ECON 4350: Growth and Investment
Lecture note 3
Department of Economics, University of Oslo
Made by: K˚are Bævre (kare.bavre@econ.uio.no)
Exploited by: Miroslav Hlouˇsek (hlousek@econ.muni.cz)
5 Cross-country studies and human capital
Required reading: Mankiw, Romer and Weil (1992), Klenow and RodriguezClare
(1997), BSiM: 1.2.10-1.2.11, 10.1-10.2,10.5
Secondary reading: Young (1995), Bil and Klenow (2000), Hsieh (1999),
Pritchett (2001)
5.1 Growth econometrics: Why and how?
• We desire theories that are able to teach us something of relevance to
the real world (E.g. the questions in section 1.2.3)
• Remember: We are addressing large and complex questions. Should
we expect our theories to be true? What other criteria should we then
use?
• We can adopt several perspectives when confronting empirical data
1. Testing theories (Is it ‘true’?)
2. Assess the explanatory power of the theory (How much do we
explain?)
3. Checking stylized facts
4. Explore data for regularities suggesting new theories
5. Calibrate our models
• There are good reason to be eclectic when it comes to choice of methods
1
5.2 Levels regressions
• A fundamental question in studies of growth is: Why are some countries
so rich, and some so poor?
• It is therefore natural to investigate empirically how good we can explain
variation in income levels across countries by different explanatory
variables.
• A common framework for such analysis is the so-called level regressions
ln((Y/L)i) = a + b1x1,i + b2x2,i + . . . + i (1)
where xj,i are explanatory variables for country i, and is an error
term.
• From the production function we know that Y/L depends on K/L and
H/L, or in the CD-case:
ln(Y/L) = α ln(K/L) + η ln(H/L)
However, we can not readily enter the capital intensities as explanatory
variables (xj,i) in (1) because they are endogenous (that is, themselves
dependent upon Y/L).
• Estimation of an equation like (1) will give biased estimates unless we
are able to control for this problem. From econometrics we know that
this can be achieved by use of instrumental variables. However, suitable
instruments are hard to come by.
• Levels regressions are therefore usually formulated somewhat differently,
and based on the assumption that all countries are in their steady
state.
• This is obviously a crude approximation, but it is still useful for putting
the model to a first test.
• We remember that the augmented Solow-model tells us that savings
rates (sk, sh) and population growth (n) affects the steady state level
of GDP per capita ((Y/L)∗
). Or as we have seen:
ln((Y (t)/L(t))∗
) = ln(T(t))+
α
1 − α − η
ln(sk)+
η
1 − α − η
ln(sh)−
α + η
1 − α − η
ln(n+x+δ)
(2)
2
• By assuming that the observed measure (Y/L)i corresponds to the
steady state level for country i, we can therefore motivate the regression
equation
ln((Y/L)i) = a + b1 ln(sk,i) + b2 ln(sh,i) − b3 ln(ni + x + δ) + i
(3)
• Note that we do not include country specific data on depreciation (δ)
since these are too hard to measure/not available, and probably do not
vary that much across countries anyway.
• The approach adopted by MRW effectively ignores differences in technology
between countries. This is a deliberate choice, based on the
following three considerations: 1) Since technology is exogenous and
unexplained in the Solow-model, they want to see how far they can
get without including technology as an explanatory variable. 2) The
perspective they adopt insists that knowledge flows quite freely across
countries, this should result in absence of large and persistent technology
differences across countries. 3) It is not possible to readily observe
x and T(t) for individual countries.
• MRW therefore use average values of x + δ (like 0.02 + 0.03 = 0.05),
the exact levels are not very important for our results.
• Notice however, that use of a common intercept a for all countries
amounts to assuming that
Ti(t) = a + εi, (4)
where a = T(t) is the average level of technology at time t, and εi is
a random country specific deviation from this level that is included in
the total error term i for this country.
• In econometric terminology the theoretical equation (2) is a structural
equation, with equation (3) as its reduced form.
• From theory/the structural equation we are lead to believe that
b1 =
α
1 − α − η
(5)
b2 =
η
1 − α − η
(6)
b3 =
α + η
1 − α − η
(7)
3
which in turn implies
b1 + b2 = b3 (8)
which is a testable restriction on the reduced form parameters b1, b2, b3
• Imposing condition (8) we are left with two independent equations from
which we can solve out for the structural parameters α and η based on
the estimated values of the reduced form parameters (MRW call these
the ‘implied’ α and β)
• Though this framework allows for a nice way of exploring the relationship
between theory (the augmented Solow-model) and the data, there
are important problems:
– What if deviations from steady-sate are important?
