Policy Research Working Paper                      9786




      Human Capital Accumulation at Work
 Estimates for the World and Implications for Development

                                 Remi Jedwab
                                  Paul Romer
                                   Asif Islam
                               Roberto Samaniego




Middle East and North Africa Region
Office of the Chief Economist
September 2021
Policy Research Working Paper 9786


  Abstract
 In this paper, the authors: (i) study wage-experience profiles                     simple accounting framework to find that the contribution
 and obtain measures of returns to potential work experi-                           of work experience to human capital accumulation and
 ence using data from about 24 million individuals in 1,084                         economic development might be as important as the contri-
 household surveys and census samples across 145 countries;                         bution of education itself; and (iv) employ panel regressions
 (ii) show that returns to work experience are strongly cor-                        to investigate how changes in the returns over time correlate
 related with economic development—workers in developed                             with several factors such as economic recessions, transitions,
 countries appear to accumulate twice more human capital                            and human capital stocks.
 at work than workers in developing countries; (iii) use a




 This paper is a product of the Office of the Chief Economist, Middle East and North Africa Region. It is part of a larger
 effort by the World Bank to provide open access to its research and make a contribution to development policy discussions
 around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The
 authors may be contacted at aislam@worldbank.org.




         The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development
         issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the
         names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those
         of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and
         its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent.


                                                       Produced by the Research Support Team
                  Human Capital Accumulation at Work:
 Estimates for the World and Implications for Development


       Remi Jedwab, Paul Romer, Asif Islam and Roberto Samaniego∗




JEL: O11; O12; O15; O47; E24; J11; J31

Keywords: Returns to Experience; Returns to Education; Human Capital; Economic
Development; Labor Markets; Development Accounting




   ∗
    Corresponding author: Remi Jedwab: Department of Economics, George Washington University,
jedwab@gwu.edu. Asif Islam: The World Bank, aislam@worldbank.org. Paul Romer: NYU Stern,
paul@paulromer.net and tracey@paulromer.net. Roberto Samaniego: Department of Economics, George
Washington University, roberto@gwu.edu. An early version of this paper served as a basis for the fifth
chapter of the 2019 World Development Report of the World Bank on “The Changing Nature of Work”. We
thank Simon Gilchrist, three anonymous referees, Barry Chiswick, Carmel Chiswick, Simeon Djankov,
Eric Hanushek, James Heckman, Aart Kraay, David Lagakos, Tommaso Porzio, Federica Saliola and
Todd Schoellman and seminar audiences at George Washington, the Inter-American Development Bank,
Johns Hopkins, Pacifico, Syracuse, UIUC, WALES, William & Mary and the World Bank for their helpful
comments. We gratefully thank Claudio Montenegro, David Newhouse and Minh Nguyen for their help
with the I2D2 database. We gratefully acknowledge the generous support of the World Bank (Office of
the Senior Vice-President and Chief Economist and Social Urban Rural and Resilience Global Practice),
the Cities Program of the International Growth Center (Grant number 89408), the GWU Institute for
International Economic Policy and the GWU Center for International Business Education and Research.
The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors.
They do not necessarily represent the views of the World Bank and its affiliated organizations, or those of
the Executive Directors of the World Bank or the governments they represent.
                                                                                      1


1.      Introduction
Human capital - regarded as the “stock of skills and productive knowledge embodied
in people” (Rosen, 2008) - is the main determinant of modern economic development
(Lucas, 1988; Romer, 1990; Jones and Romer, 2010; Jones, 2019). Work is a form of
school - it teaches new skills, updates old ones, and gives purpose. However, its
contribution to human capital is underappreciated. Since human capital accumulation
has a long-lasting impact, a pertinent query is whether returns to work could serve as a
measure of current and/or future economic development, especially given that it does
not require tracking quantities and prices as GDP per capita does. Returns to work may
focus on aspects of welfare improvements that may capture whether an economy is
“truly” developing instead of, for example, seeing its GDP increase as a result of natural
resource exploitation. We explore this by first measuring the returns to potential work
experience and then by investigating how they behave across several dimensions. These
include the level of development, economic transitions, and recessions.
      Since the seminal works of Schultz (1961), Becker (1964, 1975) and Mincer (1974),
there is a broad consensus that human capital is obtained during childhood, at school,
and at work (World Bank, 2018b). Heckman (2019) hypothesizes that the returns to
early childhood investments are likely higher than the returns to education, which
are in turn possibly higher than the returns to work experience. Subsequently, the
literature has focused on estimating the returns to education or the returns to early
childhood investments. While a substantial literature exists on the returns to experience,
the contributions of experience to human capital and economic development have
attracted comparatively less attention (for exceptions, see Lagakos et al. (2018b,a)).1
      However, the respective contributions of ECIs, education and experience to the total
stock of human capital, and thus economic development, are a function of both their
respective returns and the period over which these returns are accumulated. Indeed,
most children in the world do not attend primary school before age 6, and the average
person in the world accumulates schooling for 10 years but works up to 50 years. In
addition, schools and firms may “teach” different skills (e.g., more theoretical skills
vs. more applied skills), and the skills learned at work may be more directly relevant
  1
    Banerjee and Duflo (2005) survey the literature on the returns to human capital but do not mention
experience. Likewise, excellent surveys provided by Caselli (2005), Flabbi and Gatti (2018) and Rossi
(2020) focus on education.
           2


to the operations of firms. Therefore, it could be that work experience contributes
significantly to human capital in a way that deserves more attention.
   Few studies have consistently estimated returns to experience and explored how
they vary across a sample of developing and developed countries (Lagakos et al.,
2018b,a). Even less is known about the patterns of returns to experience for a global
sample, and how they are affected by policy. Human capital accumulation at work could
have important implications for long-run growth and the divergence in income between
rich and poor countries. Since output per worker is the main contributor of output per
capita, if workers see their earnings increase due to higher returns to experience in their
economy, then tracking the evolution of these returns over time could provide valuable
information regarding the ability of developing economies to eventually catch up to the
income level of developed economies. In this paper, we provide additional evidence
that work experience is a key factor of economic development.
   First, we study wage-experience profiles, and obtain returns using data from 1,084
household surveys and census samples across 145 countries (1990-2016). We find
average population-weighted returns to experience of 2% for the world as a whole.
Returns to education are 4 times higher, at 8%. However, returns to education do not
vary much across income levels, whereas returns to experience are strongly correlated
with economic development. We find returns of 1.7% for developing economies and
3.2% for developed economies, which suggests that workers in developed economies
accumulate almost twice as much human capital at work than workers in developing
economies.     In our preferred specifications, returns are then higher in middle-
income countries than in low-income countries, but only significantly so for a few
specifications. Significantly higher returns to experience in all specifications are only
observed when an economy achieves high-income status. These patterns could support
our conjecture that the evolution of a country’s returns to experience could proxy for
this country’s ability to “catch up” with the most developed economies.
   Second, we explore how returns to experience vary across several economic and
policy-relevant dimensions. Exploiting the fact that we are able to estimate returns to
experience for more than 100 countries and multiple years of data during the period
1990-2016, we employ panel regressions to investigate how changes in the returns to
experience over time correlate with several important time-varying factors. We find
                                                                                            3


that returns are lower for individuals that experienced a recession in the initial years of
their career, showing how recessions can have permanent effects via experience losses.
We find that economies that transitioned out of communism suffer the obsolescence
of past experience, also exhibiting lower returns to experience going forward for many
years. We also find that returns to experience are not correlated with initial stocks of
experience nor education. This suggests that experience may not suffer from decreasing
aggregate returns, and that experience may not be easily substituted by more education.
       Third, our accounting framework shows that the contribution of experience to
economic development might be as important as the contribution of education.
Indeed, education and experience both account for about one third of the income gap
between developing and developed economies. Likewise, the long run impact is similar
whether developing countries attain the same distribution of education as developed
economies or achieve the returns to experience of developed economies. However,
raising returns to experience has a faster impact. This is notable as significant reform to
the schooling system is likely costly and slow. On the other hand, if reforms that improve
the returns to experience are institutional, implementation costs may be lower.
       Our study makes several contributions to the literature. We build on a relatively
sparse literature on the relationship between experience and economic development
(Manuelli and Seshadri, 2014; Lagakos et al., 2018b,a; Dauth et al., 2019; Islam et al.,
2019; Deming, 2021).2 We first produce global estimates of the returns and confirm
the results of Lagakos et al. (2018b) who study how wages vary with experience for
125 samples in 18 countries, however for a much larger sample of 1,084 samples in
145 countries.3 Our sample includes 38 high, 81 middle and 26 low income countries,
which allows us to examine whether the returns vary non-linearly with economic
development. Second, we account for several empirical challenges in estimating the
returns. We use two methods inspired by Heckman and Robb (1985) to attempt to
reduce the experience-time-cohort problem when estimating returns to experience.4
Our large data sets allow us to include decadal cohort fixed effects or country-specific

   2
     See Rubinstein and Weiss (2006), Rosen (2008) and Polachek (2008) for surveys of the literature.
   3
     Lagakos et al. (2018b) use for robustness checks an expanded set of 263 samples in 35 countries. In a
related paper, Islam et al. (2019) use the same 1,084 samples as in the present study to investigate within-
country gaps in the returns to experience between sectors and locations.
   4
     Due to collinearity, experience, time and cohort effects cannot be simultaneously estimated. If year
fixed effects are added, the estimated experienced effects may then also capture cohort effects.
              4


period dummies based on important historical events that affected individuals during
their lifetime. Third, we highlight various stylized facts regarding the determinants
of the returns to experience. Fourth, we study the implications of these returns for
aggregate income. We build on Lagakos et al. (2018a) that uses returns to experience
accumulated in an immigrant’s birth country before migrating to the U.S., by relying on
estimates of the returns directly estimated in each country. More generally, similar to
Lagakos et al. (2018b,a), we attempt to provide a clearer picture of the contribution of
experience to human capital and economic development, and its determinants.
       Our study is part of a long literature that has explored the determinants of human
capital accumulation, whether at the individual or country level.                      A few studies
have provided a variety of global estimates of the returns to education from 1950
onwards (Psacharopoulos, 1994; Psacharopoulos and Patrinos, 2004; Banerjee and
Duflo, 2005; Caselli et al., 2016b; Psacharopoulos and Patrinos, 2018a). These earlier
studies compiled estimates from different studies that use different methodologies
given that the explosion and harmonization of micro-data sets is a relatively new
phenomenon. We expand on these studies by directly estimating returns using a
consistent methodology and for the recent period. We also carefully address specific
econometric challenges related to the measurement of these returns.5
       A sizeable literature has then used cross-country regressions (Romer, 1989; Mankiw
et al., 1992; Hanushek and Kimko, 2000; Hanushek and Woessmann, 2012) or
development/growth accounting analyses (Klenow and Rodriguez-Clare, 1997; Hall and
Jones, 1999; Bils and Klenow, 2000; Caselli, 2005; Hanushek and Woessmann, 2012;
Schoellman, 2012; Caselli and Ciccone, 2013, 2019; Jones, 2014, 2019; Manuelli and
Seshadri, 2014; Hanushek et al., 2017; Hendricks and Schoellman, 2018) to quantify the
contributions of human capital to economic development. We add to these studies by
showing, with our consistently estimated global returns, that experience might directly
explain as much as a third of cross-country differences, which is about as much as

   5
     Banerjee and Duflo (2005) argue that the conclusions of the works of Psacharopoulos and co-
authors might depend on the “quality” of the estimated returns considered. When excluding low-quality
observations (see their Table 2), their cross-sectional regressions suggest that returns to education are
lower in countries with a higher mean number of years of education as well as in richer countries. Using
our consistently estimated returns and panel analyses, we find that returns might possibly be lower in
countries with more education. However, poorer and richer countries do not appear to have significantly
different returns to education. We then explore the same relationships for the returns to experience.
                                                                                         5


schooling. Interestingly, one quarter of the impact of schooling on income is because
schooling indirectly raises returns to experience.
      One important caveat of this study is that we do not uncover the exact mechanisms
by which experience is associated with higher earnings.                       We assume that the
measured differences in returns to human capital reflect differences in human capital
accumulated in different countries, as in Lagakos et al. (2018b).                     An alternative
perspective is that agents with the same amount of human capital are rewarded
differently in different countries due to technological, institutional, organizational
or other factors (Caselli and Ciccone, 2019). Labor market frictions in developing
economies could for example prevent workers from moving up the ladder to better jobs
that fit their skills profile, so that experience generates lower earnings (Lagakos et al.,
2018b). As explained by Caselli and Ciccone (2019), one issue then is that we likely over-
estimate the contribution of human capital to the variation in income across countries.
Their reasoning should apply to both education and experience, and without further
research we cannot assess how this issue differentially affects across income levels the
respective contributions of education and experience to economic development.
      Likewise, human capital could be accumulated at work passively or actively. We
could first assume passive accumulation in the form of learning-by-doing. There is a
long tradition in the growth literature where final output and productive resources are
jointly produced.6 By operating the production technology, agents learn more about it
and improve their ability to operate it. Jovanovic and Nyarko (1996) provide a micro-
foundation for this process by modeling the accumulation of experience as an agent
gradually learning the value of a hidden parameter through Bayesian updating. An
alternative would be to assume that human capital requires investment and that agents
are only compensated for hours worked, as in Ben-Porath (1967) and Rosen (1972). We
find this model to be natural for thinking about education, but less so about experience
since formal (or informal) job training typically only takes a tiny share of the total
number of hours worked in a year. Indeed, Rosen (2008) writes that “Only a small
part of the overall concept is included in formal training programs, apprenticeships
and the like. The greater part is associated with learning from experience.” Still, to
the extent that some of the workers’ reported hours are spent in active training rather

  6
      See for example the “learning by doing” models of Arrow (1962), Uzawa (1965) and Lucas (1988).
             6


than in production, this leads to the underestimation of early-career wages, implying an
overestimate of the returns to experience due to learning-by-doing only, particularly in
developed economies where more time is spent in training. Given that we do not have
information on how much time each worker in our data spends working or training,
we cannot directly assess the contribution of each mechanism. However, indirect
robustness checks based on cross-country data on the number of hours spent in formal
or informal training suggest that our results are more consistent with learning-by-doing
than the learning-versus-doing model of Ben-Porath (1967).

