Labor Markets and Economic Growth

Robert Topel
University of Chicago
Thanks to Kevin Murphy, Canice Prendergast, and Pete Klenow for helpful discussions;

the usual disclaimer applies. Support from the Lynde and Harry Bradley Foundation and

the Sarah Scaife Foundation as administered by the George J. Stigler Center for the

Economy and the State is gratefully acknowledged.

1. Introduction

This chapter is motivated by the recent resurgence of interest in the economics of

growth. Among macroeconomists, the shift of research effort is near total, eclipsing the

business-cycle focus that had dominated the field for decades. Behind this is a recognition

of the enormous welfare implications of sustained economic growth, and a renewed desire

to understand the vast differences in living standards among countries, which dates back at

least to Smith. What some have called the “neoclassical revival” in growth economics has

come to dominate macroeconomic research.

Developments in this area should be of particular interest to labor economists

because much of the revival of growth economics builds on the theory of human capital.

Because human capital is, by definition, embodied skills and knowledge, and because

advances in technical knowledge drive economic growth, it follows that human capital

accumulation and economic growth are intimately related. Indeed, many of the issues of

modern growth economics involve questions that are familiar to labor economists. How is

human capital produced and distributed? What are the private and social returns to human

capital investment, and how do people respond to those returns? How do labor markets

operate during the development process? Most of the growth-related work on these

topics is carried on by macroeconomists; traditional labor economists are conspicuous by

their absence, even in empirical work. It shouldn’t be that way.

This paper reviews recent developments in growth economics, with a particular

focus on labor market and human capital issues. My openly confessed motive is to interest

C:\TOPEL\PAPERS\growth.doc 2
labor economists in problems of economic growth, and especially to motivate empirical

research. The paper has three substantive sections, and it unfolds as follows.

Section 2 surveys models of endogenous economic growth based on the

accumulation of human capital, beginning with Uzawa (1965) and Lucas (1988). This

survey of theory is in no way exhaustive, or even a modestly complete review of the field,

but it serves as a template for understanding the major empirical issues in growth

economics as they apply to labor markets. I briefly cover the theory’s predictions about

transitional dynamics for economies that are away from their long run growth paths, the

role of human capital in producing new human capital, and the relation between economic

growth and inequality. This section closes with a summary of empirical implications.

Section 3 turns to the data, reviewing both empirical methodologies and the state

of evidence. Of particular interest is the contribution of education – as a measurable

component of human capital – to economic growth. While richer countries are generally

more educated, it is difficult to isolate the channel through which education affects

aggregate prosperity. Remarkably, existing empirical literature finds virtually no

relationship between changes in the education of a country’s labor force and changes in

output per worker. Instead, the level of education in a country does seem to affect

growth. I re-evaluate this evidence, using panel data on output per worker and

educational attainment for 111 countries over a 30-year period. Unlike previous

literature, I find social returns to investments in education that are as large as, or perhaps

larger than, the estimates of private returns that are generally found in micro-data on

individual wages and earnings. Using within-country changes in education and

productivity, I find that a 1-year increase in average years of schooling for a country’s

C:\TOPEL\PAPERS\growth.doc 3
workforce raises output per worker by between 5 and 15 percent. The level of schooling

also affects growth in this analysis, so it appears that the social returns to education are at

least as large as the private returns.

Section 4 takes up the “operation” of labor markets during development. A

famous hypothesis of Kuznets (1955) posits that wage inequality first rises and then falls

as development progresses. I provide a simple model of this process that incorporates

many of the stylized “facts” about labor markets during periods of rapid economic growth.

In the model, export-driven demand for industrial output raises the demand for skilled

labor. In turn, investment in human capital responds to differences in wages between

skilled and unskilled labor. Wage inequality spurs investment in human capital and more

rapid economic growth, but increased relative abundance of skills serves to reduce

inequality. One of the open questions of this and related models is the impact of

investment in human capital on the relative price of skills. Factor price equalization

indicates that this effect should be negligible, but empirical evidence suggests that a rising

relative supply of skilled labor reduces its relative wage. In spite of trade theories, factor

prices in most countries appear to depend on factor ratios.

Section 5 summarizes and concludes.

2. Labor Markets and Economic Growth

This section reviews basic models of economic growth, as a basis for thinking

about data. I make no attempt to be exhaustive, or even to cover models in all of their

technical detail. The goal is to set out the broad outlines of growth models in a way that

C:\TOPEL\PAPERS\growth.doc 4
will be useful to labor economists. For technical surveys of growth theory see Barro and

Sala-i-Martin (1995) or Aghion and Howitt (1998).

2.1 Background

The last 10 years have witnessed a resurgence of economic growth as a field of

study by macroeconomists. Behind this renewed interest is the enormous impact that

changes in growth can have on economic well being. For example, real per capita income

in the United States in 1950 was $8605, the highest in the world. By 1990, this figure

stood at $18,258 (still the highest), for an average annual growth rate of 1.9 percent. In

contrast, 1950 per capita income in Canada – the third richest country at the time -- was

$6112, which grew to $17,070 by 1990; a growth rate of 2.6 percent per year. If the

United States had achieved the same rate of growth as did Canada, the effect of a 0.7

percent higher growth rate – cumulated over 40 years – would have raised per capita

income in the U.S. in 1990 to $24,033, a gain of $5775 per person. At 5 percent interest

(which is probably high for this calculation), this represents a hypothetical gain in

discounted lifetime wealth of over $100,000 per person.1 Are there changes in institutions

or government policies that could deliver such gains? As a more extreme example, are

there changes in policies or institutions that would transform a growth laggard, like India,

into an Asian “miracle”, like South Korea? Even the remote prospect of gains like these

1
Perhaps it is infeasible for the richest country in the world (the U.S.) to raise its growth rate by 0.7 points
per year, since the richest countries are presumably at the frontier of available technologies and productive
knowledge. So consider a less developed country like the Phillipines. Suppose Phillipine incomes had
grown at the Canadian rate of 2.6 percent instead of its actual rate of 1.6 percent. By 1990, per capita
income in the Phillipines would have been $2507 instead of its actual value of $1519; a 65 percent
difference.

C:\TOPEL\PAPERS\growth.doc 5
has led some economists to call economic growth “the part of macroeconomics that really

matters.”2

It is nearly tautological that the process of economic growth is driven by a

society’s accumulation of knowledge and the ability, or skills, needed to apply it. We

expect to be wealthier in the future because we will know how to do more things than at

present. Seemingly supportive evidence comes from Denison (1985), who estimated that

changes in schooling accounted for about 25 percent of growth in U.S. per-capita income

after 1929, and Shultz (1960), who estimated that investment in schooling grew much

more rapidly than investment in physical capital after 1910.3 Indeed, one view of the

growth process is that differences in per-capita incomes across countries reflect

differences in the ability to apply technologies that are, in a general sense, already

broadly known (Lucas, 1988). Then the wealth of a society is determined by its stock of

human capital, and economic growth is the process of human capital accumulation at the

level of an economy. This means that growth is supported by human capital investment

decisions that are made in labor markets. As it turns out, the role of the labor market in

modern growth theory is not much deeper than that.

2.2 Growth Theory and Human Capital

2
Barro and Sala-i-Martin, 1995. They conclude that if economists can have “even small effects on the long
term growth rate, then we can contribute much more to improvements in standards of living than has been
provided by the entire history of macroeconomic analysis of countercyclical policy and fine tuning.” This
is no doubt true.
3
Estimates of the contribution of labor – including human capital – to economic growth are all over the
map. Dougherty (1991) puts labor’s share of U.S. growth at 41 percent for the 1960-90 period, but the
conformable estimate for Germany is -8.1 percent. Christianson, Cummings, and Jorgensen (1980)
estimate essentially zero contribution from human capital in Germany for the years 1947-73.
 One of the key “facts” about economic growth is that most countries have

experienced sustained growth over long periods of time (Kaldor, 1963). For example, the

annual rate of growth in per-capita income in the U.S. has averaged about 1.75 percent

since the beginning of the 20th century. Similarly, the capital-output ratio is remarkably

stable across countries, both rich and poor (see Figure 1). To accommodate these facts,

modern growth models introduce some additional form of non-physical capital that offsets

diminishing returns to physical capital: In Solow’s (1956) original contribution, an

exogenous rate of labor-augmenting technical change offsets the effects of diminishing

returns to capital. For example, with a constant returns Cobb-Douglas aggregate

production function and zero labor force growth, output is

(1) Y ( t ) = K ( t )α ( A( t ) L )1−α

•
where A denotes the state of labor augmenting technical progress, which grows at rate a

= dlog(A(t))/dt. Equation (1) implies that output per worker is

(2) Y ( t ) / L = A( t )[ K ( t ) / A( t ) L]α

Assuming a constant saving rate, s, under perfect competition the per-worker rates of

•
output, capital, and consumption grow at the steady-state rate a . 4 Since capital and

4
It is straightforward to endogenize the savings rate by modelling intertemporal optimization by
consumers. For present purposes, this only makes the analysis more complicated, without adding new
insights.

C:\TOPEL\PAPERS\growth.doc 7
output grow at a common rate, the capital-output ratio is constant in the steady state.

This correspondence with the data is the motive for specifying technical change as labor-

augmenting.

This model of growth has the unsatisfying feature that technical change is both

exogenous (non-behavioral) and ill-defined, literally an unobserved residual that “explains”

growth after the contributions of other, observable, factors are taken into account. This

fact led T.W. Shultz (1961) and other development economists to reinterpret the residual

in terms of human capital, on the argument that technical progress is hard to distinguish

from advancement of knowledge.5 The idea was formalized by in a modern growth model

by Uzawa (1965) and later Lucas (1988), who interpret A(t) as the average stock of

human capital, or skills, embodied in workers, so H=AL.6 In Lucas’ influential

formulation of the problem, output and the law of motion for the accumulation of human

capital are

(3) Y ( t ) = K ( t )α ( uH ( t ))1−α

(4) H = BH (1 − u) − δH

5
I.M.D. Little (1982) contains a brief intellectual history of the connection between the technical progress
and human capital in growth theory and growth accounting.
6
See also Jones and Manuelli (1990), Rebelo (1991), and Stokey (1991). Others model human capital
accumulation as learning by doing (Romer, 1986; Stokey 1988; and Young, 1991) or as knowledge
accumulation through R&D (Romer, 1990; Grossman and Helpman, 1991; Aghion and Howitt, 1992).

C:\TOPEL\PAPERS\growth.doc 8
where 1-u is the portion of time devoted to production of new human capital, similar to

Ben-Porath’s (1967) model of human capital accumulation for an individual.7 Then (2)

postulates that average productivity depends on the ratio of the stocks of physical and

human capital used in production K/uH. In the Lucas-Uzawa framework, workers

embody productive skills that are accumulated through endogenous, wealth maximizing

investment decisions – schooling, training, and learning-by-doing – that sacrifice present

consumption in order to raise future productivity and income. In the steady-state

equilibrium of the model, the economy’s stocks of physical capital and human capital grow

at the same endogenous rate, which sustains economic growth in the long run. Human

capital investment decisions involve no distortions, and there are no externalities, so

human capital accumulates at the socially efficient rate. Economic growth is efficient, and

there is no role for government interventions in the process.

The conclusion that competitive growth is efficient is an artifact of the

technology (3), in which human capital produces no externalities, as well as the

assumption that privately-financed investments in human capital maximize individual

wealth. Yet education is almost always publicly financed to some (usually large) degree,

and governments often subsidize post-schooling training and apprenticeship programs as

well. (German apprenticeship programs are the latest popular example, alleged to improve

German economic performance compared to the U.S.). This positive role for government

can be rationalized when individual decisions to acquire human capital create external

7
The constant returns assumption in (4) is key. If investment is subject to diminishing returns then
human capital cannot grow indefinitely at a constant rate, so sustained growth is impossible.

C:\TOPEL\PAPERS\growth.doc 9
benefits for others.8 For example, it is plausible that an individual’s human capital is more

productive when other members of society are more skilled. Lucas (1988) analyzes an

extension of (3) in which the output of each firm depends on the human capital of its

workers, say h, as well as the average value of human capital per worker in the economy,

say ha. With this technology, decentralized decisionmaking yields too little investment in

human capital, as individual decisions to invest do not take into account the effect on

others’ productivity. Steady state output is too low relative to the social optimum, and

growth is too slow.9

While models like Lucas’ show that human capital accumulation can sustain

growth, they don’t go far in detailing the role of the labor market and individuals’

investment decisions in this process. Second-generation models have enriched the basic

approach, adding refinements such as finite individual horizons, overlapping generations,

transference of human capital across generations, and various externalities in the

production and utilization of human capital. Yet for labor economists and others

interested in applied work, the message of growth theory does not go far beyond a

statement that “human capital is important.” The theory provides no guidance about why

Singapore has grown faster than, say, India, except perhaps the accounting answer that

people in Singapore have accumulated more skills (and other factors that go with them).