– Are the right-hand variables exogenous? Or do also sk, sh and n
depend on Y/L?
– Are sk, sh and n independent of country-specific level of technology
(i.e. Ti(t))? If not, our inclusion of εi in the error term
i makes it correlated with the explanatory variable. It is well
known from econometrics that this is an important source of bias
in estimates of b1, b2, b3 (Omitted variable bias)
– It is hard to include additional explanatory variables (e.g. political
variables) in levels regressions because they are likely to be
endogenous to the income level.
5.3 A broader base for the accumulated factor saves
the Solow model? MRW (1992)
5.3.1 The text-book Solow model
• In MRW, Sect II, the authors conduct a level-regression based on the
text-book Solow-model (i.e. with only accumulation of physical capital)
• Their results (Table I) are qualitatively in accordance with what we
expect from theory.
• However:
1. The explanatory power is modest.
2. The implied value of α is much higher than what we would suspect
from estimates of this elasticity based on capital’s share of income.
4
• This is yet another indication that we should reconsider the role of
capital, and also include human capital.
• Note that an alternative way of formulating (2) is
ln(Y (t)/L(t)) = ln(T(0))+xt+
α
1 − α
ln(sk)−
α
1 − α
ln(n+x+δ)+
η
1 − α
ln(h∗
)
(9)
• If (9) is the true model, it is thus as if they have neglected the term
with h when they estimate the textbook model. Since h∗
is probably
positively correlated with sk, this introduces a bias. More specifically:
b1 = α
1−α
will be biased upwards (why?), and hence we will get a too
high implied value of α.
5.3.2 The augmented Solow-model
• In Section III the authors conduct a levels-regression based on the
augmented Solow-model.
• They prefer to estimate (2) rather than (9). Implicitly they are arguing
that sh is easier to measure than h∗
. Another argument is that sh is
probably less likely to suffer form endogeneity problems.
• MRW use enrollment rate to secondary schooling as a proxy for sh.
• Their results are impressive. The augmented model can describe:
1. Much of the cross-country income differences in the data (R2
=
0.77 in the intermediate sample)
2. The values of α and β implied by the reduced form coefficients
are both 1/3, which is close to what we should suspect from other
considerations.
• That the implied value of β makes sense can be seen from a crude
guestimate
β ≈ (1 − α)(1 −
minimum wage
average wage in maunfacturing
) ∈ (1/3, 1/2)
• All in all, the results in MRW is very good news for the Solow-model.
It is in accordance with the data, and capital accumulation seems to go
a long way towards answering our question about why some countries
are rich, and some are poor.
5
• Quite impressive of such a simple model!
• MRW was a major contribution to a ’neo-classical’ revival.
5.4 Education and the measurement of human capital
• The conclusion of MRW has been significantly modified by later re-
search
• The criticism basically goes along two lines:
1. Methodological shortcomings. i) Endogeniety, ii) Neglecting differences
in Ti
2. That their measure of sh is poor.
• There has been suggested few remedies of the first set of problems
within the framework of levels-regression. We will have more to say on
these issues when we turn to growth-regressions.
• We therefore focus on problems connected to measuring human capital.
• We follow Klenow and Rodrigez-Clare (1997)
• Their approach is somewhat different. They do not want to estimate α
and η due to the methodological problems, and likely biases. Instead
they set out to use ‘independent’ evidence on these parameter values,
and based on these calculate the contributions of the inputs directly.
I.e. with
Y
L
= A
K
Y
α/(1−α−η)
H
Y
η/(1−α−η)
= AX (10)
they measure the contributions K
Y
α/(1−α−η)
and H
Y
η/(1−α−η)
directly.
They use X = K
Y
α/(1−α−η) H
Y
η/(1−α−η)
to denote the total contribution
from accumulated factor. The remainder, A, is attributed to the
level of technology.
• Note that they end up using MRW’s estimates in lack of better alternatives
(for η) and for sake of comparison.
• Instead of running regressions they use a method called variance de-
composition
6
1 =
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))
• Note that cov(ln(X), ln(A)) is distributed equally on the two factors.
• The reported covariances are equal to the coefficients from univariate
regressions of ln(X) and ln(A) respectively, on ln(Y/L).