2.      Data Used to Estimate the Returns
Sources. The data source for the analysis is the International Income Distribution
Database (I2D2) of the World Bank.           The database consists of a large number of
individual-level surveys and census samples. The surveys include household surveys
and labor force surveys. The data was initially compiled by the World Bank’s World
Development Report unit between 2005 and 2011. The original version was used by
Montenegro and Hirn (2009) to study the labor market characteristics of developing
countries for the 2009 World Development Report of the World Bank. The database
has since been expanded and used by some of the members of the original World
Development Report unit.7 We use the expanded December 2017 vintage of the I2D2
database.    Only select members of the research team or individuals in charge of
harmonizing the data can access the database. However, a significant number of the
surveys can be accessed online or after entering an agreement with the countries’
respective statistical office. Finally, the surveys are nationally representative and large
enough (i.e., typically have more than 10,000 observations) for our purpose.
Sample Size. The version of the I2D2 database that we use includes about 1,500
survey/census samples. However, wages are only reported for about two thirds of them.
In addition, we restrict our analysis to workers aged between 18 and 67 with data on age
and education as well as samples with at least 10 observations in each experience bin
(see details below) to ensure results are not driven by small samples.
      The baseline sample that we obtain includes 24,437,020 individuals from 1,073
surveys and 11 censuses in 145 countries from 1990-2016 (median number of samples
  7
    The I2D2 database has been used to study labor markets or returns to education (Montenegro and
Patrinos, 2014; de Hoyos et al., 2015; Gindling and Newhouse, 2014; Gindling et al., 2016).
                                                                                      7


per country = 6; mean = 7.5; min = 1; max = 44). We then include surveys from the 1990s
to make sure we increase the number of countries in our sample. In particular, data is
available for 176 survey/census samples in 63 countries in 1990-1999, 630 samples in
129 countries in 2000-2009 and 278 samples in 107 countries in 2010-2016. Note that
several survey samples may be available for a same country-year.8 Finally, the data is
highly representative of the world economy. The 145 countries comprise about 95% of
the world’s population today. The data is representative for both developed countries
(76% of their population) and developing countries (98%).
Main Variables. For the 24,437,020 individuals, we know their monthly wage, age,
and number of years of education. For a substantial subsample, we know the hourly
wage and the number of hours worked. Lastly, we know their employment status (e.g.,
self-employed or paid employee) their sector (ISIC-1: agriculture, industry, services),
their occupation (ISCO-08: managers, professionals, technicians, clerks, crafts, sales,
machinists, elementary occupations, skilled agriculture, other), and whether they work
in the formal sector or live in urban areas. Finally, note that I2D2 mostly calculated the
wage of self-employed individuals as the amount of salary taken from the business.
Data Limitations. While I2D2 is particularly useful in terms of external validity, the
fact that all these surveys/censuses had to be harmonized implies that few variables
are available (about 70). Also, the I2D2 team does not provide details on how the
harmonized variables were created for each survey, and thus the level of consistency
across surveys. However, given the wide use of the data across flagship World Bank
reports, we feel compelled to trust that the I2D2 team did a good enough job that the
results generated in our study do not capture statistical artefacts. Having said that, the
only significant issue comes from the potential mismeasurement of wages. There could
be issues with how self-employed individuals calculate and self-report the amount of
salary taken from the business. As a result, we will examine the robustness of our results
when excluding self-employed individuals. Next, 99% of these samples are surveys
because censuses rarely include data on wages. In addition, we measure potential work
experience as I2D2 does not include direct measures of work experience. Ability (e.g.,
test scores) is also not measured. Finally, the data are not panelized.

  8
   Web Appx. Fig. A.1 shows the number of countries with available data for each year from 1990-2016.
The Web Appendix is available on the website of Remi Jedwab: www.remijedwab.com.
             8


Data Quality. In Web Appx. Section A.1,9 we verify that the I2D2 database generates
global patterns that are broadly consistent with patterns observed when using other
global databases such as the World Development Indicators (WDI) database of the World
Bank. More precisely, we find that our samples are globally representative in terms of
per capita incomes, age structure, education, and self-employment.10

3.     Methodology for the Estimated Returns
We calculate potential work experience – which we call “experience” in the rest of the
analysis – as follows: (i) For individuals with at least 12 years of education, we assume
children start school at age 6 and calculate experience as age - years of education - 6; (ii)
For individuals with less than 12 years of education, we assume that experience before
age 18 is inconsequential and calculate experience as age - 18. With these calculations,
the mean (population-weighted) number of years of experience differs little between
developing economies – 19 – and developed economies – 22. Indeed, while workers in
developed economies study longer on average, they also tend to live longer.
Estimated Returns. For individual i and sample t, we use OLS to estimate the following
model for each country one by one:
                                          7
                              lnWit =         βe expite + γeduit + θt + εit                           (1)
                                        e=1
where the dependent variable is the log of monthly wages (lnWit ).                       Experience is
categorized into seven bins (expite ). The bins are [5-9 years] (which we call 5), [10-14]
(10), [15-19] (15), [20-24] (20), [25-29] (25), [30-34] (30), and [35+] (35). The omitted bin
is [0-4] (0). We control for the number of years of education (eduit ). We include sample
fixed effects (θt ) to capture country-year-sample unobservables.11
Profiles. For illustration, we use the coefficients of the experience dummies to construct
wage-experience profiles for the United States and Bangladesh (see Figure 1). In the
U.S., a worker with 30 years of experience earns almost twice more than a worker
with zero experience. The steepness in the profile is attributed to the first 25 years,
particularly the first 15 or even 5. In Bangladesh, the profile is flatter. The same patterns
   9
     The Web Appendix is available on the website of Remi Jedwab: www.remijedwab.com.
  10
     For example, we find that log mean earnings in I2D2 are strongly correlated across country-years with
log per capita GDP in WDI. We show that wage variance is in line with what is expected. Finally, we verify
that the resulting Mincerian R-squared are not too different across various groups of countries.
  11
     Most countries having several samples, we use individual weights divided by the size of the sample.
                                                                                           9


are observed if we widen the scope by replacing the United States with all developed
countries and Bangladesh with all developing countries (developed countries are high-
income countries in 2017 according to the World Bank classification). To obtain these
profiles, we obtain the average wage differential for each country (1990-2016) and
construct the mean wage differentials for each group of countries weighted by the
population of each country in 2017. The “median” developed (developing) country is
Belgium (Jordan).
Returns. Measures of the returns to experience should take the integral below the
profiles.12 For each bin (5, 10, etc.) one by one, we estimate an annualized return
and then take the mean of the annualized returns across the seven bins to obtain
the mean annualized return throughout the experience distribution. More precisely,
for individuals belonging to bin e, we obtain the bin-specific annualized return as
((βe + 1)(1/e) − 1) ∗ 100, with βe being the estimated coefficient for bin e in equation
(1). For this subgroup of individuals in the society, it tells us by how many percentage
points wages increased on average for each extra year of experience. We then take the
average of these seven bin-specific annualized returns so that each bin is represented.13
       We find average population-weighted returns of 2.0% for the world. In contrast,
returns to education, measured by the coefficient of schooling, are 4 times higher, at
8.2%. Our specification has several advantages over the Mincerian specification that
includes a quadratic function of experience: (i) With seven coefficients, it is more
flexible and informative, especially given that 34 countries have non-monotonic wage-
experience profiles; (ii) We construct returns that are independent of the distribution of
experience in the country. Since we keep samples with at least 10 observations in each
bin, and then give each bin the same weight, the constructed returns are less affected
by the facts that developed countries and developing countries have more experienced
workers and more inexperienced workers, respectively;14 and (iii) We can construct
measures using only specific parts of the experience distribution (e.g., 0-25 or 0-15),

  12
      Lagakos et al. (2018b) calls this the “sum of heights”.
  13
      In our view, it is important to use information from the seven experience coefficients. However, we
acknowledge that there is no right or wrong way to use this information to construct the returns.
   14
      A related issue with the Mincerian specification is how to calculate returns from the coefficients of
experience (β1 ) and its square (β2 ). It is common to calculate the returns as β1 + β2 × the median number
of years of experience. However, this number varies across countries. This method then captures the local
return at the median experience level, not for the full distribution of experience in the country.
              10


which will prove important when dealing with changes in a country’s economic system
and selection effects in and out of the workforce or selection effects related to mortality.

4.       Returns for Developed versus Developing Economies
4.1.      Baseline Results

Figure 2 shows the strong positive relationship between returns to experience and log
per capita GDP for the mean year in the data for each country (slope = 0.33***).15
The main sample of Lagakos et al. (2018b) excludes countries from the former Soviet
Union and Sub-Saharan Africa. Some ex-USSR countries (e.g., UZB for Uzbekistan) have
relatively low returns and some African countries (e.g., UGA for Uganda) have relatively
high returns for their income level.16 Regarding the former, a change in the economic
system made experience obsolete (details will be provided in Section 5.3.). Regarding
the latter, high mortality may lead to a selection effect where late-age workers are more
able in unobservable ways (Section 4.6.). As a result, the returns of poorer countries
may be over-estimated, causing the returns gap between developed and developing
countries to be under-estimated. However, a downward bias is less consequential than
an upward bias for our analysis. Finally, we will examine in Section 4.6. below how
results vary between low-, middle- and high-income countries.
       For 145 countries c, we then regress the returns (RetExp) on a dummy equal to one
if the country is classified as high-income by the World Bank (Dev ):
              Returns to Experiencec = α + β Deveveloped Country Dummyc +                          (2)

Row 1 of Col. (1) in Table 1 shows the baseline results. The first subcolumn shows the
constant (α), i.e. the mean return for developing countries (G ). The second subcolumn
shows the coefficient (β ) of the developed (high-income) country dummy, i.e. the
difference between the mean returns of developed countries and developing countries.
(D-G ). Finally, the mean return for developed countries (D ) is obtained by adding the
constant and the coefficient of the developed country dummy and is shown in the third
subcolumn. The mean returns for developing countries and developed countries are
1.7*** and 3.2***, respectively. The difference between the two is significant (1.5***).

