COHEN & SOTO (2007) GROWTH AND HUMAN CAPITAL: GOOD DATA, GOOD RESULTS
Presented by: Andrew Chen Max Brüning Romaric Sodjahin 1
1) Introduction
2) Data
3) Regression Results
4)Criticism
POLICY IMPLICATIONS? "If you think education is expensive, wait until you see how much ignorance costs in the 21st century.“ Barack Obama
“Education is the most powerful weapon which you can use to change the world.” Nelson Mandela
“I may cut Department of Education.”
Donald Trump
1) Introduction
2) Data
POLICY IMPLICATIONS?
3) Regression Results
4)Criticism
1) Introduction
2) Data
3) Regression Results
4)Criticism
WHY GOOD DATA IS IMPORTANT: MEASUREMENT ERROR 𝑌 = ß0 + ß1 𝑥1 + ⋯ +
ß𝑘 𝑥𝑘∗
+𝑣
1st case: Classical Errors-in-Variables (CEV)
𝑥𝑘∗ = true value 𝑥𝑘 = observed value 𝑒𝑘 = measurement error
𝐶𝑜𝑟𝑟 𝑒𝑘 , 𝑥𝑘 = 0 → 𝐶𝑜𝑟𝑟(𝑒𝑘 , 𝑥𝑘∗ ) ≠ 0 Estimation of all parameters biased towards 0 and inconsistent (attenuation bias) 2nd case 𝐶𝑜𝑟𝑟(𝑒𝑘 , 𝑥𝑘∗ ) = 0 → 𝐶𝑜𝑟𝑟 𝑒𝑘 , 𝑥𝑘 ≠ 0 Estimates are consistent, but variance is greater Necessary Assumptions: E(𝑒𝑘 ) = 0; 𝑒𝑘 uncorrelated with 𝑣 and the other explanatory variables 𝑥𝑗≠𝑘
1) Introduction
2) Data
3) Regression Results
4)Criticism
DATA SOURCES Data used by Cohen & Soto (2007) OECD : data on educational attainment; UNESCO: census and surveys published; data on school
enrollment; National statistical agencies’ censuses on schooling;
Other Data Sources Mitchell(1993, 1998a,b): International historical statistics (data on school enrollment); Barro and Lee (1993, 2001) data; Penn World Table (PWT) mark 5.6 by Summers and Heston; Easterly and Levine (2001) for data on physical capital.
1) Introduction
2) Data
3) Regression Results
VARIABLES Dependent variable : Either annualized change in log(GDP per capita) Or annualized change in log(GDP per worker) Independent variables capital per worker years of schooling Labor Initial value of Years of schooling Initial value of Labor.
4)Criticism
1) Introduction
2) Data
3) Regression Results
SAMPLE OF THE PAPER 95 countries
8 from Middle East and North Africa (MENA); 26 from Sub-Saharan Africa (SSA); 23 from Latin American and Caribbean; 8 from East Asia and the Pacific; 3 from South Asia; 4 from Eastern Europe and Central Asia; 23 high income countries
Data available + Estimated covers 1960-2000.
4)Criticism
BUILDING THE EDUCATION VARIABLE Assuming that there is census data available for years of schooling by age group at year t : 1) For each group age g (g=1 for 15-19 age group, g=2 for 20-24 group and so on), calculate years of schooling of group g noted ystg , as a weighted average of the duration of each level of education j (j=primary, secondary, higher education) by the share of population of group g having attained education level j; 2) Calculate yst , the years of schooling of population aged 15 and above as a weighted average of ysg by the share of group g in the population aged 15 and above. For each year t=1960, 1970, 1980, 1990, 2000 and 2010: we want to build estimates of years of schooling: -
g g 1 If no census data is available for a date before t ---> Assumption that yt 5 yt for group 25-29 until before last group;
-
For groups 15-19 and 20-24 and the last group (eg. 60-64), estimation are made from enrollment data.
-
Then, years of schooling can be computed for date t-5 (or t+5) through a formula.
