, 2009a) for the SCF, and Rosenmerkel and Wahl (2011) and Johnson et al.
(2013) for matched estate-income tax data.
2
The analysis of the distribution of household wealth since 1913 yields two main findings.
First, wealth inequality is making a comeback, with the top 0.1% wealth share almost as
high in 2012 as in the 1916 and 1929 peaks and three times higher than in the late 1970s.
Despite population aging, however, the rich are younger today than half a century ago: in the
1960s, top 0.1% wealth holders were older than average, which is not the case anymore today.
The key driver of the rapid increase in wealth at the top is the surge in the share of income,
in particular labor income, earned by top wealth holders. Income inequality has a snowballing
effect on the wealth distribution: top incomes are being saved at high rates, pushing wealth
concentration up; in turn, rising wealth inequality leads to rising capital income concentration,
which contributes to further increasing top income and wealth shares. Our core finding is that
this snowballing effect has been sufficiently powerful to dramatically affect the shape of the US
wealth distribution over the last 30 years. Due to data limitations we cannot provide yet formal
decompositions of the relative importance of self-made vs. dynastic wealth, and we hope our
results will motivate further research in this area.3
The second key result involves the dynamics of the bottom 90% wealth share. There is a
widespread view that a key structural change in the US economy has been the rise of middle-
class wealth since the beginning of the twentieth century, in particular because of the rise of
pensions and home ownership rates. And indeed our results show that the bottom 90% wealth
share gradually increased from 20% in the 1920s to a high of 35% in the mid-1980s. But in a
sharp reversal of past trends, the bottom 90% wealth share has fallen since then, to about 23%
in 2012. Pension wealth has continued to increase but not enough to compensate for a surge
in mortgage, consumer credit, and student debt. The key driver of the declining bottom 90%
share is the fall of middle-class saving, a fall which itself may partly owe to the low growth of
middle-class income, to financial deregulation leading to some forms of predatory lending, or to
growing behavioral biases in the saving decisions of middle-class households.
Our results confirm some earlier findings using different data but contradict some others.
We provide a detailed reconciliation with previous studies. First, our results are consistent with
Forbes Magazine data on the wealth of the 400 richest Americans. Normalized for population
growth, the wealth share of the top 400 has increased from 1% in the early 1980s to over 3%
in 2012-3, on par with the tripling of our top 0.01% wealth share. Second, the SCF—a high
quality survey that over-samples wealthy individuals—displays a top 10% wealth share very
close in level and trend to the one we find, but smaller increases in the top 1% and especially
top 0.1% shares. Several factors explain this discrepancy: By design, the SCF excludes Forbes
3See Piketty et al. (2013) for such an analysis on French data.
3
400 individuals; aggregate wealth in the SCF and the Flow of Funds differs (Henriques and
Hsu, 2013); and the unit of observation in the SCF (the household) is larger than the one
we use (the family as defined by the tax code). After adjusting for these factors, the SCF
displays a substantial increase in top wealth shares from 1989 to 2013, although still not as
large as by our estimates. We also find that the SCF under-estimates the increase in capital
income concentration from 1989 to 1998 (less so afterward). Although the SCF uses a rigorous
sampling methodology—it itself relies on the capitalization method to determine the sample
of households to be surveyed—it is always difficult for surveys to capture perfectly the very
top groups (Kennickell, 2009a).4 Last, the top 0.1% wealth share estimated by Kopczuk and
Saez (2004) from estate tax returns is remarkably close in level and trend to the one we obtain
up to the late 1970s, but then hardly increases from 1976 to 2000. Estate-based estimates are
obtained using the mortality multiplier technique, whereby individual estates are weighted by
the inverse of the mortality rate (conditional on age and gender) to capture the distribution
of wealth among the living. The estate-based estimates of Kopczuk and Saez (2004) assume
a constant mortality differential between the wealthy and the overall population. We show
that the mortality differential has in fact sharply increased since the late 1970s, explaining why
estate-based estimates fail to uncover the recent surge in top wealth shares.5
Despite our best effort, we stress that we still face limitations when measuring wealth inequal-
ity. The development of the offshore wealth management industry, changes in tax optimization
behaviors, indirect wealth ownership (e.g., through trusts and foundations) all raise challenges.
Because of the lack of administrative data on wealth, none of the existing sources offer a defini-
tive estimate. We see our paper as an attempt at using the most comprehensive administrative
data currently available, but one that ought to be improved in at least two ways: by using
additional information already available at the Statistics of Income (SOI) division of the IRS
as well as new data that the US Treasury could collect at low cost. A modest data collection
effort would make it possible to obtain a better picture of the joint distributions of wealth,
income, and saving. In turn, this information would be of great relevance to evaluate proposals
for consumption or wealth taxation.
The remainder of the paper is organized as follows. Section 2 discusses our definition and
aggregate measure of wealth. In Section 3 we analyze the distribution of taxable capital income
4Systematically comparing our estimates with SCF estimates is a useful input for further improving the SCF
sample representativity so we view our approach as complementary to the extremely valuable SCF surveys.
