Transcript Empirical Evidence on Growth: A Closer Look on Cross

Empirical Evidence on Growth:
A Closer Look on CrossCountry Regressions
Presentation by
Dejan Jasnić and Philipp Wahlen
24. November 2008
Regression Analysis
Characteristics of relationship between dependent
variable and independent variables y = f(x)
 strength, direction, type

Simple linear regressions:
y = a + bx

Multivariate regressions:
y =  + 1x1 + 2x2 +… + nxn
24.11.2008
Cross-Country Regressions
2
Regression Analysis – scatter plot
What does it tell us about strength, direction and type of the
relationship?
80
80
78
78
76
76
74
74
72
72
70
70
68
68
66
66
64
64
62
62
60
60
60
60
24.11.2008
5,060
5,060
10,060
10,060
15,060
15,060
Cross-Country Regressions
20,060
20,060
25,060
3
Regression Analysis – growth regressions
Multivariate regressions: y =  + 1x1 + 2x2 +… + nxn
Complexity
What is the right model? Right variables?
More major shortcomings:
 parameter heterogeneity, outliers, measurement error, endogenity
 assumption that economic growth operates according to universal
laws across countries through the time
24.11.2008
Cross-Country Regressions
4
Cross-Country Regressions


Recall:
By regressing
annual growth on
political and economic indicators
across countries
researchers attempt to find generally valid
driving forces of economic growth.
What might be problematic with this
approach?
24.11.2008
Cross-Country Regressions
5
Basic problems
Levine and Zervos (2001) identify the following problems:
 Statistical entries are sometimes inconsistently or inaccurately
measured
 Construction of proxies that measure policy actions is difficult
 Sala-i-Martin (1997) gives example of human capital – how to
measure it?
 Regression analysis presupposes that observations are drawn
from a distinct population
 Growth is often averaged over 30 years – including business
cycles, policy changes and political disturbances – which makes
interpretation of coefficients conceptually difficult
 Cross-country regressions do not address causal issues per se.
 Relationships might be discontinuous or non-linear
24.11.2008
Cross-Country Regressions
6
Further problems
Also McCartney (2006) brings forward fundamental critique:
 Cross-country regressions (averaging growth rates) typically do not
consider structural breaks



Interactions between policy measures cannot be feasibly
incorporated into econometric cross-country models
Simple cross country regressions do not account for so-called
hysteresis effects


Growth patterns in development countries are discontinuous, periods of
fast growth are succeeded by those of slow growth. Growth
accelerations are common (see also Hausmann et al. 2005).
Temporary economic shocks that have a permanent impact on economic
growth; threshold effects; virtuous and vicious circles
Static regressions do not consider dynamic effects

Does a policy have impact during one or a few business cycles only or
on long term growth?
24.11.2008
Cross-Country Regressions
7
Two problems to focus on
All considered papers name at least one of these two
difficulties and make suggestions how to deal with them:
 Growth theories are not explicit about what variables should be in a
growth regression





Take “A”, the “level of technology” in many growth models)
Many empirical economists are tempted to try around with indicators
without a theoretical backup
Most variables are significant in some combinations and insignificant
in others
The problem is mentioned by Sala-i-Martin (1997), McCartney (2006),
Brunetti (1997), and Levine and Zervos (2001)
Government policy is probably endogenous and not random


24.11.2008
Governments adopt certain policies, because they are optimizing over
policies in some objective function. Then the estimated coefficients will
be biased.
The problem is mentioned by Rodrik (2005) and Brunetti (1997)
Cross-Country Regressions
8
Regressors and robustness of results (1)
Two papers, Levine and Zervos (2001) and Sala-i-Martin (1997)
are specifically concerned with the issue:
 Levine and Zervos cite papers by Levine and Renelt (1992) and
Levine and Zervos (1993) that have used Extreme Bounds
Analysis (EBA) to identify reliable determinants of growth
regressor of interest
growth rate = β1I + β2M + β3Z
set of fixed regressors




3 out of 7 possible regressors
log of secondary school enrolment rate in 1960
log of initial real GDP per capita
population growth
Levine and Renelt
(1992)
ratio of investment to GDP
24.11.2008
Cross-Country Regressions
9
Regressors and robustness of results (2)
Both cited papers take 1960-1989 average growth in a cross-section
of over 100 countries as regressand
EBA checks, whether a variable is positive and significant in all
regression variants; then it is called robust
Hardly any relationship is found to be robust in this approach



regressor of interest
growth rate = β1I + β2M + β3Z
set of fixed regressors




3 out of 7 possible regressors
log of secondary school enrolment rate in 1960
log of initial real GDP per capita
population growth
Levine and Renelt
(1992)
ratio of investment to GDP
24.11.2008
Cross-Country Regressions
10
Regressors and robustness of results (3)



