Growth And Human Capital: Good Data, Good Results

1
Growth And Human Capital Good Data, Good Results

• Daniel Cohen
• Marcelo Soto
• 2007

2
The Role of Human Capital In Economic Growth

• Lucas (1988) and Romer (1990)
• HC generates long-term sustained growth
• Mankiw, Romer, Weil (1992)
• HC is ordinary input cannot generate endogenous
  growth
• Bils and Klenow (2000)
• Role of HC in economic growth has been overstated

3
Why has role changed over time?

• Problem Measurement of Human Capital
• Conceptually - No clear approach
• MRW - Fraction of GDP diverted to raising HC
• Mincerian - HC exponential function of years of
  schooling
• Empirically - Poor data quality
• Existing HC data unreliable for 21 OECD countries

4
This Papers Goal

• Raise the quality of the data on human capital
• Analyze performance of new data in standard
  growth regressions
• Estimate simple production function to
  incorporate Mincerian approach to HC

5
New Data Methodology

• Two new features
• 1. Heterogeneity between age groups mortality
  rates
• i.e. older people, who on average have lower
  education, have higher mortality rates
• 2. Avoids censuses based on different
  classification systems of education

6
Data Sources

• Main Source
• OECD education censuses
• Secondary Source
• Mitchell (1993, 1998a,b) and UNESCO Statistical
  Yearbook
• Enrollment Data in school

7
Years of Schooling Equation

• Weighted average of different age groups
• Lgt represents the population share of group g in
  population 15 and above
• g1 is the 15-19 age group g2 is 20-24 age,
  etc.
• yst is the number of years of schooling of group
  g

8
Backwards and Forwards Data Extrapolation

• Assume education level unchanged for ages 25
• Backwards Extrapolation
• Forwards Extrapolation

9
Filling In The Missing Data Frog DNA

• Enrollment Data
• Net Intake Rate
• Calculate ratio of new entrants into primary
  school first grade to the 6-year-old population
• Ex 60-65 year olds education level in 1980, so
  look at net intake rate 1922-1926
• Data corrected for Repeaters, Dropouts and Pupil
  Growth

10
Methodology Caveats

• 1. Mortality rate distributed homogenously WITHIN
  each age group
• ysgt-5 lt ysg1t and ysgt5 gt ysg-1t
• 2. Migration
• Assume immigration pop. same educational level as
  home pop.

11
Data Overview

• 95 countries divided into 7 groups
• Relative growth higher than absolute growth
• MENA biggest increase
• SSA not catching up

12
Accuracy Check

• Real vs. Predicted
• Predictions very accurate
• No Forward or Backwards Bias
• Mean Error lt1

13
Accuracy Check

• Real vs. Predicted
• Predictions very accurate
• No Forward or Backwards Bias
• Mean Error lt1

14
Cohen-Soto Vs. Barro-Lee

• Cohen and Soto
• 1. OECD Data
• 2. Heterogenous mortality rates
• 3. Avoids weird results
• 4. Better information in first differences

• Barro and Lee
• 1. UNESCO Data
• 2. Homogenous mortality rates
• 3. Weird results
• 1960 - Bolivia YS French YS
• 4. Worse information in first differences

15
Data Comparison

• Table 3 YS for All
• CS YS numbers higher across board
• Homogenous mortality rate means forward
  underestimate
• Table 4 YS for VEN
• Same overall change from 1960 to 1980
• CS have better first differences information!

16
Data Comparison

• Table 3 YS for All
• CS YS numbers higher across board
• Homogenous mortality rate means forward
  underestimate
• Table 4 YS for VEN
• Same overall change from 1960 to 1980
• CS have better first differences information!

17
Data Comparison

• Table 3 YS for All
• CS YS numbers higher across board
• Homogenous mortality rate means forward
  underestimate
• Table 4 YS for VEN
• Same overall change from 1960 to 1980
• CS have better first differences information!

18
Data Comparison

• Table 3 YS for All
• CS YS numbers higher across board
• Homogenous mortality rate means forward
  underestimate
• Table 4 YS for VEN
• Same overall change from 1960 to 1980
• CS have better first differences information!

