Lecture 3: The Solow Model and the Data

Econ 602 Spring 2023

Ibn Haldun University

February 9, 2023

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 1 / 58
 Mapping the Model to Data Introduction

Solow Growth Model and the Data

Use Solow model or extensions to interpret both economic growth

over time and cross-country output differences.
Focus on proximate causes of economic growth.

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 2 / 58
 Mapping the Model to Data Growth Accounting

Growth Accounting I

Aggregate production function in its general form:

Y (t ) = F [K (t ) , L (t ) , A (t )] .

Combined with competitive factor markets, gives Solow (1957)

growth accounting framework.
Continuous-time economy and differentiate the aggregate production
function with respect to time.
Dropping time dependence,

Ẏ F A Ȧ FK K K̇ F L L̇
= A + + L . (1)
Y Y A Y K Y L

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 3 / 58
 Mapping the Model to Data Growth Accounting

Growth Accounting II

Denote growth rates of output, capital stock and labor by g ≡ Ẏ /Y ,

gK ≡ K̇ /K and gL ≡ L̇/L.
Define the contribution of technology to growth as

FA A Ȧ
x≡
Y A
Recall with competitive factor markets, w = FL and R = FK .
Define factor shares as αK ≡ RK /Y and αL ≡ wL/Y .
Putting all these together, (1) the fundamental growth accounting
equation
x = g − αK gK − αL gL . (2)
Gives estimate of contribution of technological progress, Total Factor
Productivity (TFP) or Multi Factor Productivity.

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 4 / 58
 Mapping the Model to Data Growth Accounting

Growth Accounting III

Denoting an estimate by “ˆ”:

x̂ (t ) = g (t ) − αK (t ) gK (t ) − αL (t ) gL (t ) . (3)

All terms on right-hand side are “estimates” obtained with a range of

assumptions from national accounts and other data sources.
If interested in Ȧ/A rather than x, need further assumptions. For
example, if we assume

Y (t ) = F̃ [K (t ) , A (t ) L (t )] ,

then
Ȧ 1
= [g − αK gK − αL gL ] ,
A αL
But not particularly useful,the economically interesting object is x̂ in
(3).
Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 5 / 58
 Mapping the Model to Data Growth Accounting

Growth Accounting IV

In continuous time, equation (3) is exact.

With discrete time, potential problem in using (3): over the time
horizon factor shares can change.
Use beginning-of-period or end-of-period values of αK and αL ?
Either might lead to seriously biased estimates.
Best way of avoiding such biases is to use as high-frequency data as
possible.
Typically use factor shares calculated as the average of the beginning
and end of period values.
In discrete time, the analog of equation (3) becomes

x̂t,t +1 = gt,t +1 − ᾱK ,t,t +1 gK ,t,t +1 − ᾱL,t,t +1 gL,t,t +1 , (4)

gt,t +1 is the growth rate of output between t and t + 1; other growth

rates defined analogously.
Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 6 / 58
 Mapping the Model to Data Growth Accounting

Growth Accounting V
Moreover,
αK (t ) + αK (t + 1)
ᾱK ,t,t +1 ≡
2
αL (t ) + αL (t + 1)
and ᾱL,t,t +1 ≡
2
Equation (4) would be a fairly good approximation to (3) when the
difference between t and t + 1 is small and the capital-labor ratio
does not change much during this time interval.
Solow’s (1957) applied this framework to US data: a large part of the
growth was due to technological progress.
From early days, however, a number of pitfalls were recognized.
Moses Abramovitz (1956): dubbed the x̂ term “the measure of our
ignorance”.
If we mismeasure gL and gK we will arrive at inflated estimates of x̂.
Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 7 / 58
 Mapping the Model to Data Growth Accounting

Growth Accounting VI

Reasons for mismeasurement:

what matters is not labor hours, but effective labor hours
important—though difficult—to make adjustments for changes in the
human capital of workers.
measurement of capital inputs:
in the theoretical model, capital corresponds to the final good used as
input to produce more goods.
in practice, capital is machinery, need assumptions about how relative
prices of machinery change over time.
typical assumption was to use capital expenditures but if machines
become cheaper would severely underestimate gK

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 8 / 58
 Mapping the Model to Data Regression Analysis