• Since cov(ln(X), ln(A)) = 0 by construction in MRW, their (unique)
decomposition is equal to R2
, 1 − R2
.
• Methodological differences aside, the main point in the article is showing
how sensitive the MRW results are to modifications of how the role
of human capital is measured.
• Their first major finding is that if we also include primary and tertiary
enrollment rates, human capital seems to play a much smaller role
(MRW3/4 in Table 1).
• In particular, primary enrollment rates varies much less between countries.
MRW’s focus on only secondary enrollment rates can be misleading
since it exaggerates the true variation in the explanatory variables.
• A more fundamental difference in methodology is their use of estimates
based on micro-evidence on the returns to schooling.
• Mincer regressions:
ln(ws) = ln(w0) + rs
where ws is the wage of person with s years of schooling. Microeconomic
studies also control for a lot of other factors, so r is in principle
the partial effect of schooling when all else is equal.
• In general, this functional form fits the data well. The estimates of r are
surprisingly stable around 0.1 across different studies. We can use this
as a rule of thumb/first approximation. (More precisely: r = 0.095)
• Note that the return rate, r, is fixed, implying that an additional year
of higher education is more valuable than an additional year of lower
education. Thus we put different weights on the contributions of the
different types of education.
7
• Based on measures constructed in this way, they find results which
once again modifies the original MRW results (see Table 2). However,
the effects are not by far as drastic as those suggested by their MRW4
estimation.
• Based on these new results, it appears that MRW’s results attribute
too much to the role of factor accumulation.
• As a conclusion we can say that the results indicate that both factor
accumulation and technological differences are important (roughly
50/50).
5.5 Why is education privately profitable but with weak
macro-effects?
• A r = 0.1 in the Mincer-regressions estimated on micro-data suggests
substantial private returns to education.
• Pritchett (2001) also estimates human capital based on Mincer-regressions.
• His estimates of the growth-growth equation:
γy = a + b1γk + b2γh
yields a negative b2, i.e. a negative effect of increased aggregate education
levels on growth in aggregate production.
• The results appear somewhat extreme, but strongly suggest a conflict
between private profitability of education and weak (if any) effects on
aggregate production.
• Why is this so? In sections II and III (which you should read) Pritchett
(2001) emphasizes:
1. Substantial differences between countries.
2. How is the human capital put to use? Piracy vs. chemical engineer.
The role of private opportunities.
3. The demand for educated labor has grown too slowly in many
countries (declining r).
4. Education has failed, non-productive.
5. We need to rethink the role of education and human capital in the
production function. Human capital is not just another form of
capital. The production of human capital is substantially different
from production of output/physical capital.
8
5.6 Does schooling cause growth or the other way around?
• MRW’s finding of a positive correlation between enrollment rates and
output remains.
• Can this be due to a reverse causality?
• Higher expected growth can induce more schooling because it will put
more weight on how an individual values future human capital relative
to the current (alternative) costs of education.
• The benefit of having human capital is proportional to the level of T
while working in the future. The cost of investing in human capital
is proportional to the level of T while in school (today). Increased
expected growth increases the former relative to the latter.
• This is the intuition in Bils and Klenow (2000). You should read the
introduction and section I (the model). It is, however, recommended
that you postpone this until we have gone through the Ramsey model
(Topic 7).
5.7 Growth accounting and development accounting
5.7.1 Growth accounting
• Consider a general production function
Y = F(T, K, L)
• Simple differentiation with respect to time gives
˙Y
Y
=
FT T
Y
·
˙T
T
+
FKK
Y
·
˙K
K
+
FLL
Y
·
˙L
L
• The different terms can be seen as capturing the contribution of the
growth in the inputs T, K, L to growth in Y .
• Here, ˙Y /Y , ˙K/K, and ˙L/L are directly observable (that is not to say
that they are always perfectly measured). In addition, FKK/Y and
FLL/Y can be set equal to the respective factor shares, sK and sL.
• This leaves us with
g =
FT T
Y
·
˙T
T
as the unobserved contribution to growth from technological progress.
9
• We can estimate g as the so-called Solow-residual
˜g =
˙Y
Y
− sK
˙K
K
− sL
˙L
L
• We also often refer to the estimate ˜g as growth of Total Factor Productivity
(TFP).
• Note that the estimate is a true accounting residual, and we can not
warrant that TFP growth is due to technological progress as we think
of it in the model (i.e. ˙T/T = x). Anything we are not able to explain
from changes in K and L will fall in this category (ranging from the
effects of a war, to some types of measurement errors).