19
Other Data Tests

• Correlation tests
• BL about 90 in levels, but only 10 in first
  differences
• Stronger correlation with De La Feunte and
  Domenech
• BL coeff. lower and DD coeff. higher
• DD assume full completion rate - upward bias
• Reliability Ratio tests
• CS perform systemically better than BL
• Both CS and DD have high reliability ratios
• DD does not use enrollment data for missing census

20
Growth and Human Capital

• Analyze the performance of years of schooling
  data in growth regressions
• Another test of quality of new data
• Low informational content in first differences
  could explain lack of statistical significance
  for human capital variable

21
Regression Equation

• q is output per worker
• k is physical capital per worker
• h is human capital per worker
• X is a set of additional variables intended to
  capture convergence or endogenous growth
• Initial levels of income, physical capital, labor

22
Measuring Human Capital

• Problem How do we measure human capital?
• 1. Benhabib and Spiegel (1994)
• Linear relationship between years of schooling
  and HC
• Diminishing returns to additional years of
  education
• 2. Mincer (1974)
• Based on estimated wage regressions
• Each extra year of education boosts human capital
  and wages by the same percentage

23
Testing Models With BL Data

• Table 7 uses Barro and Lee (1993) YS data
• HC variable systematically not significant
• Table 8 uses Barro and Lee (2001) YS data
• HC variable systematically not significant

24
Testing Models With BL Data

• Table 7 uses Barro and Lee (1993) YS data
• HC variable systematically not significant
• Table 8 uses Barro and Lee (2001) YS data
• HC variable systematically not significant

25
Testing Models With CS Data

• Runs same regressions with Cohen and Soto data
• Mincerian HC variable statistically
  significant!!!
• Point estimate (9.6) in line with labor studies

26
Possible Criticisms

• Several variables that may affect growth are
  omitted
• i.e. investment rate and international trade
• Response Regressions same as previous literature
• Outliers potential cause of HC significance
• Response Perform robustness (Least Trimmed
  Squares) - still significant results

27
Income and Human Capital

• Panel Data Estimation
• Solves several drawbacks from previous section
• 1. OLS regressions and No instrumental variables
  mean estimated coefficients likely biased
• 2. Cross-country regressions, so time dimension
  not exploited

28
Panel Regression Equation

• q is aggregate income
• k/q is capital-output ratio
• Ratio avoids collinearity problems
• h is human capital (Mincerian approach)
• Country and time specific effects
• TFP represented as the sum of a fixed effect, a
  time dummy, and a time varying residual
• Assumes common expected tech. growth across
  countries
• Strong assumption!!!

29
Panel Data Regression Results

• Fixed Effect regression
• Capital-output downward biased
• Both CS and BL schooling variable significant
• CS twice as large
• Less measurement error

30
Panel Data Regression Results

• Fixed Effect regression
• Capital-output downward biased
• Both CS and BL schooling variable significant
• CS twice as large
• Less measurement error

31
Panel Regression Results Ctd.

• GMM estimator regression (inst. var.)
• Joint estimation of eq. using levels and first
  differences
• Capital-output larger coefficients
• Both CS schooling significant at 5 level
• Neither BL schooling significant
• One additional year of schooling 12 income
  increase

32
Panel Regression Results Ctd.

• GMM estimator regression (inst. var.)
• Joint estimation of eq. using levels and first
  differences
• Capital-output larger coefficients
• Both CS schooling significant at 5 level
• Neither BL schooling significant
• One additional year of schooling 12 income
  increase

33
Conclusion

• Good data - Cohen and Soto introduce better
  method of estimating years of schooling data
• Account for age structure of population - HMR
• Avoid sources with different classifications
• Good Results - CS data performs better than
  BL in cross-country growth regressions and panel
  data regressions
• CS statistically significant for HC variable
  BL not
• Education has positive and significant long-term
  effect on growth of income per capita!!!

34
Discussion Questions

• CS find that 1 extra year of schooling results
  in 12 increase in income. Is this plausible?
• What about potential breaks in education like
• The HC capital variable was only statistically
  significant when using the Mincerian approach to
  human capital. Does this undermine their
  improved data argument?
• CS and DD use same source and are highly
  correllated. Is the low correlation between BL
  data and CS data due to mostly to different
  sources (OECD vs. UNESCO), or is CS data method
  just better?