Solow Model and Regression Analyses I

Another popular approach of taking the Solow model to data: growth

regressions, following Barro (1991).
Return to basic Solow model with constant population growth and
labor-augmenting technological change in continuous time:

y (t ) = A (t ) f (k (t )) , (5)

and
k̇ (t ) sf (k (t ))
= − δ − g − n, (6)
k (t ) k (t )

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 9 / 58
 Mapping the Model to Data Regression Analysis

Solow Model and Regression Analyses II

Differentiating (5) with respect to time and dividing both sides by

y (t ),
ẏ (t ) k̇ (t )
= g + ε f (k (t )) , (7)
y (t ) k (t )
where
f ′ (k (t )) k (t )
ε f (k (t )) ≡ ∈ (0, 1)
f (k (t ))
is the elasticity of the f (·) function.
ε f (k (t )) is between 0 and 1 follows from Assumption 1. For
example, with Cobb-Douglas ε f (k (t )) = α, but generally a function
of k (t ).

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 10 / 58
 Mapping the Model to Data Regression Analysis

Solow Model and Regression Analyses III

First-order Taylor expansion of (6) with respect to log k (t ) around k ∗
(and recall that ∂y /∂ log x = (∂y /∂x ) · x):
sf (k ∗ )
 
k̇ (t )
≃ −δ−g −n
k (t ) k∗
 ′ ∗ ∗
f (k ∗ )

f (k ) k
+ − 1 s (log k (t ) − log k ∗ ) .
f (k ∗ ) k∗
≃ (ε f (k ∗ ) − 1) (δ + g + n) (log k (t ) − log k ∗ ) .
First term in the first line is zero by definition of the steady-state
value k ∗ .
Also used definition of ε f (k (t )) and the fact that
sf (k ∗ ) /k ∗ = δ + g + n.
Substituting into (7),
ẏ (t )
≃ g − ε f (k ∗ ) (1 − ε f (k ∗ )) (δ + g + n) (log k (t ) − log k ∗ ) .
y (t )
Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 11 / 58
 Mapping the Model to Data Regression Analysis

Solow Model and Regression Analyses IV

Define y ∗ (t ) ≡ A (t ) f (k ∗ ); refer to y ∗ (t ) as the “steady-state level

of output per capita” even though it is not constant.
First-order Taylor expansions of log y (t ) with respect to log k (t )
around log k ∗ (t ):

log y (t ) − log y ∗ (t ) ≃ ε f (k ∗ ) (log k (t ) − log k ∗ ) .

Combining this with the previous equation, “convergence equation”:

ẏ (t )
≃ g − (1 − ε f (k ∗ )) (δ + g + n) (log y (t ) − log y ∗ (t )) . (8)
y (t )

Two sources of growth in Solow model: g , the rate of technological

progress, and “convergence”.

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 12 / 58
 Mapping the Model to Data Regression Analysis

Solow Model and Regression Analyses V

Latter source, convergence:

Negative impact of the gap between current level and steady-state level
of output per capita on rate of capital accumulation (recall
0 < ε f (k ∗ ) < 1).
The lower is y (t ) relative to y ∗ (t ), hence the lower is k (t ) relative to
k ∗ , the greater is log (kt ) − log (kt∗ ), and this leads to faster growth in
the effective capital-labor ratio.
Speed of convergence in (8), measured by the term
(1 − ε f (k ∗ )) (δ + g + n), depends on:
δ + g + n : determines rate at which effective capital-labor ratio needs
to be replenished.
ε f (k ∗ ) : when ε f (k ∗ ) is high, we are close to a
linear—AK —production function, convergence should be slow.

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 13 / 58
 Mapping the Model to Data Regression Analysis

Example: Cobb-Douglas production function and

convergence I

Consider Cobb-Douglas production function

Y (t ) = A (t ) K (t ) α L (t )1− α .
Implies that y (t ) = A (t ) k (t )α , ε f (k (t )) = α. Therefore, (8)
becomes
ẏ (t )
≃ g − (1 − α) (δ + g + n) (log y (t ) − log y ∗ (t )) .
y (t )

Enables us to “calibrate” the speed of convergence in practice

Focus on advanced economies
g ≃ 0.02 for approximately 2% per year output per capita growth,
n ≃ 0.01 for approximately 1% population growth and
δ ≃ 0.05 for about 5% per year depreciation.
Share of capital in national income is about 1/3, so α ≃ 1/3.

