PeterJ.

Klenowand AndresRodriguez-Clare
GRADUATE SCHOOL OF BUSINESS, UNIVERSITYOF CHICAGO

The

Neoclassical

in

Growth

Too

Far?

Revival

Economics:

Has

It

Gone

1. Introduction
Theories endogenizing a country's technology, such as Romer (1990) and
Grossman and Helpman (1991), arose from the desire to explain the
enormous disparity of levels and growth rates of per capita output across
countries. The belief was that differences in physical and human capital
intensity were not up to the quantitative task. This belief has been
shaken by a series of recent empirical studies. Mankiw, Romer, and Weil
(1992) estimate that the Solow model augmented to include human capital can explain 78% of the cross-country variance of output per capita in
1985. Alwyn Young (1994, 1995) finds that the East Asian growth miracles were fueled more by growth in labor and capital than by rising
productivity. And Barro and Sala-i-Martin (1995) show that the augmented Solow model is consistent with the speed of convergence they
estimate across countries as well as across regions within the United
States, Japan, and a number of European countries.1
In our view these studies constitute a neoclassicalrevival.2They suggest
We are gratefulto Ben Bemanke, MarkBils, V. V. Chari, Chad Jones, Greg Mankiw,Ed
Prescott,David Romer,JulioRotemberg,JimSchmitz,Nancy Stokey,and Alwyn Youngfor
helpful comments.
1. Mankiw, Romer, and Weil, and Barroand Sala-i-Martindo not explain the source of
countrydifferencesin investmentrates. Chari,Kehoe, and McGrattan(1996)arguethat
distortionssuch as tax rates, bribes, risk of expropriation,and corruptioncontributeto
an effective tax rate which, if it varies in the right (stochastic)way acrosscountries,can
explain the levels and growth rates of income observedin the Summers-Heston(1991)
panel.
2. We are indebted to Alwyn Youngfor this phrase.

74 *KLENOW& RODRIGUEZ-CLARE
that the level and growth rate of productivity is roughly the same across
countries, so that differences in output levels and growth rates are
largely due to differences in physical and human capital. Romer (1993),
in contrast, argues that "idea gaps" are much more important than "object gaps." In terms of a simplified production function Y = AX, where A
is productivity and X encompasses physical and human capital, this
debate is over the relative importance of A and X.
This debate matters because the positive and normative implications of
the A view can differ dramatically from those of the X view. Technologybased models of A exhibit scale effects because of the nonrival nature of
technology creation and adoption. And they suggest that openness, perhaps though its effect on technology diffusion, can have first-order effects on living standards and growth rates (without requiring big differences in rates of return to capital). These implications of openness could
be positive in all countries, as in Rodriguez-Clare (1997), or positive in
some and negative in others, as in Stokey (1991) and Young (1991). These
implications are not shared by the basic neoclassical growth model,
which has the same technology everywhere.
In this paper we offer new evidence relevant to this debate on the
importance of productivity vs. physical and human capital in explaining
international differences in levels and growth rates of output. In Section
2 we reexamine Mankiw, Romer, and Weil's (hereafter MRW) methodology for estimating human capital. We update their data and add data on
primary and tertiary schooling which have become available since their
study. Because primary school attainment varies much less across countries than secondary school attainment does, the resulting estimates of
human capital vary much less across countries than the MRW estimates.
We also incorporate evidence suggesting that the production of human
capital is more labor-intensive and less physical capital-intensive than is
the production of other goods. This further narrows country differences
in estimated human capital stocks.
In Section 3 we incorporate evidence that pins down the human capital intensity of production and the relative importance of primary vs.
secondary schooling. We exploit information contained in Mincer regressions, commonly run in the labor literature, on the amount of human
capital gained from each year of schooling. For a cross section of workers, Mincer (1974) ran a regression of worker log wages on worker years
of schooling and experience. Such regressions have since been run for
many countries (see Bils and Klenow, 1996, for citations for 48 countries).
We combine this evidence with data on schooling attainment and estimates of school quality to produce measures of human capital for 98
countries. Using these Mincer-based estimates of human capital, we find

The NeoclassicalRevival in GrowthEconomics* 75

that productivity differences account for half or more of level differences
in 1985 GDP per worker.3
In Section 4 we carry out the same analysis as in Section 3, but here the
objective is to produce 1960-1985 growth rates rather than 1985 levels.
We find that differences in productivity growth explain the overwhelming majority of growth rate differences. These results seem at odds with
Young (1994, 1995), so in Section 5 we compare and contrast our findings
for 98 countries with his careful findings for four East Asian countries.
Hall and Jones (1996) also follow Bils and Klenow (1996) in using
Mincer regression evidence to construct human capital stocks. As we do,
they find that productivity differences are important in explaining international income variation.4 Their main objective is different from ours,
however, in that they are interested in finding correlates (such as language and climate) with productivity differences. In contrast, we focus
on examining how human capital should be measured and how international productivity differences depend on how human capital is measured. We want to know, for instance, whether our measure of human
capital is more appropriate than the one used by MRW, and we want to
know how adding experience and correcting for schooling quality affects
the results on productivity. Our paper also differs from Hall and Jones
(1996) in that we study growth-rate differences as well as level differences, whereas they concentrate on level differences.5
3. OLS Mincer-equation estimates of the wage gain from each additional year of schooling
might be too generous because of oft-cited ability bias (more able people acquire more
education). In the NLSY over 1979-1993, we ran a Mincer regression of log (deflated
wage) on schooling, experience, and experience squared and found a schooling coefficient of 9.3% (s.e. 0.1%). When we included the AFQT score as a proxy for ability, the
estimated wage gain from each year of schooling fell to 6.8% (s.e. 0.2%). We stick with
the standard estimates such as 9.3% for two reasons: first, since we will find that the role
of human capital is smaller than MRW found, we prefer to err on the side of overstating
variation in human capital across countries; second, the AFQT score could be a function
of human capital investments in the home, and the overstated return may crudely
capture how these investments tend to be higher when attainment is higher.
4. Hall and Jones reach quantitatively similar conclusions to ours because of two offsetting
differences. First, as we will describe in Section 2 below, we take into account the natural
effect of higher TFP on the capital-labor ratio (which increases to keep the return on
capital at its steady-state equilibrium level) and therefore attribute the whole effect (i.e.,
higher TFP plus resulting higher capital-labor ratio) to higher productivity. Hall and
Jones attribute only the direct effect to TFP, ignoring the indirect effect on the capitallabor ratio. Second, we estimate country differences in the quality of schooling that
reinforce differences in the quantity of schooling attainment across countries. Hall and
Jones look only at the quantity of schooling.
5. Bosworth, Collins, and Chen (1995) also estimate TFP growth rates with human capital
stocks constructed using a methodology that is at some points close to ours. Instead of
exploring the importance of differences in TFP growth rates in explaining international
growth variation, these authors use such estimates of TFP growth to run cross-country
TFP growth regressions.

76 *KLENOW& RODRiGUEZ-CLARE

2. Reexamining
MRW
In this section we first describe and comment on MRW's methodology
for attributing differences in output to differences in productivity vs.
differences in capital intensity. We then update MRW's estimates and
make a series of modifications, such as incorporating primary school data
and a more labor-intensive technology for producing human capital.
MRW specify the production technology
Y = C + IK + IH = KaHI(AL)'- a-,

(1)

where Y is output, K and H are stocks of physical and human capital, A

is a productivity index, and L is the number of workers. H = hL, where h
is human capital per worker. Implicit is an infinite-lived representative
agent whose time enters production through dual components, human
capital H and raw labor L.6 As shown in (1), MRW specify the same
technology for producing human and physical capital. Time subscripts
are suppressed, as are the standard accumulation equations for K and H.
MRW assume that both stocks depreciate geometrically at a rate of 3%
per year.
If one adds competitive output and input markets and constant relative risk aversion utility, then it is well known that higher A will induce
proportionate increases in K and H. Given this fact, rather than the usual
accounting exercise that assigns output or its growth rate to contributions from K, H, L, and A, we think MRW rightly rearrange (1) to yield
Y

L =A

/ K a/(ia-)

H\/(1-a-)
Y.

= AX,

(2)

where X is a composite of the two capital intensities. We concur with

MRW's adoption of (2) for two reasons. First, Y/L is the object of interest
rather than Y, since we want to understand why output per worker
varies across countries, leaving aside how a country's number of workers L is determined.7 The A-vs.-X debate really has nothing to do with
the determination of L. Second, (2) gives A "credit" for variations in K
and H generated by differences in A. The contributions of K and H
6. Formally,the efficiency units contributedby a worker with human capital h are e =
h(l -a)A(-a-)/(1-a),

and the production function is Y = KEl-", where E = eL.

7. Parente,Rogerson,and Wright(1996)arguethat hours workedin themarket(as opposed

to home production)by the average worker varies a lot across countries, making the
number of workers a poor measure of marketlaborinput. If, as these authors argue,
markethours per worker are much higher in richercountries, it should contributeto
higher A in richercountries.