  15
     The GDP is in PPP terms and constant 2011 USD. For countries with multiple years of data, the mean
year is the average sample year using as weights the number of observations in each sample.
  16
     Web Appx. Fig. A.2 shows a world map of the returns. Web Appx. Table A.7 lists all the returns.
                                                                            11


4.2.   Robustness when Attempting to Control for Cohort Effects

Without retrospective data on individual employment, experience is necessarily
estimated as age minus schooling. Time is then equal to age plus the year of birth.
Thus, one cannot simultaneously include schooling, experience, age, time fixed effects
and cohort fixed effects in the same estimation (Heckman and Robb, 1985, p.137), in
our case equation (1). Therefore, we cannot include year of birth fixed effects and we
also cannot distinguish the effects of experience and age. See Web Appx. Section A.2 for
details on what is commonly called the experience-time-cohort problem.
HLT Approach. The “HLT” approach is coined by Lagakos et al. (2018b) to describe
their main empirical strategy as it is inspired by Heckman, Lochner and Taber (1998).
As described in Lagakos et al. (2018b), some theories of life cycle wage growth claim
that there should be “no effect of experience on wages near the end of the life cycle”.
One can then “follow a fixed cohort across multiple cross sections for the last years
of their working life” and attribute any wage changes for them to time effects. Once
time effects are recovered, “it is straightforward to estimate the experience and cohort
effects of workers who are not near the end of the life cycle”. However, wage profiles
could decrease at the end of the life cycle because of human capital depreciation. Thus,
“this approach requires assumptions about two main parameters: first, the number of
years at the end of the life cycle for which there are no experience effects and, second,
a number for the depreciation rate”. They consider 5 or 10 years and 0 or 1%. While
this approach has clear merits, it posits that there are no experience effects near the end
of the life cycle. Also, selective early retirement in developed countries and selective
mortality in developing countries complicate the analysis, since estimated time effects
near the end of the life cycle could capture unobservable compositional changes in the
population of the older cohorts across years.
Our Approach. Heckman and Robb (1985, p.145) propose to include cohort effects
consisting of “a sequence of adjacent years (e.g., Depression or 1950s youth, etc.)”.
In Col. (2), we include decadal cohort fixed effects in the country-specific regressions
estimating the returns, but only for 122 countries with at least two years of data. Since
we focus on workers aged 18-67 in samples from 1990-2016, we study individuals born
between 1923 and 1998. We thus include 8 cohort dummies (1920s, ..., 1990s). This
approach is only valid if cross-sections are fewer than ten years apart. As explained
               12


in Web Appx. Section A.3, this issue only affects a few older individuals in the earliest
surveys of 12 countries. Using returns based on the 0-25 bins, as we estimate later, also
eliminates the issue. Finally, results hold if we exclude the 12 countries or exclude only
the problematic surveys that number 12 in total and involve 7 countries.17
       An alternative approach is to construct for each country periods based on important
years, e.g. for the United States, 1923-1928, 1929-1932, 1933-1940, 1941-1944, 1945-
1953, 1954-1961, 1962-1963, 1964-1967, 1968-1973, 1974-1979, 1980-1989, 1990-1993
and 1994-2000, due to important events in 1929 (Great Depression), 1933 (New Deal),
1941 (Pearl Harbor), 1945 (World War II ends), 1954 (ends of racial segregation in
schools), 1962 (Cuban Missile Crisis), 1964 (Civil Rights Act), 1968 (Martin Luther King
assassinated), 1974 (Watergate), 1980 (Reagan), 1990 (Gulf War starts, Cold War ends),
1994 (NAFTA), 2001 (9-11), and 2007 (Great Recession).18 We identify important events
using information from standard sources such as Wikipedia. For each country, we read
the top of the Wikipedia webpage of the country (in English as well as in the main
language of the country) as well as the “contemporary/modern period” and “colonial
period” sections.19 We cross-check this information using other sources, e.g. the BBC
Profile of each country.20 We then include in the regressions estimating the returns
country-specific period dummies equal to one if the individual was aged between 18
and 67 during the period(s), for countries with at least two years of data. Doing so
captures the fact that each person was affected by multiple events during her lifetime.
For example, an American person born in 1931 was 18 in 1949 and 67 in 1998. For
her, the 1945-1953, 1954-1961, ..., 1990-1993 and 1994-2000 period dummies are equal
to one. Of course, blindly following Wikipedia-BBC may lead us to include events
that might not be that “important” for workers. At the same time, it is better if we
  17
      Web Appx. Section A.3 and Web Appx. Table A.8 show that effects do not vary much if we increase the
number of years of data required for a country to be considered or use 20-year cohort fixed effects. The 5-
year cohort fixed effects preserve the difference between developed and developing countries. However,
returns are lower across all countries, since collinearity increases as one uses more refined cohort fixed
effects.
   18
      For Kenya: 1920 (Crown Colony), 1940 (Italian attack), 1944 (Kenyan African Union), 1947 (Kenyatta
becomes KAU leader), 1952 (Mau Mau rebellion), 1956 (rebellion ends), 1960 (state of emergency ends),
1963 (Independence), 1969 (Mboya assassinated), 1978 (Kenyatta dies, rule of Moi), 1982 (one-party
state), 1987 (opposition suppressed), 1992 (elections, conflict), 1996 (constitution amended), 2002 (Moi’s
reign ends), 2007 (electoral violence), 2010 (constitution, East African market), 2013 (terrorist attacks).
   19
      See https://en.wikipedia.org/wiki/United States for the United States. Accessed 03-15-2019.
   20
      See https://www.bbc.com/news/world-us-canada-16761057 for the United States. Accessed 03-15-
2019.
                                                                               13


blindly follow their selection of important events rather than us cherry-picking the
events ourselves. Selecting too many events by including events that might not appear
so “important” should also lead to more conservative estimates because the included
cohort fixed effects are then more refined, i.e. lump together fewer years than if we were
not using these events. However, in the rest of the analysis, we will privilege the results
with decadal cohort fixed effects.
   Both approaches reduce returns for both developing countries and developed
countries, to 0.8-1.1*** and 2.6-2.8*** respectively (N = 122; row 1). They accentuate
the gap between developed and developing countries (to 1.7***-1.8***). However, while
these two approaches help capture cohort effects, they do not fully solve the experience-
time-cohort problem since there may still be yearly cohort effects.
   Web Appx. Section A.3 and Web Appx. Table A.8 show results hold if: (i) periods
are constructed based on ages 18-40 or 18-30, in case important events are more
consequential in the first decades or years after high school; (ii) we add as a control
the country’s average per capita GDP growth rate for those years during which the
individual was aged between 18 and 67, since the identified important years may
not capture well uneventful periods of growth or recessions; (iii) add country-specific
period dummies equal to one if the individual was aged between 0 and 17 during the
period(s), and the individual’s corresponding growth rate, in case important events are
consequential during childhood; (iv) combine the decade fixed effects, events fixed
effects and growth rate approaches; and (v) use Lagakos et al. (2018b)’s HLT approach.
Like them, we consider four versions of the HLT approach, based on the last 5 or 10
years of experience and using a depreciation rate of human capital of 0 or 1% (see rows
12-15).
4.3.      Other Main Results
Hourly Wages vs. Hours Worked. Results hold if we use returns based on hourly wages
(row 2 of Table 1). Indeed, monthly wages can be decomposed into hourly wages and
hours worked (N = 131; 105; 105). Hours worked slightly increases with development
when ignoring cohort effects (Col. (1) in row 3). Yet, since this effect is very small, hourly
wages, and thus hourly productivity, explains most of the returns based on earnings.
However, using monthly wages allows us to use the full sample of 145 countries.
Selected Bins. Results hold if we use returns (N = 145; 122; 122): (i) Based on the 0-25
             14


(row 4) or 0-15 bins (row 5), to minimize selection issues due to changes in economic
systems and selective mortality or early retirement disproportionately affecting more
experienced workers; (ii) Based on the 5-35 bins (row 6).21
Excluding Selected Types of Workers. Results hold for returns excluding female, part-
time, self-employed and public sector workers, considering both monthly wages (row
7; N = 124; 101; 101) and hourly wages (row 8; N = 122; 98; 98).22 Web Appx. Table A.14
(rows 2-5) shows results hold if we exclude each group of workers at a time. Excluding
female workers slightly raises the returns in both developed and developing countries. If
anything, the developed-developing gap slightly increases. Excluding part-time workers
lowers the returns in developing countries when cohort fixed effects are included, thus
also widening the developed-developing gap. Excluding self-employed workers does
not substantially affect the returns. The developed-developing gap slightly increases.
Finally, excluding public sector only slightly reduces the returns, and the gaps.
Weights. Results hold when country populations in 2017 are employed as weights, thus
giving more weight to larger countries (row 9; N = 145; 122; 122).
Complementarities.          Rows 10-13 examine if returns vary across education levels.
According to UNESCO (2012), individuals finish high school after 12-13 years of
education, thus 13 years is a reasonable cut-off. In developed countries, returns are
slightly lower for more educated individuals (rows 10-11; 3.4*** vs. 3.2***; N = 143-122-
122 in row 10, 124-98-98 in row 11). In developing countries, returns are higher for
more educated individuals (2.1*** vs. 1.7*** in col. (1)). However, we lose about 20-25
countries because we do not have at least ten individuals for each experience-education
bin. To remedy that issue, we use country population circa 2017 as weights (rows 12-
13; N = 143-122-122 in row 12, 124-98-98 in row 13), making compositional changes due
to smaller countries less consequential. Results are unchanged.23
Main Returns. In our simulations we use population-weighted education-specific

  21
     Web Appx. Table A.9 shows that local returns are higher for lower-experience bins whereas
Web Appx. Table A.12 shows that results hold for different combinations of local returns.
  22
     Lagakos et al. (2018b) exclude: (i) female and part-time workers because potential experience is a
better proxy of actual experience for male and full-time workers; (ii) self-employed individuals because
their wages may also include the returns to the capital investments they make. In our case, I2D2 calculates
their wage as the amount of salary taken from the business; and (iii) public sector workers because they
receive nonwage compensation and their wages may not reflect the full payment for their labor.
  23
     Results hold if we use 12 or 14 years or information about high school graduation
(Web Appx. Table A.13).
                                                                                               15


returns based on the 0-25 bins.               Returns are higher but the previously discussed
differences are maintained, with the corresponding returns being 4.4*** and 4.2*** for
developed countries and 2.0*** and 2.8*** for developing countries (rows 14-15; N = 143-
122-122 and 124-98-98). These results hold if exclude female, part-time, self-employed
and public sector workers (rows 16-17; N = 122-100-100 and 91-72-72).

4.4.      Robustness Checks

Estimates of Returns. Web Appx. Section A.4 and Web Appx. Table A.14 show results
hold if we use country-specific ages of entry in school or drop samples exhibiting
age heaping.24 Teenagers and children then disproportionately work in developing
economies.        We may thus mismeasure experience in developing countries.                               This
could affect our results if we under-estimate the returns of developing countries, thus
wrongly asserting how large the gap in the returns is between developed and developing
countries. We indeed find that including potential teen labor experience (from age
15) or potential child labor experience (from age 13) raises the returns slightly more
in developing countries – from 1.7 to 1.9-2.0 (see rows 7-8 of Web Appx. Table A.14)
– than in developed countries – from 3.2 to 3.4 (ibid.).25 The gap in the returns then
either remains the same, at 1.5, or slightly decreases to 1.3 (ibid.). If we consider
teen/child labor for developing countries only, the gap is 1.3-1.2. Also controlling for
cohort effects, the gaps also remain high, at 1.5-1.3 and 1.7-1.4. Overall, the returns and
the developed-developing gap remain essentially similar, at least enough not to change
the main implications of our results.26
       Results then generally hold if we (see Web Appx. Section A.4 and Web Appx. Table
A.14): (i) Restrict the sample to post-2000 samples to minimize issues arising from ex-
USSR countries; (ii) Use three- or seven-year bins; (iii) Keep workers for which the wage
was reported for a period of at least a month since earnings may be mismeasured for
workers paid hourly, daily or weekly; (iv) Include the wage of any secondary occupation
(available for fewer countries); (v) Use mean wages, experience and education in the