5The recent increase in the mortality differential by life-time earnings and education levels has been carefully
documented (see, e.g., Waldron, 2004, 2007). The differential mortality estimates by wealth class we compute
could be used to improve the estate multiplier method. Hence, the capitalization method is also a useful
complement to the estate multiplier method.
4
and present our method for inferring wealth from income. Section 4 discusses the pros and
cons of the capitalization method and provides a number of checks suggesting that it works
well in practice. We present our results on the distribution of household wealth in Section 5
and we analyze the relative importance of changes in income shares, saving rates, and capital
gains in the dynamics of US wealth inequality in Section 6. Section 7 compares our estimates to
previous studies. Section 8 concludes. The key steps of the analysis are presented in the text,
while complete tabulations of results with detailed methodological notes are posted in a set of
online Excel files on the authors’ websites.
2 What is Wealth? Definition and Aggregate Measures
2.1 The Wealth Concept We Use
Let us first define the concept of wealth that we consider in this paper. Wealth is the current
market value of all the assets owned by households net of all their debts. Following international
standards codified in the System of National Accounts (United Nations, 2009), assets include
all the non-financial and financial assets over which ownership rights can be enforced and that
provide economic benefits to their owners.
Our definition of wealth includes all pension wealth—whether held on individual retirement
accounts, or through pension funds and life insurance companies—with the exception of Social
Security and unfunded defined benefit pensions. Although Social Security matters for saving
decisions, the same is true for all promises of future government transfers. Including Social
Security in wealth would thus call for including the present value of future Medicare benefits,
future government education spending for one’s children, etc., net of future taxes. It is not clear
where to stop, and such computations are inherently fragile because of the lack of observable
market prices for this type of assets. Unfunded defined benefit pensions are promises of future
payments which are not backed by actual wealth. The vast majority (94% in 2013) of unfunded
pension entitlements are for Federal, State and local government employees, thus are conceptu-
ally similar to promises of future government transfers, and just like those are better excluded
from wealth. According to the Flow of Funds, unfunded defined benefit pensions represent the
equivalent of 5% of total household wealth today, down from 10-15% in the 1960s-1970s. Treat-
ing them as household wealth would reinforce our finding of an inverted-U shaped evolution of
the bottom 90% wealth share, as unfunded pensions are relatively equally distributed.6
Our wealth concept excludes human capital, which contrary to non-human wealth cannot
6Recall that we treat all funded defined benefit pensions as wealth, just like defined contribution pensions.
5
be sold on markets. Because the distributions of human and non-human capital are shaped
by different economic forces (savings, inheritance, and rates of returns matter for non-human
capital; technology and education, among others, matter for human capital), it is necessary to
start by studying the two of them separately. In Section 5 we investigate how the labor income of
top wealth-holders has evolved, and we refer to Aaberge et al. (2014) for a more comprehensive
analysis of the joint distribution of human and non-human capital.
We also exclude the wealth of nonprofit institutions, which amounts to about 10% of house-
hold wealth.7 The bulk of nonprofit wealth belongs to hospitals, churches, museums, education
institutions and research centers, and thus cannot easily be attributed to any particular group of
households. It would probably be desirable to attribute the wealth of some private foundations
(e.g., Bill and Melinda Gates) to specific families. But this cannot always be done easily, as in
the case of foundations created long ago (like the Ford or MacArthur foundations). The wealth
of foundations is still modest compared to that of the very top groups—it amounts to 1.2% of
total household wealth in 2012—but it is growing (it was 0.8% in 1985).8
Last, we exclude consumer durables (about 10% of household wealth) and valuables from
assets. Durables are not considered as assets by the System of National Accounts and there is
no information on tax returns about them.9
2.2 Aggregate Wealth: Data and Trends
With this definition in hand, we construct total household wealth—the denominator we use
when computing wealth shares—as follows. For the post-1945 period, we rely on the latest
Flow of Funds (US Board of Governors of the Federal Reserve System, 2014). The Flow of
Funds report wealth as of December 31 and we compute mid-year estimates by averaging end-
of-year values. For the 1913-1945 period, we combine earlier estimates from Goldsmith et al.
(1956), Wolff (1989), and Kopczuk and Saez (2004) that are based on the same concepts and
methods as the Flow of Funds, although they are less precise than post-1945 data.
For our purposes, the Flow of Funds data have two main limitations. First, they fail to
capture most of the wealth held by households abroad such as the portfolios of equities, bonds,
and mutual fund shares held by US persons through offshore financial institutions in Switzerland,
the Cayman Islands, and similar tax havens, as well as foreign real estate. Zucman (2013, 2014)
7See Appendix Tables A31 and A32 for data on nonprofit institutions’ wealth and income.
8See Appendix Table C9. Note that Forbes Magazine does not include the wealth transferred to private
foundations in its estimates of the 400 richest Americans either.
9According to the Survey of Consumer Finances, cars, which represent the majority of durables wealth, are
relatively equally distributed so adding durables would reduce the level of wealth disparity but may not have
much impact on trends.