The NeoclassicalRevival in GrowthEconomics* 77

variations that are not induced by A are captured by variations in capital

intensity X. This decomposition was also adopted by King and Levine
(1994), albeit for a setting with physical capital but not human capital.
We offer two caveats to the decomposition in (2), both related to A
being endogenous, say resulting from technology adoption decisions.8
First, one would expect that many country policies affect both A and X.
Weak enforcement of property rights in a country, for example, is likely
to decrease both A and X. We think the decomposition in (2) is still
useful, however, because there are some policies that could affect one
factor much more than the other (e.g., education policies). Thus, finding
that high levels of output per worker are explained mostly by high levels
of H/Y would suggest that differences in education policies are an important element in explaining international differences in output per
worker. Similarly, finding that differences in K/Y are important in accounting for the international variation in output per worker would
point towards capital taxation or policies that affect the relative price of
investment goods.
The second related caveat concerning the decomposition in (2) is that,
just as K and H are affected by A, A itself may be affected by capital
intensity X. If A is determined by technology adoption, for instance,
8. We do not list the embodimentof technologyin physical capitalas problematicfor the
decompositionin (2). Suppose that productivityentirelyreflectsthe qualityof physical
capital,and that all countriesinvest in the highest qualitycapitalgoods availablein the
world in the period of investment. Then differencesin countryproductivitylevels could
be due to differencesin the vintage or age of a country'scapitalstock, i.e. unmeasured
differencesin the qualityof a country'scapitalstock.In this situation,one mightthink(2)
would attributeto productivitydifferenceswhat in realityshould be attributedto differences in physicalcapitalintensity,say becausecountrieswith high investmentratesand
high capitalintensityareusing youngerand thereforebetterequipment.If so, ourresults
would understate the role of capitalintensity in explaininginternationaloutput differences. This concernturns out to be unfounded along a steady-stategrowthpath. Tosee
why, suppose the true,quality-adjustedcapitalstockevolves accordingto AKt= BtIt- 8Kt,
where A is the first-differenceoperatorand B is an exogenouscapital-embodiedtechnology index which grows at the constant rate g and (recall)is the same for all countries.
Imagine also that the measuredcapital stock evolves according to AKMt= It - 8KMtso that it

does not reflectimprovementsin qualitycomingfromembodiedtechnology.In this case

one can show that if Y = K"L1l-,then along a steady-stategrowthpath with constantIIY

and KM/Yone has Y/Lt = (cBt)'(1-a)(KM/Y)a'/-a)

with c a constant which depends on g, a,
and 8, and with KM/Y= (IIY)/(g+n+8), where n is the exogenous growth rate of L. Thus,

along a steady-stategrowthpath, a country'sTFPis independentof its investmentratein

physicalcapital.Thatis, the investmentratedoes not affectthe TFPresidual,which in this
case is equal to ln(Y/Lt) - a/(1 - a) In(K/Y), which in turn is equal to [1/(1 - a)] ln(cB,).

The intuitionfor this resultis thata higherinvestmentratereducesthe averageage of the

capitalonly temporarily,along the transitionpath. When the new steady-statepath is
reachedwith higher capitalintensity,the age distributionof the capitalstock-which is
synonymous with the qualitydistribution-is the same as the distributionwith a lower
capitalintensity.Thus a countrywith a permanentlyhigherIIYthananothercountrywill
have no younger or bettercapitaland thereforeno higherTFP.

78 KLENOW& RODRIGUEZ-CLARE
then it is likely that higher schooling (i.e., higher H/Y) leads to a higher
level of A. Once we obtain estimates of A and X, below we will actually
use the decomposition of equation (2) to study this issue by looking at
the correlation between A and the capital-intensity variables K/Y and H/
Y. A hypothetical example will illustrate the usefulness of this approach.
Imagine that, using this decomposition, we find that almost all of the
international variation in levels of Y/L is accounted for by international
differences in A. But imagine that we also find a strong positive correlation between A and H/Y. This would be consistent with-but not necessarily proof for-the view that human capital explains differences in Y/L,
albeit indirectly through its effect on A (say through economywide technology adoption as in the model of Ciccone, 1994). On the other hand,
finding no correlation between A and H/Y would suggest that differences in schooling are not important in explaining international output
differences.
With these preliminaries out of the way, we now proceed to updating
and modifying MRW's estimates. For Y/L MRW use the SummersHeston GDP per capita in 1985. For K/Y for each country they use
K
K
Y

IK/Y
IKY

g+

+n

(3)

where IK/Y is the average Summers-Heston investment rate in physical

capital over 1960-1985, g is 0.02 (an estimate of the world average
growth rate of YIL), 8 is 0.03 (a rather low depreciation rate, but none of
the results in their paper or ours are sensitive to using 0.06 instead), and
n is the country's average rate of growth of its working-age population
(15- to 64-year-olds) over 1960-1985 (UNESCO yearbook). Expression (3)
is derived as the constant (or steady-state) K/Y implied by the capital
accumulation equation given a constant IIYand constant growth rates of
Y/L and L.9 For H/Y MRW use the average 1960-1985 investment rate in
human capital divided by the same sum:
H
Y

'"/Y .
_- IHIY
g+

+ n

(4)

(4)

9. In Section 3 below we relaxthis assumptionthat in 1985KIYis at its steady-statevalue.

We make varying assumptions about 1960 K/Ylevels for each country, then use the
accumulationequation and data on IIYand Y over 1960-1985to calculatethe 1985 K/Y.
We find that the results vary little depending on the assumed 1960value of K/Y.As for
the H/Ylevels, in Section3 we use data on schoolingattainmentin 1960and 1985.

*79
TheNeoclassical
Revivalin GrowthEconomics
MRW use the average 1960-1985 ratio of secondary-school students to
the working-age population (UNESCO yearbook) as an estimate of the
average investment rate in human capital:
IH

= (secondary enrollment rate)

population 12-17

population 15-64
15-64
Lpopulation

Ignoring nonsecondary schooling (which we will find matters), the ratio

is LHIL,the fraction of worker time spent in the human-capital sector.'0
Since in (1) the production technology for human capital is the same as
that for other goods, equating the ratios of marginal products of labor
and physical capital across sectors entails LH/L= KH/K,whereL = LH + Ly
and K = KH+ Ky.l
With their 1985 levels of Y/L, K/Y, and H/Y for 98 countries, MRW
regress ln(Y/L) on ln(K/Y) and ln(H/Y). They obtain an R2 of 0.78, and
their estimated coefficients are consistent with production elasticities of
a = 0.30 for physical capital and /3 = 0.28 for human capital (their Table
II, restricted regression). The high R2 is the basis of Mankiw's (1995, p.
295) conclusion that "Put simply, most international differences in living
standards can be explained by differences in accumulation of both human and physical capital."
Even assuming H/Y is measured properly, we are deeply uncomfortable with estimating a and 83from an OLS regression of ln(Y/L) on ln(K/
Y) and ln(H/Y). Consistency of such estimates requires that ln(X) be
orthogonal to ln(A). Yet countries with policies discouraging capital accumulation may also tend to have policies discouraging activites (such as
technology adoption) which contribute to higher A. In Rodriguez-Clare
(1997), one of us develops a quality ladder model wherein higher tariffs
on imported capital goods result in both lower X (by reducing the investment rate) and lower A (by increasing the average distance between the
quality of goods imported and the highest quality of goods available in
the world).
Given the possibility that true X and true A are correlated, our preference is to use independent evidence to determine appropriate values of
a and /, and then use them to construct X and A. This being said, a =
0.30 is actually in the ballpark of estimates obtained using national income accounts (see Gollin, 1996, for evidence on 31 countries with care10. The implicit infinite-lived representative agent is simultaneously teacher (H) and stu-

dent (L)!

11. HHH = LHL,because the representative worker's human capital and labor components
are supplied jointly.

80 *KLENOW& RODRIGUEZ-CLARE
ful treatment of proprietors' income). But we have no cause for comfort
with p = 0.28. Studies such as Jorgenson (1995) and Young (1995) look at
compensation of workers in different education and experience categories, thereby bypassing the need to choose a single share going to human capital. In other words, the share going to labor input is 1 - a, and
workers with more education and experience receive larger subshares of
this 1 - a. We will do something similar in Section 3 below by looking at
Mincerian estimates of wage differences across workers with different
education and experience levels. For the rest of this section we keep f =
0.28, but we discuss at the end how the results are affected by considering higher values.
We now show how the MRW results are sensitive to several modifications that we deem necessary. We keep a = 0.30 and P = 0.28 for comparison purposes. Given that our modified estimates of X and A will be
correlated, there will not be a unique decomposition of the variance of
ln(Y/L) into the variance of ln(X) and the variance of ln(A).12 We think an
informative way of characterizing the data is to split the covariance term,
giving half to ln(X) and half to ln(A). This means we decompose the
variance of ln(Y/L) as follows:
var ln(Y/L)
var ln(Y/L)

cov(ln(Y/L),ln(Y/L))
var ln(Y/L)

cov(ln(Y/L),ln(X)) + cov(ln(Y/L),ln(A))
var ln(Y/L),

or
1=

_cov(ln(Y/L),ln(X))

var ln(Y/L)

cov(ln(Y/L),ln(A))
var ln(Y/L)

This decomposition is equivalent to looking at the coefficients from independently regressing ln(X) and ln(A), respectively, on ln(Y/L). Since
ln(X) + ln(A) = ln(Y/L) and OLS is a linear operator, the coefficients sum
to one. So our decomposition amounts to asking, "When we see 1%
higher Y/Lin one country relative to the mean of 98 countries, how much
higher is our conditional expectation of X and how much higher is our
conditional expectation of A?" The first row of Table 1 gives the answer
for MRWO,the original MRW measure. Since the covariance term is zero
by construction for MRWO, the breakdown is precisely their 78% ln(X)
and 22% ln(A).
12. Since MRW construct ln(X) by regressing In(Y/L) on ln(H/Y) and ln(K/Y) with ln(A) as
the residual, their ln(X) and ln(A) are orthogonal by construction, and the unique
variance decomposition is R2, 1 - R2.