• One should therefore be careful not to interpret to much about causality
into growth accounting. (See also BSiM 10.5)
• Note that growth accounting differs from the regression approaches
since we use estimates of factor shares taken from separate sources
instead of estimating these as elasticities.
• The challenge for getting good growth accounting estimates is to measure
changes in inputs carefully. One would like to capture both quantitative
and qualitative changes.
• Note in particular that changes in human capital is usually entered as
changes in the quality of labor (educational attainment, age composition
etc).
• In practice it is often convenient to work with a so-called trans-log
production function (e.g. Young (1995)), which is basically a generalization
of the Cobb-Douglas. The main advantage is that it makes it
easy to work with disaggregation of inputs into sub-categories.
• Growth accounts usually find that a substantial part of growth is due
to factor-accumulation, but also finds that TFP-growth is substantial.
• Particular attention has been devoted to the East-Asian growth mira-
cles.
• The most famous study is that by Young (1995). His estimates attributes
the high growth rates to exceptionally strong factor accumulation.
TFP growth in these countries is at levels comparable to other
countries.
• Young’s results was a main contribution to the neo-classical revival.
10
5.7.2 Dual growth accounting
• Under the standard assumptions we have
Y = RK + wL
• From this relationship, differentiation gives us
˙Y
Y
= sK( ˙R/R + ˙K/K) + sL( ˙w/w + ˙L/L)
• Hence we get the dual formulation of the estimate of TFP growth:
˜g =
˙Y
Y
− sK
˙K
K
− sL
˙L
L
= sK
˙R/R + sL ˙w/w
That is, estimated based on changes in factor prices.
• The most important aspect of this approach is that it allows us to
estimate TFP-growth based on alternative data, which are perhaps
easier to measure and more reliable. At least it provides us with an
important check of estimates based on the primary approach.
• Hsieh (1999) used the dual approach to redo Young (1995) analysis of
South East Asia.
• The most important difference is for Singapore. He argues that this is
because the national accounts overstates capital growth, while rental
prices have been fairly stable.
5.7.3 Development accounting
• Let the production function be Cobb-Douglas
Y = TKα
Hη
L1−α−η
• In this special case, the (primary) growth account was based on looking
at the differentiated form
˙Y
Y
=
˙T
T
+ α
˙K
K
+ η
˙H
H
+ (1 − α − η)
˙L
L
• The growth account looks at changes in a given country over time.
11
• It is also worth considering development accounting. I.e. for countries
i and j we should have
Yi
Yj
=
Ti
Tj
Ki
Kj
α
Hi
Hj
η
Li
Lj
1−α−η
• Based on this relationship one use different criteria to judge how much
of differences across countries in Y (i.e. the left hand side) is due to
differences in TFP (i.e. T) or inputs (K, H and L).
• Different authors have used different criteria. The main thing to notice,
however, is that this approach is based on calibration by using estimates
of α and η derived from independent sources. I.e. the parameters are
not estimated from the data under consideration, as it is done in the
cross-country regressions.
• Klenow and Rodriguez-Clare (1997) use a variant of development ac-
counting.
6 Convergence and growth regressions
Required reading: BSiM: 11 (11.6-11.9 cursory reading), 12 (12.5 cursory)
Secondary reading: Islam (1995), Jones (1997)
6.1 Convergence to the steady state
• A fundamental property of the Solow- model is convergence to a steady
state, characterized by the fundamental equation
˙ˆk
ˆk
=
sf(ˆk)
ˆk
− (n + x + δ) (11)
• Along the transition to the steady-state growth is stronger the further
we are from the steady state (growth is positive if we are below, negative
if we are above).
12
• This is easily seen from the graph
• More formally we can easily show show (BSiM 1.30) that (11) can be
reformulated to
˙ˆk
ˆk
= (n + x + δ)


ˆk
ˆk∗
α−1
− 1


• To characterize the convergence it is however often more convenient to
instead work with the linearization
˙ˆk
ˆk
= β(ln ˆk∗
− ln ˆk)
where
β = (1 − α)(n + x + δ) (12)
is the speed of convergence. (See exercises for seminar 2)
• We also know that the dynamics of ˆk translates directly to that of ˆy,
so we get
˙ˆy
ˆy
= β(ln ˆy∗
− ln ˆy) (13)
• Or, solving the differential equation (13), we get:
ln ˆy(t) − ln ˆy(0) = (1 − e−βt
) ln ˆy∗
− (1 − e−βt
) ln ˆy(0)
• Dividing throughout by t gives a growth rate on the left hand side
(hence the choice of linearizing in ln ˆy rather than ˆy)
ln ˆy(t) − ln ˆy(0)
t
= b1 ln ˆy∗
− b2 ln ˆy(0) (14)
where b1 = b2 = (1 − e−βt
)/t.