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 14 / 58
 Mapping the Model to Data Regression Analysis

Example: Cobb-Douglas production function and

convergence II

Thus convergence coefficient would be around 0.054 (≃ 0.67 × 0.08).

Very rapid rate of convergence:
gap of income between two similar countries should be halved in little
more than 10 years
At odds with the patterns we saw before.

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 15 / 58
 Mapping the Model to Data Regression Analysis

Solow Model and Regression Analyses VI

Using (8), we can obtain a growth regression similar to those

estimated by Barro (1991).
Using discrete time approximations, equation (8) yields:

gi,t,t −1 = b 0 + b 1 log yi,t −1 + ε i,t , (9)

ε i,t is a stochastic term capturing all omitted influences.

If such an equation is estimated in the sample of core OECD
countries, b 1 is indeed estimated to be negative.
But for the whole world, no evidence for a negative b 1 . If anything,
b 1 would be positive.
I.e., there is no evidence of world-wide convergence,
Barro and Sala-i-Martin refer to this as “unconditional convergence.”

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 16 / 58
 Mapping the Model to Data Regression Analysis

Solow Model and Regression Analyses VII

Unconditional convergence may be too demanding:

requires income gap between any two countries to decline, irrespective
of what types of technological opportunities, investment behavior,
policies and institutions these countries have.
If countries do differ, Solow model would not predict that they should
converge in income level.
If countries differ according to their characteristics, a more
appropriate regression equation may be:

gi,t,t −1 = bi0 + b 1 log yi,t −1 + ε i,t , (10)

Now the constant term, bi0 , is country specific.

Slope term, measuring the speed of convergence, b 1 , should also be
country specific.
May then model bi0 as a function of certain country characteristics.
Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 17 / 58
 Mapping the Model to Data Regression Analysis

Solow Model and Regression Analyses VII

If the true equation is (10), (9) would not be a good fit to the data.
I.e., there is no guarantee that the estimates of b 1 resulting from this
equation will be negative.
In particular, it is natural to expect that Cov bi0 , log yi,t −1 > 0:

economies with certain growth-reducing characteristics will have low

levels of output.
Implies an upward bias in the estimate of b 1 in equation (9), when the
more appropriate equation is (10).
With this motivation, Barro (1991) and Barro and Sala-i-Martin
(2004) favor the notion of “conditional convergence:”
convergence effects should lead to negative estimates of b 1 once bi0 is
allowed to vary across countries.

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 18 / 58
 Mapping the Model to Data Regression Analysis

Solow Model and Regression Analyses VIII

Barro (1991) and Barro and Sala-i-Martin (2004) estimate models
where bi0 is assumed to be a function of:
male schooling rate, female schooling rate, fertility rate, investment
rate, government-consumption ratio, inflation rate, changes in terms of
trades, openness and institutional variables such as rule of law and
democracy.
In regression form,
′
gi,t,t −1 = Xi,t β + b 1 log yi,t −1 + ε i,t , (11)
Xi,t is a (column) vector including the variables mentioned above
(and a constant).
Imposes that bi0 in equation (10) can be approximated by Xi,t′ β.

Conditional convergence: regressions of (11) tend to show a negative

estimate of b 1 .
But the magnitude is much lower than that suggested by the
computations in the Cobb-Douglas Example.
Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 19 / 58
 Mapping the Model to Data Regression Analysis

Drawbacks of Growth Regressions I

Regressions similar to (11) have not only been used to support

“conditional convergence,” but also to estimate the “determinants of
economic growth”.
Coefficient vector β: information about causal effects of various
variables on economic growth.
Several problematic features with regressions of this form. These
include:
Many variables in Xi,t and log yi,t −1 , are econometrically
endogenous: jointly determined gi ,t,t −1 .
May argue b 1 is of interest even without “causal interpretation”.
But if Xi ,t is econometrically endogenous, estimate of b 1 will also be
inconsistent (unless Xi ,t is independent from log yi ,t −1 ).

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 20 / 58
 Mapping the Model to Data Regression Analysis

Drawbacks of Growth Regressions II

Even if Xi,t ’s were econometrically exogenous, a negative b 1
could be by measurement error or other transitory shocks to
yi,t .
For example, suppose we only observe ỹi,t = yi,t exp (ui,t ).
Note

log ỹi ,t − log ỹi ,t −1 = log yi ,t − log yi ,t −1 + ui ,t − ui ,t −1 .