The NeoclassicalRevival in GrowthEconomics* 81

Table 1 THE ROLES OF A AND X IN 1985 PROSPERITYa
cov[ln(Y/L), In (Z)]/var ln(Y/L)
Sourcea
MRWO
MRW1
MRW2
MRW3
MRW4

Z= (

Z= X

Z=

.29
.27
.31
.29
.29

.49
.49
.47
.11
.04

.78
.76
.78
.40
.33

Z=A
.22
.24
.22
.60
.67

aMRWO:from MRW(uses their data appendix).MRW1:MRW0 but with Ky/Yinstead of K/Y.MRW2:

MRW1but with L = workerinsteadof working-agepopulation,14 countriesin/out. MRW3:MRW2but

with all enrollmentratherthan just secondaryenrollment.MRW4:MRW3but with (K,H, L) sharesof
(0.1, 0.4, 0.5), not (0.20, 0.28, 0.42), in H production.

is to recognize that,
Our first modification of MRW's methodology
measures
of output do not
to
national
income
(1),
contrary
accounting
include the value of student time-an
important component of human
investment.13
To
see
how
capital
important this might be, we consider
the extreme case in which none of the human capital investment is
measured as part of total output. To do this we replace K/Y and H/Y in
equation (2) with Ky/Y and Hy/Y, since only Ky and Hy are used in the
production of Y when Y does not include human capital investment. It
turns out that the MRW measure of IH/Y,namely LHIL,is also appropriate
for Hy/Y when all human capital investment goes unmeasured. The same
is not true for physical capital intensity, for which we must use Ky/Y =
(KIY)(Ly/L). As shown by the MRW1 row of Table 1, this modification
results in a 76% ln(X) vs. 24% ln(A) breakdown, so this distinction does
not appear to be quantitatively important.
The MRW2 row of Table 1 reproduces the MRW1 row, only with updated data and a set of countries for which we have all the necessary
schooling attainment data for the remainder of this paper. Like MRW,
13. MRWcontend that this slippage between model and data is not quantitativelyimportant. Parenteand Prescott(1996)disagree,contendingthat unmeasuredhuman capital
investment must be implausiblylarge for the combined share of capital to be about
two-thirds. Parente, Rogerson, and Wright(1996)illustratethat unmeasuredinvestment would have to be 25-76%of GDP.We are in closeragreementwith MRW,since,
accordingto Kendrick(1976),about half of schoolinginvestmentconsists of education
expenditures(teachers,facilities)which areincludedin measuredoutput. Accordingto
the 1996 Digest of Education Statistics published by the U.S. Department of Education

(1996), education expenditures averaged 7% of GDP over 1960-1990. Back-of-theenvelope calculationssuggest unmeasuredinvestmentmight thereforebe only 13%of
GDP.

82 * KLENOW & RODRiGUEZ-CLARE

we have a sample of 98 countries.14 We use the latest Summers-Heston
data (Mark 5.6) and use output per worker, whereas MRW used output
per capita. The measure we use for Hy/Y here is the same as that used by
MRW, except we use Barro and Lee's (1993) data on secondary enrollment rates in 1960, 1965, . . . , 1985 and United Nations (1994) population data by age groups to compute the value of
-= secondary
Y

enrollment

rate

population
population

15-19
15-64 J

for each country's investment rate in human capital in 1960, 1965, . ..

1985. MRW used population aged 12-17 in the numerator of the fraction
in brackets, due to data availability. As shown in the MRW2 row of Table
1, if we see 1% higher Y/L, we expect 0.78% higher X and 0.22% higher
A. Since the results are very similar to MRW1, we can now incorporate
data on primary schooling enrollment, etc., for our sample of 98 countries without fear that the change in sample obscures the comparison.
Using data from Barro and Lee (1993), the MRW3 row of Table 1 uses
Hy/Y calculated with all three enrollment rates. The results are striking.
Conditional on 1% higher Y/L in a country, we now expect only 0.40%
higher X and fully 0.60% higher A. As suggested by these results, primary enrollment rates do not vary as much across countries as secondary enrollment rates do. The MRW3 measure of ln(Hy/Y) has only about
one-fourth the variance of the MRW0 measure of ln(Hy/Y).15 Moreover,
the correlation of the MRW3 measure of ln(Hy/Y) with ln(Y/L) is only .52,
as opposed to .84 for the MRW0 measure. This is not to say that primaryare unproductive compared to other schooling
schooling investments
our
for
investments,
methodology assumes that they are productive.16 It
14. In our sample but not in MRW's:Gambia,Guinea-Bissau,Lesotho, Swaziland,Barbados, Guyana,Iran,Iraq,Taiwan,Cyprus,Iceland,Malta,Yugoslavia,and Fiji.In MRW's
sample but not in ours: Angola, BurkinaFaso, Burundi,Chad, Egypt, Ethiopia,Ivory
Coast, Madagascar,Mauritania,Morocco,Nigeria, SierraLeone, Somalia,and Sudan.
15. One might think that adding primaryschooling enrollmentrates to secondaryenrollment rates will not lower the variance, since adding a constant does not affect the
variance,
variance of a random variable. But since we are looking at the percentage
adding the relativelystable primaryschool enrollmentindeed lowers the variance.
16. In the next section we discuss Mincer regression evidence consistent with primary
schoolingindeed being productive.Specifically,each additionalyearof primaryschooling in poor countries is associatedwith roughly 10%higher wages, suggesting important human capital investment is going on in primary schools that should not be
ignored. There remainsthe issue of whether a year of enrollmentin secondaryschool
involves more investment in human capitalthan a year in primaryschool, so that the
two enrollment rates should not simply be added together as we have done in the
current section. The Mincer evidence that each additional year of schooling raises
wages about 10%suggests that more absoluteinvestmentin human capitalis occurring
in secondaryschool than in primaryschool.

*83
TheNeoclassical
Revivalin GrowthEconomics
says, rather, that primary schooling does not vary anywhere near as
much with Y/L across countries as secondary schooling does. By focusing only on secondary schooling, one overstates the percentage variation
in human capital across countries and its covariance with output per
worker.
A further objection we have to the MRW measure of the human capital
stock is that, as shown in (1), its construction assumes the same technology for producing human capital as for producing consumption and
physical capital. Kendrick (1976) presents evidence that the technology
for producing human capital is more intensive in labor than is the technology for producing other goods. He estimates that about 50% of investment in human capital in the-United States represents the opportunity
cost of student time. The remaining 50% is composed of expenditures on
teachers (human capital) and facilities (physical capital). According to
the 1996 Digest of EducationStatistics, expenditures on teachers represent
about 80% of all expenditures. These figures suggest factor shares of
10%, 40%, and 50% for physical capital, human capital, and raw labor in
the production of human capital, as opposed to the 30%, 28%, and 42%
shares MRW use for the production of consumption goods and new
physical capital. If we let
IH = K --H

(ALH) ,

(5)

this evidence suggests 4) = 0.4 and A = 0.5. Combining (2), (4), and (5)
yields17
Hy

L/ L

n+g+8

\1/[1-K+A,I/(1-a-f3)]

/K)[1-4-

(1-fP)/(1-a-3)]/[1-

+A3/(1-a-)()]

Y~L^-JI~~~~~

~(6) U~

When the two sectors have the same factor intensity (4)= 3 and A = 1 a - 3), this reduces to Hy/Y = (LH/L)(n+g+8), which MRW used and we
used above. But with ) = 0.40 and A = 0.50 the powers in (6) are 1.07 on
the first fraction and -0.28 on the second. Because human capital production is more human capital-intensive than is the production of Y
(0)>,3), a large share of labor devoted to human capital accumulation has
a more than proportionate effect on Hy/Y. And because human capital
production is less physical capital-intensive than is the production of Y
(1 - 4 - A<a), a high rate of investment in physical capital raises Y more
17. Hy/Y = (Ly/L)(HIY) = (Ly/L)(IHIY)/(n+g+8) = [(Ly/L)/(n+g+8) (KH/Y)1'--A(HH/Y)t(ALH/
Y). The expression in the text can be obtained by substituting for A using (2) with Hy/Y
and Ky/Y and by using KHIY= (LH/Ly)(Ky/Y) and HH/Y = (LH/Ly)(Hy/Y) (ignoring
multiplicative constants).

84 *KLENOW& RODRiGUEZ-CLARE
than H, thereby reducing Hy/Y. As shown in the MRW4 row of Table 1,
using ) = 0.40 and A = 0.50 results in a split of 33% ln(X) vs. 67% ln(A).
Comparing MRW4 with MRW3, we see that lowering the capital intensity of human capital production modestly lowers the variation of Hy/Y
across countries.
As shown by comparing MRW0 with MRW4 in Table 1, the cumulative effect of these modifications is to remove the linchpin of the neoclassical revival: MRW's original (78%, 22%) decomposition has given way
to a (33%, 67%) decomposition. Can one restore MRW's results with a
higher 3?Doubling 3 from 0.28 to 0.56 yields a (51%, 49%) division. As
,3 rises toward 23, the decomposition approaches 60% vs. 40%. Thus a
sufficiently high 3 does generate results that, although not as dramatic
as those of MRW, still have the major part of international income
variation explained by differences in levels of physical and human capital per worker. But what is the right value for 3? Unfortunately, we
know of no independent estimates of "the" share of human capital.
Fortunately, in the next section we are able to exploit wage regressions
to measure human-capital stocks in a way that does not depend on the
value of 3. This regression evidence also appropriately weights primary
schooling attainment relative to secondary schooling attainment, rather
than lumping them together with equal weight as we have done in the
preceding.