It does, however, provide some foundation for empirical studies of differences in growth

rates across countries, and a number of empirical implications that can be confronted with

8
Liquidity constraints will also do the trick, though I don’t analyze them in any detail. These may be
relevant, for the usual reason that human capital provides no collateral against which to finance
investments.

C:\TOPEL\PAPERS\growth.doc 10
data (see section 2.5). Even so, the only tested implications have to do with transitional

dynamics for economies that may be off of their equilibrium growth paths (see below) and

the issue of whether measures of human capital – like education – raise productivity at all.

On the latter point, the connection of human capital to growth has proven surprisingly

resistant to empirical confirmation, which to some economists (e.g. Klenow and

Rodriguez-Clare, 1997) calls the entire enterprise into question. At the least, there is

substantial debate over the channel through which human capital may affect growth. I

take up this issue in Section 3.

2.3 Transitional Dynamics

The human capital interpretation of (2) yields interesting transitional dynamics

for economies that are away from the steady state ratio of physical to human capital for

one reason or another. For example, consider an economy that “loses” capital in a war,

leaving the stock of human capital intact. The stock of physical capital is too low relative

to the stock of human capital. Then the path back to the steady state involves higher

growth and more rapid investment in physical capital. This is consistent with the actual

performances of Germany and Japan in the decades following World War II. Symmetrical

dynamics are implied for a country that finds itself with “too little” human capital per

worker, say because of past policy mistakes. The returns to human capital investment are

high – due to diminishing returns – and so output grows faster than in the steady state.

These are examples of what has come to be called “conditional convergence”: an economy

9
Romer (1986) studies a similar model, in which aggregate capital enters each firm’s production
function, because of spillover effects in R&D. As in the model with human capital externalities,
competitive growth is inefficient.

C:\TOPEL\PAPERS\growth.doc 11
invests more and grows faster when its current ratio of physical to human capital is

different than its steady state value. Note that this effect is different than the idea that

countries with greater stocks of human capital have an advantage in growing because

human capital is an aid to innovation. Conditional convergence follows solely from

neoclassical properties of production, together with optimal investment going forward.

One of the puzzles of economic growth is that some countries suddenly rise

from underdevelopment, accumulating human (and physical) capital along a path of rapid

output growth, while other countries seem to trapped in a low growth state. Becker,

Murphy, and Tamura (BMT, 1990) and Azariadis and Drazen (1990) model this as a

problem of multiple growth equilibria, where the needed non-convexity comes from the

technology for producing human capital. Both of these papers argue that human capital

begets the production of more human capital: education and other sectors that produce

human capital are intensive users of skilled (e.g. educated) labor. Within a country, this

means that rates of return on investment in human capital may initially rise instead of fall

as the stock of human capital increases, because the large stock makes it cheaper to

produce more. Comparing countries, this means that differences in initial conditions can

lead to different long run growth paths. The result is multiple steady states, one with low

output, little human capital investment, and (in BMT) high fertility; the other with higher

returns, greater investment, skills, and growth, and lower fertility. BMT argue that the

circumstances that push an economy from one steady state to another may be largely a

matter of “history and luck,” and “accidents and good fortune,” while Azariadis and

Drazen see a role for government policy in getting the ball rolling. In their model of

C:\TOPEL\PAPERS\growth.doc 12
overlapping generations with a threshold externality in the production of human capital, a

one-time intervention will do the trick.

Luck and accidents aside, this type of model can help us to understand a key

feature of human capital investment in the development process. In the Asian miracles like

Taiwan (Lu, 1993) and Korea (Kim and Topel, 1995) and in some Latin American

economies (Robbins, 1996), successive cohorts of the young acquire human capital in

larger and larger numbers. Yet empirical evidence discussed below suggests that

increased stocks of educated labor cause the returns to human capital to fall. With

constant costs of educating the young, declining returns should reduce investment. But a

technology in which the existing stock of human capital raises the productivity of current

investment can generate declining costs of adding to the stock. Then investment can rise

in spite of declining returns to human capital.

2.4 Human Capital and Aggregate Inequality

Beginning with Kuznets (1955), a long tradition in development economics is

concerned with the effects of growth on wage and income inequality. Human capital

models of endogenous growth typically abstract from this issue by treating H as a

homogeneous aggregate, so that the distribution of H among workers is has no bearing on

the growth rate of output. Yet human capital investment affects inequality in at least two

ways. First, it affects the distribution of the stock of human capital, which could either

increase or decrease inequality depending on where in the distribution of skills the new

investments occur. For example, at the initial stages of economic development human

capital investment in the form of education may be concentrated among a privileged elite.

C:\TOPEL\PAPERS\growth.doc 13
This would tend to raise inequality. Later investment may be concentrated on the least

skilled, especially with diminishing returns to investment at the individual level, so

inequality may eventually fall. This pattern is consistent with the “Kuznets Curve”

hypothesis that inequality first rises and then falls as development proceeds.

Glomm and Ravikumar (1992) and Benabou (1996) analyze different structures for

school finance, and how differential access to human capital among individuals can affect

inequality and growth. In Glomm and Ravikumar (1992) a spillover externality in public

school finance makes individual human capital accumulation more productive when the

average human capital of the population is higher. This effective “subsidy” of the less

skilled causes inequality of human capital to die out over time. In contrast, privately

financed schooling tends to make inequality persist. Benebou (1996) analyzes the effects

on growth of schooling when students of heterogeneous abilities can either be segregated

or mixed together. In the short run, segregation may increase growth because talented

people are complements in producing new human capital. In the long run, however,

segregation leaves intact the overall heterogeneity of skills in the economy, which is a drag

on productivity growth. This perpetuates inequality in the long run, and can reduce

growth. This has implications for school finance. If schools are financed locally, in

communities that are sorted on talent or resources, then expenditures on education will

tend to perpetuate inequality and, perhaps, reduce long run growth. Greater funding

equity – say through centralized taxation to finance schools and reduced segregation on

talent – leads to lower long run inequality and higher growth. In this model, centralized

C:\TOPEL\PAPERS\growth.doc 14
financing and a national curriculum – along the lines of some European countries – may

provide a long run advantage relative to a decentralized system.10

The second effect of human capital accumulation on inequality occurs because

human capital investment affects factor proportions, which should impact relative wages.

As human capital accumulates, the aggregate share of skilled labor rises so that the relative

price of skills may fall. As Leamer (1995) argues, this force is mitigated by Stolper-

Samuelson effects of unimpeded trade. If output prices are fixed on international markets,

and if sectoral production functions exhibit constant returns, then factor price equalization

implies that relative wages of different skill groups are independent of their factor shares

within a particular country. (The technical conditions for this are discussed in greater

detail in Section 4). With constant returns, an increase in the labor force share of skilled

(educated) workers can be accommodated by shifting labor from low-skill to high-skill

sectors, leaving factor proportions in each sector (and thus relative wages) unchanged.

Empirical evidence from a number of countries appears to reject this prediction, however.

Increases in the aggregate share of educated labor do not simply increase the size of skill-

intensive sectors. Within-sector shares of educated labor rise as well (Murphy and Welch,

1993; Topel, 1994; Kim and Topel, 1995, Robbins, 1996), which indicates that the relative

“price” of skilled labor will fall as it becomes more abundant. Empirical evidence on this

issue is taken up in Section 4, below.

2.5 Empirical Implications of Neoclassical Growth Theory

10
For example, Swedish schools are centrally financed, and funding equity is strictly adhered to.
Curriculum is uniform across schools.

C:\TOPEL\PAPERS\growth.doc 15
 Models that base sustained growth on human capital accumulation have a number

of important, and testable, empirical implications. Most obvious is that accumulation of

human capital increases economic growth. As discussed below (Section 3), this central

prediction has proven surprisingly resistant to empirical confirmation, at least in the form

that the theory implies. Secondary and more subtle predictions are (i) rising returns to

skill should spur investment and, therefore, growth; (ii) the private and social returns to

human capital may differ when spillover effects are important; and (iii) economies that are

initially below their steady state values of physical or human capital will experience faster

growth. Complementarity between the existing stock of human capital and new

investment implies that: (iv) investment in human capital may rise, even while the returns

are declining; (v) countries with little initial human capital may be “trapped” in a low

growth, low income state; and (vi) the distribution of human capital can affect investment,

and hence growth. Some of these predictions are taken up in Sections 3 and 4.

2.6 Alternative Models of Human Capital and Growth

In the models outlined above, human capital drives growth because it is an input

to the production of goods and services, as in (3). Then growth in human capital per

worker is equivalent to growth in output per worker; human capital simply earns its

private marginal product. Nelson and Phelps (1966) offer an alternative view. In their

analysis, growth is driven by the stock of human capital because skilled workers are more

likely to innovate new technologies and – for countries that are not at the technological

frontier – more able to adopt existing technologies. In this analysis, a greater level of

human capital at time t raises subsequent growth by producing technical change. A

C:\TOPEL\PAPERS\growth.doc 16
number of microeconomic studies of the role of education in production, beginning with

Welch (1966), find empirical evidence for the idea that educated workers are more likely

to adopt new productive technologies. For example, in a study of Indian farmers, Foster

and Rosenzweig (1996) find that more educated farmers are the first to adopt new seed

technologies.

At the aggregate level the most obvious empirical implication of this view is

that changes in the rate of output can depend on the level of human capital, rather than

simply on the change in human capital as implied by standard growth models. This

prediction is consistent with empirical results of Barro and Sala-i-Martin (1995) and

Benhabib and Spiegel (1994), who estimate models of economic growth on a cross-

section of countries. They find little evidence that growth of human capital is associated

with growth of output, but a higher level of education per worker (measured by average

years of schooling in the population) is associated with a higher rate of economic growth.

In Barro and Sala-i-Martin’s analysis, average years of secondary education have a

stronger effect than years of primary education, which may also reflect greater ability to

innovate and adopt technologies among more skilled workers. Benhabib and Spiegel find

that the level of education has a stronger effect on growth for relatively low income

countries, which may indicate a role for education in “catching up” to technological

leaders. The next section provides a more detailed discussion of these and other empirical

results.

2.7 Summary: Human Capital, Education, and Growth

C:\TOPEL\PAPERS\growth.doc 17
 The recent revival of growth theory is built on the idea that human capital is

central to growth. Yet there is little consensus on what is the channel of causality leading

from human capital investment to economic growth. Following Lucas (1988),

neoclassical models treat human capital as a produced input to a standard technology, so

that growth of human capital and growth of output are nearly synonymous. An alternative

theory, with support in some recent empirical work, is that the level of human capital

affects growth through greater innovation and adoption of technologies. As pointed out

by Aghion and Howitt (1998), the theories have starkly different implications for the

effects of human capital investment on long run growth. Narrowly interpreted,

neoclassical models imply that current investment leads to a one-time surge in output as

new human capital is applied in production. In contrast, models like that of Nelson and

Phelps (1966) imply that current investment – by raising the level of human capital – has a

permanent effect on technical change and hence growth.

It is plausible that both theories of the role of human capital are true. Growth of

human capital may increase output and set the stage for subsequent growth. Yet even

then, the differences between the theories is more semantic than real. Neoclassical

theorists define human capital broadly, so that accumulation of human capital

encompasses the accumulation of knowledge and the ability to apply it in productive ways.