Since measured growth is

g̃i ,t ,t −1 ≈ log ỹi ,t − log ỹi ,t −1 = log yi ,t − log yi ,t −1 + ui ,t − ui ,t −1 ,
when we look at the growth regression

g̃i ,t ,t −1 = Xi′ ,t β + b 1 log ỹi ,t −1 + ε i ,t ,

measurement error ui ,t −1 will be part of both ε i ,t and

log ỹi ,t −1 = log yi ,t −1 + ui ,t −1 : negative bias in the estimation of b 1 .
Thus can end up negative estimate of b 1 , even when there is no
conditional convergence.
Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 21 / 58
 Mapping the Model to Data Regression Analysis

Drawbacks of Growth Regressions III

Interpretation of regression equations like (11) is not always

straightforward
Many of the growth regressions include investment rate as part of the
vector Xi ,t
However in the Solow model, differences in investment rates are the
channel by which the potential determinants included in Xi ,t will
influence economic growth.
Thus conditional on investment rate, coefficients of other variables no
longer measure their full impact on economic growth.
Hence estimates of growth equations with investment like variables are
difficult to link to theory.
Equation for (8) is derived for closed Solow economy.

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 22 / 58
 Mapping the Model to Data Regression Analysis

Drawbacks of Growth Regressions III

This discussion does not imply that growth regressions are
uninformative.
At the basic level they provide important information on salient
correlations in the data.
Knowing these correlations are an important input to the process of
formulating empirically plausible theories.
A complementary regression framework for investigating correlations
in the data is
′
log yi,t = α log yi,t −1 Xi,t β + δi + µt + ε i,t
Here δi are the country fixed effects. B including these, the regression
equation takes out country fixed factors that might be simultaneously
affecting economic growth and the right-hand side variables.
Therefore panel data regressions are more informative about the
statistical relation between a range of factors and income per capita.
Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 23 / 58
 The Solow Model with Human Capital Human Capital

The Solow Model with Human Capital I

Labor hours supplied by different individuals do not contain the same
efficiency units.
Focus on the continuous time economy and suppose:

Y = F (K , H, AL) , (12)

where H denotes “human capital”.

Assume throughout that A > 0.
Assume F : R3+ → R+ in (12) is twice continuously differentiable in
K , H and L, and satisfies the equivalent of the neoclassical
assumptions.
Households save a fraction sk of their income to invest in physical
capital and a fraction sh to invest in human capital.
Human capital also depreciates in the same way as physical capital,
denote depreciation rates by δk and δh .
Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 24 / 58
 The Solow Model with Human Capital Human Capital

The Solow Model with Human Capital III

Assume constant population growth and a constant rate of
labor-augmenting technological progress, i.e.,
L̇ (t ) Ȧ (t )
= n and = g.
L (t ) A (t )
Defining effective human and physical capital ratios as
K (t ) H (t )
k (t ) ≡ and h (t ) ≡ ,
A (t ) L (t ) A (t ) L (t )
Using the constant returns to scale, output per effective unit of labor
can be written as
Y (t )
ŷ (t ) ≡
A (t ) L (t )
 
K (t ) H (t )
= F , ,1
A (t ) L (t ) A (t ) L (t )
≡ f (k (t ) , h (t )) .
Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 25 / 58
 The Solow Model with Human Capital Human Capital

The Solow Model with Human Capital IV

Law of motion of k (t ) and h (t ) can then be obtained as:

k̇ (t ) = sk f (k (t ) , h (t )) − (δk + g + n ) k (t ) ,
ḣ (t ) = sh f (k (t ) , h (t )) − (δh + g + n ) h (t ) .

Steady-state equilibrium: effective human and physical capital ratios,

(k ∗ , h∗ ), which satisfiy:

sk f (k ∗ , h∗ ) − (δk + g + n ) k ∗ = 0, (13)

and
sh f (k ∗ , h∗ ) − (δh + g + n ) h∗ = 0. (14)

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 26 / 58
 The Solow Model with Human Capital Human Capital

The Solow Model with Human Capital V

Focus on steady-state equilibria with k ∗ > 0 and h∗ > 0 (if

f (0, 0) = 0, then there exists a trivial steady state with k = h = 0,
which we ignore it).
Can first prove that steady-state equilibrium is unique. To see this
heuristically, consider the Figure in the (k, h ) space.
Both lines are upward sloping, but proof of next proposition shows
(14) is always shallower in the (k, h ) space, so the two curves can
only intersect once.
Proposition In the augmented Solow model with human capital, there
exists a unique, globally stable steady-state equilibrium
(k ∗ , h ∗ ).