3. UsingMincerRegressionEvidenceto EstimateHuman
CapitalStocks
In this section we exploit evidence from the labor literature on the wage
gains associated with more schooling and experience. For a cross section
of workers, Mincer (1974) ran a regression of worker log wages on
worker years of schooling and experience. He chose this specification
because it fit the data much better than, say, a regression of the level of
wages on the years of schooling and experience. To incorporate this
evidence into the technology for producing human capital, we abandon
the infinite-life construct in favor of a life cycle in which people first go to
school full time and then work full time. We specify the following technology for human capital:
h = (KH/LH)1 A(hT)(AeY/A)s)A,

(7)

where hs is the human capital of somebody with s years of schooling, KH

is the capital stock used in the education sector, LH is the number of

TheNeoclassical
Revivalin GrowthEconomics
?85
students, and
leads to18
Hy

(/

\/[1-,+

hT is

the human capital of each teacher. Manipulating (7)

A,3/(l-a-3)]

(Ky)[l-,-A(l-3)/(1-a-3a)]/[-1

+ A3/(1-a-f3)]

--^'M~~~~~(7)

<(8)

Bils and Klenow (1996) look at Mincer regression studies covering 48

countries and find that the wage gain associated with an additional year
of education averages 9.5% across the 48 countries and ranges from 5%
to 15% for 36 of the 48 countries. Based on technologies (1) and (7), the
percentage wage gain to a representative agent from one more year of
schooling is /8y/(1-a). Therefore, to match an estimated wage gain of
9.5% we set y = 0.095(1-a)/,3.
Table 2 presents results based on (7). The rows are labeled BKn because
(7) is from Bils and Klenow (1996). As with MRW4 above, we use a = 0.30,
f3 = 0.28, 4 = 0.4, and A = 0.5. For years of schooling s, row BK1 uses the
level implied by the enrollment rates used in MRW3 and MRW4: s = 8 ?
primary + 4 *secondary + 4 *tertiary. As the BK1 row shows, conditional
on 1% higher Y/L we expect 0.60% higher X and 0.40% higher A. So
switching from (6) used for MRW4 to (8) used for BK1 dramatically shifts
the breakdown from (33%, 67%) to (60%, 40%). The exponential form of
(7) implies that the higher the level of schooling, the bigger is the absolute
amount of human capital obtained from the next year of schooling. The
exponential form therefore puts more weight on secondary school enrollment than on primary school enrollment, moving us back toward MRW's
78%-vs.-22% breakdown.
One concern we have about BK1, as well as all of Table 1, is the
assumption that in 1985 K/Y and H/Y are at steady-state levels. The data
show lots of movement in country growth rates of Y, L, and YILand in
country investment rates in physical and human capital, suggesting that
country K/Y and H/Y levels change over time. To estimate an off-steadystate 1985 K/Y, we use the accumulation equation and data on IIY and Y
18. In steady state h, = hT = h, so that the human capital of each student entering the
workforce is the same as that of each teacher or worker. Using this fact, h = HHILH,so
that H = hL = L(KH/ILH)1-~-A(HH/LH)(Ae(/)s)^, where HHis the total human capital of
teachers. Expression (8) can then be obtained much as expression (6) was above. There
are two (offsetting?) shortcomings in our treatment: First, we are assuming the
student-teacher ratio is the same in each country (we fix it at one, but the level does
not affect cross-country variance analysis). This ratio is presumably lower in richer
countries. Second, our setup assumes that teacher education varies as much across
countries as average worker education does (hT = h). In reality teacher schooling may
vary less than average worker schooling, say if in every country high-school teachers
must have at least a high-school diploma.

86 *KLENOW& RODRiGUEZ-CLARE
Table2 THEROLESOF A AND X IN 1985PROSPERITY
cov[ln(Y/L),In (Z)]/varln(Y/L)

Sourcea

Z
Z== X

Z =A

BK1

.29

.31

.60

.40

BK2
BK3
BK4

.23
.23
.23

.33
.31
.11

.56
.53
.34

.44
.47
.66

"BK1:uses (7), i.e. Mincerevidence. BK2:calculatesyears of schoolings from Barro-Lee1985stocks

insteadof 1960-1985flows. BK3:adds averageyearsof experience.BK4:BK3but with (K,H, L)sharesof
(0, 0, 1) instead of (0.1, 0.4, 0.5) in H production.

over 1960-1985. Unfortunately, direct estimates of the 1960 K/Y are not
available for most countries. We therefore set, for each country,
/K

IK_ IK/Y
J 196=g+8+n

with the investment rate IK/Y,the growth rate of YIL(g), and the population growth rate (n) equal to the country's averages over either 1960-1965,
1960-1970, or 1960-1985, and 8 either 0.03, 0.05, or 0.07. We also followed
a procedure akin to King and Levine (1994) where we set g in the denominator equal to a weighted average of own-country and world growth. The
results were not at all sensitive to which way we calculated the 1960 K/Y,
so we report the results with 1960 K/Ycalculated using 8 = 0.03 (as in Table
1) and the country's own averages over 1960-1970 for g and n. To construct the 1985 H/Y, we use Barro and Lee's (1993) data on average years of
schooling attained by the 25-64-year-old population in each country in
1985. We report the results of using this approach to obtain 1985 levels of
K/Y and H/Y in the BK2 row of Table 2. Conditional on 1% higher Y/L in
one country in 1985, we expect 0.56% higher X and 0.44% higher A in that
country. These results are not far from the (60%, 40%) breakdown in BK1
with the steady-state assumption for K/Y and H/Y.19
We now modify (7) to incorporate human capital acquired through
experience:
hs = (KH/LH) -- A(hT)4(Ae(ls+

Y2exp+y3exp2)/)

(9)

19. The close similarity between H/Y calculated in BK1 and in BK2, i.e. between school
attainment implied by enrollments and measures of years of schooling attained, suggests that differences in the duration of primary, secondary, and higher education from
our assumed 8, 4, and 4 years are not quantitatively meaningful.

The NeoclassicalRevival in GrowthEconomics* 87

where exp = (age - s - 6). The average experience level among workers
was estimated using United Nations (1994) data in combination with
Barro and Lee's schooling attainment data. For each country experience
was calculated as the population-weighted average of (age - s - 6) at
ages 27, 32, . . ., 62 for the groups 25-29,

30-34,

. ... , 60-64 in 1985.

Surprisingly, we find that the correlation between average years of experience of 25-64-year-olds and ln(Y/L) in 1985 is -0.67. Richer countries
have older workforces, but slightly less experienced ones because they
spend more years in school. As above, y, = 0.095(1-a)//3. Bils and
Klenow (1996) report average estimated coefficients on exp and exp2
across 48 countries of 0.0495 and -0.0007. Based on these, we set 72 =
0.0495(1 -a)//3 and y3 = 0.0007(1 -a)c//. The consequence of adding experience can be seen in the BK3 row of Table 2. As compared to the (56%,
44%) split in BK2, the split in BK3 is (53%, 47%).
Underlying this breakdown of 53% ln(X) vs. 47% ln(A) is the supposition that the quality of schooling is much higher in richer countries. In
richer countries, students enjoy better facilities (higher KH/LH)and better
teachers (higher HHILH). From (9), the quality of schooling is
-h OAI .
quality of schooling = (KH/L-H)1
Using this formula, for BK3 the elasticity of quality with respect to a
country's Y/L is 0.95%.20This means a country with 1% higher Y/L has
0.95% higher quality education. Note that higher quality of this type
does not raise the percentage wage premium from education, but instead
raises the base (log) wage for anyone in the country receiving some
education. It should affect the intercept of the Mincer regression for a
country, but not the coefficient on schooling.
Is an educational quality elasticity of 0.95% reasonable? Is it plausible
that, like GDP per worker, the quality of education varies by a factor of
about 34 across countries in 1985? An independent estimate of the quality elasticity can be gleaned from the wages of U.S. immigrants.21 Using
1970 and 1980 census data on the U.S. earnings of immigrants from 41
countries, Borjas (1987) estimates country-of-origin-specific intercepts in
a Mincer regression of log wages on immigrant years of education and
experience. He finds that immigrants with 1% higher per capita income
in their country of origin exhibit a 0.116% higher wage intercept (stan20. Calculatedas cov[ln(quality),ln(Y/L)]/var
ln(Y/L).
21. Incidentally,the enormous pressure for migrationfrom poor to rich countriesis itself
consistent with substantialdifferencesin productivityacrosscountries.However, this
pressure could be entirely explained by higher physical capital-output ratios and
greaternonpecuniarybenefits of living in richercountries.