When we think of new ways to do things, human capital has increased.11

If this is so, then why do some empirical studies – like Barro and Sala-i-Martin

(1995) and Benhabib and Spiegel (1994) – find that the level of human capital, as

measured by average years of schooling, raises growth? An answer is that human capital

C:\TOPEL\PAPERS\growth.doc 18
is an input to its own production, a fact that is central to many growth models, and that

schooling is only one form of human capital. Other forms of human capital accumulation

– like on the job training, acquisition of knowledge outside of formal schooling, and

learning-by-doing – are unmeasured. Empirically, this means that the level of schooling

will be correlated with growth because countries with more education invest more in other

forms of human capital. A related point is that countries with more schooling may have

lower costs of investing in other forms of human capital, so schooling is simply a proxy for

unobserved heterogeneity in the costs of investment.12

3. Empirical Evidence

3.1 Background

As I noted above, the role of labor markets in modern endogenous growth

theory does not go much beyond the idea that human capital should be important to

sustained economic growth. The empirical questions are (i) How important is it?; and (ii)

What are the channels through which human capital affects growth? Does growth of

broadly-defined human capital “account” for what we would otherwise call productivity

growth, as suggested by Lucas (1988) and others? If so, would government policies that

encourage human capital investment improve welfare, especially among less developed

countries that might be able to “catch up” with more advanced countries, which are closer

to the technological frontier? Are some policies and institutions, such as income

11
In this sense, I think that Aghion and Howitt (1998) greatly exaggerate the difference between
neoclassical and “Shumpertarian” models of human capital and growth.
12
In earnings data for individuals, age-earnings data for more educated workers are steeper for more
educated workers. The standard explanation is that education reduces the costs of subsequent, on-the-job

C:\TOPEL\PAPERS\growth.doc 19
redistribution and centralized wage setting, a hindrance or boon to human capital

investment and growth? These are key empirical issues for which we have few good

answers.

There are two main strands of empirical research on economic growth. Both

attempt to measure the effect of input differences, or accumulation, on productivity and

per-capita incomes. Growth accounting divides output growth among changes in

measurable input quantities – physical and human capital -- and a residual called “total

factor productivity” (TFP). The art in this approach lies in measuring inputs, which is

especially difficult when the input in question is an abstract stock like “human capital.”

The other main body of research is more regression-oriented, estimating cross-sectional

and panel models of the determinants of countries’ incomes. Our main interest in this

literature will stem from what can be learned about the empirical relationship between

education and economic growth.

This section also provides some new evidence on the effects of schooling on

economic growth. I find that returns to schooling estimated from aggregate data on

country growth rates are generally as large, or larger than, the returns estimated by labor

economists from micro-data on individuals’ wages and earnings.

3.2 Growth Accounting

Following Solow (1957), suppose that aggregate output is produced using

investment in human capital. Heterogeneity of talent has the same implication: those with more education
have lower costs of investing, so we expect them to invest more in other forms of human capital.

C:\TOPEL\PAPERS\growth.doc 20
physical capital (K) and human capital (H) via F(K,AH), where A denotes the state of

labor augmenting technical progress. Assuming constant returns to scale and competitive

factor markets, the rate of change of output for country i at date t is given by

• • • •
(5) y it = α it k it + (1 − α it ) hit + pit

• • • •
where, y , k , h and p refer to the proportional rates of change of output, physical

capital, human capital, and TFP, respectively, and α is capital’s share of national income.

With the exception of p, all quantities in (5) are measurable, at least to some degree,

which leaves TFP as the part of output growth that remains unexplained after taking

account of the growth rates of physical and human capital. Hence the estimate of TFP is

commonly called the Solow residual.

Original applications of (5), such as Solow (1957) and Denison (1962, 1967)

treated raw labor as the human capital input, and did not account for changes in the

quality of capital, so that a large portion of growth was attributed to TFP. Later work by

Jorgensen and Griliches (1967) and Jorgensen, Gollop, and Fraumeni (1987) showed that

a substantial portion of the Solow residual could be accounted for by changes in input

quality. For our purposes, the quality of the human capital input has increased in most

countries because of improvements in health and in the quantity and quality of schooling

among working age populations. This means that subcategories of the labor force (years

of schooling and experience, gender, and so on) should be weighted by their marginal

C:\TOPEL\PAPERS\growth.doc 21
products (wages) in forming a human capital aggregate.13 Then accumulation of human

capital means that H grows faster than the labor force, which accounts for some of

productivity growth.

The most recent applications of this method are in three influential papers by

Young (1992, 1994, 1995). He studies the growth experience of the four “Asian tigers:”

South Korea, Hong Kong, Taiwan, and Singapore. As shown in Table 1, between 1966

and 1990 output per worker in these economies grew at average annual rates of between 4

and 5 percent, far above the 1.4 percent rate achieved by the U.S. over this period. Before

Young’s work, many observers attributed this remarkable growth record to technical

improvements, driven perhaps by government “industrial policies” that encouraged the

growth of certain industries and technologies. By carefully measuring the quantities of

physical and human capital in these countries, Young concludes that their rapid growth is

due to input accumulation (and utilization, in the case of labor), while TFP growth was

not unusually high by world standards. In fact, for Singapore, Young finds that TFP

growth contributed essentially nothing to income growth over this period. All of the

growth in output can be accounted for by changes in the quantity and quality of capital,

sharply increased labor force participation, and increased years of schooling of workers.

The implication is that the remarkable growth record of these economies is unlikely to be

sustainable, since input utilization cannot increase indefinitely.14

13
The production function is y = F(K,Σ ni Hi ), where ni is the number of workers in group i and Hi is
average human capital of those in the group. Equating marginal products to wages for each group, we get
Hi /Hj = wi /wj, so Hi = (wi /wj)Hj. For example, an increase in the number of high school graduates relative
to elementary school graduates, holding population fixed, will raise the measured stock of human capital
in proportion to the college/high-school wage ratio.
14
In Table 1, the differences between the growth rates of GDP and GDP per worker are largest for the
four “Asian Tigers.” Much of GDP growth in these economies was accomplished by increased labor force
particiapation.

C:\TOPEL\PAPERS\growth.doc 22
 The growth accounting literature suggests an important role for the labor market

in economic growth. Consider Young’s results for Korea. Beginning in 1966, growth in

labor input (including quality) “contributed” a breathtaking 4.4 percent per year, for a 25

year period, to growth in aggregate output. As shown in Table 1, almost all of this effect

was due to increased labor utilization rather than to any increase in measured human

capital per worker. Over this period, the Korean non-agricultural labor force grew at an

annual rate of 5.4 percent per year (!), while population grew at only 1.6 percent. The

difference reflects increased labor force participation and a wholesale migration out of

agriculture. A growth accounting measure of human capital per Korean worker grew at

an annual average rate of .007/(1-.32) = 1 percent per year, faster than for any country in

the table save Singapore. This increase was driven by a massive investment in public

education that reduced the share of workers with a primary education from over 60

percent in 1970 to less than 30 percent by 1990 (Kim and Topel, 1995). Yet rising human

capital per worker was swamped by the concomitant rise in the capital/labor ratio, which

“accounted” for 61 percent of the increase in Korean labor productivity. By this method,

human capital growth accounted for only 14 percent of the growth in output per worker.

Indeed, for the countries in Table 1, human capital never accounts for the major portion of

economic growth. Does this mean that human capital isn’t so important after all?

3.3 Limitations of Growth Accounting

The obvious answer is “no.” Growth accounting is mainly descriptive, treating

human and physical capital in virtually identical ways. It has nothing to say about how or

C:\TOPEL\PAPERS\growth.doc 23
why factor accumulation took place, or whether human capital accumulation is essential

for growth. Three limitations of this approach seem particularly relevant.

The first point has to do with what it means to measure a factor’s contribution to

growth. Consider again the estimate that human capital contributed 0.7 percentage points

per year to Korea’s productivity growth. This figure is simply an average of marginal

contributions of labor, along the actual path of physical and human capital

accumulation. It doesn’t say that output per worker would have grown at 0.7 percent had

capital remained fixed. Even so, it may vastly understate the importance of human capital

to the growth process. Suppose for the sake of argument that a Lucas-Uzawa style model

is an appropriate description. Their theory is that human capital is the whole story. In the

steady state, the proportional rate of growth of physical capital is equal to the proportional

•
rate of growth of human capital, given by a = B(1 − u ∗ ) − δ (see equation (4)). The ratio

of physical to human capital is constant in the steady state, so that output per capita also

•
grows at rate a . Growth is driven by human capital accumulation, but a growth

•
accounting exercise attributes the product of capital’s share and a to capital.

More generally, the quantity and type of physical capital investments that actually

occurred may depend on the quality of human capital that is available to work with it.

Without large investments in human capital, particularly in education, Korea may not have

adopted the existing technologies that fueled its growth. In this sense, investments in

human capital, such as education, may be essential to the growth process. Then growth

accounting is uninformative about the importance of human capital accumulation.

C:\TOPEL\PAPERS\growth.doc 24
 A second limitation is that changes in human capital are poorly measured. A virtue

of studying developing nations is that changes in the amount of human capital employed in

production may be well measured by changes in observable quantities like the number of

workers and their years of schooling and experience. Think of the typical Korean worker

who, let’s say, now enters the labor market with a secondary instead of a primary

education. The things he learned in those additional years are “common knowledge,” like

arithmetic and grammar, well inside the frontier of ideas. In this case, the change in the

quantity of human capital per worker may be well approximated by increased years of

schooling. Now think of workers in a developed economy, like the U.S, where average

years of schooling has changed by less. The additional knowledge that workers bring to

the labor market largely consists of new knowledge, like how to use a computer. Their

human capital is greater, but no observable measure picks this up. Conceptually, the

contribution of human capital is the same in both economies, but in Korea the increase in

human capital is more accurately measured by observables. In the U.S., where observable

quantities did not change by much, more of the contribution of human capital is attributed

to “total factor productivity.”

More broadly, any measure of human capital for growth accounting is based on

changes in observable quantities – such as education or experience – and the relative

prices that those observables command. If school quality improves at all levels, or if post-

schooling investment in human capital becomes more widespread or productive, growth

accounting measures are unlikely to capture the change.

The third limitation of growth accounting is that it is silent about how the labor

market actually operates during economic growth. In rapidly growing Asian (and other)

C:\TOPEL\PAPERS\growth.doc 25
economies, we know that industrial expansion was fueled by migration of workers from

agriculture. We also know that public investments in education sharply raised average

schooling levels. What market forces supported this? A market-driven scenario is that the

relative price of skilled labor, needed for industrial production, was initially quite high

because skilled labor was scarce. Expansion of public education increased opportunities to

invest in skills, and wage inequality provided the incentive for young workers to do so.

As successive cohorts of young workers acquired more schooling, and migrated to

industrial employment, the relative price of skilled labor fell, which further fueled growth.

This description of events gives a prominent role to a smoothly operating labor market,

with market-determined wages, and fairly elastic responses of investment in education, in

supporting growth. Indeed, in this scenario, educational opportunities start the ball

rolling. The modern macroeconomics of growth provides little evidence on whether this

or any other model is true.

3.4 Measuring the Social Returns to Human Capital

In growth accounting, the marginal return to a unit of human capital is assumed to

be at least equal to the private return. Human capital is measured by a wage-weighted sum

of labor inputs, which is simply multiplied by 1-(capital’s share) to get an estimate of the

contribution of human capital to national income. A growing econometric literature takes

a less constrained approach, seeking direct evidence on whether various measures of

human capital actually raise aggregate output.

Suppose that aggregate output, Y = F(K,AH), of country i is Cobb-Douglas with

constant returns. Then output per worker satisfies

C:\TOPEL\PAPERS\growth.doc 26
 ( 6) ln(Yit / Lit ) = α i ln( K it / Lit ) + (1 − α i ) ln(hit ) + (1 − α i ) ln( Ait )

where hit is average human capital per worker. If we assume α i = α , then an

unconstrained form of (6) is

(7 ) ln(Yit / Lit ) = β i + β k ln( K it / Lit ) + β h ln(hit ) + ε it

The parameters β k and β h represent the contributions of physical capital and human

capital to aggregate productivity, and β i allows for differences in total factor productivity

across countries. Assume that adequate measures of output per worker and physical and

human capital are available for a large sample of countries at a point in time. Then with

appropriate assumptions about the distribution of β i across countries (it should be

orthogonal to physical and human capital intensities) or with appropriate instruments

(good luck), β k and β h can, in principle, be estimated from cross-sectional data. This is

the basic approach taken in empirical studies by Mankiw, Romer, and Weil (1992) and

Klenow and Rodriguez-Clare (1997). Alternatively, with panel data equation (7) can be

differenced over time to obtain an empirical model of economic growth:

C:\TOPEL\PAPERS\growth.doc 27
(8) ∆ ln(Yit / Lit ) = β k ∆ ln( K it / Lit ) + β h ∆ ln(hit ) + ∆ ε it

Variants of equation (8) underlie empirical growth studies by Benhabib and Spiegel

(1994), Pritchett (1997), and (to a lesser extent) Barro and Sala-i-Martin (1995). Notice

that unlike the growth accounting approach, models (7) and (8) treat β k and β h as free

parameters, which adds a layer of testability to the theory.