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 27 / 58
 The Solow Model with Human Capital Human Capital

Figure: Dynamics of physical capital-labor and human capital-labor ratios in the

Solow model with human capital.

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 28 / 58
 The Solow Model with Human Capital Example

Example: Augmented Solow model with Cobb-Douglas

production function I

Aggregate production function is

Y (t ) = K (t )α H (t ) β (A (t ) L (t ))1−α− β , (15)

where 0 < α < 1, 0 < β < 1 and α + β < 1.

Output per effective unit of labor can then be written as

ŷ (t ) = k α (t ) h β (t ) ,

with the same definition of ŷ (t ), k (t ) and h (t ) as above.

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 29 / 58
 The Solow Model with Human Capital Example

Example: Augmented Solow model with Cobb-Douglas

production function II
Using this functional form, (13) and (14) give the unique steady-state
equilibrium:
  1− β   β ! 1−α1−β
∗ sk sh
k = (16)
n + g + δk n + g + δh
 α  1−α ! 1−α1−β
sk sh
h∗ = ,
n + g + δk n + g + δh

Higher saving rate in physical capital not only increases k ∗ , but also
h∗ .
Same applies for a higher saving rate in human capital.
Reflects that higher k ∗ raises overall output and thus the amount
invested in schooling (since sh is constant).
Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 30 / 58
 The Solow Model with Human Capital Example

Example: Augmented Solow model with Cobb-Douglas

production function III

Given (16), output per effective unit of labor in steady state is

obtained as
  1−αα−β   1−αβ−β
∗ sk sh
ŷ = . (17)
n + g + δk n + g + δh

Relative contributions of the saving rates depends on the shares of

physical and human capital:
the larger is α, the more important is sk and the larger is β, the more
important is sh .

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 31 / 58
 Regression Analysis A World of Augmented Solow Economies

A World of Augmented Solow Economies I

Mankiw, Romer and Weil (1992) used regression analysis to take the
augmented Solow model, with human capital, to data.
Use the Cobb-Douglas model and envisage a world consisting of
j = 1, ..., N countries.
“Each country is an island”: countries do not interact (perhaps
except for sharing some common technology growth).
Country j = 1, ..., N has the aggregate production function:

Yj (t ) = Kj (t )α Hj (t ) β (Aj (t ) Lj (t ))1−α− β .

Nests the basic Solow model without human capital when β = 0.

Countries differ in terms of their saving rates, sk,j and sh,j , population
growth rates, nj , and technology growth rates Ȧj (t ) /Aj (t ) = gj .
Define kj ≡ Kj /Aj Lj and hj ≡ Hj /Aj Lj .

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 32 / 58
 Regression Analysis A World of Augmented Solow Economies

A World of Augmented Solow Economies II

Focus on a world in which each country is in their steady state
Equivalents of equations (16) apply here and imply:
  1− β   β ! 1−α1−β
s k,j s h,j
kj∗ =
nj + gj + δk nj + gj + δh
 α  1−α ! 1−α1−β
s k,j s h,j
hj∗ = .
nj + gj + δk nj + gj + δh
Consequently, using (17), the “steady-state”/balanced growth path
income per capita of country j can be written as
Y (t )
yj∗ (t ) ≡ (18)
L (t )
  1−αα−β   1−αβ−β
sk,j sh,j
= Aj (t ) .
nj + gj + δk nj + gj + δh
Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 33 / 58
 Regression Analysis A World of Augmented Solow Economies

A World of Augmented Solow Economies II

Here yj∗ (t ) stands for output per capita of country j along the
balanced growth path.
Note if gj ’s are not equal across countries, income per capita will
diverge.
Mankiw, Romer and Weil (1992) make the following assumption:

Aj (t ) = Āj exp (gt ) .

Countries differ according to technology level, (initial level Āj ) but

they share the same common technology growth rate, g .