  24
     Results hold if we drop samples with a rough Whipple index. The index measures to what extent
individuals disproportionately report their age as a round number ending in 0 or 5.
  25
     For teen (child) labor, for individuals with at least 9 (7) years of education, experience is age - years of
education - 9 (7). For individuals with less than 9 (7) years of education, experience is age - 15 (13).
  26
     Web Appx. Figure A.4 replicates Figure 1 but add the respective wage-experience profiles of developed
countries and developing countries when including teen labor (“TL” in the figure) and child labor (“CL”).
              16


household to account for intra-household optimization. This is particularly important
for households where women do not work or for multi-generational or multi-family
households; and (vi) Keep samples with at least 100 observations per experience bin.
       Additionally, we focus on workers, which causes a selection problem in samples
with skewed unemployment/non-labor participation (NLP) rates. In the context of
low unemployment/NLP, non-workers may have specific unobservables. That is also
true in contexts with high unemployment/NLP where competition for few jobs selects
the “best”.     That is only an issue for us if unemployment/NLP is correlated with
development.       However, we find no such correlation and results hold if we (Web
Appx. Section A.4 and Web Appx. Table A.14): (i) Add controls for selection, i.e. sector
fixed effects, occupation fixed effects, an informal sector dummy, an urban dummy, and
the square of education, to compare workers that are similar except in their experience
level; (ii) Keep or drop samples with unemployment or NLP rates below the 25th
percentile (7% and 35%, respectively) or 50th percentile (10% and 35%, respectively)
in the sample; and (iii) Use the Heckman correction with the selection equation also
including marital status and the number of children in the household.
Developing vs. Developed. The developed-developing gap holds after conducting the
following checks (Web Appx. Section A.5 and Web Appx. Table A.15): (i) Add seven World
Bank region fixed effects; (ii) Drop each developing region one by one or ex-Communist
countries; (iii) Drop the top and bottom 10% outliers in the returns, i.e. countries that
may have abnormally high or low returns, either because these returns are real or
because of measurement error; (iv) Add country-level controls for factors generating
cohort effects: whether a country had a Communist regime,27 past child and adult
mortality rates (including for HIV),28 and past per capita GDP or population growth
rates29 for each country-decade from 1950 to 2010 (source: United Nations (2019)).
       In addition, if agriculture exhibits lower returns to experience and given that
developing countries have a higher share of their workers in agriculture (37% vs. 5%
  27
     These include a dummy equal to one if the country ever had a Communist regime, the number of
years the country has had a Communist regime, and the year the country transitioned to capitalism
(source: Wikipedia (2019)). Since we are controlling for the country-specific mean year in the data, this
also controls for how many years ago the country transitioned to capitalism.
  28
     These include average child mortality rates (for 0-5 year-olds) and adult mortality rates (15-45) for
each country-decade from 1950 to 2010 (source: United Nations (2019)) and average HIV infection rates
for each country-decade from 1990 to 2010 (sources: World Bank (2018a); data unavailable before 1990).
  29
     These include average annual per capita GDP growth rates (source: see footnote 15).
                                                                                        17


for developed countries), then their lower returns could be due to lack of structural
change. For 111 countries with sectoral data, we obtain the wage-experience profiles
of workers currently belonging to the agricultural sector (Ag) or the non-agricultural
sector (Non-Ag).30 A number of observations are provided in Web Appx. Fig. A.5. First,
for both developed and developing countries, the Non-Ag profile is above the Ag profile.
Second, the Non-Ag profile of developing countries is below the Ag profile of developed
countries. Lastly, the gap in the profile for Ag appears to be similar to the gap in the
profile for Non-Ag. Relatedly, the D-G gap in the Ag returns (1.5***) is similar to the gap
in Non-Ag (1.6***) as well as the aggregate gap (1.5***). Therefore, structural change (or
lack thereof) does not explain our patterns.31
       Finally, standard errors (SE) in eq. (2) ignore the fact that the coefficients of the
experience dummies may be imprecisely estimated in eq. (1). However, 75-92% out of
the 145 x 7 = 1,015 coefficients have a p-value below 1%. Furthermore, the medians of
the mean and max SEs across the seven coefficients in our sample are low, at 0.02.32

4.5.      Returns to Education and Complementarities
Returns to Education. We use the coefficient of schooling as a proxy for the returns
to education. Table 2 replicates our main Table 1 except that we use as our dependent
variable the estimated returns to education. Row 1 shows that returns to education are
not significantly different between developed and developing countries, whether one
ignores cohort effects – ≈ 8.7-9.0*** – or attempt to control for them – ≈ 8.4-8.8*** –
(N=145, 122, 122). Web Appx. Fig. A.3 also shows a statistically insignificant relation
between returns to education and log per capita GDP (PPP) for the mean year in the data
for each country. Rows 2-3 then show most of the returns in terms of monthly wages are
driven by differences in hourly wages (row 2, N=131, 105, 105). Education increases the
number of hours worked in developed countries, but the developed-developing gap is
only significant when no cohort fixed effects are included (row 3, ditto).
       Next, mean returns are lower for both developing and developed economies, and
the developed-developing gap is positive and significant, if we restrict the sample to
  30
     We do not know sectoral employment history. The returns are thus estimated using the current sector.
  31
     More precisely, we obtain the following (population-weighted) returns to experience for Ag and Non-
Ag: 1.2 and 2.0 for developing countries and 2.7 and 3.6 for developed countries (not shown).
  32
     Results hold if we only use countries whose mean or max SE is below the median of the mean/max
SE in the sample or bootstrapped SE 10,000 times (Web Appx. Section A.5 and Web Appx. Table A.15).
               18


the type of workers studied by Lagakos et al. (2018b). However, this is only true if we
consider monthly wages (row 4, N=123, 96, 96), not hourly wages (row 5, N=121, 93, 93).
In row 4, the gap is also relatively small compared to the absolute values of the returns in
both groups of countries. Indeed, the mean return to education for developed countries
is only 20-25% higher than the mean return to education for developing countries
(vs. 90-225% for experience based on the results of row 1 in Table 1).
       The returns to education of developing countries slightly decrease and the returns
to education of developed countries slightly increase when country populations in
2017 are employed as regression weights (row 6; N = 145; 122; 122). The developed-
developing gap is significant in col. (1), but the difference remains small relative to the
absolute values of the returns. Indeed, the mean return to education for developed
countries is only 27-30% higher than the mean return to education for developing
countries (vs. 100-150% for experience based on the results of row 9 in Table 1).
       We find relatively higher returns for individuals with more than 13 years of education
than for individuals with less than 13 years. When ignoring cohort effects and using
population weights (col. (1)), these are 13.0*** vs 6.7*** in developed countries and
13.8*** vs. 6.8*** in developing countries (rows 9-10; see rows 7-8 for the equivalent
returns when not using weights). If we try to account for cohort effects (col. (2)-(3)),
the returns to education are similar, at 12.7-13.0*** vs. 6.6*** and 13.1-13.4*** vs. 6.8***
in developed and developing economies, respectively. If we exclude female, part-time,
self-employed and public sector workers, results are similar (rows 11-12).33
Literature. How do our returns to education compare with existing estimates from the
literature? Earlier studies such as Psacharopoulos and Patrinos (2018b) (henceforth,
PP18) and Caselli et al. (2016a) (henceforth, CPR16) compile estimates from different
studies. PP18 reports estimates of the returns for 72 countries during the same period
as this study (1990-2016).          Using population weights, the estimated PP18 average
returns are 9.9 and 9.0 for developed countries (N = 5) and developing countries (67),
respectively. We obtain respective returns of 10.3 and 7.9 in our data (145; row 6 of Table
2 Col. (1)). The correlation between our data and their data is about 0.6 (62; see Web

  33
     These returns are not necessarily causal. They are nonetheless consistent with causal studies (Jedwab
et al., 2020). Furthermore, if returns are over-estimated because better able individuals study longer, the
causal returns would mean a lower relative contribution of education to economic development than
what we obtain, reinforcing our result that experience significantly matters.
                                                                                        19


Appx. Fig. A.6). CPR16 add their own estimates from different studies to the estimates
provided by PP18. The CPR16 database consists of 178 estimates defined c. 1995 or 2005.
Using population weights, the estimated CPR16 average returns are about 9.0 for both
developed observations (N = 48) and developing observations (130), respectively. The
correlation between our data and their data is about 0.4 (117; see Web Appx. Fig. A.6).
Comparability is likely limited because both PP18 and CPR16 rely on different studies
that use different methodologies (e.g., 89 in CPR16) whereas we directly estimate our
returns using a consistently harmonized global dataset and a consistent methodology.

4.6.      High-Income vs. Middle-Income vs. Low-Income Countries
In this subsection, we show that, using our best specification and accounting for various
issues, the relationship between returns to experience and economic development
is monotonic.       While the returns are twice higher in developed countries than in
developing countries, among developing countries returns are only slightly higher
in middle-income countries than in low-income countries, which explains why the
relationship is not fully monotonic. Returns to experience thus only greatly increases
as economies become developed, which suggests that returns to experience may proxy
for an economy’s ability to converge and catch up to high-income economies.
Graphical results. Figure 2 suggests that the relationship between returns to experience
and economic development is U-shaped, with low-income countries (most of them in
Africa) having higher returns than middle-income countries. We examine if this U-
shape pattern holds across the various specifications investigated previously or whether
it is a statistical artefact resulting from selective mortality raising the returns of poorer
countries and transitions to capitalism having made experience more obsolete in ex-
Communist countries which are now majoritarily middle-income countries.34
       In Web Appx. Table A.16, we examine the Spearman rank-order correlation between
the returns and log per capita GDP (PPP; constant 2011 USD; for the mean year in
the data for each country). The baseline correlation is 0.38 (N = 145). Including
decadal cohort fixed effects, using returns based on the 0-25 bins, and employing
country populations (2017) as weights increases the value to 0.60 (specification 1). If
  34
    Since most low-income countries are in sub-Saharan Africa and to ensure that our results are not
driven by a lower quality of surveys from the region, we verify that the patterns that we obtain for sub-
Saharan Africa in I2D2 are consistent with the patterns observed when using other global databases such
as the World Development Indicators database of the World Bank (see Web Appx. Section A.1).
             20


we additionally exclude female, part-time, self-employed and public sector workers, the
value becomes 0.58 (specification 2). Thus, the relationship appears to be moderately
monotonic.
     Excluding countries in the top 5% in past child or adult mortality rate (averages
in the 1950s-2010s) or HIV prevalence rate (1990s-2010s) increases the values for the
two specifications to 0.63 and 0.60, respectively. If we also exclude ex-Communist
countries, the correlations become 0.75 and 0.57, respectively, indicating a strongly
monotonic relationship. The relationship is still not fully monotonic, which suggests
non-linearities. These come from developing countries, not developed countries.
     If we focus on specifications 1 and 2 and exclude ex-Communist and high-mortality
countries, the correlations respectively increase to about 0.8 and 0.6 if we additionally
exclude low-income countries (not shown).           If we exclude high-income countries
instead, the correlations respectively decrease to about 0.3 and 0 (ibid.). Thus, the
relationship is barely monotonic among developing countries, which we discuss next.
Income Group-Specific Returns. We modify equation (2) to obtain the mean returns
to experience for low-income countries, middle-income countries and high-income
countries in 2017 according to the classification of the World Bank.           As seen in
Web Appx. Table A.17, with specification 1 or specification 2, returns are slightly
higher for middle-income countries than for low-income countries (at 1.8-2.0 vs. 1.6-
0.8), but not significantly so. Simultaneously excluding high-mortality/HIV countries
and ex-Communist further increases the gap between low-income and middle-income
countries. However, the difference is only significant in specification 2.

5.     Variation in Returns to Work Experience Over Time
Returns to experience are available for 951 country-years for the time period 1990-
2016. Using this data, we investigate using panel regressions how the returns vary with
economic development, recessions and economic transitions as well as existing stocks
of human capital. While estimates derived are not causal, the aim is to highlight novel
stylized facts about how returns to experience vary across countries over time.
     More precisely, for countries c and years t, we estimate the following model:

                    RetExpc,t = α + β Var. of Interestc,t + λc + θt + µc,t            (3)

with Var. of Interest being the variable explored as a potential predictor of the returns
                                                                                          21


to experience. We focus on 122 countries with returns available for at least two years.
Using only the first year and last year available for each country allows us to estimate
the long-difference effect of changes in the variable of interest on the returns.35
       We also construct a five-year panel centered around the years 1994, 1999, 2004, 2009
and 2014.36 Standard errors are clustered at the country level. One advantage of the
five-year panel is that the sample is larger and the periodicity fixed. However, they only
capture short-term effects and returns may be too “slow moving” with limited variation.
       Our main analyses use returns to experience constructed using 0-25 bins, hence 25
years. In a panel framework limited to a period of 26 years (1990-2016), using fewer bins
is warranted given that returns are mechanically more correlated over time. However,
returns based on only one bin (0-5) are also limited as they are based on 5 years only. We
choose an intermediate approach and study returns constructed using the 0-15 bins.37

5.1.      Economic Development and Returns to Work Experience

In col. (1) of Panels A and C in 3, we report the long-difference and five-year correlations
between the returns to experience and economic development proxied by a dummy
variable equal to one if the country was “developed” (had reached high-income status)
by year t, respectively. The correlations are quite high, at 1.03**-1.11**, and not that
much lower than the baseline cross-sectional effect – 1.5*** – estimated in column (1)
of row 1 in Table 1. In contrast, for the returns to education, the correlations are small
and not significant (Web Appx. Table A.19), broadly consistent with the cross-sectional
analysis.
       In col. (2), we report the correlations between the returns to experience and two
dummy variables that are equal to 1 if a country reached high-income status or middle-
income status by year t respectively. The coefficient of the middle-income status
dummy variable is positive but not significant. The coefficient of the high-income
status dummy in Panel A is now larger. The statistically insignificant results for middle-
income status are consistent with the weak results obtained from the cross-sectional
  35
     Since there are countries for which the first year and last year are only separated by a few years, we
use as weights the number of years between the two years.
  36
     For each year (e.g., 2014), we use the year’s return when available. If not, we use the mean return for
the year before (2013) and/or after (2015). If still not available, we use the mean return for the two years
before and after (2012, 2016). For 1994, if needed, we use three (1991) and four (1990) years before.
  37
     Results generally hold if we use other combinations (not shown, but available upon request).
               22


regressions. There is then no clear pattern between the returns to education and the
income status of the country (Web Appx. Table A.19).38
       Thus, developed economies may have features that increase the value of experience,
hence workers accumulate more human capital at work in developed countries.