3.5 Empirical Results

Mankiw, Romer, and Weil (MRW, 1992) reach a similar conclusion to that of

Young – input accumulation explains prosperity -- but on a much broader sample of 98

countries. They study the cross-country distribution of output per capita in 1985, using a

Solow-type model that is extended to account for differences in the quality of human

capital across countries. To measure stocks, they capitalize investment flows using the

average 1960-85 flow of investment in physical capital (for K) and the 1960-85 secondary

school enrollment rate (for H). They find that input differences – especially human capital

differences – “account” for nearly 80% of the cross-country variance in income.15 Only

about 1/5 of the variance is due to unobserved productivity differences. Thus Mankiw

(1995) concludes that “most international differences in living standards can be explained

by differences in accumulation of both physical and human capital.”

This conclusion clearly rests on the dubious assumption that physical and human

capital intensities are orthogonal to productivity differences across countries. If more

15
Formally, using data on 98 countries, they estimate a model of the form ln(Y/L) = b1ln(K/Y) + b2ln
(H/Y) + e. The R2 from this regression is .78, with elasticities of b1 = .30 for capital and b2 = .28 for labor.

C:\TOPEL\PAPERS\growth.doc 28
productive (higher A) countries are also more intense users of physical and human capital,

the causal contribution of observed inputs will be overstated by MRW’s regression

approach. And if it isn’t “A” that drives things, this research also leaves open the question

of why countries with similar technological and other opportunities end up with

dramatically varying stocks of physical and human capital. Do poor countries experience

decades of sub-optimal investment because of policy mistakes, like excessive taxes and

inadequate investments in public schooling, and inefficient institutions?16 This possibility

might give economists real value as policy advisors (“Stop doing that. Invest.”). Or do

observed stocks of physical and human capital reflect optimal responses to other, country-

specific constraints? The empirical literature on economic growth leaves this basic

question unanswered.

MRW’s conclusion that inputs account for income differences has also been

criticized on more basic, empirical, grounds. Klenow and Rodriguez-Clare (1997a,

1997b) argue that MRW misstate the contribution of human capital by calculating human

capital stocks from international differences in secondary school enrollment rates. By

adding primary school enrollments in the construction of H, Klenow and Rodriguez-Clare

find a substantially smaller contribution of human capital to international income

differences, and correspondingly larger contributions of unmeasured technology.17

According to their estimates, human capital stocks vary less across countries when

primary enrollments are included in the flow of investment. As importantly, in cross-

16
See Chari, Kehoe, and McGratten (1996), who argue that such inefficiencies can explain international
income differences.
17
Klenow and Rodriguez-Clare also account for differences in shapes of age-earnings profiles, based on
standard Mincerian regression techniques. Implicitly, then, their analysis also accounts for international
differences in post-schooling investments in human capital and learning-by-doing.

C:\TOPEL\PAPERS\growth.doc 30
sectional data differences in output per worker are more strongly correlated with

differences in secondary enrollments than with differences in primary enrollments. They

interpret their findings as favoring the notion of technological “catch-up” rather than

simple input accumulation.

An alternative interpretation is that different education categories are imperfect

substitutes in aggregate production, and that between-country differences in levels of

secondary schooling have larger impacts on income than do differences in primary

schooling. Regression estimates of the effects of schooling on economic growth, reported

below, support this interpretation.18 The usual method of aggregating skill groups simply

weights the number of worker hours in each group by its relative wage, resulting in an

estimate of “H.” This method assumes that human capital of high school graduates (for

example), measured in efficiency units, is a perfect substitute for the human capital of

college graduates. A long list of country studies of relative wages rejects this assumption

(e.g. Freeman, 1980, 1986; Katz and Murphy, 1990; Kim and Topel, 1992; Edin and

Holmlund, 1992; Freeman and Needles, 1993).

Barro and Sala-i-Martin (BSM, 1995) summarize a number of regression-based

studies of international differences in economic growth, based mainly on the Summers-

Heston (1995) international data set. For our purposes, they study two main issues. First,

do international and other data contain evidence that would favor convergence of

incomes? That is, do low-income countries (or regions) grow faster than high-income

18
The growth regressions of Barro and Sala-i-Martin, discussed below, are consistent with this. They find
that a higher initial of stock of secondary school graduates raises a country’s growth rate, but that the
stock of primary graduates has no effect. Klenow and Rodriguez-Clare also use United Nations data on
the share of each country’s population at various ages to construct an experience measure. Their measure

C:\TOPEL\PAPERS\growth.doc 30
ones? Using data on European regions, U.S. states, and Japanese prefectures, they find

evidence that strongly favors the convergence hypothesis. For example, the poorest U.S.

states in 1880 had the highest rates of per-capita income growth over the 1980-1990

period. The underlying assumption of these regressions is that different areas have similar

institutions and access to the same basic technology, so that income differences reflect

deviations from steady-state values. Again, however, we are left with little understanding

of why different areas started with different incomes and, correspondingly, different levels

of human and physical capital. Even so, the findings are important and are consistent with

related work on changing quality of inputs. For example, Smith and Welch (1986) and

Card and Krueger (1992), among others, have documented the long-term improvement in

educational quality (and years of schooling) in the American South during the 20th century.

Surely this is a contributor to the long-term relative economic growth of the region. Yet,

given the costs of investing in physical and human capital, the growth process itself is

disturbingly long. After a century of convergence, with largely identical legal and

economic institutions, per-capita income in Mississippi remains less than half of that in

Connecticut or New Jersey.

BSM also seek to estimate the contribution of human capital, measured by

schooling and health, to economic growth. Their concern with issues of convergence

leads them to bypass a formal specification like (8). Instead they found their empirical

analysis on variants of (BSM, 1995, p384)

(9) Gi(0,t) = β i - (Yi0 - Yi*)β

β 1 + uit

of H is then based on returns to schooling and experience derived from a standard cross sectional earnings

C:\TOPEL\PAPERS\growth.doc 31
where G denotes country i’s average annual rate of growth between time 0 and t, Yi0 is the

log of initial per-capita income, Yi* is the log of steady state per-capita income, and β i is a

steady-state growth rate for country i. The parameter β 1 indexes the average “speed of

convergence” over [0,t]. In light of (9), BSM estimate models of the form

(10) Gi(0,t) = β 0 - Yi0β 1 + Hi0β 2 + Hi0Yi0β 3 + Xiβ 4 + ui

where H is a vector of human capital measures, and X is a vector of controls for political

stability, terms of trade, and the like. The hypothesis of convergence (which doesn’t

concern us much here) is that β 1 > 0: rich countries grow slower than poor ones, other

things equal.

The interpretation of human capital measures in (10) is ambiguous. One

interpretation is that human capital is a proxy for steady-state income: conditional on

current income per-capita, a country with a more skilled workforce “should” be richer. So

we expect β 2 > 0 from the convergence hypothesis. In this case human capital raises the

steady state income of country i without affecting steady state growth. Alternatively,

human capital can affect the growth rate itself. There are three possibilities. First,

education can be a boon to technical change, as a more educated workforce is more likely

to think of and implement new ways of doing things (Nelson and Phelps, 1966). This

raises steady state income growth, even for economies that are at their current steady state

income level. Again, this implies β 2 > 0. Alternatively, a more skilled workforce may be

regression.

C:\TOPEL\PAPERS\growth.doc 32
better at adopting existing technologies (Welch, 1966). For example, South Korea’s post-

1970 expansion of secondary and higher education may have positioned it for more rapid

subsequent growth, by making existing technologies of developed countries easier to

adopt (Kim and Topel, 1995). This effect raises the speed of convergence for economies

that are below their steady state income level. This might yield β 3 < 0: additional human

capital has a smaller impact on growth when initial income is high, and there is less to

adopt from abroad.

The third possibility is that β 2 < 0: countries with low initial stocks of human

capital have greater opportunities to grow. In fact, this is implied by conditional

convergence. Further, for less developed countries much of the growth process is likely

to be “catching up” by accumulating knowledge from abroad. Education, particularly at

low levels, is simply the transference of knowledge that has already been produced and

used somewhere else. Other things equal, a country with low initial educational

attainment may have lower costs of growing. To me, this means that little can be learned

from a model like (10) about the effects of initial human capital on growth. Both β 2 > 0

and β 2 < 0 are consistent with the idea that human capital investment is a boon to growth

and development.19

These points aside, what do the data reveal about the relationship between

initial human capital and growth? BSM estimate models of long term (1960-85) growth,

with controls for H that include the time-0 average years of primary, secondary, and

higher education, public expenditures on education as a proportion of GDP, and life

19
Using BSM’s educational data for 1960-1990, a regression of the growth of average years of schooling
on initial schooling and initial log output per worker yields a coefficient of -.005 (t=2.9) on initial

C:\TOPEL\PAPERS\growth.doc 33
expectancy at birth. Consistent with the findings of MRW, above, BSM find that initial

educational attainment at the primary level is unrelated to country differences in

subsequent economic growth, but that secondary higher educational attainment are related

to growth. For men, they find that a one standard deviation increase in average years of

secondary education (about .9 years) raises the average annual growth rate by 1.5

percentage points per year. A one standard deviation increase in average years of post-

secondary education (.2 years) raises growth by 1.0 points per year. For women, BSM’s

estimates imply that greater educational attainment reduces growth. For example, a one

standard deviation increase in years of secondary schooling for women (.9 years) reduces

annual growth by 0.8 points per year. This is consistent with the argument for β 2 < 0

stated above; countries with low educational attainment for women may have greater

opportunities to grow, because they have an untapped source of potentially productive

human capital.

These results are suggestive of important effects of human capital on growth –

education seems to do something – but it is hard to take them seriously for any sort of

calibration or policy purposes, even ignoring the negative effects of female schooling.

Consider what we might expect to from standard estimates of the private returns to

schooling and from a Solow-type model augmented to include human capital. Let the

human capital stock be H=hL, where L is the labor force and h is human capital per

worker. Then steady-state log per-capita income in country i is (assuming Cobb-Douglas

production and labor-augmenting technical progress):

schooling. Over a 30 year period, a 2 standard deviation increase in initial schooling (5.3 years) reduces

C:\TOPEL\PAPERS\growth.doc 34
(11) ln(Y / N ) i = α ln( K / N ) i + (1 − α )[ln( L / N ) i + ln hi ] + (1 − α ) ln Ai

Human capital models of endogenous growth imply that the capital/output ratio is

constant, which seems to supported by the data (see Young, 1992, and Figure 1, above,

for evidence on this). With this condition we can rewrite (11) as:

(12) ln(Y / N ) i = φi + ln( L / N ) i + ln hi + ln Ai

where φi is the constant log capital/output ratio for country i. Equation (12) says that

increases in human capital per worker result in equal proportionate increases in per capita

income, as in the endogenous growth models reviewed above. Now specify:

(13) ln hi = θ i + θ S S i + θ X X i

where Si is average years of completed schooling and Xi includes other determinants of

average human capital per worker such as experience, on-the-job training, and the like.

Equation (13) can be interpreted as an aggregate form of human capital earnings functions

that are commonly estimated on micro-data, which assume that wages are proportional to

human capital supplied.20 Notice that the form of (13) implies that an additional year of

cumulative growth in schooling by .75 years.

20
The common form of human capital earnings functions implies that lnhij = θi0 + θZij for person j in
country i, where Z is a vector of human capital controls such as schooling, experience, and so on. As
pointed out by Heckman and Klenow (1997), if individual wages are log normal then average human
capital per worker is hi = exp[θi0 + θZ i + θ’Vi θ] where Zi is the mean of Zij and V is the covariance
matrix of Z. Thus the discussion in the text ignores the cross sectional variance of human capital in the

C:\TOPEL\PAPERS\growth.doc 35
schooling raises human capital by a constant percentage amount, θ S , independent of the

level of hi. Typical estimates of the returns to schooling from micro-data yield a effect of

a year of additional schooling on log wages in the range of .06-.10, depending on country

and time period under study. For the sake of argument, put this value at .08. To gauge

this impact on labor’s marginal product against θ S , divide by labor’s share of aggregate

output, which we can put at about .60 (see Table 1). This yields an approximate value for

θ S of .13, so an additional year of schooling for the average worker should raise per-

capita income by about 13 percent if the private and social returns to schooling are equal.