88 *KLENOW& RODRiGUEZ-CLARE
dard error 0.025). This implies a quality elasticity of only 0.12%, suggesting that the elasticity embedded in BK3 is very aggressive.22'23
Borjas's evidence suggests an alternative, namely that teachers and
class facilities affect school quality through the schooling coefficient y.
In this event we would expect to see higher Mincer schooling coefficients in richer countries. Bils and Klenow (1996) find the opposite: each
additional year of schooling brings roughly 10% higher wages in a
country where the average worker has 5 years of schooling, compared
to only about 5% higher wages in a country where the average worker
has 10 years of schooling. Perhaps 's are higher in richer countries, but
the effect on the education premium is more than offset by a lower
relative marginal product of human capital in richer countries. This
could arise because of imperfect substitutability of workers with different education levels combined with abundance of human capital in
richer countries. Indeed, the Cobb-Douglas technology in (1) implies
unit elasticity of substitution between human capital and raw labor and
therefore a falling education premium with a country's H/Y, holding y
constant.
It is interesting to explore the possibility that the Mincer coefficients
already capture the effect of education quality combined with imperfect
substitutability. This would correspond to the extreme case when teachers and class buildings affect only the /s, so that 4 = 0 and A = 1. It
would be ideal to do this exercise using Mincer coefficients for each
country, but unfortunately we do not have such data for all countries.
Here we use the average Mincer coefficient of 9.5% instead. (The reader
should note that, since the Mincer coefficient is actually declining with
income per worker, this biases the results against a large role for A.) As
we report in the BK4 row of Table 2, without Mincer-intercept-type
variations in school quality, human capital contributes much less to Y/L
variation. The (ln(X), ln(A)) division shifts from (53%, 47%) in BK3 to
(34%, 66%) in BK4.
How do we choose between BK3 and BK4? Recall that the school
22. Immigrantsmay be more able than the average person in their country of origin.
Borjas'sregression controls for observabledifferencesin immigrants'ability such as
age, years of schooling, and English proficiency.With regard to unobservables,the
estimate of education qualitydifferenceswould be biased downward if positive selection (in Borjas'sterminology)were greaterthe poorerthe countryof origin. As Borjas
notes, the opposite may be true, since income inequalitytends to be greaterin poorer
countries.
23. Furthercause for concern is the difficulty researcherssuch as Hanushek (1986)and
Heckman, Layne-Farrar,and Todd (1996)have encountered in correlatingschooling
outcomes with teacherinputs. A (virtuallycontrolled)experimentin Tennessee, however, found that the group of students placedin smallerK-3 classes performedsignificantlybetter on standardizedtests (Mosteller,1995).

*89
Revivalin GrowthEconomics
TheNeoclassical
quality elasticity implied by BK3 is 0.95%, whereas BK4 implies no variation in school quality (of the intercept type) across countries. The evidence from Borjas (1987) suggests that BK3's elasticity is much higher
than the truth; the zero elasticity in BK4 seems closer. But if BK3 comes
from data on the shares of capital and teachers in the U.S. education
sector, why might it deliver wrong results? There are three possibilities.
First is the reason we just gave, namely that these inputs affect the /s.
Second, it could be that human capital varies much less across education
sectors than across other sectors; i.e., international differences in human
capital may be smaller for teachers than for other workers. Finally, it
could be that productivity A is not as significant in the education sector
as it is in other sectors. In the extreme case when A does not enter the
education sector at all, we find that the parameter values 4 = 0.19 and A
= 0.81 generate a quality elasticity matching Borjas's 0.12. In this case
we find a (42%, 58%) breakdown, in between BK3 and BK4 but a little
closer to the latter.
We conclude that richer countries tend to have higher K/Y,higher H/Y,
and higher A, with a dominant role for A, a large role for K/Y, and a
modest-to-large role for H/Y. To us this says that theorizing about international output differences should center at least as much on differences in
productivity as on differences in physical or human capital intensity.
Figures 1 and 2 display ln(A) and ln(X) using MRWOand BK4, respectively.24 In the MRWOworld, research should focus on explaining why
ln(X) varies so much; in the BK4 world a greater priority is to understand
differences in ln(A).
In contrasting MRWOand BK4, it is instructive to look at the correlations among output per worker (Y/L), capital intensity (K/Y and H/Y),
and productivity (A). Table 3a shows the correlation matrix for the case
where HIY is measured according to MRWO;Table 3b does the same for
the case where H/Y is measured according to BK4. In both cases YIL,K/Y,
and H/Y are highly positively correlated with each other. The difference
arises in the correlation between these variables and A. In particular,
Table 3a shows no correlation between A and K/Y or between A and H/Y
(both by MRW construction), whereas the corresponding correlations
are quite positive in Table 3b. In thinking about a theory of endogenous
A, it is hard to imagine that policies discouraging K/Y and H/Y-such as
high tax rates-would not also discourge A. The positive BK4 correlations seem much easier to generate theoretically.
A possible reason for the 0.57 correlation between ln(A) and ln(H/Y) is
24. These figures display visually what we try to convey by splitting the covariance terms
in Tables 1 and 2.

90 *KLENOW& RODRiGUEZ-CLARE
Figure 1 1985LEVELS:MRWO
2.0
1.5 -

1.0 -

0.5

* t

* *

0.0 -

o
*0

* *~
v

-0.5

* *
* *

-1.0
-1.5
-2.0
-2.5
-2.5

....I

-2.0

-1.5

-1.0

-0.5

0.0

0.5

1.0

1.5

2.0

In(X)
that high H/Y, say due to generous education subsidies, facilitates technology adoption. Ciccone (1994) presents a model with this feature: the
larger the economy's stock of human capital, the more profitable it is for
a firm to spend the fixed costs of adopting a given technology, and
therefore the higher the economy's A relative to the world frontier. Note
that this story links an economy's A to an economy's HIY,not an individual worker's A to an individual worker's h. We stress that the Mincer
evidence deployed in this section should capture any link between the
schooling of an individual worker and the level of technology (e.g. equipment quality) that worker can use. This is because technology adoption
that is linked to the individual should show up in the private wage gain
to more education.

The NeoclassicalRevival in GrowthEconomics?91

Figure 2 1985 LEVELS:BK4
1.5

1.0 -

*
* **
* * *'
*
*2 - *

*,
*

0.5 -

* S*
**%

**
4,

0.0 -

C
-0.5

*
*

*~~~
-1.0

-1.5 -

-2.0
-2.0

** *

-1.5

-1.0

-0.5

0.0

0.5

1.0

1.5

In(X)
Table 3 CORRELATION MATRICES
a.

With MRWOmethodology
ln(Y/L)

ln(K/Y)
ln(H/Y)
ln(A)
b.

.77
.84
.47

With BK4 Methodology

ln(Y/L)

ln(K/Y)
ln(H/Y)
ln(A)

.59
.60
.93

ln(KY)

ln(H/Y)

.67
.00

.00

ln(K/Y)

ln(H/Y)

.02
.28

.57

92 *KLENOW& RODRIGUEZ-CLARE
At this point several robustness checks are in order. We first look at the
potential importance of imperfect substitutability. For 45 countries in Bils
and Klenow's (1996) sample, Barro and Lee (1996) report percentages of
the 25-64-year-old population in seven educational attainment categories: none, some primary, completed primary only, some secondary, completed secondary only, some tertiary, and completed tertiary. We treat
"some" as half-completed, and assume the durations are 8, 4, and 4 years
for primary, secondary, and tertiary schooling. We assume the first three
categories are perfectly substitutable "primary equivalents," and that the
last four are perfectly substitutable "secondary equivalents":
Y = K(Hl-1/t
lla)
sec /
pnm Hl-1/1o)(1-a)/(1-

with

Hpnm=

E ALseY,
s=0,4,8

Hs, =

>
s=10,12,14,16

ALAey,

where Ls is the number of working-age people in schooling group s.

Note that this specification follows BK4 in eschewing Mincer-interceptaffecting school quality differences. We use nonlinear least squares to
estimate or and y78/(1-a) using Barro and Lee's Ls data and Bils and
Klenow's data on the estimated education premium for the 45 countries.
The resulting estimates are y = 0.09(1-a)/f, and ar = 65.25We then use
these estimates to construct H aggregates for the 84 of our 98 countries
for which Barro and Lee (1996) have the necessary schooling attainment
data. The resulting breakdown is (40%, 60%), tilted a little toward ln(X)
relative to the BK4 row, which uses y = 0.095(1- a)/l and ar = 1 (albeit
with human capital vs. raw labor rather than primary equivalents vs.
secondary equivalents). We conclude from this exercise that allowing for
imperfect substitutability (and incorporating heterogeneity in schooling
attainment within each country) does not significantly affect the results.
Our next robustness check concerns the size of f. In the previous
section we found that raising the value of this parameter boosted the role
of human capital in explaining international income variation. This does
not happen here. Here we choose the coefficients y in (7) and (9) so that
the implied wage gain for each additional year of education, which is
25. This degree of substitutability is very high compared to the 1.5 estimated by Katz and
Murphy (1992) for high-school vs. college equivalents in the United States. We have
imposed a common y, however, so our high estimated substitutability may be capturing the combination of less substitutability and higher y's in richer countries.

TheNeoclassical
Revivalin GrowthEconomics
?93
given by /3y/(l-a), matches the Mincer evidence discussed in Bils and
Klenow (1996). Thus changing 3 results in an offsetting adjustment in y
to preserve the equality 3y/(l-a) = 9.5%. In other words, there is no
doubling of the importance of human capital from doubling f to 0.56,
since the coefficient y must be halved at the same time. Indeed, there is
zero effect. The intuition is that the Mincer estimates pin down the
combined effect of translating schooling into human capital and translating human capital into output. Thus a larger elasticity of output with
respect to human capital requires a smaller elasticity of human capital
with respect to schooling in order to maintain consistency with the
Mincer regression evidence.
An objection to the Mincer evidence is that the coefficient on schooling
captures only private gains from schooling. Productive benefits of economywidehuman capital, as proposed by Lucas (1988), would be absorbed
in the Mincer intercept. Lucas (1990) argues that human capital externalities can explain the large differences in TFP that Krueger (1968) found
across 28 countries even after adjusting for human capital per worker
(measured much as in BK4). Leaving aside the nature of these externalities, it is illuminating to ask how big they have to be in order to
restore MRW's 78%-vs.-22% breakdown. For BK3, with its substantial
variation in education quality, we find that the social Mincer coefficient
on schooling would have to be 15.6%, as opposed to the 9.5% or so
typically found. For BK4, the social education premium would have to be
29%. Since the evidence on school quality favors BK4, it appears that
external benefits of schooling would need to be larger than the private
benefits! In any case, entertaining externalities leads to questions about
their exact nature and transmission. To us, this supports our call for more
research into the source of productivity differences across countries.