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 34 / 58
 Regression Analysis A World of Augmented Solow Economies

A World of Augmented Solow Economies III

Using this together with (18) and taking logs, equation for the
balanced growth path of income for country j = 1, ..., N:
 
∗ α sk,j
ln yj (t ) = ln Āj + gt + ln (19)
1−α−β nj + g + δk
 
β sh,j
+ ln .
1−α−β nj + g + δh

Mankiw, Romer and Weil (1992) take:

δk = δh = δ and δ + g = 0.05.
sk ,j =average investment rates (investments/GDP).
sh,j =fraction of the school-age population that is enrolled in secondary
school.
nj = population growth rates

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 35 / 58
 Regression Analysis A World of Augmented Solow Economies

A World of Augmented Solow Economies IV

Even with all of these assumptions, (19) can still not be estimated
consistently.
ln Āj is unobserved (at least to the econometrician) and thus will be
captured by the error term.
Most reasonable models would suggest ln Āj ’s should be correlated
with investment rates.
Thus an estimation of (19) would lead to omitted variable bias and
inconsistent estimates.
Implicitly, MRW make another crucial assumption, the orthogonal
technology assumption:

Āj = ε j A, with ε j orthogonal to all other variables.

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 36 / 58
 Regression Analysis A World of Augmented Solow Economies

Cross-Country Income Differences: Regressions I

MRW first estimate equation (19) without the human capital term for
the cross-sectional sample of non-oil producing countries
α α
ln yj∗ = constant + ln (sk,j ) − ln (nj + g + δk ) + ε j .
1−α 1−α

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 37 / 58
 Regression Analysis A World of Augmented Solow Economies

Cross-Country Income Differences: Regressions II

Estimates of the Basic Solow Model
MRW Updated data
1985 1985 2000

ln(sk ) 1.42 1.01 1.22

(.14) (.11) (.13)

ln(n + g + δ) -1.97 -1.12 -1.31

(.56) (.55) (.36)

Adj R2 .59 .49 .49

Implied β .59 .50 .55

No. of observations 98 98 107

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 38 / 58
 Regression Analysis A World of Augmented Solow Economies

Cross-Country Income Differences: Regressions III

Their estimates for α/ (1 − α), implies that α must be around 2/3,

but should be around 1/3.
The most natural reason for the high implied values of α is that ε j is
correlated with ln (sk,j ), either because:
1 the orthogonal technology assumption is not a good approximation to
reality or

2 there are also human capital differences correlated with ln sk ,j .
Mankiw, Romer and Weil favor the second interpretation and
estimate the augmented model,
α α
ln yj∗ = cst + ln (sk,j ) − ln (nj + g + δk )(20)
1−α−β 1−α−β
β β
+ ln (sh,j ) − ln (nj + g + δh ) + ε j .
1−α−β 1−α−β

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 39 / 58
 Regression Analysis A World of Augmented Solow Economies

Estimates of the Augmented Solow Model

MRW Updated data
1985 1985 2000

ln(sk ) .69 .65 .96

(.13) (.11) (.13)

ln(n + g + δ) -1.73 -1.02 -1.06

(.41) (.45) (.33)

ln(sh ) .66 .47 .70

(.07) (.07) (.13)

Adj R2 .78 .65 .60

Implied β .30 .31 .36

Implied α .28 .22 .26
No. of observations 98 98 107
Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 40 / 58
 Regression Analysis A World of Augmented Solow Economies

Cross-Country Income Differences: Regressions IV

If these regression results are reliable, they give a big boost to the
augmented Solow model.
Adjusted R 2 suggests that three quarters of income per capita
differences across countries can be explained by differences in their
physical and human capital investment.
Immediate implication is technology (TFP) differences have a
somewhat limited role.
But this conclusion should not be accepted without further
investigation.

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 41 / 58
 Regression Analysis Challenges to Regression Analyses

Challenges to the Regression Analyses I

Technology differences across countries are not orthogonal to

all other variables.
Āj is correlated with measures of sjh and sjk for two reasons.
1 omitted variable bias: societies with high Āj will be those that have
invested more in technology for various reasons; same reasons likely to
induce greater investment in physical and human capital as well.
2 reverse causality: complementarity between technology and physical or
human capital imply that countries with high Āj will find it more
beneficial to increase their stock of human and physical capital.
In terms of (20), implies that key right-hand side variables are
correlated with the error term, ε j .
OLS estimates of α and β and R 2 are biased upwards.