Unless the human capital externalities suggested by Lucas (1988) and others are truly

grand, this puts an approximate upper bound on the impact of schooling on output per

worker.

Now compare this value to BSM’s estimates of the impact of human capital on

growth. Under one interpretation, human capital raises steady-state income, Yi*, in (9). If

a year of additional schooling raises steady state income by 13 percent – using the private

returns – and the rate of convergence is on the order of β 1 = .03 per year, then the effect

of additional human capital on growth should be about .03x.13 = .0039 per year of

additional schooling. Thus the BSM estimates – which suggest effects of well over .01

per year of schooling—are vastly too big for the model they purport to estimate. The

alternative interpretation, that an additional year of average schooling raises an economy’s

steady state growth rate by over 1% per year, does not have a well-defined benchmark

from micro-data. Yet at any reasonable rate of interest this effect on growth implies a

interpretation of equation (13). To the extent that variance terms are relatively stable within a country,
they are removed by the fixed effects estimator employed here.

C:\TOPEL\PAPERS\growth.doc 36
huge rate of return.21 The conclusion that seems warranted is that countries with high

levels of educational attainment also have other, unmeasured, attributes (such as

subsequent investment) that cause growth. Thus, it is impossible to interpret BSM’s

estimates as the effect of human capital on economic growth.

Benhabib and Spiegel (1994), Pritchett (1997), and Bils and Klenow (1997)

study the impact of growth in imputed human capital on growth in output per worker,

using a form of equation (8). Each of these studies finds minor, or even negative, effects

of growth in imputed human capital on growth in output, though Benhabib and Spiegel

confirm BSM’s finding that the level, as opposed to the change in the stock of educational

capital is correlated with growth. Like BSM’s estimates, however, the magnitude of the

effect of education on growth is vastly too large to be interpreted as a causal force.22 In

short, the empirical growth literature does not lend much support to the idea that human

capital, at least as represented by measured educational attainment, is a key element of

economic growth. My own examination of the data leads me to be less pessimistic,

however.

3.6 New Evidence from Old Data

To examine the relationship between education and economic growth, I use the

Summers-Heston Mark 5.6 (1995) data, combined with the Barro-Lee (1993) data on

educational attainment. The Barro-Lee data record various measures of educational

attainment for a panel of 118 countries, at 5-year intervals from 1960 to 1990. The

Summers-Heston data record various measures of output, investment, and living standards

21
At 5% real interest and a .01 effect on growth, the returns are roughly 4 times the costs.

C:\TOPEL\PAPERS\growth.doc 37
for 152 countries beginning in the 1950’s. Merging these two sources yields an

unbalanced panel of 111 countries, with usable data on education and output, beginning in

1960. Most previous efforts with data like these examine the determinants of long-term

changes in output, typically over the period from 1960-85. Instead I look for a closer

connection between the timing of input and output changes, using the data recorded at 5-

year intervals from 1960 to 1990. Combining (6) and (13), write output per worker in

country i at time t as:

(14) ln( y it ) = α ln( k it ) + (1 − α )[θ i + θ S S it + θ X X it ] + (1 − α )vit

where v includes the state of unobserved technology in country i as well as unobserved

components of human capital. Data on capital per worker are fairly sparse, which leads

me to two alternative strategies. First, using estimates of capital per-worker constructed

by Klenow and Rodriguez-Clare (1997) for 1965 and 1985, I impute estimated

capital/labor ratios in other years by linear interpolation. Then rearrange (14) to obtain

(15) [ln( y it ) − α ln( k it )] /(1 − α ) = θ i + θ S S it + θ X X it + vit

Application of (15) requires an assumption about capital’s share, α, to measure the left-

hand side, along with the assumption that this value does not differ across countries.

22
They find that an additional year of average schooling raises the 1965-85 growth of rate by about .13

C:\TOPEL\PAPERS\growth.doc 38
Alternatively, if we accept the evidence that motivates endogenous growth models and

treat the capital/output ratio as a country-specific constant, then (15) becomes

(16) ln( y it ) = φ i + θ i + θ S S it + θ X X it + vit

So long as capital/output ratios or average levels of unobserved human capital differ

across countries, both (15) and (16) involve country-specific fixed effects that should be

accounted for in estimating the model. These effects can be eliminated by using either a

fixed-effects estimator or by differencing the data over time.

I estimate these models in several different ways. Table 2 shows estimates of

equations (15) and (16) when the only measured determinant of human capital is average

years of schooling. The data on output and schooling are recorded at 5 year intervals, on

an unbalanced panel of 111 countries, yielding 719 observations. All models contain

country effects to account for the terms φi + θi, so the estimates are generated solely by

within-country variations in output and educational attainment. The first panel of

estimates (columns 1-4) assumes that the capital-output ratio is fixed within a country, so

no adjustment for capital intensity is needed.23 The estimated social returns to schooling

are remarkably large. Omitting year effects in columns (1) and (2), the effect of an

additional year of schooling on average productivity exceeds 20 percent, with slightly

larger returns to secondary than to primary education (column 2).24

23
A within-country regression of (imputed) log capital per worker on log output per worker has a
coefficient of 1.17 (se = .05).
24
Heckman and Klenow (1997) report cross-sectional estimates of a regression of GDP per-capita on
average years of schooling, also using the Summers-Heston data, for 1960, 1985, and 1990. Their

C:\TOPEL\PAPERS\growth.doc 39
 Some care should be taken in interpreting these, and the following effects. The

effect of .226 in column (1) means that an additional year of schooling raises productivity

by this amount, allowing for the endogenous response of capital that keeps the capital-

output ratio fixed. It should be multiplied by labor’s share to gauge how schooling affects

the marginal product of labor, which is then comparable to estimates of the private returns

taken from individual data. For example, if labor’s share is .6 then the effect of schooling

on the log average wage is .6x.226 = .135 per year of additional schooling. This exceeds

the typical private return.

When year effects are added to the model in column 3, the estimated unconditional

return to an additional year of schooling is falls to 10 percent. Accounting for year effects

(column 4), when average years of schooling are broken down into primary and secondary

components, the estimated returns to secondary schooling are more than double the

returns to primary schooling, though both are different than zero by the usual criterion.

Columns 5-8 of the table drop the assumption of a constant capital/output ratio

by adjusting the productivity data for differences in estimated values of ln(K/L). In

columns 5-6 I assume that capital’s share of aggregate output is .35 – assumed to be fixed

across countries and over time – while the estimates in columns 7-8 assume α = .50. The

latter estimate is probably at the upper end of what can be deemed reasonable – see Table

1 above. The estimated returns to schooling fall as capital’s assumed share rises, and

estimated standard errors rise as well. Even so, the implied returns in columns 5 and 7 of

the table are not unreasonable in light of the effects of schooling typically found in micro

estimates, generated by between country variation in average school attainment, also show returns in the
.2-.3 range.

C:\TOPEL\PAPERS\growth.doc 40
data. Again, the division between years of primary and years of secondary schooling

indicates larger returns for secondary education.

The estimated returns in Table 2 neglect other measurable determinants of

human capital. This is not a concern if neglected elements of human capital are fixed

within countries, as these differences are absorbed by the fixed effects. But measured and

unmeasured innovations to human capital (or other inputs that are complementary with

human capital) might well be correlated, which would lead to an overestimate of the

returns to schooling. I considered two other measurable correlates of human capital. The

U.S. Census Bureau compiles international statistics on the age distribution of the

economically active population for various years and for most of the countries used in the

estimation procedure, using census and other data from each country. I used these data to

construct average age and experience (age - schooling - 6) for the working aged

population. The Census also reports average life expectancy at birth for most countries

and years, using age-specific mortality rates. As shown in Table 3, after controlling for

year effects these variables have no substantial impact on productivity.

The finding that changes in life expectancy at birth do not have a substantial

effect is not too surprising, since much of increased life expectancy in developing

countries is accomplished through reductions in infant mortality. These changes may have

little to do with improvements in the human capital of the working age population.25

Further, any effects that do emerge from the least-squares estimates – as in columns 1 and

25
For example, in Gineau-Bissau in 1980 the average age of the economically active population was 38.7
years, while life expectancy was just 44.4 years. In the same year, the average age of the economically
active in Israel was 38.1 years, but life expectancy was 75 years. There is remarkably little variation in
average ages of the economically active. In 1990, the mean age across countries was 36 years, with a
standard deviation of 1.79.

C:\TOPEL\PAPERS\growth.doc 41
2 – may reflect the effect of economic growth on health. The negligible impact of average

age (and thus of experience) is more surprising in light of evidence from micro data on the

private returns to experience. The panel data on average age of the economically active

population are fairly meager, however. And even accurate measurements would be

affected by improving health status of successive cohorts of workers, which could reduce

average age while raising productivity.

While the estimates in Tables 2 and 3 are generated by within-country changes

in schooling and productivity, the fixed-effects estimator is not an explicit empirical model

of growth. Fixed effects can also be removed by first-differencing the data, so the

variables of interest are expressed as changes over time:

(17) ∆ ln( y it ) = θ S ∆S it + θ X ∆X it + ∆vit

Estimates based on (17) are more comparable to the empirical growth literature, especially

contributions by Barro and Sala-i-Martin (1995), Pritchett (1997), and Benhabib and

Speigel (1994). Before proceeding to the estimates, it is worth noting the effect of

differencing in magnifying the effects of measurement error in recorded schooling.

Assume that average years of recorded schooling measures true schooling with classically

distributed measurement error, SM = S + e. Then the asymptotic bias of least squares

applied to (17) follows from (ignoring the role of other regressors):

C:\TOPEL\PAPERS\growth.doc 42
 θS
(18) p lim(θ S ) =
σ2
1+ 2 e
σ S (1 − ρ )

where ρ is the correlation between St and St-1. Thus serial correlation in St increases the

noise-to-signal ratio in differenced data, magnifying the downward bias caused by classical

measurement error. This suggests an econometric tradeoff in analyzing the determinants

of economic growth: more frequent observations increase sample size, but frequent

observations are less informative about the effects of interest in the presence of

measurement error and serial correlation.26

This point is demonstrated in Table 4. I calculated average annual growth rates

of output per worker based on intervals of 5, 10, 15, and 20 years, along with the average

annual change in years of schooling. Columns 1, 4, 7, and 10 in Table 4 simply regress

growth of output on growth in educational attainment and year effects. Notice that when

the 5-year average growth rate is used, the effect of measured schooling on productivity is

only .028 per year, which is barely significant by conventional standards and which implies

an effect on wages that is well below the private returns to schooling estimated in micro-

data. This estimate rises, however, as the interval for calculating growth lengthens. At

intervals greater than 20 years the effect of an additional year of schooling on log output

per worker is .167, which implies an effect of .10 of schooling on wages. This rising

impact of schooling as the length of the growth interval is lengthened may reflect the

C:\TOPEL\PAPERS\growth.doc 43
impact of measurement error, or the effect of accumulating complementary inputs in the

longer run.

Columns 2, 5, 8, and 11 of Table 4 repeat the exercise, but add initial years of

schooling and initial log output per worker to the growth regressions. An additional year

of initial schooling raises subsequent growth by about 0.4 percent per year, independent of

the length of the interval used to gauge growth27. While this is smaller than the effects of

initial schooling found in Barro and Sala-i-Martin (1995), in combination with the effect of

initial log output per worker it is still too large to be interpreted as the effect of schooling

alone on steady-state growth. The effects of lagged output show what is typically

interpreted as evidence of convergence: at any point in time, lower income countries tend

to grow (slightly) faster. Conditioning on these initial values tends to raise the effects of

changes in schooling on growth.28

The last column of estimates under each growth interval adds an interaction

between growth in average schooling and initial log income per worker. Initial income is

deviated from its year-specific mean, so the reported main effect is the impact of additional

schooling at the mean of the distribution of initial productivity. At least for the shorter

growth intervals, there is some evidence that additional years of schooling have a larger

impact at low levels of initial output per worker. The fact that the interaction dies out as

the growth interval is lengthened indicates that most of the effect comes from within-

26
More generally, differencing exacerbates the effect of measurement error if serial correlation in
schooling exceeds serial correlation in measurement error.
27
Benhabib and Spiegal (1994) and Barrro and Sala-i-Martin (1995) also find that initial schooling raises
growth. Benhabib and Spiegal interpret this effect as reflecting the ability of more educated workers to
adopt existing technologies.
28
I do not report similar regressions which allow for separate effects of primary and secondary schooling.
Tests of the restriction that primary and secondary schooling have identical effects cannot be rejected for
these models.