4. FromDevelopment
Accountingto GrowthAccounting
Whereas Tables 1 and 2 were concerned with development accounting
(King and Levine's felicitous 1994 phrase), Table 4 is about growth accounting. For Table 4 we constructed K/Yand H/Y for each country in 1960
so that we could compute 1960-1985 growth rates. We did this for BK2
through BK4 (one cannot do it under the steady-state assumptions used
for MRWs and for BK1). For H/Y we used Barroand Lee's (1993) schooling
stocks in 1960 and, when necessary to construct experience levels, the
United Nations (1994) population data for 1960. We estimated the 1960 Kl
Y's as described in the previous section, and the results here are not at all
sensitive to the various ways we tried to estimate 1960 K/Y's.
Table 4 presents the results of 1960-1985 growth accounting. When a

94 *KLENOW& RODRiGUEZ-CLARE
Table4 THEROLESOF A AND X IN 1960-1985GROWTH
cov[Aln(Y/L),A In (Z)]/varA ln(Y/L)
Source
BK2
BK3
BK4

Z=

YZ
.03
.03
.03

.12
.12
.06

Z=X

Z= A

.15
.14
.09

.85
.86
.91

aBK2:calculatesyears of schoolings fromBarro-Lee1985stocksinsteadof 1960-1985flows. BK3:adds

averageyears of experience.BK4:BK3but with (K, H, L) sharesof (0, 0, 1) insteadof (0.1, 0.4, 0.5) in H
production.

country's 1960-1985 growth rate of output per worker is 1% faster than

average, growth in physical capital intensity typically contributes about
0.03%. For BK2, which includes only the schooling contribution to human capital, H/Y growth contributes 0.12% more, the share owing to A
being 0.85%. Adding experience (BK3) does not change the calculus.
Letting education quality enter through the Mincer coefficients, as in
BK4, boosts the contribution of A to 91%.
The consistent outcome in Table 4 is that differences in growth rates of
Y/L derive overwhelmingly from differences in growth rates of A.26 Figure 3 plots A growth against X growth (based on BK4) to demonstrate
this visually. The small role we find for growth in the human-capital
stock is not new. Benhabib and Spiegel (1994) and Pritchett (1995) report
that the growth rate of schooling attainment is virtually uncorrelated
with growth in output per worker across countries over 1960-1985.
Barro and Sala-i-Martin find the same thing (1995, Chapter 12), but mark
it down to measurement error.
Table 4 suggests that Chapters 1 to 4 in Barro and Sala-i-Martin (1995)
and studies such as Chari, Kehoe, and McGrattan (1996) that emphasize
transition dynamics of the neoclassical growth model ignore the major
source of differences in country growth rates. Our results call for greater
emphasis on models of technology diffusion and policies that directly
affect productivity.27
26. For the 98-country sample, the unweighted average Y/Lgrowth across the 98 countries
is 2.24%. Using (2), this can be broken down into 0.77% from KIYgrowth, 0.44% from

is distinct
H/Ygrowth, and 1.03%from A growth. This average-world-growth
accounting
from the country-variation-in-growth-rates
accountingthat we focus on above. Unlikethe

country-variation-in-growth rates, which are dominated by variation in the growth rate

of A, the trend growth in Y/Lin the world (2.24%) owes more to X growth (1.21%) than
to A growth (1.03%).
27. Some of the examples offered by Chari, Kehoe, and McGrattan (1996) as contributors to
their effective tax rate may affect productivity A directly. Regulations and corruption
would be expected to hinder firms' ability to translate K and H into Y.

*95
TheNeoclassical
Revivalin GrowthEconomics
Figure3 1960-1985GROWTHRATES:BK4
5%
4% -

2% -

:t
E 1%
e

'

#* *.*
*

0%-

.C

* A * .
*. **
2%-3% -

-5%
-5%

-4%

-3%

-2%

-1%

0%

In(X) growth

1%

2%

3%

4%

5%

rate

The Appendix contains 1985 levels and 1960-1985 growth rates of Y/L,
KIY,HIY, and A (the latter two for BK4). For ease of interpretation the
1985 level variables are given relative to the United States. Many countries surprisingly come out higher than the United States in our estimates for A. Perhaps we have been aggressive in our estimates of the
return to human capital (e.g. making no attempt to adjust for ability
bias), but we prefer to err on this side, given our conclusion that human
capital's importance has been seriously overstated in previous research.
As we mentioned at the beginning of Section 2, the fact that A is not
exogenous implies that the growth rate of A could be affected by the
growth rate of K/Y and H/Y. Increasing levels of capital intensity and
schooling could thus be responsible for high growth rates indirectly, by
allowing for a faster growth of A. To examine this possibility Table 5
shows the correlation matrix for the 1960-1985 growth rates of Y/L, K/Y,

96 *KLENOW& RODRIGUEZ-CLARE
Table5 CORRELATION
MATRIX(BK4GROWTHRATES)
A ln(Y/L)
A ln(K/Y)
A ln(H/Y)
Aln(A)

.04
.28
.87

A ln(K/Y)
-.50
-.42

A ln(H/Y)

.34

H/Y, and A according to BK4. The 0.34 correlation between the growth
rates of A and H/Y suggests that countries with high growth in A have
had unusually high growth rates of schooling. Thus it could be that high
growth in economywide schooling attainment powerfully boosts growth
through its effect on technology adoption. In contrast, the negative correlations between the growth rate of KIYand the growth rates of, respectively, H/Y and A are puzzling.28
An alternative way to think about the role of factor accumulation and
total productivity factor (TFP) growth in explaining differences in economic performance across countries is to look at what has happened to
the standard deviations of Y/L, KIY, H/Y, X, and A (as logarithms)
across time. Table 6 compares the standard deviations of these variables
in 1960 and 1985. As is well known, the standard deviation of the
logarithm of output per worker increased somewhat during this period
(i.e., o-divergence). We find that o-convergence occurred for K/Y, but
not for H/Y, X, and A. Thus the lack of a-convergence in Y/L does not
stem from, say, A-convergence combined with X-divergence.29

5. Do Young'sFindings ContradictOurs?
The debate over whether fast rates of growth in some countries stem
from accumulation of capital or from technology catch-up has been heavily influenced by the East Asian miracles. It was initially thought that
these countries had very high TFP growth rates, pointing to technology
catch-up as the heart of the story. Then came the careful work of Alwyn
Young showing that these countries grew mostly through input accumulation (Young, 1995), and that their TFP growth rates were not extraordinarily high (Young, 1994). Singapore, for instance, was shown to have
28. The negative correlation between the growth rates of A and K/Y could indicate an
overstatement of the contribution of K/Y to outut per worker. One reason for this may
be that public investment (which is part of the data on investment that we used to
generate K/Y) is less efficient than private investment in generating efficiency units of
capital. If this is true, then the role of A is even larger than shown in our results.
29. We also did our variance decomposition on the 1960 numbers and obtained exactly the
same breakdown (34% ln(X) vs. 66% ln(A)) for BK4 that we did for 1985 in Table 2.

*97
TheNeoclassical
Revivalin GrowthEconomics
Table6 STANDARDDEVIATIONS(BK4LEVELS)
Quantity

1960

1985

ln(Y/L)
ln(K/Y)
ln(H/Y)
ln(X)
ln(A)

0.95
0.73
0.28
0.46
0.71

1.01
0.55
0.28
0.44
0.72

virtually no productivity growth over the last decades. As a result of this

work, many people have concluded that the East Asian episodes illustrate the importance of neoclassical transition dynamics rather than technology catch-up.
We do not think this interpretation of Young's results is correct. First, as
we argued above, we think the debate is over whether capital accumulation or technology catch-up explains growth in output per worker,not
growth in output. Neither hypothesis tries to explain the growth rate of
employment. Second, as we also argued above, growth in physical capital induced by rising productivity should be attributed to productivity
[Barro and Sala-i-Martin also make this point (1995, p. 352)]. A higher
level of productivity raises the marginal point of capital, thereby stimulating investment and capital accumulation that would not have occurred
without the higher level of productivity. The role of capital accumulation
over and above that stimulated by productivity growth can be measured
by the growth rate of the capital-output ratio. Table 7a reports a few
calculations from Young's (1995) tables to illustrate the quantitative importance of these considerations. The annual growth rates of output and TFP,
respectively, were 7.3% and 2.3% in Hong Kong, 8.7% and 0.2% in Singapore, 10.3% and 1.7% in South Korea, and 9.4% and 2.6% in Taiwan. So
growth in output clearly came primarily from input accumulation. But
the growth rates of output per workerand adjustedTFP-TFP raised to 1/
(1 - capital's share) because of its effect on capital accumulation-were
as follows: 4.7% and 3.7% in Hong Kong, 4.2% and 0.3% in Singapore,
4.9% and 2.5% in South Korea, and 4.8% and 3.5% in Taiwan. So in three
of the four East Asian miracles growth in output per worker came mostly
from productivity gains.
In any case, the debate should not focus entirely on the miracle countries of East Asia. Although our data is much less detailed than the data
Young compiled for each of the four Asian tigers, our hope is that by
covering 98 countries we get a sense of whether Young's results are typical of the sources of growth differences in the world as a whole. We are
particularly interested, therefore, in whether our results for the Asian

98 *KLENOW& RODRiGUEZ-CLARE
Table7
a. AlwynYoung'sresults
Y Growth
Country
Hong Kong
Korea
Singapore
Taiwan
Country
Hong Kong
Korea
Singapore
Taiwan

TFPGrowth

7.3
10.3
8.7
9.4

2.3
1.7
0.2
2.6

Y/LGrowth

A Growth

4.7
4.9
4.2
4.8

3.7
2.5
0.3
3.5

b. Young'sA growthvs. ours

A growth

Country
Hong Kong
Korea
Singapore

Taiwan

Young's

Ours

Whydifferent?