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 42 / 58
 Regression Analysis Challenges to Regression Analyses

Challenges to the Regression Analyses II

β is too large relative to what we should expect on the basis of

microeconometric evidence.
The working age population enrolled in school ranges from 0.4% to
over 12% in the sample of countries.
Predicted log difference in incomes between these two countries is

β
(ln 12 − ln (0.4)) = 0.7 × (ln 12 − ln (0.4)) ≈ 2.38.
1−α−β

Thus a country with schooling investment of over 12 should be about

exp (2.38) − 1 ≈ 10.8 times richer than one with investment of
around 0.4.

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 43 / 58
 Regression Analysis Challenges to Regression Analyses

Challenges to the Regression Analyses III

Take Mincer regressions of the form:

ln wi = Xi′ γ + ϕSi + ui , (21)

Microeconometrics literature suggests that ϕ is between 0.06 and

0.10.
Can deduce how much richer a country with 12 if we assume:
1 That the micro-level relationship as captured by (21) applies identically
to all countries.
2 Wages can be expressed as wi = ϕ̃(Si ) ≈ exp (ϕSi ).
3 That there are no human capital externalities.

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 44 / 58
 Regression Analysis Challenges to Regression Analyses

Challenges to the Regression Analyses IV

Suppose that each firm f in country j has access to the production

function
yfj = Kf1−α (Aj Hf )α ,
Suppose also that firms in this country face a cost of capital equal to
Rj . With perfectly competitive factor markets,
 −α
Kf
Rj = ( 1 − α ) . (22)
Aj Hf

Implies all firms ought to function at the same physical to human

capital ratio.
Thus all workers, irrespective of level of schooling, ought to work at
the same physical to human capital ratio.

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 45 / 58
 Regression Analysis Challenges to Regression Analyses

Challenges to the Regression Analyses V

Another direct implication of competitive labor markets is that in

country j,
−(1−α)/α
wj = α (1 − α)(1−α)/α Aj Rj .
Consequently, a worker with human capital hi will receive a wage
income of wj hi .
Next, substituting for capital from (22), we have total income in
country j as
−(1−α)/α
Yj = (1 − α)(1−α)/α Rj Aj Hj ,
where Hj is the total efficiency units of labor in country j.

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 46 / 58
 Regression Analysis Challenges to Regression Analyses

Challenges to the Regression Analyses V

Implies that ceteris paribus (in particular, holding constant capital

intensity corresponding to Rj and technology, Aj ), a doubling of
human capital will translate into a doubling of total income.
So in the absence of human capital externalities: a country with 12
more years of average schooling should have between
exp (0.10 × 12) ≃ 3.3 and exp (0.06 × 12) ≃ 2.05 times the stock of
human capital of a county with fewer years of schooling.

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 47 / 58
 Regression Analysis Challenges to Regression Analyses

Challenges to the Regression Analyses VI

Thus holding other factors constant, this country should be about 2-3
times as rich as the country with zero years of average schooling.
Much less than the 11 fold difference implied by the
Mankiw-Romer-Weil analysis.
Thus β in MRW is too high relative to the estimates implied by the
microeconometric evidence and thus likely upwardly biased.
Overestimation of β is, in turn, most likely related to correlation
between the error term ε j and the key right-hand side regressors in
(20).

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 48 / 58
 Regression Analysis Calibrating Productivity Differences

Calibrating Productivity Differences I

Suppose each country has access to the Cobb-Douglas aggregate

production function:

Yj = Kj1−α (Aj Hj )α , (23)

Each worker in country j has Sj years of schooling.

Then using the Mincer equation (21) ignoring the other covariates
and taking exponents, Hj can be estimated as

Hj = exp (ϕSj ) Lj ,

Does not take into account differences in other “human capital”

factors, such as experience.

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 49 / 58
 Regression Analysis Calibrating Productivity Differences

Calibrating Productivity Differences II

Let the rate of return to acquiring the Sth year of schooling be ϕ (S ).

A better estimate of the stock of human capital can be constructed as

Hj = ∑ exp {ϕ (S ) S } Lj (S )
S

Lj (S ) now refers to the total employment of workers with S years of

schooling in country j.
Series for Kj can be constructed from Summers-Heston dataset using
investment data and the perpetual inventory method.

Kj (t + 1) = (1 − δ) Kj (t ) + Ij (t ) ,

Assume, following Hall and Jones that δ = 0.06.

With same arguments as before, choose a value of 2/3 for α.