C:\TOPEL\PAPERS\growth.doc 44
country variation in growth. At a 20-year interval, the estimated impact of a year of

schooling on average productivity rises to .246, evaluated at the mean level of initial

productivity.

As in Table 3, these estimates do not control for other elements of human

capital, ∆Xit , which may be correlated with changes in schooling. Given the quality of the

data, direct measurement of these is infeasible. Further, average growth in schooling may

be correlated with technical change, which is also unmeasured. An alternative is to

assume that unobserved components of human capital and technical progress evolve at a

constant rate within a country, so that θx∆Xit = λi. Then λi is a fixed effect in the growth

model (17), which can be eliminated (in panel data) by standard methods. The resulting

difference-in-differences estimator is unaffected by the correlation between innovations to

average schooling and unmeasured factors in λi. A limitation of this approach is that the

fixed effects estimator is best suited to “short” growth intervals – say 5 or 10 years in the

panel length available here – which increases the number of observations per country. But

this also increases the importance of measurement error in recorded schooling, as

indicated above.

With this limitation in mind, Table 5 reports estimates of the effects of

schooling on economic growth, controlling for fixed country effects, in productivity data

measured at 5 and 10 year intervals. Despite issues of measurement error, there remains a

positive (and reasonable) effect of schooling in the 10-year growth data. At the mean

level of initial productivity, the estimates in column 4 of the table indicate that an

additional year of schooling raises average productivity by nearly 9 percent. This estimate

C:\TOPEL\PAPERS\growth.doc 45
may understate the true returns to the extent that measurement error is exacerbated by the

difference-in-differences estimator.

The estimates in Tables 2-5 indicate that increases in average years of schooling

of the workforce do raise productivity and contribute to economic growth. Taken as

estimates of the contribution of average years of schooling to the stock of human capital,

the low end of the range of these estimates – say 7-10 percent per year of schooling -- is

consistent with comparable estimates of the private returns to schooling derived from

micro-data. The upper end of the range of estimates suggests social returns to an

additional year of schooling that may be larger than traditional estimates of private returns.

This possible excess of social over private returns is consistent with growth models that

incorporate human capital externalities in the production of output, such as Lucas

(1988).29 Then private decisions lead to too little investment in human capital, and too

little growth, compared to the social optimum. The confirmation that the level of

schooling also affects growth, though with a smaller impact than in previous literature,

suggests that more than simple input accumulation is at work in generating growth.

The finding that investments in schooling raise productivity and growth is

different from the conclusions of Pritchett (1997) and Benhabib and Spiegel (1994).

Using data on long term (20 years or more) growth for roughly the same sample of

countries examined here, they find that changes in the measured stock of human capital

29
Heckman and Klenow (1997) make a similar point.

C:\TOPEL\PAPERS\growth.doc 46
are unrelated to changes in the average product of labor.30 Why do they find negligible

effects of contemporaneous investments in education?

Pritchett’s (1997) results are due to the way he measures human capital. Unlike

equation (13), which is the aggregate analogue of the usual human capital earnings

function, Pritchett’s measure assumes that an additional year of schooling raises the stock

of human capital by a larger proportional amount in less educated countries than in more

educated ones. When this form for human capital is used in the data for tables 2-5, I also

find no effect of schooling on growth. This restriction is rejected by the data, and it is

inconsistent with widely accepted evidence on the form of human capital earnings

functions.31 Benhabib and Spiegel (1994) obtain their measure of human capital from

Kyriacou (1991), who imputed average years of schooling from a linear regression of

schooling on past enrollment rates. Their regressor is the log change in imputed average

years of schooling for a country; thus, like Pritchett (1997) they assume that a year of

schooling has a larger proportional impact on the stock of human capital in low-education

countries. The models estimated here assume that each additional year of schooling raises

the stock by a constant proportional amount, as is implied by the standard form of human

capital earnings functions applied to individual data.

3.7 Summary: What Do We Know about Human Capital and Growth?

30
As noted above, Benhabib and Spiegel (1994) do find that the initial stock of measured human capital
raises subsequent growth, which they attribute to the ability of a more educated workforce to adopt
existing technologies from abroad.
31
More formally, Pritchett assumes that lnh = H0 + ln(exp(θsSi ) - 1). Then dlnh/dS = θs/(exp(θsSi - 1),
which →∞ as S → 0.

C:\TOPEL\PAPERS\growth.doc 47
 As this discussion indicates, the empirical literature connecting human capital

investment to aggregate productivity and economic growth is inconclusive. Results from

cross-country comparisons of the levels of productivity often hinge on the particular way

that human capital is measured. Thus in the parlance of “ models, Mankiw, Romer,

and Weil (1992) attribute cross-country differences in productivity to differences in “K”

(measured inputs), while Klenow and Rodrigues-Clare (1997) give greater weight to “A”

(differences in factor productivity). Empirical models of economic growth generally

conclude that human capital – especially schooling – plays a role, but the particular

channels through which schooling affects growth are open to debate. Several studies find

that initial levels of schooling raise subsequent growth though, as noted above, these

effects appear suspiciously large. And there is no well-articulated theory of how these

effects come about. A more direct connection follows from Solow-style models of

economic growth, which predict that changes in the stock of human capital should drive

changes in output. This relationship appears to hold in the data examined above, and the

magnitude of the estimated effect appears reasonable in light of prior knowledge of the

impact of schooling on wages.

The overwhelming evidence from studies on micro-data is that human capital

investment raises productivity. Though signalling models of schooling (Spence, 1974)

imply that the private returns to schooling can exceed the social returns, empirical

evidence for important signalling effects is at best meager. In my view, the weight of

evidence from micro-data yields a strong prior that rising educational attainment of the

labor force should spur economic growth. Further, this evidence on private returns

provides a fairly precise range for how big the effects should be. The key empirical issue

C:\TOPEL\PAPERS\growth.doc 48
is not whether schooling raises aggregate output – evidence to the contrary should be

regarded with great suspicion, especially given the quality of data that are used in

aggregate growth studies. Rather, the significant open question is whether the social

returns to human capital investment substantially exceed the private returns (Heckman and

Klenow, 1997). Then public expansion of education may be a key ingredient in economic

growth.

4. Growth, Investment, and Relative Wages

4.1 Background

In his presidential address to the American Economic Association, Simon Kuznets

(1955) attempted to characterize the development process in terms of a few common

themes. Unlike the balanced growth models in vogue today, he viewed rapid economic

development of a country as a transition from a rural, agricultural base – with a relatively

unskilled labor force – to modern industrialism. He was particularly interested in the

effects of growth on income distribution, hypothesizing that wage and income inequality

increased in the early stages of rapid growth, but later fell. This inverted-U relation of

inequality to development came to known as the “Kuznets Curve.”

In Kuznet’s description of events, rising demand for industrial labor is the precursor

of growth. How does this come about? A plausible candidate is trade liberalization. As

described in Tsiang (1979), early economic policies in less developed countries emphasized

the protection of domestic industries through “import substitution.” Experiences of

Taiwan and other Asian economies discredited this approach, and the opening to trade led

to rising demand for manufactured exports. In this description of events, rising export

C:\TOPEL\PAPERS\growth.doc 49
demand raises the demand for skilled industrial workers, leading to rising investment in

human capital and an exodus of labor from agriculture. The Kuznets Curve occurs because

rising industrial wages draw the small number of skilled workers to that sector, raising

inequality, but later migration and investment in human capital changes overall factor

proportions. Skill intensive sectors expand during the development process, but wage

inequality eventually declines as skilled workers become less scarce.

To focus on this process, consider an economy with two labor types, skilled (S)

and unskilled (U) and two sectors. Output in sector j is Yj = Fj(Sj,Uj), with constant

returns to scale. Assume that sector 1 is relatively skill intensive; think of it as the

“industrial” sector and sector 2 as “agriculture.” Then aggregate output at date t is

(19) Yt = At1 F 1 ( S t1 ,U t1 ) + At2 F 2 ( S t2 ,U t2 )

The stock of skilled labor can be augmented by investments in human capital, which

transforms type U labor into type S. The technology for this is

•
(20) S t = Bf (vt , S t )S t − δS t

In (20), v is the proportion of the stock of skilled labor that is devoted to training, and B is

the number of unskilled workers who obtain training per unit of v. The function f satisfies

fvv < 0, so there are diminishing returns to raising v. I also assume complementarity

between the stock of skilled labor and the productivity of time devoted to training, fvS >

0. The model is closed by labor supply constraints for each labor type:

C:\TOPEL\PAPERS\growth.doc 50
(21) (1 − vt ) S t = S t1 + S t2

(22) U t = U t1 + U t2 + Bvt S t

(23) Nt = St + Ut

An efficient allocation maximizes the present value of social output by allocating labor

among production and training. In addition to the usual marginal conditions equating the

marginal products of skilled and unskilled labor across sectors, an interior solution for

efficient investment in human capital must satisfy:

∞
(24) Wt s + BWbu = Bf v (vt , S t ) ∫ (Wτs − Wτu )e −( r +δ − g ( v , S ))(τ −t ) dτ
t

Where g(v,S) > 0 is the difference between the average and marginal products of f.

Equation (24) is the condition of interest for the following discussion. The left hand side

is the opportunity cost of human capital investment: Each unit of skilled labor devoted to

training has an opportunity cost of Ws – the wage of skilled labor – and the student-

teacher ratio of B requires that B unskilled workers be withdrawn from production as well.

The right side represents the present value of human capital produced by raising vt, where

Ws - Wu is the wage premium commanded by skilled labor. Equation (23) has several

important implications.

First, suppose that Ws and Wu are fixed through time. This means that relative

wages in this economy are independent of factor proportions, as would occur under factor

C:\TOPEL\PAPERS\growth.doc 51
price equalization for a small open economy. With constant returns to scale in each

sector, a sufficient condition for this Stolper-Samuelson result is that the prices of each

sector’s output are fixed, say on international markets. Given this, (24) presents two

interesting cases. First, if St is sufficiently low then the marginal product of new

investments may be too small to satisfy (23). As in Becker, Murphy, and Tamura (1990)

or Azariadis and Drazen (1990) the economy is stuck in a low growth “trap” because the

marginal cost of investing in skills is too high. This is an outcome of the fact that it takes

human capital to produce more, and the stock has a spillover effect in f, so poor initial

conditions can block subsequent economic growth.

The second case occurs when S is sufficiently large to spur investment and

growth. It is obvious from (24) that, for any S sufficiently large, v is increasing with the

wage premium Ws - Wu. Then S, aggregate output, and output per capita all rise over

time, and labor migrates from low-skill sector 2 to high-skill sector 1. Relative wages are

unchanged, however. With constant returns and fixed output prices in each sector the

increase in the aggregate ratio of skilled to unskilled labor is absorbed by migration of

labor to the skill-intensive sector, 1. The skill ratio in each sector is unchanged, which

leaves marginal products unchanged. But because S is rising, so is the proportion of S

devoted to human capital investment. In other words, development begins slowly because

S is initially low, but it accelerates as the stock of skilled labor rises over time.

The dynamics of this model incorporate a sort of “Kuznets Curve” in that the

relative proportions of high and low-wage labor change over time. If the share of skilled

labor in the labor force is initially low, then measures of wage dispersion can rise, and

C:\TOPEL\PAPERS\growth.doc 52
 then fall, as “development” (skill upgrading) proceeds.32 While this is much of what

Kuznets had in mind, the more interesting case occurs when factor prices adjust to the

increased skill intensity of the labor force. Then the wage distribution may narrow

because investment makes skilled labor relatively less scarce. Indeed, in this case

investment in human capital serves to offset wage inequality. When will this occur?

It obviously cannot occur if Stolper-Samuelson conditions are satisfied, for then

factor price equalization (FPE) nails down the skill premium in wages. But the

conditions for FPE are so strong that it would be surprising if factor proportions did not

affect relative wages. FPE will fail if (i) there are diminishing returns in production; or

(ii) output prices decline with quantities produced – the economy’s goods are imperfect

substitutes for others on international markets; or (iii) there are more labor types than

sectors with constant returns and fixed output prices. If any of these conditions are true,

then the relative price of skilled labor will fall as skills become more abundant.

In this case, a rising stock of skilled labor reduces the present value of wage

premiums on the right side of (24). If the initial value of S is too low, the model implies a

“low growth” state with low investment, low growth, and a large skill premium in wages.