13
6.2 Absolute and conditional convergence
• So far we have been looking at convergence of a given economy to its
steady state.
• It is also an important question whether we should expect convergence
across countries.
• We now use (14) to discuss this property. Note that we talk about poor
and rich countries (i.e. as measured by y(t) = Y (t)/L(t) even if (14)
involves ˆy(t) = y(t)/T(t). The role of T(t) in the denominator is not
important here, since we assume equal technological progress across
countries. We will deal with this more carefully soon.
• The Solow-model predicts the following: Countries that are poor will
grow faster than rich ones provided they have the same steady state
(and are both below this steady state).
• This is conditional convergence, which is the same as predicting
that b2 in (14) is positive, so the partial impact of ln ˆy(0) on the growth
rate is negative.
• In (14) we have conditioned on the steady state by the term ln ˆy∗
.
• Note that there is one further complication: A factor like population
growth affects both the steady state for the country and the β for that
country (see (12)). Practical implementations usually ignore the latter
effect, and sets β equal across countries.
• Note that the Solow-model does not predict absolute convergence,
i.e. that a poor country will always grow faster than a rich country.
• This can easily be demonstrated in the familiar graph
14
6.3 β and σ-convergence
• It is important to be careful about how we think about convergence.
• The two types of convergence we discussed above are usually referred to
as β-convergence. That is, they have to do with the b2 above (which
is derived from β, hence the name). We speak respectively of absolute
and conditional β-convergence.
• In addition we can define σ-convergence: A group of economies are
converging in the sense of σ if the dispersion of their real per capita
GDP levels (or equivalent measure of interest) tends to decrease over
time. That is, if
σt2 < σt1
when t2 > t1 and σt is the standard deviation of ln(y(t)) across the
economies at time t.
• It is easy to realize that β-convergence is necessary for σ-convergence.
• However, β-convergence does not necessarily imply σ-convergence. (Galton’
fallacy).
• This can be illustrated by the graphs:
• See BSiM 11.1 for a more detailed discussion.
15
6.4 Growth regressions
• Most of the empirical literature (at least the early one) is primarily
concerned with β-convergence, and obtaining estimates of the speed of
convergence.
• We have derived the following log-linear equation for growth rates of ˆy
ln ˆy(t) − ln ˆy(0)
t
= b1 ln ˆy∗
− b2 ln ˆy(0) (15)
where b1 = b2 = (1 − e−βt
)/t.
• This suggest that we should run regressions of growth rates on initial
level of income per capita. This will enable us to derive an estimate of
β from b2
• Importantly, we must also condition upon ˆy∗
, i.e. we must consider
conditional convergence.
• We can distinguish between two approaches
1. Theory driven (MRW): We use the theoretical expression for ˆy∗
,
i.e. as a function of savings rate(s) and population growth. For
the augmented model we insert from equation (2) in L.N. 3, and
solve out for y instead of ˆy and get the structural equation
ln(y(t)) − ln(y(0))
t
=
(1 − e−βt
)
t
α
1 − α − η
ln(sk) +
(1 − e−βt
)
t
η
1 − α − η
ln(sh)
−
(1 − e−βt
)
t
α + η
1 − α − η
ln(n + g + δ) −
(1 − e−βt
)
t
ln(y(0))
+
(1 − e−βt
)
t
ln T(0) + xt
2. Barro-regressions: We run regressions of the type
ln y(t) − ln y(0)
t
= b0 + B1X − b2 ln(y(0)) (16)
where X is a vector of variables characterizing each economy which
are assumed to be of influence for the steady state level.
16
6.5 Empirical evidence
• Initial results on absolute convergence where spurious, and due to ex
post selection bias (only the success stories had data/had joined the
OECD) and hence convergence followed from the choice of sample.
• MRW: Not absolute convergence. Evidence of conditional β-convergence.
• Speed of convergence: 0.01 − 0.02. The short run is long!