3.7
2.5

4.4
2.5

L data
O.K.

0.3

3.3

L data; K share

3.5

3.0

O.K.

tigers are not too far off from Young's numbers. Table 7b reports our BK4
1960-1985 A growth rates for the Asian tigers alongside Young's estimates. Our estimate for South Korea matches Young's (2.5%), and our
estimate for Taiwan actually falls below Young's (3.0% vs. his 3.5%). For
Hong Kong our estimate is higher (4.4% vs. 3.7%), and for Singapore
our estimate is much higher (3.3% vs. 0.3%). Young uses census data for
L rather than Summers-Heston data, and the census data show faster
growth of L for Hong Kong and Singapore.30 Faster growth of L translates into slower growth of Y/L and A (with growth in H/Y and K/Y
unaffected). This explains the entire difference in our estimates for Hong
Kong. For Singapore Young used a physical-capital share of 0.49, as
opposed to the 0.30 we used. Combined with the difference in L growth,
Young's higher capital share explains almost all of the gap in our Singapore estimates, since K/Y grew sharply there.
We close by stressing that our results share two important features
30. For 1966-1990,Young'sworker/populationratiosrose from0.38 to 0.49 for Hong Kong
and from 0.27 to 0.51 for Singapore.The comparableSummers-Heston figures were
0.54 to 0.65 and 0.34 to 0.48.

TheNeoclassical
Revivalin GrowthEconomics?99
with those of Young (1994, 1995). First, we find a very modest role for
growth in human capital per worker in explaining growth (Young's adjustments for labor quality are a few tenths of a percent per year). Second, we find that TFP growth accounts for most of the growth of output
per worker in Hong Kong, South Korea, and Taiwan. And we stress that
this relative importance of TFP growth for three of the four Asian tigers
generalizes to our sample of 98 countries: we find that roughly 90% of
country differences in YIL growth are attributable to differences in A
growth. Combining these growth results with our findings on levels, we
call for returning productivity differences to the center of theorizing
about international differences in output per worker.

Appendix.Data
Y/L= 1985RGDPWin Summers-Heston PWT5.6.
K/Y= 1985physical-capital-to-outputratio (see Section4 for K/Yused for BK2BK4).
H/Y = 1985human-capital-to-outputratio (see Section4 for H/Yused for BK4).
A = 1985level of productivity[see equation (2)]. Note: Levels are relativeto the
United States.
g(Z) = 1960-1985annual growth rate of series Z.
Y/L KIY H/Y
Algeria
Benin
Botswana
Cameroon
CentralAfr.R.
Congo
Gambia
Ghana
Guinea-Biss
Kenya
Lesotho
Liberia
Malawi
Mali
Mauritius
Mozambique
Niger
Rwanda
Senegal
South Africa

0.40
0.07
0.20
0.11
0.04
0.20
0.05
0.07
0.04
0.06
0.06
0.07
0.03
0.05
0.22
0.04
0.03
0.05
0.08
0.29

0.79
0.47
0.49
0.27
0.60
0.29
0.39
0.43
1.23
0.63
0.41
0.74
0.57
0.67
0.42
0.18
0.79
0.18
0.38
0.84

0.33
0.34
0.47
0.56
0.30
0.48
0.38
0.47
0.22
0.40
0.51
0.31
0.39
0.27
0.57
0.52
0.25
0.55
0.45
0.45

g(Y/L) g(K/Y) g(H/Y) g(A)

(%)
(%)
(%)
(%)

0.99
2.89
0.25
0.78
0.55
6.85
4.22
0.43
0.12
0.33
0.81
4.06
0.18
1.37
0.20
0.36
0.10
1.52
0.15
1.31
0.18
5.03
1.22
0.18
0.10
1.70
0.16
0.45
0.59
0.90
0.22 -1.18
0.10
0.79
0.23
1.89
0.27
0.87
0.57
1.82

0.71
1.37
2.20
1.16
-0.79
-1.81
4.32
-0.60
-0.01
-0.10
4.28
-0.82
1.87
-2.28
-1.09
2.98
1.82
1.70
-0.64
1.23

0.64
0.02
0.71
0.63
0.53
0.30
-0.70
1.51
0.34
1.03
-0.55
0.92
0.04
0.77
1.43
-0.39
-0.64
-0.02
0.62
0.12

1.96
-0.21
4.81
2.97
0.54
5.16
-1.25
-0.22
1.29
0.70
2.34
1.19
0.34
1.57
0.72
-3.05
-0.09
0.68
0.92
0.86

100 - KLENOW & RODRIGUEZ-CLARE

g(YIL) g(K/Y) g(H/Y)
(%)
(%)
(%)

g(A)
(%)

0.41
0.08
0.12
0.81
0.17
0.14
0.12
0.26

2.96
2.08
2.60
3.22
0.06
0.42
-0.42
1.50

2.87
1.80
3.02
-0.94
0.96
4.19
-0.07
-0.96

0.35
0.05
0.28
1.35
0.17
-0.19
1.43
0.74

0.67
0.77
0.26
2.99
-0.74
-2.45
-1.32
1.70

0.77
0.85
0.58
0.51
0.51
0.40
0.39
0.46
0.39
0.53
0.47
0.62
0.66
1.00
0.51
0.42
0.40
0.53
0.51
0.53
0.37
0.58
0.54
0.48
0.49

0.70
1.05
0.64
0.55
0.48
0.62
0.22
0.38
0.28
1.29
0.47
0.60
1.55
1.00
0.64
0.35
0.77
0.47
0.68
0.58
0.17
0.50
0.47
0.43
1.07

2.39
1.88
1.17
2.16
0.95
1.32
0.96
1.41
0.34
2.33
0.56
3.00
1.65
1.30
1.11
2.11
2.73
0.44
2.10
3.07
-1.80
2.23
1.02
0.17
-0.43

1.55
0.64
1.68
2.33
1.94
1.24
2.23
0.21
1.23
1.27
2.35
1.45
1.57
0.56
1.38
0.57
0.32
0.73
0.33
0.78
1.58
2.10
1.45
1.34
2.46

0.46
0.94
0.50
0.33
0.53
0.47
-0.07
0.93
0.56
0.70
0.28
0.62
0.52
1.27
0.57
0.48
0.36
0.52
0.84
1.12
-0.18
0.09
1.05
0.87
0.82

0.97
0.80
-0.36
0.28
-0.79
0.12
-0.58
0.64
-0.91
0.95
-1.31
1.55
0.18
0.04
-0.25
1.38
2.26
-0.43
1.31
1.77
-2.82
0.67
-0.72
-1.36
-2.74

0.52
0.74
0.38
0.45
0.39
0.41
0.79
0.60
0.61
0.76

0.55
0.88
0.20
0.32
0.97
1.26
0.84
0.63
1.39
0.54

1.73
5.49
1.74
3.89
1.29
0.85
3.27
5.30
5.00
5.37

-1.11
0.53
-0.70
1.88
3.35
5.70
0.79
2.01
2.34
2.32

0.90
1.08
1.07
0.95
0.68
0.05
1.02
0.51
0.95
1.77

1.92
4.39
1.53
1.91
-1.55
-3.26
2.03
3.53
2.69
2.54

Y/L

K/Y

H/Y

Swaziland
Tanzania
Togo
Tunisia
Uganda
Zaire
Zambia
Zimbabwe

0.15
0.03
0.04
0.26
0.04
0.03
0.07
0.10

0.53
0.59
0.76
0.46
0.19
0.24
1.31
0.61

0.47
0.37
0.32
0.42
0.57
0.54
0.34
0.38

Barbados
Canada
Costa Rica
Dominican Rep.
El Salvador
Guatemala
Haiti
Honduras
Jamaica
Mexico

0.36
0.92
0.27
0.21
0.16
0.22
0.06
0.14
0.14
0.50
0.17
0.30
0.76
1.00
0.44
0.17
0.32
0.29
0.27
0.28
0.11
0.18
0.24
0.30
0.54

0.51
0.97
0.50
0.48
0.42
0.54
0.42
0.50
0.93
0.49
0.51
0.58
0.54
1.00
1.11
0.81
0.70
0.91
0.53
0.67
1.34
0.42
0.70
1.22
0.74

0.13
0.49
0.08
0.13
0.41
0.47
0.65
0.56
0.46
0.31

0.24
0.57
0.71
0.59
0.73
0.58
0.88
1.35
0.34
0.58

Nicaragua
Panama
Trinidad & Tobago
United States
Argentina
Bolivia
Brazil
Chile
Colombia
Ecuador
Guyana
Paraguay
Peru
Uruguay
Venezuela
Bangladesh
Hong Kong
India
Indonesia
Iran
Iraq
Israel
Japan
Jordan
Korea, Rep.