But if S is above a threshold that generates investment, the model implies rising per-

capita output along with: (i) migration of labor from the low-skill to the high-skill sector;

(ii) skill upgrading (increasing Sj/Uj) in each sector; and (iii) a steadily declining wage

differential, Ws - Wu, between skilled and unskilled labor, as the real wages of both skill

groups rise. As envisioned by Kuznets (1955), economic development is an increase in

32
Here, the variance of wages is rising with S so long as the skilled share of the labor force is smaller than
.5.

C:\TOPEL\PAPERS\growth.doc 53
 the relative abundance of skills, which is a force toward greater wage and income

equality.

The assumption of complementarity between the stock of skills, S, and the portion

of that stock devoted to investment, v, delivers a final implication. Suppose the contrary,

that fvS = 0, so abundant skills do not affect the productivity of training. Then a rising

stock of skills reduces the present value of wage premiums on the right side of (24).

With diminishing returns, this means that the portion of the stock devoted to new

investment declines over time, as the marginal returns on investment are falling. From

(19) this means that the rate of growth of S is falling as well. But fvS > 0 implies that

productivity rises over time, so the rate of growth in the stock of skilled labor need not

decline with development.

4.2 Wage Inequality and Development: Evidence

Evidence on the relationship between wage or income inequality and the process of

development is fairly limited. There are two main empirical issues. First, in a model like

the one outlined above, wage inequality can increase growth by raising the returns to

human capital investment. So the first question is: Do greater returns to skill increase

human capital investment? The second question has to do with how relative wages

respond to investment: Does investment, by changing factor proportions, reduce the

relative price of skilled labor and thus reduce wage inequality? In effect, we need to

separate (i) the effects of skill prices on the flow of new investment from (ii) the effects of

human capital stocks (cumulative investment) on relative skill prices.

Evidence on the effect of wage differences on human capital investment requires time

series data on measurable investment activity, and corresponding data on the returns.

C:\TOPEL\PAPERS\growth.doc 54
Edin and Topel (1996) and Topel (1997) examine the one dimension of human capital

investment that is directly observable and measurable, schooling, in Sweden and the U.S.

Figure 2, taken from Edin and Topel (1996), graphs the relation between the estimated

returns to a college education in Sweden on the left hand scale (measured as the difference

in log wages between workers with 16 years of schooling and workers with 12 years), and

the proportion of Swedish men aged 20-24 who are enrolled in school on the right-hand

scale.33 Between 1968 and 1984, the log wage differential for a college graduate fell from

.59 to .20, with most of the drop occurring in the six years 1968-74. At its nadir, the

returns to a college education in Sweden were less than half of the lowest returns

observed in U.S. data. This decline corresponds to an overall compression of the

Swedish wage distribution during this period. How did young people respond? The

figure shows that the proportion of young men attending college fell from 13 percent in

1968 to only 8 percent in the early 1980’s. The correspondence between enrollments and

returns suggests that investment in human capital is sensitive to its price.

Figure 3 shows corresponding evidence for American men. Here enrollments

are measured as the fraction of men aged 20-24 with some college. Again, the

correspondence in the two series is striking. As the college wage premium rose through

the 1960’s, the fraction of young men with some college climbed, peaking at 44 percent in

the early 1970’s, when the returns to college were higher than ever before. As the college

wage premium fell in the 1970’s (why? See below), so did school attendance, reaching a

low of 37 percent in 1980. Then both the returns to college and college attendance

trended up, the latter reaching an all time high of 46 percent in 1992, when the returns

33
The wage figures are for workers with 1-9 years of labor market experience.

C:\TOPEL\PAPERS\growth.doc 55
were also at a record high. Similar patterns hold for young women. From 1979 to 1993,

school attendance rates for young women rose from .30 to .41. As in Sweden, the

American evidence is that the supply of human capital rises with the relative price of skill.

Greater investment increases the stock of human capital, and we expect the

rental price to fall with the stock. This underlies Katz and Murphy’s (1990) explanation of

the time series shape of the college wage premium in the U.S., shown in Figure 3.34 They

assume that the relative demand for college graduates grows at a steady pace (trend) –

presumably because of skill-biased technical change – and they show that changes in the

relative supply of college graduates is inversely related to the college premium. The glut

of college graduates in the 1970’s, driven by baby boom cohorts, caused the college

premium to decline. Katz and Murphy’s implied elasticity of substitution between college

and high school graduates (about 1.4) is consistent with other labor demand studies

(Hamermesh, 1993).

This basic approach has been replicated in other countries, which have experienced

even larger changes in the relative supplies of educated labor (see Katz, Loveman, and

Blanchflower, 1995, for comparisons of the U.S., Japan, France, and the U.K.). Edin and

Holmlund (1995) show that the rapid decline in the college wage premium in Sweden

coincides with an increase in the labor force share of college graduates, which more than

doubled between 1971 and 1985. Their estimate of the elasticity of substitution between

college and high school graduates (2.9) is substantially larger than what has been found in

other labor demand studies, however (Freeman, 1986). Table 6 provides illustrative

34
Freeman (1980) is a related analysis of the changing returns to a college education.

C:\TOPEL\PAPERS\growth.doc 56
estimates of the effects of factor proportions on relative wages for the small number of

countries where formal studies have been carried out.

While the college premium in western economies increased during the 1980’s, it

fell dramatically in some developing countries. In Korea, the college wage premium fell

by 25 log points between 1976 and 1989. By the end of the 1980’s, the return to a college

education in was Korea only two-thirds as large as in the U.S.; it had been double the U.S.

return in 1979. Kim and Topel (1995) argue that this compression was driven by a rapid

upgrading of educational attainment in the Korean labor force. The labor force share of

elementary school graduates fell from 60 to 30 percent in only 19 years, while the shares

of college and high school graduates soared. Consistent with this, the relative wage of

university graduates fell and the relative wage of those with an elementary education rose

sharply (see Figure 4). The result was a decline in wage inequality in Korea – shown in

Table 7 – that ran against the trend toward greater inequality in developed economies.

How do these results square with the notion that free trade equalizes factor

prices, so that factor proportions in any particular country should not affect that country’s

relative wages? As noted above, the conditions for factor price equalization are, at best,

extreme. In terms of observable quantities, a key implication of FPE is that factor

proportions in individual industries are independent of overall factor proportions for the

economy as a whole. Expansion of skill-intensive sectors absorbs the rising supply of

skilled labor without affecting relative wages. This prediction is at odds with evidence

from a number of countries, most notably the U.S., where rising educational attainment of

the labor force has resulted in skill upgrading in virtually all industries (Murphy and

Welch, 1993). In terms of actual responses of relative wages in developing economies

C:\TOPEL\PAPERS\growth.doc 57
Robbins (1996) summarizes a number of country studies of the effects of trade

liberalization and changes in relative supplies of skills on relative wages. His main concern

is with the applicability of FPE to wage distributions in developing countries. As in the

other studies surveyed above, he finds a general pattern that relative wages are inversely

related to relative supplies, even in apparently open economies where FPE might hold.

What can we take from this evidence? If we accept the postulate of modern

growth theory, that human capital accumulation is the engine of economic growth, the

data tell us that there is a trade-off between growth and income redistribution. Policies

that compress wage or after-tax income differences across skill groups reduce human

capital investment and, in so doing, reduce steady state growth and long-run prosperity.

Many critics of redistributional policies recognize this, at least implicitly, and some

governments pursue offsetting policies.35 For example, as college enrollments fell in

Australia and Sweden – where centralized wage setting serves to compress skill premiums

in wages (Gregory and Vella, 1995; Edin and Topel, 1997) – policymakers went beyond

traditional subsidies to education by paying students to stay in college. Of course,

education is only one component of human capital investment. Policies that artificially

compress the wage distribution also reduce the return on post-schooling investment, such

as on the job training. Indeed, this may be the larger effect.

A related point is also important. Increased inequality is widely thought to

represent an important social problem. Schooling is the one dimension of human capital

investment that is directly observable and measurable, and the evidence is that it responds

35
Lindbeck et. al. (1994) are explicit: “Empirical studies indicate that slower accumulation of physical
capital can only explain the Swedish fall in productivity growth in the 1970’s and 1980’s to a limited

C:\TOPEL\PAPERS\growth.doc 58
to changes in returns. As importantly, evidence from several countries indicates that

changes in the relative quantity of skilled labor, driven by investment, causes inequality to

fall. To the extent that rising inequality is driven by increased scarcity of skilled labor,

investment in human capital is the long run solution.36 In the long run, policies that

compress wages or incomes may exacerbate the underlying economic forces that cause

inequality, while retarding economic growth.

5. Concluding Remarks

I have offered a selective survey of the economics of growth, with an obvious

bias toward issues of interest to labor economists. The recent “growth of growth” as an

area of economic research has been dominated by theory, and there has been little

participation by labor economists. This is lamentable, as the questions raised by growth

models are enormously important and ripe for applied analysis.

But barriers to entry by applied researchers are not low. The fact that growth

theory is far ahead of empirical research has much to do with the quality and quantity of

data. Much of applied research in economic growth is an extension of growth accounting,

refining the measurements of inputs in Solow-style models of aggregate output. Statistical

analyses of the empirical determinants of economic growth have been largely constrained

to a single (valuable) data set, covering roughly 100 countries and a few standard

measures of inputs and output. These data seem to confirm a connection between human

capital (mainly measured by schooling) and economic growth, though the channels

extent....It is then tempting to pinpoint the accumulation of human capital. The private return on
education and on-the-job training has indeed been quite low in Sweden for a long time.”
36
Topel (1997) contains a more detailed discussion.

C:\TOPEL\PAPERS\growth.doc 59
through which these effects operate is open to debate. Barring a specification no one has

thought of, there isn’t much more to be learned from these data.

The most fruitful path may be closer to “development” economics than to the

kinds of empirical research that has been carried out thus far. By this I mean detailed

empirical studies of the operation of labor markets and the impact of policies and

institutions within individual countries. Emerging economies of Asia and Latin America

offer on-going laboratories in which to study labor markets during periods of rapid

economic growth, and many collect the kinds of detailed data that labor economists are

known to relish. It is only through this kind of tedious but rewarding empirical work that

we will come to understand the role of labor markets in the growth process.

C:\TOPEL\PAPERS\growth.doc 60
 References

Aghion, P. and Howitt, P. (1992) ‘A model of growth through creative destruction’,

Econometrica, 60:323–351.

Aghion, P. and Howitt, P. (1998) Endogenous growth theory, Cambridge, MA: M.I.T
Press.

Azariadis, C. and Drazen, A. (1990) ‘Threshold externalities in economic development’,

Quarterly Journal of Economics, 105:501-526.

Barro, R. and Lee, Jong-Wha (1993) ‘International comparisons of educational

Journal of Monetary Economics, 32:363–394.

Barro, R. and Sala-i-Martin, X. (1995), Economic growth. New York: McGraw-Hill.

Becker, G., Murphy, K.M. and Tamura, R. (1990) ‘Human capital, fertility, and economic
Journal of Political Economy, 98:S12–S37.

Benabou, R. (1996) ‘Heterogeneity, stratification, and growth: macroeconomic

implications of community structure and school finance’, American Economic
Review 86:584-609.

Benhabib, J. and Spiegel, M. (1994), ‘The role of human capital in economic

development: evidence from aggregate cross-country data’, Journal of Monetary
Economics, 34:143–74.

Ben-Porath, Y. (1967). ‘The production of human capital and the life cycle model of labor
supply’, Journal of Political Economy 75:352–65.

Card, D. and Krueger, A. (1992) ‘Does school quality matter? Returns to education and
the characteristics of public schools in the United States’, Journal of Political
Economy 100:1-40.

Chari, V.V., Kehoe, P., and McGratten, E. (1996) ‘The poverty of nations: a quantitative
exploration’, NBER Working Paper #5414, Cambridge, MA.

Christensen, L., Cummings, D., and Jorgensen, D.W. (1980) ‘Economic growth, 1947–
1973: an international comparison’, in: John W. Kendrick and Beatrice Vaccara,
eds., Developments in productivity measurement and analysis. Chicago:
University of Chicago Press.

Denison, E.F. (1985) Trends in American economic growth, 1929–1982. Washington,

D.C.: Brookings Institution.
Denison, E.F. (1962) ‘Sources of growth in the United States and the alternatives before
us’, Supplement paper no. 13, New York: Committee for Economic Development.

Denison, E.F. (1967) Why growth rates differ. Washington, D.C.: Brookings Institution.