5.2.      Economic Recessions and Returns to Work Experience
Is the wage growth of cohorts entering labor markets in turbulent times permanently
affected? Since we focus on returns based on the 15 years after leaving school, we
measure labor market turbulence in col. (3)-(5) as the share of years during their first
five years of work experience with a per capita GDP growth rate (PPP; constant USD)
below a particular threshold: below -1%, -3% and -5%. For individuals with 0-15 years of
experience in 1990-2016, data on the country’s growth rate is needed for each year from
1975 to 2001. In our data, one third of the country-years experienced a recession.39
       The long-difference effects (Panel A) of recessions are negative on the returns. If
we control for whether the country has reached high-income status by year t, in order
to capture the independent effect of labor market turbulence 10-15 years ago rather
than underdevelopment more generally, the effects remain unchanged (Panel B). The
five-year difference effects are slightly lower, and not significant for -1% (Panel C). As
expected, the effect are stronger and more significant for deeper recessions (col. (5)
vs. col. (3)). More precisely, for each extra year of contraction above 5%, the share
increases by 0.2 (since we consider five years). Consequently, returns might be 0.2*2.24
= 0.45 percentage points lower, given a mean return of 3.2% in that sample.
       Could wages be lower after 15 years because recessions led to the unemployment
of some individuals for some time, thus reducing their years of actual experience?
The mean unemployment rate in the world is 10%. Assuming each deep recession
year doubles the unemployment rate to say 20% for a full year, after one year average
experience among individuals who were typically employed (the 100 - 10 = 90%)
increases by (1 - 0.10/0.90) = 0.89 instead of 1. Thus, actual experience is 11% lower
on average. Returns to experience are then lower by 0.45/3.2 = 14%. The numbers
  38
     These countries are geographically dispersed. 18 countries reached high-income status. 10 of them
are in Europe, 4 in Latin America and the Caribbean (LAC), 2 in Asia, 1 in the Middle-East and 1 in Africa.
19 countries reached middle-income status. 7 are in Africa, 6 in East Asia, 5 in South Asia and 1 in LAC.
  39
     Among observations with a negative rate, -1%, -3% and -5% correspond to the 75th percentile, the
median, and the mean in the sample, respectively.
                                                                                          23


are broadly consistent with the possibility that the effect of labor market turbulence
is mostly due to missed experience for individuals who would have worked otherwise.
       In Web Appx. Table A.18, we use the same econometric framework to study how the
returns (based on 15 years) correlate with the mean unemployment rate (UR) or labor
force participation rate (LFPR) during the cohorts’ first five years of experience. We tend
to find significant, negative correlations with UR or LFPR variables for youth (15-24) but
not for the generally active population (15+) (cols. (1)-(4)). However, UR and LFPR being
correlated and the sample being small, we cannot assess how much of the negative
impact of recessions is coming from a higher UR and/or a lower LFPRh (cols. (5)-(8)).40
       Alternatively, we can study whether recession experience has lower returns than
non-recession experience. For each respondent, we separate their potential work
experience into experience during recession years and experience in non-recession
years, again using the -1%, -3% and -5% thresholds. We re-run the main regression
(eq. (1)) using 5-year bins for each type of experience. We plot the estimated coefficients
in Web Appx. Fig. 7(a), finding no wage growth effects of recession experience.
       Finally, we also find significant negative effects of recessions on the returns to
education (Web Appx. Table A.19). Once we account for the fact that the returns to
education are on average three times higher than the returns to experienced based
on the 0-15 bins, the effects of recessions on experience and education are broadly
similar.41

5.3.      Economic Transitions and Returns to Work Experience
40 countries had a communist regime at one point. We have information on the
time when their transition out of communism began.42 Since the transition involved
structural or institutional changes that may have made pre-transition experience
obsolete, we hypothesize that countries that recently transitioned out of communism
should have lower returns than non-communist economies and this difference should
  40
     Note that data (source: World Development Indicators of the World Bank) is available for fewer
country-years than before since URs and LFPRs were not systematically measured before the 1990s.
  41
     Kahn (2010) find large, negative wage effects of graduating in a bad economy in the United States
whereas we focus on the global effects of recessions on the returns to experience, i.e. to what extent more
experienced workers are paid more than less experienced workers. We do not study the effects on wage
levels.
  42
     We obtain this information from Wikipedia (2019) and the Wikipedia webpage of each country. For
example, we use 1975 for China, 1986 for Vietnam, 1989 for Poland, 1991 for Russia, and 1992 for Albania.
              24


attenuate over time as new cohorts accumulate post-transition experience.
       Column (6) shows that the returns to experience are correlated with the number
of years since the transition out of communism began. Ten years after the transition
is initiated, returns are 10 * 0.08 = 0.8 percentage points higher, a large effect given
the mean of 3.2%. This correlation holds if one controls for whether the country has
achieved high-income status by year t (Panel B), suggesting the effect of transitioning
out of communism is not purely due to economic development.
       Figure 3(a) shows the effects of transitioning out of communism in a more flexible
manner. To obtain these effects, we run panel regressions for all 951 country-years
available in 1990-2016 (country fixed effects and year fixed effects included) and include
dummies corresponding to different groupings of years since a transition. “3” includes
1-3 years, “6” includes 4-6 years, ..., and “27” includes 25-27 years. “30” (28-+) is
the omitted category (standard errors clustered at the country level). The effects are
generated for the returns based on the 0-15 bins but also for the returns based on
the 0-25 bins and the 0-5 bin. As seen, countries that just began their transition start
with much lower returns. Fast convergence is observed for groupings 6 to 12 (hence
between 4 and 12 years after a transition). Returns continue converging after that but
at a slower pace. Returns are still 1 percentage point lower after more than 25 years. As
expected, returns adjust much faster when considering less experienced bins, showing
how the patterns are driven by new cohorts accumulating post-communist experience.
The results on fast convergence echo the results of Dauth et al. (2019) who study the
determinants of wage convergence between East and West German workers. 43
       Alternatively, we can study whether communist experience has lower returns than
post-communist experience. For each respondent in the 40 ex-communist countries,
we separate their potential work experience into experience before the transition (incl.)
and experience strictly after the transition. We re-run the main regression (eq. (1)) using
5-year bins for each type of experience. We plot the estimated coefficients in Figure 3(b).
The results show that an individual with 20 years of post-communist experience earns
about 50% more than an individual with no experience. For communist experience,
the differential is at least -25%, implying that post-communist experience has negative
  43
    We also find significant negative effects of transitioning out of communism on the returns to
education (Appx. Table A.19). Again, once we account for the relatively higher returns to education, the
effects of transitioning out of communism on experience and education are similar.
                                                                                            25


effects post-transition. The profile for post-communist experience is similar to the
profile for the “average” (market-oriented) developing economy.44

5.4.     Are Returns to Experience Decreasing with its Stocks?
While returns to education correlate negatively with the stock of education (e.g., Card
and Lemieux, 2001; Psacharopoulos and Patrinos, 2018b), there is less evidence on
the correlation between the returns to experience and the stock of experience. Table
4 replicates Table 3 with the exception that the variables of interest considered are
measures of stocks of education and experience. For the long-difference specifications
(col. (1) of Panels A-B), we indeed find a negative correlation between the returns to
education and the mean number of years of education in the same year. However,
the five-year specification does not show a statistically significant negative correlation
(Panel C). In addition, developed countries do not have significantly lower returns
to education despite the fact that workers are far more educated on average.                               In
the simulation exercises below, we will thus ignore any possible feedback effect of
increasing the quantity of schooling on the returns to education.                         The returns to
experience are not particularly correlated with the mean number of years of experience
(col. (2)), again suggesting the lack of a feedback effect.45

5.5.     Are Experience and Education Substitutes?

Returns to experience may decrease with the stock of education or the returns to
education, possibly indicating that work experience and education are substitutable.
However, if education and experience are positively related or unrelated, this may
imply that experience is important and the lack of it might not be compensated by
investments in education. We find evidence in line with the latter. Indeed, across
  44
     With most transitions taking place c. 1990 and the latest surveys taking place c. 2010, we do not have
enough countries with a post-communist experience of 25 years or more. We also do not have enough
observations with pre-communist experience in bin 35. We do not report estimates for these bins.
  45
     The distinct long-difference results between education and experience warrants some consideration.
First, experience may not have the same signaling role as education, and thus returns to experience may
not decrease as much when the stock of experience increases. Second, education and experience may
provide different skills for which the price elasticity of demand may also differ. Education systems may
teach students skills that are not as valued by current labor markets. As the quantity of skills taught by
education systems increases, prices may fall. Conversely, firms teach skills that are directly relevant for
their operations and may also be updated to reflect new advancements in the field. Third, education
systems focus on teaching hard basic skills. Firms also “teach” workers soft skills. If soft skills are deeply
valued and scarce, increasing their quantity via more experience may not reduce their “price” as much.
              26


the three panel specifications we do not find any correlation between the returns to
experience and the mean number of years of education (col. (3)). We find, if anything,
a positive correlation between the returns to experience and the returns of education
(col. (4)) or its product with the mean number of years of education (col. (5)).46

6.       Macroeconomic Impact
We now perform counterfactual experiments to further explore the relative importance
of education and experience as factors of economic development. We require an
accounting framework that tracks the accumulation of human capital and aggregates its
returns across individuals. For this exercise, we assume that the measured differences
in the returns mostly reflect differences in human capital accumulated in different
countries, as in Lagakos et al. (2018b,a). However, Caselli and Ciccone (2019) argue
that the estimated skill premia could reflect technological, institutional, organization
and other differences across countries, in which case we over-estimate the contribution
of human capital to economic development. Without further research, it is difficult to
assess whether this disproportionately affects our interpretations of the education or
experience premia, in developed countries or developing countries. In the rest of the
analysis, we abstract from this issue and note that our analysis provides an upper bound
of the contributions of both education and experience to economic development.
       We also do not distinguish between human capital at work that is accumulated
actively (via investment) or passively (via learning-by-doing). In the former case, we
could imagine that less experienced workers accept initially lower wages in order to
benefit from the training offered by their work environment (as, for example, for medical
residents). Eventually, once their level of human capital increases, their wages also
increase. We view human capital accumulated through experience as primarily a by-
product of working, a form of learning-by-doing. One might ask whether this blurs
the distinction between human capital and technology. We do not think so: indeed,
early attempts to develop models of endogenous economic growth build on the idea
that durable resources (such as knowledge, which is embodied in human capital) may
  46
     Education and experience may not be substitutes for a number of reasons. One possibility is that both
education and work experience teach skills that are not easily substitutable, in particular more theoretical
skills vs. more applied skills that can only be learned on the job. Furthermore, education may teach hard
skills vs. soft skills that one can only learn when working in a company. Soft skills may then become
disproportionately important for productivity, and wages, to increase over time.
                                                                                           27


be generated as a by-product of operating a production technology, and that these
resources could in future periods be used as inputs into the production technology: see
Arrow (1962), Uzawa (1965) and Lucas (1988). For example, knowledge about how to
operate a machine is distinct from the machine itself, and is embodied in a worker’s
human capital not in the machine. In the original formulation of Arrow (1962), the
knowledge generated by production is proportional to output, in which case it would
be no surprise to find that countries that produce output more efficiently also build
human capital faster. That said, it is not known whether the production function
parameters that produce output and those that produce knowledge are as tightly linked
as such a formulation might suggest. The positive correlation between development
and returns to experience suggests that such parameters are indeed linked, albeit with
noise, suggesting that further work on this topic could be valuable for understanding
the joint determinants of productivity in output and in learning-by-doing. Finally, our
framework assumes that different levels of human capital are perfect substitutes.47

6.1.       Accounting Framework
                      t
Each year t, bt = b0 gb agents are born. Agents of age a die with probability δ (a). The
age structure of the model is a stationary distribution where the population of each age
group rises over time by the factor gb .
       Schooling begins at age 6. Agents in school with schooling sit ≥ 0 proceed to sit + 1
with probability πs (sit ). Agents who leave school do not return.
       Agents not in school work with probability πe (ait ) where ait is age. If they work, they
generate earnings and accumulate experience. Earnings wit of a working agent i at date
t are log wit = hit where hit is their human capital measured in units of the return it
generates. While in school,
                                          hit = hi,t−1 + rs (sit )

       where rs (sit ) is the return to schooling level sit . When not in school,

  47
    Our model assumes that workers of different schooling levels, or different human capital levels, are
perfect substitutes even if their marginal contribution to output is different. If schooling groups are
imperfect substitutes and the higher schooling groups are more plentiful in richer countries, then the
Mincer returns may understate the importance of schooling to human capital in rich countries relative to
in poorer countries (see Jones (2014, 2019) and Caselli and Ciccone (2019) for a discussion). That said, we
would expect that, to the extent that different levels of schooling are poor substitutes, then the returns to
schooling should not decline with the stock, for which the evidence in the previous Section is mixed. In
particular, we found no such evidence in the case of human capital experience.
            28


                           
                           
                            hi,t−1 + re (sit , pit ) with probability πe (ait )
                           
                   hit =
                           
                                     hi,t−1                 otherwise
                           
                           

   where re (sit , pit ) is the return to experience, which may depend on schooling sit and
on potential experience pit ≡ ait − sit .
   Finally, GDP per person pcGDPt equals total earnings divided by the population.
   We parameterize this framework for two countries: a representative developed
country and a representative developing economy. Then we ask:
   1. How much does output differ between the two economies?
   2. What is the impact on the developing economy of changing the returns to
education and/or experience and/or the schooling probabilities to those in the
developed economy? We perform these experiments with three goals. One will be to
see the potential long run impact on pcGDPt . Another will be to examine the rate of
transition. A third will be to gauge the impact on welfare.