The NeoclassicalRevival in GrowthEconomics? 101

g(YIL) g(K/Y) g(H/Y)
(%)
(%)
(%)

g(A)
(%)

Y/L

K/Y

H/Y

Malaysia
Myanmar
Nepal
Pakistan
Philippines
Singapore
Sri Lanka
Syria
Taiwan
Thailand

0.31
0.04
0.07
0.13
0.13
0.53
0.17
0.51
0.38
0.14

0.68
0.39
0.35
0.48
0.53
0.93
0.35
0.37
0.51
0.53

0.51
0.42
0.38
0.38
0.66
0.41
0.69
0.56
0.73
0.55

0.63
0.14
0.27
0.40
0.26
1.03
0.45
1.52
0.75
0.33

3.74
2.64
2.25
2.96
1.41
5.11
1.87
4.42
5.30
3.70

1.30
0.36
1.36
-0.31
2.18
2.40
0.63
0.97
1.76
0.87

1.21
0.46
0.09
0.76
0.76
0.16
0.84
1.35
1.52
0.62

2.00
2.08
1.21
2.68
-0.65
3.29
0.85
2.83
3.03
2.66

Austria
Belgium
Cyprus
Denmark
Finland
France
Germany, W.
Greece
Iceland
Ireland

0.71
0.81
0.41
0.71
0.70
0.80
0.81
0.48
0.69
0.57
0.80
0.46
0.85
0.85
0.34
0.63
0.78
0.88
0.21
0.68
0.34

1.53
1.43
1.11
1.49
1.80
1.47
1.85
1.25
1.10
1.07
1.51
0.95
1.28
1.49
1.21
1.28
1.54
2.03
0.79
1.23
1.52

0.45
0.64
0.54
0.73
0.60
0.45
0.53
0.50
0.59
0.62
0.43
0.56
0.61
0.73
0.34
0.42
0.65
0.55
0.37
0.64
0.48

0.88
0.84
0.58
0.66
0.65
1.04
0.79
0.65
0.91
0.75
1.04
0.70
0.98
0.79
0.60
0.93
0.77
0.80
0.48
0.79
0.41

3.20
2.59
4.12
1.91
2.87
2.79
2.69
4.60
2.46
3.31
3.60
4.71
2.05
2.80
3.40
3.80
1.69
1.57
3.19
1.77
3.97

0.74
0.63
0.49
0.58
0.15
1.16
0.51
1.41
-0.42
0.93
0.44
-1.08
0.68
0.18
0.93
1.91
0.57
1.03
1.23
0.21
1.44

1.42
0.73
1.33
0.26
0.97
1.02
0.30
0.94
1.20
0.47
0.85
1.18
1.45
2.35
0.83
0.71
0.77
0.86
0.41
0.38
1.25

1.72
1.65
2.88
1.32
2.11
1.29
2.12
2.97
1.95
2.33
2.72
4.69
0.59
1.10
2.18
1.96
0.77
0.26
2.04
1.37
2.11

0.86
0.29
0.77
0.10

1.33
0.68
1.25
1.08

0.74
0.61
0.94
0.26

0.85
0.53
0.68
0.23

1.63
1.03
0.81
1.59

0.46
0.25
0.38
2.77

0.48
0.86
1.01
-0.54

0.98
0.28
-0.14
-0.03

Italy
Malta
Netherlands
Norway
Portugal
Spain
Sweden
Switzerland
Turkey
United Kingdom
Yugoslavia
Australia
Fiji
New Zealand
Papua N. Guinea
REFERENCES

Barro, R. J., and J.-W. Lee. (1993). International comparisons of educational

attainment. Journalof Monetary Economics32(3):363-394.
. (1996). International measures of schooling years and
,and
schooling quality. American Economic Review, Papers and Proceedings 86(2):
218-223.
, and X. Sala-i-Martin. (1995). EconomicGrowth. New York: McGraw-Hill.

102 * KLENOW & RODRIGUEZ-CLARE

Benhabib, J., and M. M. Spiegel. (1994). The role of human capital in economic
development: Evidence from aggregate cross-country data. Journalof Monetary
Economics34(2):143-174.
Bils, M., and P. J. Klenow. (1996). Does schooling cause growth or the other way
around? University of Chicago. Mimeo.
Borjas, G. J. (1987). Self-selection and the earnings of immigrants. American
EconomicReview 77(4):531-553.
Bosworth, B., S. M. Collins, and Y. Chen. (1995). Accounting for differences in
economic growth. Washington: Brookings Institution. Brookings Discussion
Paper 115.
Chari, V. V., P. J. Kehoe, and E. R. McGrattan. (1996). The poverty of nations: A
quantitative exploration. Cambridge, MA: National Bureau of Economic Research. NBER Working Paper 5414.
Ciccone, A. (1994). Human capital and technical progress: Stagnation, transition,
and growth, University of California at Berkeley. Mimeo.
Gollin, D. (1996). Getting income shares right: Accounting for the self-employed.
Williamstown, MA: Williams College. Mimeo.
Grossman, G. M., and E. Helpman. (1991). Innovation and Growth in the Global
Economy.Cambridge, MA: The MIT Press.
Hall, R. E., and C. I. Jones. (1996). The productivity of nations. Cambridge, MA:
National Bureau of Economic Research. NBER Working Paper 5812.
Hanushek, E. A. (1986). The economics of schooling: Production and efficiency
in public schools. Journalof EconomicLiterature24(September):1141-1177.
Heckman, J. J., A. Lyane-Farrar, and P Todd (1996). "Does money matter? The
effect of school resources on student achievement and adult success." An
examination of the earnings-quality relationship. In The Link between Schools,
Student Achievement,and Adult Success, Gary Burtless (ed.). Washington: Brookings Institution, 192-289.
Jorgenson, D. W. (1995). Productivity: Volume1, Postwar U.S. EconomicGrowth.
Cambridge, MA: The MIT Press.
Katz, L. E, and K. M. Murphy. (1992). Changes in relative wages, 1963-1997:
Supply and demand factors. QuarterlyJournalof Economics107(1):35-78.
Kendrick, J. W. (1976). TheFormationand Stocksof TotalCapital. New York:Columbia University Press for the NBER.
King, R. G., and R. Levine. (1994). "Capital Fundamentalism, Economic Development, and Economic Growth." Carnegie-RochesterConferenceSeries on Public
Policy 40:259-92.
Krueger, A. 0. (1968). Factor endowments and per capital income differences
among countries. EconomicJournal78(311):641-659.
Lucas, R. E. (1988). On the mechanics of economic development. Journal of
Monetary Economics22(1):3-42.
. (1990). Why doesn't capital flow from poor to rich countries? American
EconomicReview 80(2):92-96.
Mankiw, N. G., (1995). The growth of nations. In BrookingsPapers on Economic
Activity, G. Perry and W. Brainard (eds.). Washington: Brookings Institution,
Vol. 1, pp. 275-326.
, D. Romer, and D. N. Weil (1992). A contribution to the empirics of
economic growth. QuarterlyJournalof Economics107(2):407-437.
Mincer, J. (1974). Schooling, Experience,and Earnings. New York:Columbia University Press.

Comment 103
Mosteller,F (1995).The Tennessee study of class size in the early school grades.
TheFutureof Children5(2):113-127.
Parente, S., and E. C. Prescott. (1996).The barriersto riches. Philadelphia:University of Pennsylvania;Universityof Minnesota.Manuscript.
,R. Rogerson,and R. Wright.(1996).Homeworkin developmenteconomics: Household productionand the wealth of nations. Philadelphia:University
of Pennsylvania.Mimeo.
Pritchett,L. (1995).Where has all the educationgone? WorldBank.Mimeo.
Rodriguez-Clare,A. (1997).The role of tradein technology diffusion. University
of Chicago. Mimeo.
Romer, P. M. (1990). Endogenous technologicalchange. Journalof PoliticalEconomy98(5):S71-S102.
. (1993). Idea gaps and object gaps in economic development. Journalof
32(3):543-573.
MonetaryEconomics
Summers, R., and A. Heston. (1991). The Penn WorldTable(Mark5): An expanded set of internationalcomparisons, 1950-1988. QuarterlyJournalof Economics106(2):327-368.Mark5.6 (an update) is availableat nber.harvard.edu.
Stokey, N. L. (1991). Human capital, product quality, and growth. Quarterly
Journalof Economics
106(2):587-617.
United Nations. (1994). TheSex and Age Distributionof the WorldPopulations,the
1994Revision.New York:United Nations.
U.S. Departmentof Education.(1996).1996DigestofEducation
Statistics.Washington: U.S. GovernmentPrintingOffice.
Young, A. (1991). Learningby doing and the dynamic effects of international
trade. QuarterlyJournalof Economics
106(2):369-406.
. (1994). Lessons from the East Asian NICs: A contrarianview. European
Economic
Review38(3-4):964-973.
. (1995).The tyrannyof numbers:Confrontingthe statisticalrealitiesof the
east Asian growth experience. QuarterlyJournalof Economics
110(3):641-680.

Comment
N. GREGORYMANKIW
HarvardUniversity

Instructors of macroeconomics who teach their students about economic

growth often use Solow's version of the neoclassical growth model as
the starting point for discussion. This model shows very simply how an
economy's production technology and its rates of capital accumulation
determine its steady-state level of income per person. After presenting
this elegant theory, the instructor is left with a nagging question: So
what? Does this model really explain why some countries are rich and
others are poor? Or does this model leave most of the action unexplained
in a variable that has been called, at various times, total factor productivity, the Solow residual, and "a measure of our ignorance"?