Dougherty, C. (1991) ‘A comparison of productivity and economic growth in the G-7

countries’, Ph.D. dissertation, Harvard University.

Edin, P-A. and Topel, R. (1997) ‘Wage policy and restructuring: the Swedish labor
market since 1960’, in: Richard Freeman, Robert Topel, and Birgitta Swedenborg,
eds., The welfare state in transition. Chicago: University of Chicago Press.

Edin, P-A. and Holmlund, B. (1992) ‘The Swedish wage structure: the rise and fall of
solidaristic wage policy?’ in: Richard Freeman and Lawrence Katz, eds.,
Differences and Changes in Wage Structures. Chicago: University of Chicago
Press.

Elias, V. (1990) Sources of growth: A study of seven Latin American economies. San
Francisco: ISC Press.

Foster, A.D. and Rosenzweig, M.R. (1996) ‘Technical change and human-capital returns
and investments: evidence from the green revolution’, American Economic
Review, 86:931–953.

Freeman, R. (1981) ‘The changing economic value of higher education in developed

economies: a report to the OECD’, NBER Working Paper #820, Cambridge, MA

Freeman, R. and Needels, K. (1993) ‘Skill differentials in Canada in an era of rising labor
market inequality’, in: David Card and Richard Freeman, eds., Small differences
that matter, Chicago: University of Chicago Press for NBER.

Glomm, G. and Ravikumar, B. (1992) ‘Public vs. private investment in human capital:
endogenous growth and income inequality’, Journal of Political Economy,
100:818-834.

Gregory, R. and Vella, F. (1995) ‘Real wages, employment, and wage dispersion in U.S.
and Australian labor markets’, in: Richard Freeman and Lawrence Katz, eds.,
Differences and changes in wage structures, Chicago: University of Chicago Press
for NBER.

Grossman, G. and Helpman, E. (1991) Innovation and growth in the global economy.
Cambridge, MA: M.I.T. Press.

Heckman, J. and Klenow, P. (1997), “Human capital policy,” Working Paper, University
of Chicago, December.
Jones, L.E. and Manuelli, R.E. (1990) ‘A convex model of equilibrium growth: theory and
Journal of Political Economy, 98:1008–1038.

Jorgensen, D.W. and Griliches, Z. (1967) ‘The explanation of productivity change’,

Review of Economic Studies, 34:249–280.

Jorgensen , D., Gollop, F.M., and Fraumeni, B.M. (1987) Productivity and U.S. economic
growth. Cambridge, MA: Harvard University Press.

Kaldor, N. (1963) ‘Capital accumulation and economic growth’, in: Friedrich A. Lutz and
Douglas C. Hague, eds., Proceedings of a conference held by the London
Economics Association. London: Macmillan.

Katz, L. and Murphy, K.M. (1992) “Changes in relative wages, 1963-1987: supply and
Quarterly Journal of Economics, 107:35–78

Katz, L., Loveman, G., and Blanchflower, D. (1995), ‘Changes in the structure of wages
in four OECD countries’, in: Richard Freeman and Lawrence Katz. eds.,
Difference and changes in wage structures. Chicago: University of Chicago Press.

Kim, D. and Topel, R. (1995) ‘Labor market and economic growth: lessons from Korea’s
industrialization, 1970-1990’, in: Richard Freeman and Lawrence Katz. eds.,
Difference and changes in wage structures. Chicago: University of Chicago Press.

Klenow, P. and Rodriguez-Clare, A. (1997), ‘Economic growth, a review essay’ Working

Paper, University of Chicago.

Kuznets, S. (1955), ‘Economic growth and income inequality’, American Economic

Review, 45:1–28.

Kuznets, S. (1973) ‘Modern economic growth: findings and reflections’, American

Economic Review 63:247–258.

Kyriacou, G. (1991) ‘Level and growth effects of human capital’, Working Paper, C.V.
Starr Center, New York University.

Leamer, E. (1995) ‘A trade economist’s view of U.S. wages and “globalization”,’

Working Paper, UCLA.

Lindbeck, A., Mulander, P., Persson, T., Petesson, O., Sandmo, A., Swedenborg, B., and
Thygesen, N. (1994), Turning Sweden around. Cambridge, MA: M.I.T. Press.

Little, I.M.D. (1982), Economic development. New York: Basic Books.

Lu, H.C. (1993) ‘The structure of wages in Taiwan: the roles of female labor force
participation and international competition’, Ph.D. dissertation, University of
Chicago.

Lucas, R. (1988) ‘On the mechanics of economic development’, Journal of Monetary

Economics 22:3–42.

Mankiw, N.G. (1995), “The growth of nations,” Brookings Papers on Economic Activity,
G. Perry and W. Brainard ed., vol 1, 275-326.

Mankiw, N.G., Romer, D., and Weil, D.M. (1992) ‘A contribution to the empirics of
Quarterly Journal of Economics 107:407–437.

Murphy, K.M. and Welch, F. (1991) ‘The role of international trade in wage differentials’,
in: Marvin Kosters, ed., Workers and their wages. Washington, D.C.: AEI Press.

Nelson, R. and Phelps, E. (1966) ‘Investment in humans, technological diffusion, and

American Economic Review 56:69–75.

Pritchett, L. (1997) ‘Where has all the education gone’, Policy Research Working Paper,
Washington, D.C.: World Bank, June.

Psacharopolous, G. (1994) ‘Returns to investment in education: a global update’, World

Development, 22:1325–43.

Rebelo, S. (1991) ‘Long-run policy analysis and long-run growth’, Journal of Political
Economy, 99:500–521.

Robbins, D. (1996) “HOS meets facts, facts win: evidence on trade and wages in the
developing world,” Working Paper, Harvard Institute for International
Development, Cambridge.

Romer, P. (1990) ‘Endogenous technological change’, Journal of Political Economy,

98:S71–S102.

Romer, P. (1986) ‘Increasing returns and long-run growth’, Journal of Political

Economy, 94:1002–1037.

Schultz, T.W. (1960), “Capital formation by education,” Journal of Political Economy,

68: 571–83.

Schultz, T.W. (1961), “Investment in human capital,” American Economic Review,

Smith, J. and Welch, F. (1986) ‘Closing the Gap: Forty Years of Economic Progress for
 Blacks’, Rand Publication Series R-3330-DOL, 1986.

Solow, R. (1956) ‘A contribution to the theory of economic growth’, Quarterly Journal

of Economics, 70:65–94.

Solow, R. (1957) ‘Technical change and the aggregate production function’, Review of
Economics and Statistics, 39:312–320.

Stokey, N. (1988) ‘Learning by doing and the introduction of new goods,” Journal of
Political Economy, 96, 4, 701-17.

Summers, R. and Heston, A. (1991) ‘The Penn world table (Mark 5): An expanded set of
international comparisons, 1950–1988’, Quarterly Journal of Economics,
106:327–368.

Topel, R. (1997) ‘Factor proportions and relative wages: the supply side determinants of
Journal of Economic Perspectives 11:55–74.

Tsiang, S.C. (1984) ‘Taiwan’s economic miracle: lessons in economic development’, in:
Arnold Harberger, ed., World economic growth. San Francisco: ICS Press.

Uzawa, H. (1965) ‘Optimal technical change in an aggregative model of economic

International economic Review, 6:18–31.

Welch, F. (1966), ‘Education in production’, Journal of Political Economy 78:35–59.

Young, A. (1991) ‘Learning by doing and the dynamic effects of international trade’,
Quarterly Journal of Economics, 106:369–406.

Young, A. (1994) ‘Lessons from the East Asian NIC’s: a contrarian view,” European
Economic Review, 38:964–73.

Young, A. (1992) ‘A tale of two cities: factor accumulation and technical change in Hong
Kong and Singapore’, in: Olivier J. Blanchard and Stanley Fischer, eds., NBER
Macroeconomics Annual. Cambridge, MA: MIT Press.

Young, A. (1995) ‘The tyranny of numbers: confronting the statistical realities of the East
Quarterly Journal of Economics, 110:641–80.
 Table 2

The Effects of Education on Productivity:

Fixed Country Effects, 1960-1990

(Dependent variable is log real output per worker, measured at 5-year intervals)

Rn yit (Rn yit ! .35Rn kit)/.65 (Rn yit ! .5Rn kit)/.5

______________________________ ___________________ ________________
(1) (2) (3) (4) (5) (6) (7) (8)

Average Years .226 .102 .085 .072

of Schooling (22.67) (6.21) (4.20) (2.89)

Average Years .203 .057 .052 .047

Primary Schooling (10.28) (2.05) (1.53) (1.12)

Average Years .276 .138 .111 .092

Secondary Schooling (7.62) (5.76) (3.77) (2.54)

Year Effects no no yes yes yes yes yes yes

R2 .46 .46 .58 .59 .38 .38 .24 .24

N 719 719 719 719 664 664 664 664

__________________________________________________________________________________________________________

Note.--Data are from Summers and Heston (1991) and Barro and Lee (1993). Data on capital stocks were provided by Pete Klenow.
Absolute t-ratios are in paretheses. All models contain 111 country effects.
 Table 3

Fixed Effects Estimates of the Impact of Education on Productivity:

Controlling for Average Age and Life Expectancy

(1) (2) (3) (4)

Average Years .142 .158 .100 .062

of Schooling (9.15) (7.15) (5.65) (2.41)

Life Expectancy .027 .029 .005 .005

(6.68) (4.44) (0.98) (0.72)

Age -- !.006 -- .007

(0.51) (0.60)

Year Effects no no yes yes

R2 .489 .55 .58 .65

N 669 324 669 324

______________________________________________________________________________

Notes.--See notes to Table 2. Age and life expectancy data are available from the U.S. Census
website at http://www.census.gov.
 Table 4

The Effects of Education on Productivity:

First-Differenced Estimates at Various Growth Intervals

5-year 10-year 15-year $ 20-year

Growth Growth Growth Growth
_______________ ______________ _______________ _______________
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)

) Years of Schooling .028 .041 .058 .064 .085 .115 .120 .148 .155 .167 .252 .246
(2.02) (2.95) (3.70) (3.15) (4.26) (5.07) (4.01) (5.07) (5.23) (3.66) (6.10) (5.73)

Initial Years of Schooling .004 .004 .004 .003 .004 .003 .004 .004
(5.71) (5.57) (5.02) (4.85) (4.84) (4.59) (6.37) (5.93)

Log Inital Output per !.007 !.005 !.007 !.004 !.008 !.005 !.010 !.009
Worker: Rn(Y/L) (4.06) (2.56) (3.54) (1.56) (3.68) (1.77) (4.86) (2.26)

) Schooling × Rn(Y/L) !.036 !.060 !.041 !.025

(2.28) (2.70) (1.30) (0.57)

Year Effects yes yes yes yes yes yes yes yes yes yes

R2 .176 .218 .224 .256 .315 .332 .308 .386 .391 .133 .397 .399

N 608 608 608 290 290 290 186 186 186 101 101 101
____________________________________________________________________________________________________________

Notes.--See notes to Table 2.

 Table 5

The Effects of Education on Aggregate Productivity:

First-Difference Estimator with Country Effects

Five-Year Ten-Year
Growth Growth
(1) (2) (3) (4)

) Years of .013 .022 .058 .086

Schooling (0.87) (1.32) (2.15) (2.85)

Initial Schooling .004 .004 .009 .009

(1.21) (1.29) (2.35) (2.49)

log initial output !.044 !.043 !.050 !.047

per worker: Rn(Y/L) (6.20) (6.02) (6.45) (6.03)

) Schooling × Rn(Y/L) -- !.020 -- !.049

(1.25) (2.00)

Fixed Country Effects yes yes yes yes

R2 .285 .287 .481 .493

Observations 604 604 290 290

__________________________________________________________________

Note.--See notes to Table 2.

 Table 6

The Effects of Relative Supply of Educated Labor on Relative Wages

Sweden South Korea Taiwan U.S. Canada

College/ !.350 !.196 !.05 !.59 !.53
High School (.048) (.336) (.008) (.12) (.38)

High School/ !.784 !.09

Elementary School (.091 (.015)

Note: Estimates are coefficient from regressions of the form ln(wi/wj) = $0 + $1 ln(Li/Lj) + $2X +
u, where Li is the labor supply of individuals from skill group I.

Sources: Sweden: Edin and Holmlund (1995); South Korea: Kim and Topel (1995); Taiwan: Lu
(1993); U.S. and Canada: Freeman and Needels (1993).