6
estimates that offshore financial wealth amounts to about 8% of household financial wealth at
the global level and to about 4% in the case of the United States. We will examine how imputing
offshore wealth to households affects our estimates. Second, the Flow of Funds evaluates fixed
income claims at face value instead of market value. Changes in Federal fund rates can have
large effects on long-term bond prices (issued at a fixed interest rate) and this variation is ignored
when pricing bonds at face value. Because bonds are very unequally distributed,10 face-value
pricing means that we might under-estimate wealth concentration since the beginning of the
low interest rate period in 2008.
At the aggregate level, the key fact about US wealth is that it is growing fast. The ratio
of household wealth to national income has followed a U-shape evolution over the past century,
a pattern also seen in other advanced economies (Piketty and Zucman, 2014a).11 Household
wealth amounted to about 400% of national income in the early 20th century, fell to around
300% in the post-World War II decades, and has been rising since the late 1970s to around
430% in 2013 (Figure 2). But the composition of wealth has changed markedly. Pensions were
negligible a century ago and now amount to over 100% of national income, while there has
been a secular fall in unincorporated business assets, driven primarily by the decline of the
share of agriculture in the economy. One should not interpret the rise of pension wealth as a
proof that inherited wealth is bound to play a minor role in the future. In 2013, about half
of pension wealth is transmissible at death, namely all individual retirement accounts (IRAs),
defined contribution pensions (such as 401(k)s), and non-annuitized life insurance assets.
3 From Reported Income to the Distribution of Wealth
The goal of our analysis is to allocate the total Flow of Funds wealth depicted in Figure 2 to
the various groups of the distribution. To do so, we begin by looking at the distribution of
reported capital income. We then capitalize this income, and systematically account for wealth
that does not generate taxable income.
3.1 The Distribution of Taxable Capital Income
The starting point is the taxable capital income reported on individual tax returns. For the
post-1962 period, we rely on the yearly public-use micro-files available at the NBER that provide
10According to our estimates, the top 0.1% of the wealth distribution owns about 39% of all fixed income
claims (vs. 22% of all wealth), see Appendix Table B11.
11National income comes from the NIPAs since 1929, Kuznets (1941) for 1919-1929 and King (1930) for
1913-1919.
7
information for a large sample of taxpayers, with detailed income categories. We supplement
this dataset using the internal use Statistics of Income (SOI) Individual Tax Return Sample
files from 1979 forward. All the results using internal data used in this paper are published
in Saez and Zucman (2014).12 For the pre-1962 period, no micro-files are available so we rely
instead on the Piketty and Saez (2003) series of top incomes which were constructed from annual
tabulations of income and its composition by size of income (US Treasury Department, Internal
Revenue Service, 2012). Our unit of analysis is the tax unit, as in Piketty and Saez (2003).
A tax unit is either a single person aged 20 or above or a married couple, in both cases with
children dependents if any. Fractiles are defined relative to the total number of tax units in
the population—including both income tax filers and non-filers—as estimated from decennial
censuses and current population surveys. In 2012, there are 160.7 million tax units covering
the full population of 313.9 million US residents.13 The top 0.1% of the distribution, therefore,
includes 160,700 tax units.
Figure 3 depicts the share of reported taxable capital income earned by the top 0.1%. Capital
income includes dividends, taxable interest, rents, estate and trust income, as well as the profits
of S-corporations, sole proprietorships and partnerships, and excludes interest of municipal
securities (which is tax exempt, although it is reported on tax returns since 1987). We also
report the series including realized capital gains. The series in Figure 3 imperfectly capture the
distribution of the total economic capital income of US families, because not all of it is taxable.
But they nonetheless provide a useful starting point: they display the tax return data with no
assumption whatsoever.
Three results are worth noting. First, the concentration of taxable capital income has risen
enormously. The top 0.1% share excluding capital gains used to be 10% in the 1960s-1970s. In
2012, the latest data point available, it is 33%. Second, part of this rise occurs at the time of
the Tax Reform Act of 1986, and may thus reflects changes in tax avoidance rather than in the
distribution of true economic income. Yet the top 0.1% share including capital gains—which
12SOI maintains high quality individual tax sample data since 1979 and population wide data since 1996, with
information that could be used to refine our estimates. Our estimates use the public use files up to 1995 and the
internal files starting in 1996 (due to methodological changes in the public use files altering its representativity
at the high end starting in 1996).
13US citizens are taxable in the United States even when living abroad. In 2011, about 1.5 million non-
resident citizens filed a 1040 return (Hollenbeck and Kahr, 2014, Figure B p.143, col. 2). These families should
in principle be added to our tax units total. We ignore this issue and leave the task of accounting for the
income and wealth of non-resident citizens to future research. The total number of US citizens living abroad is
uncertain (a recent estimate of the Association of American Resident Overseas puts it at 6.3 million, excluding
government employees). The lack of exchange of information between countries makes it difficult to enforce taxes
on non-residents, so that a large fraction of them do not appear to be filing a return. Our estimates should be
seen as representative of the distribution of income among US residents rather than US citizens.
8