6.2.   Parameters
Once agents complete schooling (or reach 18 years of age, whichever is later), a
share lp of agents participate in the labor force. Those that participate face youth
unemployment with probability uy each year until they are 24. Above 24 they face
unemployment each period with probability u, until they retire at age R = 65. We
assume unemployment and participation are even across schooling and age groups.
   Schooling returns rs (·) have two values rs and r
                                                   ¯s such that:
                                         
                                         
                                          rs if s ≤ 13
                                         
                                rs (s) =                 ,
                                         
                                           ¯s if s > 13
                                          r
                                         

                                                      ¯s > rs ). Experience returns
   reflecting the higher returns to college education (r
re (s, p) have three values:              
                                            re if s ≤ 13 and p ≤ 25
                                          
                                          
                                          
                                          
                                          
                                          
                                          
                            re (s, p) =     ¯e if s > 13 and p ≤ 25 ,
                                            r
                                          
                                          
                                          
                                          
                                          
                                           0        if p > 25.
                                          
                                                                                              29


                                                             ¯e , then the returns to experience
       so that re (s, p) flatten out with experience. If re < r
also embody a college premium.
       We require values of each of these parameters for a representative developing
economy and a representative developed economy. We obtain the population growth
rates gb and the mortality functions δ (·) from United Nations (2019), corresponding to
the “More developed regions” and “Less developed regions”. We use the medium variant
forecasts over the 2020-2050 period, assuming that δ (a) = 1 for a ≥ 99.48
       The return parameters are the average estimates for each group, weighted by
population (see col. (1) of rows 9-10 in Table 2 and col. (1) of rows 14-15 in Table 1).
                                             ¯s = 13.0%, re = 4.4% and r
For the developed economy, we use rs = 6.7%, r                         ¯e = 4.2%. For
                                          ¯s = 13.8%, re = 2.0% and r
the developing economy, we use rs = 6.8%, r                         ¯e = 2.8%.
       We set the schooling transition probabilities πs (·) to match the schooling
distribution in developed and developing economies (I2D2).49 We assume agents may
only accumulate up to 25 years of schooling as some countries only record schooling
up to 25. Finally, we set the employment probabilities πe (·) using participation, youth
unemployment and adult unemployment rate averages reported by World Bank (2018a).
       See Table 5 for a summary of the values of the parameters we used.                                  See
Web Appx. Section A.6 for details of the parameterization of the accounting framework.

6.3.       Main Quantitative Findings
A typical developing economy has 24% of the GDP per capita of a typical developed
economy (World Bank, 2018a). In the simulations, we find a value of 48.5%. The gap
between 48.5 and 100 is 68% of the gap between 24 and 100. Thus, in our accounting
model, differences in education, experience and the returns thereof, plus demographic
and labor market differences, account for two thirds of the income gap.50 Demographic
and labor market differences are small, however, with the exception of labor market
participation. Assume instead that demographic and labor market factors are the same
as in the developed economy, allowing only differences in the stock of schooling and the
returns to human capital. Then, we find that relative pcGDP is 53.5 percent, accounting
  48
     See Web Appx. Fig. 8(a) for the mortality distributions.
  49
     See Web Appx. Fig. 8(b) for the distribution of schooling we use, based on the I2D2 data (1990-2016).
  50
     We compare GDP per capita in levels. An alternative approach would be to compare GDP per capita
in log levels. In that case, the gap is of 51%. In any case, the main point of our inquiry in this section is to
compare the relative contribution of education and experience, so the absolute metric is not important.
           30


for 61 percent of the income gap. Estimates in the literature range from one third or
less (Hall and Jones, 1999; Bils and Klenow, 2000; Caselli, 2005) up to four fifths (Jones,
2014), but recent work by Hendricks and Schoellman (2018) finds that human capital
accounts for 62 percent of the income gap. Lagakos et al. (2018a) find a value of 74%
when estimating returns to education and experience among immigrants in the US.
Thus, our results are broadly consistent with the recent literature.
   What is the impact of education and experience separately ?               Assume that
demographic and labor market factors are the same in the developing economy as in
the developed economy, but that the returns to experience are the same, so that the only
difference is in the quantity and returns to schooling. In this case, GDP is 75.6%, leaving
33.9% of the output gap unaccounted for. Instead suppose that demographic and labor
market factors are again the same, but that the quantity and returns to education are
the same, so that the only difference is in the returns to experience. Now GDP is
75.2%, leaving 34.4% of the output gap unaccounted for. Thus, separately, schooling
and experience each account for about a third of the income gap.
   Figure 4 displays pcGDPt in the developing economy after various reforms that might
impact the schooling distribution and/or the returns to schooling and experience. First,
in the NW panel, we increase the quantity of schooling, the returns to education and
the returns to experience in the developing economy to the levels of the developed
economy. Relative pcGDPt rises steadily for about 50 years before slowing and settling
around 87 percent. Output per worker converges to 92 percent of that in the developed
economy (not shown) – closer, because it takes account of the fact that participation is
lower in the developing economy, but still some distance away.
   The NE panel assumes the schooling distribution converges to that in the developed
economy. This raises income per head to 69.4 percent. Part of this is attributable to
the experience college premium. If we set r
                                          ¯e equal to re in the developing economy
after transition (eliminating the premium), the impact of increasing education is lower,
raising income to 65.2 percent of the developed economy. About a quarter of the jump
from increasing schooling is due to the experience college premium.
   This experiment assumed that the returns to schooling do not respond to increases
in the quantity. We repeat the experiment of increasing the distribution of schooling to
that in the developed economy, while also changing the schooling return parameters to
                                                                           31


match those in the developed economy – which are lower. See the SW panel. Changing
the returns has less impact than changing the quantity of schooling alone, increasing
output to 67.7 percent of the developed economy. Again, a portion of this is due to the
experience premium. If we remove the premium, pcGDP only rises to 63.6 percent.
   The SE panel shows what happens if we increase the returns to experience only.
This increases pcGDP to 66.4 percent. This is similar to the impact of increasing the
quantity of education. The panel also compares the impact from changing the returns
to experience to changing the quantity of education (NE panel): the impact on income
of changing the returns to experience is faster. This is notable as significant reform to
the schooling system is likely costly and slow. On the other hand, if reforms that improve
the returns to experience are institutional, costs may be lower.
Robustness. Web Appx. Section A.7 shows that our quantitative results hold when we
perform a number of robustness experiments with different parameter values. For
example, the results are robust to alterations in demographics, retirement ages, labor
market conditions or sets of returns. Furthermore, the findings are unchanged if
specific types of workers or countries are excluded, more prominence is given to youth
unemployment, experience is allowed to be accumulated before age 18, human capital
spillovers are accounted for, or the sample is divided into three income groups (low-
, middle- and high-income countries). Web Appx. Section A.8 discusses the long-
run welfare impacts of reforms. These impacts depend on the chosen discount rate,
especially as reforms to experience have a faster impact than reforms to education. The
same section then describes how our results can shed light on the long-run impact of
reforms to the retirement age (since they affect the period of time over which the returns
to experience or education pay off ) or reforms that increase youth employment.
   Finally, if human capital accumulation is not entirely a by-product of production
(i.e. some investment in time is required) then the time spent by workers at work
is reduced by the amount of time spent investing in human capital – let us call
this “training.” If training is spread over one’s working life, then our estimates of
the gap are not biased. However, as suggested Ben-Porath (1967), training is likely
concentrated in a worker’s early years. This means that workers’ earnings in their
early years are higher than measured, so that the experience slope is flatter.           If
workers in developed economies spend more time training than workers in developing
              32


economies, then this may lead estimates of the gap to be biased upwards. However,
Web Appx. Fig. 7(b) shows that the number of hours worked does not significantly
increases with experience, whether in developed or developing countries. Thus, any
training that could differentially affect the returns in developed vs. developing countries
would be at work.
       To see whether the differences in the experience premium might be accounted for
by differences in hours spent training, we draw on recent data from OECD.Stat (OECD,
2021). The data contain information on 33 developed economies and 4 developing
economies.51 Let us assume that all this time was compensated, and that none of it
involved training that was performed simultaneously with production. Assuming the
data are reasonably representative of developed and developing economies, we can use
them to explore the impact of any adjustment of hours to account for time spent in
formal or informal training on our measured gaps in the returns. To do this, first we
obtain an estimate of the share of time that workers spend in training. Then, given that
the Ben-Porath framework has human capital accumulation concentrated in the early
years, we explore different scenarios for the years in which this training is concentrated.
       First, in developed economies 59% of workers report spending time formally and/or
informally training, compared to 31% in developing economies. On the other hand,
when they do train, workers train slightly more in developing economies than in
developed economies, 154 hours compared to 120 hours. This implies that a typical
worker obtains 71 hours of training per year in developed economies, compared to 48 in
developing economies. According to I2D2 data, workers in developed economies work
43 hours per week on average compared to 50 in developing economies. Assuming three
weeks of time off, we get that workers in developed economies on average train for 3.4%
of their working time, compared to 2.0% in developing economies.52
       Now, suppose that this training is entirely concentrated among the least-
experienced workers. This is an extreme assumption perhaps, but it does serve to verify
whether our results might be sensitive to the potential issue described above.
       Suppose these workers are in the lowest two experience bins, i.e. 0-4 and 5-9. In
  51
    Colombia, Costa Rica, Mexico and Turkey
  52
    Recall that we estimate returns separately for workers with and without college. College-educated
workers do receive slightly more training than their less-educated counterparts. However, if we focus
on the college-educated we get training ratios of 3.5% and 2.2% for developed and developing countries,
respectively. Since the differences are very small, we abstract from them for clarity.
                                                                            33


this case, the coefficients on the other bins in Figure 1 are over-estimated relative to
bin 0-4. The coefficients on workers in the 5-9 year bin are not over-estimated relative
to bin 0-4, however, since both groups train. There are 8 bins so, abstracting from
mortality, training is concentrated among one quarter of the workers. This implies that
workers in the 0-4 and 5-9 experience bins spend 13% of their time training in developed
economies, compared to 8% in developing economies. What does this imply for the
returns? According to our benchmark estimates, the gap in experience returns between
developed and developing economies was 2.4% for workers without college, or 1.4%
percent for workers with college. If we adjust the returns in Figure 1 based on the above
scenario and recompute the gap, it declines to 2.2% for workers without college and 1.2%
for workers with college. These are declines of 8.3% and 14.3% respectively. Thus, while
the magnitude of our estimated gaps are somewhat sensitive to assumptions regarding
worker training early in their careers, the difference is small. See Web Appx. Table A.11.
     Suppose instead that all training is performed by workers in the lowest experience
bin, i.e. 0-4. In this extreme case, the coefficients on the other bins in Fig. 1 are over-
estimated relative to bin 0-4, including the 5-9 bin. There are 8 bins so, abstracting
from mortality, training is concentrated among one eighth of the workers. This implies
that workers in the 0-4 experience bin spend 25% of their time training in developed
economies, compared to 15% in developing economies. If we recompute the gap, it
declines to 2.2% for workers without college and 1.2% for workers with college. These are
declines of 25% and 35.7% respectively. Thus, even under a scenario where all training
is complete within 4 years, a significant gap in returns remains.