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 50 / 58
 Regression Analysis Calibrating Productivity Differences

Calibrating Productivity Differences III

Given series for Hj and Kj and a value for α, construct “predicted”
incomes at a point in time using

Ŷj = Kj1/3 (AUS Hj )2/3

1/3
AUS is computed so that YUS = KUS (AUS HUS )2/3 .
Once a series for Ŷj has been constructed, it can be compared to the
actual output series.
Gap between the two series represents the contribution of technology.
Alternatively, could back out country-specific technology terms
(relative to the United States) as
 3/2  1/2  
Aj Yj KUS HUS
= .
AUS YUS Kj Hj

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 51 / 58
 Regression Analysis Calibrating Productivity Differences

Calibrating Productivity Differences IV

Figure: Calibrated technology levels relative to the US technology (from the

Solow growth model with human capital) versus log GDP per worker, 1980, 1990
and 2000.

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 52 / 58
 Regression Analysis Calibrating Productivity Differences

Calibrating Productivity Differences V

Figure: Calibrated technology levels relative to the US technology (from the

Solow growth model with human capital) versus log GDP per worker, 1980, 1990
and 2000.

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 53 / 58
 Regression Analysis Calibrating Productivity Differences

Calibrating Productivity Differences VI

The following features are noteworthy:

1 Differences in physical and human capital still matter a lot.
2 However, differently from the regression analysis, this exercise also
shows significant technology (productivity) differences.
3 Same pattern visible in the next three figures for the estimates of the
technology differences, Aj /AUS , against log GDP per capita in the
corresponding year.
4 Also interesting is the pattern that the empirical fit of the Solow
growth model seems to deteriorate over time. Observations are further
above the 45-degree line in 2000 than in 1980. Why is the fit of the
Solow model better in 1980 than in 2000 is an unanswered question.

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 54 / 58
 Regression Analysis Challenges to Callibration

Challenges to Callibration I

In addition to the standard assumptions of competitive factor

markets, we had to assume :
no human capital externalities, a Cobb-Douglas production function,
and a range of approximations to measure cross-country differences in
the stocks of physical and human capital.
The calibration approach is in fact a close cousin of the
growth-accounting exercise (sometimes referred to as “levels
accounting”).
Imagine that the production function that applies to all countries in
the world is
F (Kj , Hj , Aj ) ,
Assume countries differ according to their physical and human capital
as well as technology—but not according to F .

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 55 / 58
 Regression Analysis Challenges to Callibration

Challenges to Callibration II

Rank countries in descending order according to their physical capital

to human capital ratios, Kj /Hj Then

x̂j,j +1 = gj,j +1 − ᾱK ,j,j +1 gK ,j,j +1 − ᾱLj,j +1 gH,j,j +1 , (24)

where:
gj ,j +1 : proportional difference in output between countries j and j + 1,
gK ,j ,j +1 : proportional difference in capital stock between these
countries and
gH ,j ,j +1 : proportional difference in human capital stocks.
ᾱK ,j ,j +1 and ᾱLj ,j +1 : average capital and labor shares between the two
countries.
The estimate x̂j,j +1 is then the proportional TFP difference between
the two countries.

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 56 / 58
 Regression Analysis Challenges to Callibration

Challenges to Callibration III

Levels-accounting faces two challenges.

1 Data on capital and labor shares across countries are not widely
available. Almost all exercises use the Cobb-Douglas approach (i.e., a
constant value of αK equal to 1/3).
2 The differences in factor proportions, e.g., differences in Kj /Hj , across
countries are large. An equation like (24) is a good approximation
when we consider small (infinitesimal) changes.

Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 57 / 58
 Conclusions Conclusions

Conclusions
Message is somewhat mixed.
On the positive side, despite its simplicity, the Solow model has enough
substance that we can take it to data in various different forms,
including TFP accounting, regression analysis and calibration.
On the negative side, however, no single approach is entirely
convincing.
Complete agreement is not possible, but safe to say that consensus
favors the interpretation that cross-country differences in income per
capita cannot be understood solely on the basis of differences in
physical and human capital
Differences in TFP are not necessarily due to technology in the
narrow sense.
Have not examined fundamental causes of differences in prosperity:
why some societies make choices that lead them to low physical
capital, low human capital and inefficient technology and thus to
relative poverty.
Econ 602 Spring 2023 (Ibn Haldun University) Lecture 3 February 9, 2023 58 / 58