Transcript Geography, Institutions, Openness, Macroeconomic Policy: What

What Really Matters for Long-term
Growth and Development?
A Re-Examination of the Deep
Determinants of Per Capita Income
Dorian Owen and Clayton Weatherston
University of Otago
EDGES ‘Roads to Riches’ Workshop
15 November 2005
Introduction
• Average living standards in richest countries
100× those in poorest countries
• Recent studies examine (very) parsimonious
models to evaluate the overall and relative
importance of hypothesized ‘deep’
determinants of economic development
• Aims
– To argue that much of this literature suffers from
problems of ‘model uncertainty’
– To outline an approach for re-examining the role
of deep determinants
– To present some preliminary results
Outline
• Brief review of the literature on ‘deep’
determinants of cross-country income levels
– Geography versus institutions
– Instruments and inference
• Criticisms focusing on model uncertainty and
evidence of mis-specification
• A general-to specific (Gets) approach
• Preliminary results
• Further work in progress
Growth Determinants – the Conventional
‘Production Function’ Approach
Inputs
Physical Capital (K)
Labour (L)
Human Capital (H)
Technology (A)
Aggregate
Production Function
Y = f(A, K, L, H, …)
Output
GDP (Y)
‘Proximate determinants’
… but what determines the proximate determinants?
Deep Determinants – the Contenders
• Geography
• Institutions
– Protecting property rights
– Coordinating/enhancing investment (K, H)
– Making governments/rulers accountable
• ‘Openness’/Integration
• Others – Culture, Ethnic/Linguistic/ Religious
Composition
• Characteristics: ‘Timescale’ criterion 
relative constancy/persistence as a measure
of ‘depth’. Not exogenous versus
endogenous.
Geography Hypothesis
• ‘Geography hypothesis’ includes direct and
indirect effects
• Geography  Development
– Climate
– Ground surface
– Geological
– Bio-geography
• Geography  Institutions  Development
– E.g., Acemoglu et al (AER 2001) – high disease
environment leads to ‘extractive’ colonies and
‘bad’ institutions, which impede long-term
development
Institutions Hypothesis
• Institutions  Development
– “institutions in society … are the underlying
determinant of the long-run performance of
economies” (North 1990)
• ‘Good institutions’: main focus on contract
enforcement, protection of property rights,
rule of law (‘market-creating’), covering broad
cross section of society
• Development of institutions:
– Legal origin
– Endowments: any effect of geography is only via
indirect effect on institutions
Measures of Deep Determinants
• Geographical variables
– Latitude, Average mean temperature, % land
area within 100km of coast, axis, frost days, etc
– Proportion of popn at risk from malaria
• Institutional variables
– ICRG survey indicators of investors’ risk
– World Bank survey assessments of govt
effectiveness (including Rule of Law)
– Polity IV – constraints on executive
Reflect ‘outcomes’ more than durable ‘constraints’,
are volatile, and increase with per capita income
(Glaeser et al, 2004)
Example study:
Rodrik et al. (J Econ Growth, 2004)
ln y = m + a INS + b INT + g GEO + e1
y = GDP per cap 1995
INS = ‘rule of law’ index
INT = ln(nominal trade/nominal GDP)
GEO = abs(latitude)
• Potentially complicated set of interlinkages
• INS and INT potentially endogenous
• Use of instrumental variables estimation
(2SLS)
INS = l + d SM + f ln(FR) + j GEO + e2
INT = q + t SM + s ln(FR) + w GEO + e3
SM = ln(settler mortality)
ln(FR) = ln(Frankel & Romer measure of
constructed trade shares)
GEO = abs(latitude) – exogenous
regressor in GDP per capita equation
• Instrumental Variables Estimation requires
‘valid’ instruments:
– Instrument relevance – variables in X need to
be highly correlated with the endogenous deep
determinant, say INS.
– Instrument exogeneity – X variables need to
be uncorrelated with the model’s error term, e –
if not, estimates are inconsistent
– Key problem – exogeneity fails if instruments
affect income via other channels or are
correlated with omitted variables
Key Instrument
• Acemoglu, Johnson and Robinson (AER,
2001): Europeans adopted different
colonisation strategies in different colonies:
‘settler’ versus ‘extractive’ colonies
Colonisation mode = f(disease environment)
High settler mortality  extractive colonies
Low settler mortality  settler colonies
(Potential) settler mortality  settlement type
 early institutions  current institutions 
current economic performance
Initial Consensus
Primacy of institutions – although geographic
conditions affect development (income per
capita) they do so only through their impact
on the development of institutions
– Acemoglu, Johnson & Robinson (AER 2001)
– Easterly and Levine (J Monetary Econ 2003)
– Rodrik, Subramanian &Trebbi (J Econ Growth
2004)
Later studies provide conflicting results
– Sachs (NBER WP 2003)
– Olsson and Hibbs (EER 2005)
Model Uncertainty
• Brock and Durlauf (2001) critique of crosscountry empirical growth literature:
– Violations of assumptions required for
estimation by OLS and interpretation as a
structural model
– ‘Open-endedness’ of theories - validity of one
causal theory does not imply falsity of
another. OK if regressors orthogonal but not
with a high degree of collinearity between
potential regressors
– ‘Model uncertainty’  likely sensitivity of
coefficient estimates and t-values to ‘other’
regressors under such conditions
• Open-endedness of growth theories also has
implications for the validity of instrumental
variable methods  predetermined variables
may not be valid instruments if correlated
with omitted variables
• Problem – don’t know which variables are
relevant, due to open-endedness of theories
and range of different feasible mechanisms
• Also, parameter heterogeneity in crosscountry samples. Cross-section estimates
best interpreted as ‘average effects’ - Temple
(JEL, 1999) but need to look out for evidence
of parameter heterogeneity
Study
Institutions
variable
HJ
GADP
(1999) (EngFrac
EurFrac
Latitude)
AJR
Exprop
(2001) (Settmort)
EL
Instit Dev
(2003) (Settmort
Latitude
Landlock)
Trade
Variable
Geog
Variable
YrsOpen Excluded
(lnFR)
Mis-spec
tests
×√×××
Excluded Latitude
×√××S
MeanTemp
Humidity
Malfal
YrsOpen Excluded
×√√√√
Study
Institutions
variable
Trade
Variable
Geog
Variable
Mis-spec
tests
Sachs
(2003)
Rule of Law
(Settmort
KGPTemp)
Excluded
×××√S
RST
(2004)
Rule of Law
(Settmort)
Trade %
of GDP
(lnFR)
% popn
close to
coast
Malfal
(ME)
Latitude
OH
(2005)
Political
Excluded
environment
×√×√×
Bio- and
×√××S
GeoConditions
Replication of Key Existing Studies
Key issues apparent in Table:
• Choice of regressors (range of proxies)
varies
• Control for openness – some do, some don’t;
other exogenous regressors also vary
• Evidence of mis-specification (tests for
RESET, normality, hetero)
• Parameter constancy
• Choice of instruments - Over-identification
tests
• Not congruent or encompassing – ‘illustrate’
rather than test competing theories
Why Use a General-to-Specific
(Gets) Approach?