7.     Conclusion
In this paper, we provide additional global evidence that experience matters.
     First, we studied wage-experience profiles and obtained returns to experience, and
education, using data from 1,084 household surveys and census samples across 145
countries. Across many specifications, we found that returns to experience are strongly
correlated with economic development and that workers in developed countries may
accumulate twice more human capital at work than workers in developing countries.
     Second, exploiting the fact that we were able to estimate returns to experience
for more than 100 countries and multiple years of data during the period 1990-
               34


 2016, we employed panel regressions to investigate how changes in the returns to
 experience over time correlated with several important factors. Returns to experience
 are positively correlated with economic development and are lower for individuals that
 experienced a recession in the initial years of their career. Economies that transitioned
 out of communism made past experience obsolete and thus exhibited lower returns to
 experience. Returns to experience had no correlation with initial stocks of experience
 or education. In particular, the latter two results suggest that experience may not have
 decreasing returns and may not be easily substituted by more education.
     Third, our accounting framework showed that the contribution of experience to
 human capital and economic development is as important as the contribution of
 education itself. Indeed, the long run impact is similar whether developing countries
 achieve the returns to experience or attain the same distribution of education as
 developed economies. However, raising returns to experience has a faster impact.
     Finally, returns to experience only greatly increase as economies become developed.
 This suggests that returns to experience might proxy for an economy’s ability to
 converge and catch up to high-income economies.


REFERENCES
Arrow, Kenneth J., “The Economic Implications of Learning by Doing,” The Review of Economic Studies, 1962,
   29 (3), 155–173.
Banerjee, Abhijit and Esther Duflo, “Growth Theory through the Lens of Development Economics,” in
   Philippe Aghion and Steven Durlauf, eds., Handbook of Economic Growth, 1 ed., Vol. 1, Part A, Elsevier,
   2005, chapter 07, pp. 473–552.
Becker, Gary, Human capital., New York: Columbia University Press., 1964, 1975.
Ben-Porath, Yoram, “The Production of Human Capital and the Life Cycle of Earnings,” Journal of Political
   Economy, 1967, 75, 352–352.
Bils, Mark and Peter J. Klenow, “Does Schooling Cause Growth?,” American Economic Review, December
   2000, 90 (5), 1160–1183.
Card, David and Thomas Lemieux, “Can Falling Supply Explain the Rising Return to College for Younger Men?
   A Cohort-Based Analysis,” The Quarterly Journal of Economics, 2001, 116 (2), 705–746.
Caselli, Francesco, “Accounting for Cross-Country Income Differences,” in Philippe Aghion and Steven
   Durlauf, eds., Handbook of Economic Growth, Vol. 1 of Handbook of Economic Growth, Elsevier, 2005,
   chapter 9, pp. 679–741.
     and Antonio Ciccone, “The contribution of schooling in development accounting: Results from a
   nonparametric upper bound,” Journal of Development Economics, 2013, 104, 199 – 211.
    and     , “The Human Capital Stock: A Generalized Approach: Comment,” American Economic Review,
   March 2019, 109 (3), 1155–74.
   , Jacopo Ponticelli, and Federico Rossi, “A New Dataset on Mincerian Returns,” in Francesco Caselli, ed.,
   Technology Differences over Space and Time, Princeton University Press, 2016.
   , , and , A New Data Set on Mincerian Returns Technology Differences across Space and Time, Princeton
   University Press, 2016.
Dauth, Wolfgang, Sang Yoon Lee, Sebastian Findeisen, and Tommaso Porzio, “Labor Reallocation and
   Convergence: Evidence from East Germany,” Technical Report 2019.
de Hoyos, Rafael, John Kennan, and Rebecca Lessem, “Returns to Investments in Human Capital, within and
   across Countries,” 2015. Mimeo.
                                                                                          35


Deming, David J., “The Growing Importance of Decision-Making on the Job,” NBER Working Papers 28733,
   National Bureau of Economic Research, Inc April 2021.
Flabbi, Luca and Roberta V. Gatti, “A primer on human capital,” Policy Research Working Paper Series 8309,
   The World Bank January 2018.
Gindling, T. H., Nadwa Mossaad, and David Newhouse, “How Large are Earnings Penalties for Self-Employed
   and Informal Wage Workers?,” IZA Journal of Labor & Development, Dec 2016, 5 (1), 20.
Gindling, T.H. and David Newhouse, “Self-Employment in the Developing World,” World Development, 2014,
   56, 313 – 331.
Hall, Robert E. and Charles I. Jones, “Why do Some Countries Produce So Much More Output Per Worker than
   Others?,” The Quarterly Journal of Economics, 1999, 114 (1), 83–116.
Hanushek, Eric A. and Dennis D. Kimko, “Schooling, Labor-Force Quality, and the Growth of Nations,”
   American Economic Review, December 2000, 90 (5), 1184–1208.
   , Jens Ruhose, and Ludger Woessmann, “Knowledge Capital and Aggregate Income Differences:
   Development Accounting for US States,” American Economic Journal: Macroeconomics, October 2017, 9
   (4), 184–224.
Hanushek, Eric and Ludger Woessmann, “Do better schools lead to more growth? Cognitive skills, economic
   outcomes, and causation,” Journal of Economic Growth, December 2012, 17 (4), 267–321.
Heckman, James, The Heckman Equation Brochure, Heckman: The Economics of Human Potential, 2019.
    and Richard Robb, “Using Longitudinal Data to Estimate Age, Period and Cohort Effects in Earnings
   Equations,” in W.M. Mason and S.E. Fienberg S.E., eds., Cohort Analysis in Social Research, Springer, New
   York, NY, 1985, pp. 137–150.
Heckman, James J, Lance Lochner, and Christopher Taber, “Explaining Rising Wage Inequality: Explorations
   with a Dynamic General Equilibrium Model of Labor Earnings with Heterogeneous Agents,” Review of
   Economic Dynamics, 1998, 1 (1), 1 – 58.
Hendricks, Lutz and Todd Schoellman, “Human Capital and Development Accounting: New Evidence from
   Wage Gains at Migration,” The Quarterly Journal of Economics, 2018, 133 (2), 665–700.
Islam, Asif, Remi Jedwab, Paul Romer, and Daniel Pereira, “Returns to Experience and the Sectoral and Spatial
   Allocation of Labor,” Mimeo 2019.
Jedwab, Remi, Asif Islam, Paul Romer, and Robert Samaniego, “Human Capital Accumulation at Work:
   Estimates for the World and Implications for Development,” Working Papers 2020-3, The George
   Washington University, Institute for International Economic Policy March 2020.
Jones, Benjamin F., “The Human Capital Stock: A Generalized Approach,” American Economic Review,
   November 2014, 104 (11), 3752–77.
   , “The Human Capital Stock: A Generalized Approach: Reply,” American Economic Review, March 2019,
   109 (3), 1175–95.
Jones, Charles I. and Paul M. Romer, “The New Kaldor Facts: Ideas, Institutions, Population, and Human
   Capital,” American Economic Journal: Macroeconomics, January 2010, 2 (1), 224–45.
Jovanovic, Boyan and Yaw Nyarko, “Learning by Doing and the Choice of Technology,” Econometrica, 1996,
   64 (6), 1299–1310.
Kahn, Lisa B., “The long-term labor market consequences of graduating from college in a bad economy,”
   Labour Economics, 2010, 17 (2), 303–316.
Klenow, Peter and Andres Rodriguez-Clare, “The Neoclassical Revival in Growth Economics: Has It Gone Too
   Far?,” in “NBER Macroeconomics Annual 1997, Volume 12” NBER Chapters, National Bureau of Economic
   Research, Inc, 1997, pp. 73–114.
Lagakos, David, Benjamin Moll, Tommaso Porzio, Nancy Qian, and Todd Schoellman, “Life-Cycle Human
   Capital Accumulation across Countries: Lessons from US Immigrants,” Journal of Human Capital, 2018,
   12 (2), 305–342.
   , , , , and , “Life Cycle Wage Growth across Countries,” Journal of Political Economy, 2018, 126 (2),
   797–849.
Lucas, Robert E., “On the mechanics of economic development,” Journal of Monetary Economics, 1988, 22
   (1), 3 – 42.
Mankiw, N. Gregory, David Romer, and David N. Weil, “A Contribution to the Empirics of Economic Growth,”
   The Quarterly Journal of Economics, 1992, 107 (2), 407–437.
Manuelli, Rodolfo E. and Ananth Seshadri, “Human Capital and the Wealth of Nations,” American Economic
   Review, September 2014, 104 (9), 2736–62.
Mincer, J., Schooling, experience and earnings, New York: Columbia University Press, 1974.
Montenegro, Claudio and Maximilian Hirn, A New Disaggregated Set of Labor Market Indicators Using
   Standardized Household Surveys from Around the World, Background Paper for the 2009 World
   Development Report (“Reshaping Economic Geography”). Washington DC: The World Bank., 2009.
Montenegro, Claudio E. and Harry Anthony Patrinos, “Comparable Estimates of Returns to Schooling around
               36


  the World,” Working Paper 7020, The World Bank September 2014.
OECD, OECD Statistics, Paris: OECD, 2021.
Polachek, Solomon, “Earnings Over the Life Cycle: The Mincer Earnings Function and Its Applications,”
  Foundations and Trends(R) in Microeconomics, 2008, 4 (3), 165–272.
Psacharopoulos, George, “Returns to investment in education: A global update,” World Development,
  September 1994, 22 (9), 1325–1343.
    and Harry Anthony Patrinos, “Returns to investment in education: a further update,” Education
  Economics, 2004, 12 (2), 111–134.
    and    , “Returns to investment in education: a decennial review of the global literature,” Education
  Economics, September 2018, 26 (5), 445–458.
    and    , “Returns to investment in education: a decennial review of the global literature,” Education
  Economics, September 2018, 26 (5), 445–458.
Romer, Paul M., “Human Capital And Growth: Theory and Evidence,” NBER Working Papers 3173, National
  Bureau of Economic Research, Inc November 1989.
Romer, Paul M, “Endogenous Technological Change,” Journal of Political Economy, 1990, 98 (5), 71–102.
Rosen, Sherwin, “Learning and Experience in the Labor Market,” The Journal of Human Resources, 1972, 7 (3),
  326–342.
  , Human Capital, London: Palgrave Macmillan UK,
Rossi, Federico, “Human Capital and Macroeconomic Development: A Review of the Evidence,” The World
  Bank Research Observer, 04 2020, 35 (2), 227–262.
Rubinstein, Yona and Yoram Weiss, “Chapter 1 Post Schooling Wage Growth: Investment, Search and
  Learning,” in E. Hanushek and F. Welch, eds., E. Hanushek and F. Welch, eds., Vol. 1 of Handbook of the
  Economics of Education, Elsevier, 2006, pp. 1 – 67.
Schoellman, Todd, “Education Quality and Development Accounting,” Review of Economic Studies, 2012, 79
  (1), 388–417.
Schultz, Theodore W., “Investment in Human Capital,” The American Economic Review, 1961, 51 (1), 1–17.
UNESCO, International Standard Classification of Education: ISCED 2011 2012.
United Nations, World Population Prospects: The 2019 Revision 2019.
Uzawa, Hirofumi, “Optimum Technical Change in An Aggregative Model of Economic Growth,” International
  Economic Review, 1965, 6 (1), 18–31.
Wikipedia, Communist State, Wikipedia, 2019.
World Bank, World Development Indicators 2018.
  , World Development Report 2019: The Changing Nature of Work, World Bank Publications, 2018.
                                                                                      37



       Figure 1: Wage-Experience Profiles for Developed vs. Developing Countries




Notes: This figure shows the average wage differential for the seven experience bins for the U.S.,
Bangladesh, developed countries (Dev’D) and developing countries (Dev’G) (using population ca. 2017
as weights). The 0 experience bin is the omitted group. Only samples from 1990 to 2016 are used.


            Figure 2: Estimated Returns to Experience and Log Per Capita GDP




Notes: This figure plots the relationship between the estimated returns to experience and log per capita
GDP (PPP; constant 2011 USD; for the mean year in the data for each country). The sample consists of
138 countries (145 countries - 7 countries in the bottom 5% in estimated returns).
                      38

       Figure 3: Effect of the Number of Years Since Transition from Communism
   (a) Overall Returns to Exp. and Communism                                          (b) Communist vs. Post-Communist Experience




Note: Subfig. 3(a) shows the effects of the number of years since the country transitioned out of
communism on the returns to experience (estimated for the 0-25, 0-15 and 0-5 bins). These results are
obtained through panel regression estimations for 951 country-years available for the period 1990-2016
(145 country fixed effects and 27 year fixed effects included). The specification includes dummy variables
corresponding to different groupings of years. The groupings are as follows: “3” includes 1-3 years, “6”
includes 4-6 years, ..., and “27” includes 25-27 years. “30” (28-+) is the omitted category. Robust SEs
clustered at the country level. Subfig. 3(b) shows for 40 ex-communist countries the (pop.-weighted)
wage-experience profiles for communist and post-communist experience separately. The figure also
shows the (pop.-weighted) profiles that we obtained for developed and developing countries.