1) Introduction
2) Data
3) Regression Results
4)Criticism
STYLIZED FACTS
Table 1 :Years of schooling (population 15-64; population weighted averages) High-Income (1) Middle and low income (2) Ratio (1)/(2) (1) – (2)
1960
1970
1980
1990
2000
2010
8.7
9.8
10.9
11.6
12.1
12.5
5.7 2.1 6.4
6.5 1.9 6
2.1 4.1 6.6
2.9 3.7 4.8 3.4 2.9 2.4 6.9 7.2 6.8 Source : Cohen and Soto (2007) + our calculations
1) Introduction
2) Data
3) Regression Results
SOURCES OF MEASUREMENT ERROR Differential Mortality More educated people could have lower mortality rates Backward extrapolation might be biased upwards
Robustness check shows that this issue probably is not important
Immigration
Immigrants could have different education levels than native persons
4)Criticism
1) Introduction
2) Data
3) Regression Results
REGRESSION SPECIFICATION Cross-Section Regression
Δ log 𝑞𝑡 = 𝜋0 + 𝜋1 Δ log 𝑘𝑡 + 𝜋2 Δlog(ℎ𝑡 ) + 𝑋𝑡 𝐵 + 𝜀𝑡
q is output per worker h human capital per worker X is intended to capture convergence or endogenous growth effect
Pritchett’s Approach .
log(ℎ𝑡) = log(𝑐) + log(𝑤0) + log(𝑒0 1𝑦𝑠𝑡 − 1) w0 wage of labor without education w0e0.1yst wage of the worker with ys year of schooling
Mincerian Approach log(ℎ𝑡) = 𝑎 + 𝑏 × 𝑦𝑠𝑡 + 𝑒𝑡
yst the number of schooling of the labor force
4)Criticism
1) Introduction
2) Data
3) Regression Results
4)Criticism
RESULTS 1: BARRO & LEE DATA Dependent variable: annualized change in log(GDP) BS
PR
KL
(1)
(2)
(3)
(4)
(5)
Δlog(k)
.532a
.594a
.595a
.538a
.642a
Δlog(ys)
.070
.045
-0.14
Δ(ys)
.061
.018
ys60
.0016
.0005
Δlog(e0.1yst-1)
Log(k60)
.010a
Log(GDP60)
-.0035
Δlog(L)
-.437a
-.0019
-.005b
Table 1. Income Growth 1960 – 1990 (Barro & Lee data)
-.016a
1) Introduction
2) Data
3) Regression Results
4)Criticism
RESULTS II: COHEN & SOTO DATA Dependent variable: annualized change in log(GDP)
BS 1 (1960–1990)
KL 1 (19601990)
BS 2 (1970-1990)
KL 2 (19701990)
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
Δlog(k)
.521a
.589a
.594a
.516a
.616a
.577a
.595a
.596a
.516a
.549a
Δlog(ys)
.120
-.071
-.085
Δlog(e0.1yst-1)
-.090 .068
.018
Δ(ys)
.096a
.049
.123a
.090a
ys60
.0014b
.0005
.0022a
.0012
Log(k60)
.009a
Log(GDP60)
-.0034
Δlog(L)
-.471a
-.0017
-.0054
-.015a
.0082a -.0018 -.305
Table 2. Income Growth (Cohen & Soto data)
-.0066a
-.015a
1) Introduction
2) Data
3) Regression Results
4)Criticism
PANEL REGRESSION Augmented Solow production function log 𝑞𝑖𝑡 = 𝜋1 log 𝑘𝑖𝑡ൗ𝑞 + 𝜋2𝑦𝑠𝑖𝑡 + 𝜂𝑖 + 𝜏𝑡 + 𝜀𝑖𝑡 𝑖𝑡 qit income per worker yst the number of schooling of the labor force
Fixed CS
Fixed BL
GMM CS 1
GMM BL 1
GMM CS 2
GMM BL 2
CO ratio
.032
0.40
.680
.945a
.700b
.953a
YS
.221a
.120a
.126b
.106
.123b
.105
Table. 3 Income level : Panel Estimation (1960-1990)
1) Introduction
2) Data
3) Regression Results
4)Criticism
CONTRIBUTIONS OF THE PAPER • Improve dataset • Include age groups and mortality heterogeneity, • Standardise census data
• Smaller measurement error
• Get the „right“ results • Previous studies found insignificant or even negative correlation between HC and growth
• Enlightens discussion of economic theory (functional form)
1) Introduction
2) Data
3) Regression Results
4)Criticism
CRITICISM I: SCHOOLING AS A PROXY FOR HUMAN CAPITAL School quality and attendance are neglected Other factors also influence human capital, eg. nutrition and health
Measure human capital directly (e.g. PISA tests)
In regressions that contain both years of schooling and a direct measure of skills, years of schooling was insignificant (Hanushek & Woessmann, 2012)
1) Introduction
2) Data
3) Regression Results
4)Criticism
CRITICISM II Regressions rather represent correlation, not causation Reverse causality: richer countries can afford to invest more into education Omitted variables: institutions, exposure to violence, …
Cohen & Soto assume that human capital does not change after age 25
Barro & Lee (2012) further updated the data set and introduced some further changes for more reliable data (e.g. including more data sources)
Thank you!