The Levine and Renelt (1992) paper is criticized by Sala-i-Martin
(1997) for employing a robustness test that coefficients cannot
survive
It is in the nature of statistical tests that they yield insignificant
results every once in a while, even if the assessed coefficient is
significant; thus, when taking into account many regressions, one
should use softer measures of robustness
Sala-i-Martin
Uses 63 variables collected from different papers
 Runs 30,856 regressions for each of 58 “variables of concern” of the
type on the previous two slides
 Weights estimated coefficients using integrated likelihoods
 Constructs a distribution of the weighted coefficients and then ranks
coefficients by the percentage of their cumulative density functions that
concentrates in either the positive or the negative quadrant
 Furthermore, for every variable he records the percentage of regressions
in which significance isCross-Country
recorded Regressions
24.11.2008
11

Regressors and robustness of results (4)
24.11.2008
Cross-Country Regressions
12
Endogeneity of government behavior (1)

Recall: Government policies are probably endogenous and not
random

Governments adopt certain policies, because they are optimizing some
objective function. Then the estimated coefficients will be biased.
Rodrik (2005) shows how cross-country growth regressions might
yield negative coefficients for policy measures, even if the
country is better off with the policy:
 Rodrik introduces the following simple model:

Growth is determined by:
g(s, θ, φ) = (1 – θ(1-s))A – φα(s) – ρ
market imperfection

policy tool
government ability
α´(s)>0,
α´´(s)>0
Government behavior is given by maximization over s:
max u = λg(s, θ, φ) + π(s)
π´(s)>0, π´´(s)<0
weight placed on growth
24.11.2008
Cross-Country Regressions
13
Endogeneity of government behavior (2)



Then, under the assumption
- that governments differ in their weights for growth, or
- or that the severity of market failures differs between the countries,
the estimated coefficient should be negative
Mathematical derivations can be studied in the paper
Intuition:



Assuming that governments can create rents for themselves from a
certain policy, they will adopt the policy to a farther extent than optimal
for growth; dg/ds < 0 evaluated at this point
Growth is decreasing in the market failure parameter; optimal extent of
policy implementation is growing in market failure parameter; this leads
to a seemingly negative effect of policy on growth in the model, dg/ds < 0
Example: Ownership of banks around the world (as in La Porta,
Lopez-de-Silanes, and Shleifer, 2002)
24.11.2008
Cross-Country Regressions
14
Endogeneity of government behavior (3)
How to deal with the problem:
 Rodrik sees no opportunities for field experiments, in which policies
would be randomized over situations; governments will not readily
allow for such experiments
 Instrumental variables are hard to find in the context of crosscountry regressions (Rodrik, Brunetti 1997)



Furthermore, they are not able to identify the success of purposeful
policy action, which we are most interested in (Rodrik)
Finally, the common use of past values of variables as instruments bears
the danger of autocorrelation (Brunetti)
Rodrik demands that before regressing and testing


a full theoretical model must be specified;
this should incorporate a likely channel through which policies might
operate
24.11.2008
Cross-Country Regressions
15
General outlooks





Rodrik (2005): See previous slide; in general very pessimistic about
the sense of general cross-country regressions
Just as well Levine and Zervos (2001); they call for better measures
of policies and for closer investigation of interaction between policies
and the interactions‘ effect on growth; thus they do not seem to
completely discard the approach
Sala-i-Martin (1997) acknowledges problems but recommends not to
be too pessimistic; quite a few indicators can be found as very
robust under the right robustness test
Brunetti (1997) claims that with the development of more advanced
policy indicators, the predictions seem to become more robust;
furthermore, he calls for the identification of IV to deal with the
endogeneity issue
McCartney (2006), however, suggests to discard the cross-country
regression approach as a whole, as it rests on the neo-classical
assumption of mechanically equal growth processes over all
countries
24.11.2008
Cross-Country Regressions
16
Used Papers





Ross Levine and Sara Zervos (2001): What We Have Learned
About Policy and Growth from Cross-Country Regressions?
Xavier Sala-i-Martin (1997): I Just Ran Four Million Regressions
Dani Rodrik (2005): Why We Learn Nothing from Regressing
Economic Growth on Policies
Aymo Brunetti (1997): Political Variables in Cross-Country Growth
Analysis
Matthew Mc Cartney (2006): Can a Heterodox Economist Use
Cross-country Growth Regressions?
24.11.2008
Cross-Country Regressions
17