                                            Figure 4: Counterfactual Experiments, Income Levels
                                             Change all parameters                                                   Change schooling
                                   90                                                                   90
            % of Developed pcGDP




                                                                                 % of Developed pcGDP




                                                                                                                           With college exp. premium
                                   80                                                                   80                 No college exp. premium

                                   70                                                                   70

                                   60                                                                   60

                                   50                                                                   50

                                        0      20         40       60       80                               0        20         40       60       80
                                               Years                                                                        Years
                                    Change schooling and ret. edu.                                               Raise returns to experience
                                   90                                                                   90
            % of Developed pcGDP




                                                                                 % of Developed pcGDP




                                                    With college exp. premium                                              Returns to experience
                                   80               No college exp. premium                             80                 Quantity of education


                                   70                                                                   70

                                   60                                                                   60

                                   50                                                                   50

                                        0      20         40       60       80                               0        20         40       60       80
                              Years                                    Years
Notes: Per capita GDP (pcGDP) in the developing economy is measured relative to the developed
economy. The dashed lines in the NE and SW panels assume that there is no experience premium to
college education. The dot-dashed line in the SE panel (“Quantity of education”)) is the same as the path
representing the reform to schooling in the NE panel (“With college exp. premium”).
                                                                                      39




               Table 1: Returns to Work Experience by Development Status


Dependent Variable:                          Estimated Returns to Potential Work Experience

Return Specification:             (1) No Cohort FE        (2) Decadal Cohort FE     (3) Events Cohort FE

2 Development Groups:            G     D-G      D         G       D-G    D           G        D-G     D

1. Baseline                     1.7*** 1.5*** 3.2***     1.1*** 1.7*** 2.8***      0.8*** 1.8*** 2.6***

2. Log Hourly Wage              1.8*** 1.3*** 3.1***     0.8*** 1.8*** 2.6***      0.4**      2.0*** 2.4***
3. Log Hours Worked             0.0    0.3*** 0.3***     0.3***   0.0   0.3***     0.4***     -0.1   0.3***

4. 0-25 Exp. Bins Only          2.0*** 1.8*** 3.8***     1.3*** 2.1*** 3.4***      1.1*** 2.1*** 3.2***
5. 0-15 Exp. Bins Only          2.4*** 2.4*** 4.8***     1.7*** 2.7*** 4.4***      1.4*** 2.8*** 4.1***
6. 5-35 Exp. Bins Only          1.5*** 1.1*** 2.6***     0.9*** 1.3*** 2.2***      0.6*** 1.4*** 2.0***

7. Lagakos et al 2018 (m)       1.9*** 1.6*** 3.5***     0.9*** 2.1*** 3.1***      0.6*** 2.2*** 2.7***
8. Lagakos et al 2018 (h)       1.9*** 1.5*** 3.4***     0.9*** 2.1*** 3.0***      0.6*** 2.1*** 2.6***

9. Cntry Pop. 2017 as Wgts      1.8*** 1.8*** 3.6***     1.5*** 1.5*** 3.0***      1.1*** 1.7*** 2.8***

10. Yrs of Educ. ≤13            1.7*** 1.7*** 3.4***     1.4*** 1.6*** 3.1***      1.4*** 1.6*** 3.0***
11. Yrs of Educ. >13            2.1*** 1.2*** 3.2***     1.7*** 1.2*** 3.0***      1.5*** 1.2*** 2.7***
12. Yrs of Educ. ≤13 & Wgts     1.7*** 2.0*** 3.7***     1.8*** 1.5*** 3.3***      1.7*** 1.7*** 3.4***
13. Yrs of Educ. >13 & Wgts     2.4*** 1.0*** 3.5***     2.2*** 1.1*** 3.3***      1.9*** 1.2*** 3.1***

14. 0-25, Educ. ≤13 & Wgts      2.0*** 2.4*** 4.4***     2.2*** 1.8*** 3.9***      2.0*** 2.0*** 4.1***
15. 0-25, Educ. >13 & Wgts      2.8*** 1.4*** 4.2***     2.5*** 1.4*** 3.9***      2.2*** 1.5*** 3.7***

16. Row 14, Lagakos et al ’18   2.2*** 2.2*** 4.4***     2.2*** 1.7*** 3.9***      2.3*** 1.6*** 3.9***
17. Row 15, Lagakos et al ’18   2.9*** 1.6*** 4.5***     2.7*** 1.3*** 4.0***      2.4*** 1.3*** 3.7***

Notes: This table shows the constant, i.e. the mean return for developing countries (“G ”), and the
coefficient of the developed country dummy, i.e. the difference between the mean returns of developed
countries and developing countries (“D-G ”). The mean return for developed countries (“D ”) is obtained
by adding the constant and the coefficient of the developed country dummy. See text for details for each
row and each column (N for row 1 = 145; 12; 122). Robust SEs: * p < 0.10, ** p < 0.05, *** p < 0.01.
               40

                        Table 2: Returns to Education by Development Status

Dependent Variable:                                                 Estimated Returns to Education

Return Specification:                   (1) No Cohort FE                  (2) Decadal Cohort FE                  (3) Events Cohort FE
2 Development Groups:                  G        D-G           D            G      D-G            D                G       D-G        D
1. Baseline                          8.7***     0.3          9.0***      8.4***    0.3          8.8***          8.4***     0.3      8.6***
2. Log Hourly Wage                   8.4***     0.4          8.8***      7.9***    0.8          8.6***          7.8***     0.6      8.5***
3. Log Hours Worked                   0.1      0.4**         0.5***       0.2*     0.3          0.5***          0.2**      0.3      0.5***
4. Lagakos et al 2018 (m)            7.0***    1.5***        8.5***      6.5***   1.6***        8.1***          6.5***    1.5***    7.9***
5. Lagakos et al 2018 (h)            7.7***     0.6          8.4***      7.3***    0.7          8.0***          7.2***     0.6      7.9***
6. Cntry Pop. 2017 as Wgts           7.9***     2.4*         10.3***     7.8***    2.1          9.9***          7.7***     2.1      9.8***
7. Yrs of Educ. ≤13                  7.2***    -1.1*         6.2***      7.2***   -0.8          6.4***          7.2***     -0.8     6.40***
8. Yrs of Educ. >13                  15.8*** -3.1*           12.8***     15.8*** -3.9**         11.9***         15.5*** -4.0**      11.6***
9. Yrs of Educ. ≤13 & Wgts           6.8***     -0.1         6.7***      6.8***   -0.2          6.6***          6.8***     -0.2     6.6***
10. Yrs of Educ. >13 & Wgts 13.8*** -0.7                     13.0***     13.4*** -0.4           13.0***         13.1*** -0.4        12.7***
11. Row 9 Lagakos et al 18           5.4***     0.6          6.0***      5.2***    0.8          6.0***          5.2***     0.8      5.9***
12. Row 10 Lagakos et al 18          15.1***-3.2***          11.8***     14.4***-3.1***         11.3***         14.2***-3.3***      11.0***
Notes: The dependent variable is the estimated return to education, i.e. the coefficient of the number of years of education.


   Table 3: Development, Recessions, Transitions and Returns to Work Experience

Dep. Var.:                 Estimated Returns to Potential Work Experience (0-15 Bins) of Country c in Year t
                               (1)                   (2)                 (3)              (4)               (5)             (6)
Var. of Interest:            High-           High- Middle-             Share 0-5 Years Exp. with Growth                  Years Since
                            Income          Income Income               ≤ -1%        ≤ -3%       ≤ -5%                   Transition
Panel A:                                    Long-Difference Panel Regression (N = 244; 226 in (3)-(5))
Var. of Interest c, t        1.11**         1.63***         0.47       -1.32**       -1.68***            -2.24***         0.08***
                             [0.45]         [0.61]         [0.41]       [0.50]        [0.52]              [0.78]           [0.02]
Panel B:                   Long-Diff. Panel Regression, Ctrl for High-Income Dummy (N = 244; 226 in (3)-(5))
Var. of Interest c, t                                                  -1.31***      -1.64***             -2.05**         0.07***
                                                                         [0.48]        [0.51]              [0.79]          [0.02]
High-Income c, t                                                        1.06***       1.00***              0.83**         0.97**
                                                                         [0.31]        [0.36]              [0.41]          [0.41]
Panel C:                                       Five-Year Panel Regression (N = 420; 390 in (3)-(5))
Var. of Interest c, t        1.03**         1.11*           0.08        -0.96         -1.15*              -1.52*          0.07**
                             [0.52]         [0.64]         [0.39]       [0.60]        [0.63]              [0.88]          [0.03]
Country FE, Year FE             Y                     Y                   Y                Y                Y                Y
Note: This table uses panel regressions to explore how returns to experience correlate with various country-level factors. These
include development status, recessions, and duration (no. of years) post communism. The sample consists of 122 countries with
returns available for several years. Panels A-B: First and last years of data only (weighted by the number of years between the
first year and last year). Panel C: Five-year panel for t = {1994 (1990-1996), 1999 (1997-2001), 2004 (2002-2006), 2009 (2007-2011),
2014 (2012-2016)}. Robust SEs (clustered at the country level in Panel C): * p < 0.10, ** p < 0.05, *** p < 0.01.
                                                                                                                41


              Table 4: Stocks of and Returns to Education and Work Experience

Dep. Var.:                       Estimated Returns to Education and Experience of Country c in Year t
                          Col. (1): Education              Col. (2)-(5): Potential Work Experience (0-15 Bins)
                                    (1)                    (2)                (3)                 (4)                  (5)
Var. of Interest:              Mean Years            Mean Years         Mean Years           Returns to           Education
                               Education             Experience         Education            Education          (Years x Returns)
Panel A:                                           Long-Difference Panel Regression (N = 244)
Var. of Interest c, t            -1.23***                 0.13              -0.08               0.11***              0.01***
                                  [0.41]                 [0.08]             [0.10]               [0.03]               [0.00]
Panel B:                     Long-Difference Panel Regression, Control for High-Income Dummy (N = 244)
Var. of Interest c, t            -1.23***                 0.12              -0.08               0.11***              0.01***
                                  [0.41]                 [0.08]             [0.10]               [0.03]               [0.00]
High-Income c, t                   0.36                  1.06**             1.10**              1.06**               1.08**
                                  [0.84]                 [0.41]             [0.45]               [0.45]               [0.45]
Panel C:                                               Five-Year Panel Regression (N = 420)
Var. of Interest c, t             -0.12                   0.07              -0.02               0.12**               0.01**
                                  [0.46]                 [0.09]             [0.10]              [0.05]               [0.01]
Country FE, Year FE                  Y                       Y                 Y                  Y                    Y
Note: This table uses panel regressions to explore how returns to education and experience correlate with various country-
level factors. These include a country’s mean num. of years of education, mean num. of years of experience and/or return to
education. The sample consists of 122 countries with returns available for several years. Panels A-B: First and last years of data
only (weighted by the num. of years between the first year and last year). Panel C: Five-year panel for t = {1994 (1990-1996),
1999 (1997-2001), 2004 (2002-2006), 2009 (2007-2011), 2014 (2012-2016)}. Robust SEs (clust. at the country level in Panel C).



                     Table 5: Parameters Used For Quantitative Experiments

Control Parameters               Parameter Developed              Developing         High (H)       Middle (M)         Low (L)
Population growth rate               gb            0.2%             0.8%              0.2%               0.6%           2.2%
Mortality function                  δ (·)         See text         See text          See text           See text       See text
Labor force participation            lp            84.0%            77.1%             84.0%              75.7%          89.6%
Youth unemployment                   uy            13.7%            13.0%             13.7%              13.8%          6.3%
Adult unemployment                    u            4.9%             3.7%              4.9%               3.8%           2.5%
Retirement age                       R               65               65                65                 65             65
Schooling                             –           See text         See text          See text           See text       See text
Mortality                           δ (·)         See text         See text          See text           See text       See text
Return Parameters                Parameter Developed              Developing         High (H)       Middle (M)         Low (L)
Ret. to exp., before college         re             4.4%             2.0%             4.4%               2.1%              1.8%
Ret. to exp., college +              ¯e
                                     r              4.2%             2.8%             4.2%               2.8%              2.8%
Ret. to educ., before college        rs            6.7%             6.8%              6.7%              8.7%            5.0%
Ret. to educ., college +             ¯s
                                     r             13.0%            13.8%             13.0%             12.8%           14.6%
Notes: This table summarizes the parameters used for the quantitative experiments.