• The implied share of the accumulated factor suggests a broad concept
of capital or externalities (α + η ≈ 0.75).
• Similar studies: There is no evidence of absolute β-convergence in samples
covering ‘the world’. There is even evidence of σ-divergence.
• When we run conditional regressions we find rather strong evidence of
both β- and σ-convergence. The rate of convergence is usually found
to be around 0.02.
• We also find strong signs of absolute convergence when we confine our
comparisons to close/similar economies such as US states, Japanese
prefectures.
6.6 Econometric issues
• A problem with the approach of MRW above is that it assumes that
the initial level of technology, T(0), is the same for all economies in the
sample.
• In reality this implies neglect of variable in the regression (because T(0)
is likely to differ).
• Once again we are left with potential biases because T(0) will be positively
correlated with y(0). This gives a bias toward zero for the parameter
b2, and hence a downward bias for β.
• Note that problem is probably more severe here than in the case with
levels-regression.
• A solution to this problem is to include a dummy for each economy in
the sample. This dummy then captures T(0).
• One must include in the model several growth rates for each country,
i.e. from subsequent time-periods.
17
• This panel data approach is adopted in most recent studies. See Islam
(1995), which you should read cursory (without caring too much about
technical details).
• The results differ quite a bit from those reported by MRW. Most notably
there appears to be a higher speed of convergence and hence a
smaller implied share of capital. In addition, note that the concept of
convergence is now even more conditional and in a sense more hollow.
• A problem with this approach is that available time series are usually
short, so that we must use growth rates for undesirably short periods.
6.7 World income distribution, Twin-Peaks and club-
convergence
• Even if we find evidence of conditional β-convergence it is worth noticing
that absolute σ-divergence is perhaps the really important phe-
nomenon.
• There also appears to be evidence of a twin-peaks phenomenon in the
world distribution of income.
• Recent research focuses much on club-convergence. I.e. how multiple
equilibria might give rise to a twin-peaks like phenomenon.
• The important message is that we have to look closer at structural
characteristics of the economies under consideration.
6.8 “Barro-regressions”
• There is a substantial literature on theory-free regressions of the type
(16).
• These are sometimes referred to as “Barro-regressions” after an influential
early study by Barro.
• Also these studies report fairly similar results about convergence.
• However, the focus of these studies is often directed more towards
exploring the empirical relationship between different variables and
growth.
18
• On these issues there is substantial variation in the results reported and
few robust findings seems to transpire. There are substantial methodological
problems connected to these type of regressions.
• A brief survey of results can be cast as follows:
– Population growth does not seem as detrimental as usually as-
sumed
– High inequality lowers growth, perhaps by raising social and political
instability
– The depth of financial intermediation seems important for growth
– Democracy in itself does not seem to be important
– Extended economic freedom and protection of property rights
seems to be important
– Results are ambiguous over the role of big government and high
taxation
– Government spending on infrastructure is important
– Openness to trade is beneficial for growth, but only under certain
circumstances that have not been fully identified
References
[1] Bils, M. and Klenow, P., Does Schooling Cause Growth?, The American
Economic Review, December 2000, 90 (5). 1160-1183.
[2] Hsieh, C. T., Productivity Growth and Factor Prices in East Asia,
American Economic Review, May 1999, 89 (2). 133-138.
[3] Islam, N., Growth Empirics: a Panel Data Approach, Quarterly Journal
of Economics, 1995, 110 (4). Pp. 1127-1170.
[4] Jones, C.I., On the Evolution of the World Income Distribution, Journal
of Economic Perspectives, 1997, 11 (3). Pp. 19-36.
[5] Klenow, P. and Rodriguez-Clare, A. The Neoclassical Revival in
Growth Economics: Has it gone too Far?, In B.S. Bernanke and J.J.
Rotemberg (eds) NBER Macroeconomics Annual 1997, Cambridge, MA:
MIT Press, 1997.
19
[6] Mankiw, N. G., Romed D. and Weil, N. D., A Contribution to
the Empirics of Economic Growth, Quarterly Journal of Economics,
May 1992, 107 (2), 407-437.
[7] Pritchett, L., Where Has All the Education Gone?, World Bank Economic
Review, 2001, 15 (3). 367-391.
[8] Young, A., The Tyranny of Numbers: Confronting the Statistical Realities
of the East Asian Growth Experience, Quarterly Journal of Economics,
August 1995, 110 (3), 641-680.
20