• Theory relatively ‘loose’ – admits a wide
range of candidate regressors, e.g., different
geographical mechanisms, interactions
• Model selection important – untested
exclusion restrictions. ‘Open-ended theory’
problem
• Impressive Monte Carlo results for overall
PcGets algorithm
• Applicable to cross-section data (Hoover &
Perez, Oxford Bulletin 2004)
General Unrestricted Model (GUM)
• ln(GDP per capita) = f(Const, PhysGeog,
Climate, BioGeog, Resources, Institutions,
Integration, Culture, e)
Vectors of different factors representing
PhysGeog, Climate, etc
PhysGeog = (Axis, Size, Land100km, Mount)
Climate =(MeanTemp, Latitude,TempRange,
Frost)
BioGeog = (Malfal, Plants, Animals)
Resources = (Crop and Mineral dummies)
Institutions = (Exprop, ExConst, Plurality)
Integration = (YrsOpen)
Culture = (EthnicFrac, LingFrac, ReligFrac,
Catholic, Muslim)
Illustrative OLS Results
GUM
Const
SIZE
AXIS
PLANTS
ANIMALS
Malfal
lc100km
MOUNT
LATITUDE
RANGE
FROST
MEANTEMP
EXPROP
EXCONST
PLURAL
YRSOPEN
oil
CATH
MUSLIM
EthFrac
ReligFrac
LangFrac
Gets ‘testimation’
Const
Malfal
MOUNT
FROST
oil
EXPROP
YRSOPEN
CATH
Coefficient t-value
t-prob reliable
Constant
6.33030 16.913
0.0000
1.0000
MOUNT
-0.01201 -3.187
0.0023
1.0000
Malfal
-0.99967 -5.888
0.0000
1.0000
FROST
0.69508
2.755
0.0078
1.0000
EXPROP
0.27445
5.398
0.0000
1.0000
CATH
0.00536
3.098
0.0030
1.0000
YRSOPEN
0.74580
3.286
0.0017
1.0000
oil
0.39362
2.507
0.0149
0.7000
R^2 = 0.84731
Radj^2 = 0.82920
N = 67
FpNull = 0.00000
FpGUM = 0.97713
value
prob
Chow(34:1)
F( 34, 25)
0.7581
0.7764
Chow(61:1)
F( 7, 52)
0.6827
0.6859
normality test chi^2( 2)
1.8437
0.3978
hetero test
chi^2( 13)
18.3188
0.1458
IV estimates – final model
Coefficient
t-value
t-prob
reliable
Constant
6.54184
0.654
0.0000
1.0000
MOUNT
-0.01227 -3.016
0.0038
1.0000
FROST
0.64326
2.177
0.0335
1.0000
CATH
0.00475
2.629
0.0109
1.0000
oil
0.40337
2.510
0.0148
0.7000
Malfal*
-1.13708 -5.603
0.0000
1.0000
EXPROP*
0.25820
2.850
0.0060
1.0000
YRSOPEN*
0.70794
2.042
0.0456
1.0000
R^2 = 0.84555
Radj^2 = 0.82722
N = 67
FpNull = 0.00000
FpGUM = 0.99766
Additional instruments: LORGFR, ME, STATEHIST,
LSETTMORT, ENGFRAC, EURFRAC, LOGFR; SIZE, AXIS,
lc100km, LATITUDE, PLANTS, ANIMALS, RANGE, MEANTEMP,
MUSLIM, EthFrac, ReligFrac, LangFrac.
Sargan test: chi^2(16) =
13.0364 [0.6701]
chi^2( 4) =
7.4498 [0.9636]
value
prob
normality test chi^2( 2)
1.3034
0.5212
hetero test
chi^2( 13)
17.9626
0.1589
Conclusions and Further Work
1. Model uncertainty and mis-specification (lack of
congruence) are problems with existing studies
2. A Gets approach can address these issues
3. Preliminary results suggest that institutions are
not all that matters and that geographical variables
as well as openness and aspects of culture exert an
independent influence on per capita income levels
4. Examining sensitivity of results to variable
definition and choice of instruments
5. Ideal would be to select instruments and
regressors simultaneously as part of the Gets
modelling process (Hendry and Krolzig, EJ 2005)