IFPRI Discussion Paper 00694

March 2007

Is the Relationship Between Aid and Economic Growth

Nonlinear?

Andros Kourtellos, University of Cyprus

Chih Ming Tan, Tufts University
and Xiaobo Zhang, International Food Policy Research Institute

Development Strategy and Governance Division

INTERNATIONAL FOOD POLICY RESEARCH INSTITUTE.
The International Food Policy Research Institute (IFPRI) was established in 1975. IFPRI is one
of 15 agricultural research centers that receive principal funding from governments, private
foundations, and international and regional organizations, most of which are members of the
Consultative Group on International Agricultural Research.

Financial Contributors and Partners

IFPRI’s research, capacity strengthening, and communications work is made possible by its financial contributors and partners.
IFPRI gratefully acknowledges the generous unrestricted funding from Australia, Canada, China, Denmark, Finland, France,
Germany, India, Ireland, Italy, Japan, Netherlands, Norway, Philippines, Sweden, Switzerland, United Kingdom, United States,
and World Bank.
 IFPRI Discussion Paper 00694
March 2007

Is the Relationship Between Aid and Economic Growth

Nonlinear?

Andros Kourtellos, University of Cyprus

Chih Ming Tan, Tufts University
and Xiaobo Zhang, International Food Policy Research Institute

Development Strategy and Governance Division

PUBLISHED BY

INTERNATIONAL FOOD POLICY

RESEARCH INSTITUTE
2033 K Street, NW
Washington, DC 20006-1002 USA
Tel.: +1-202-862-5600
Fax: +1-202-467-4439
Email: ifpri@cgiar.org

www.ifpri.org

Notices:
1
Effective January 2007, the Discussion Paper series within each division and the Director General’s Office of IFPRI
were merged into one IFPRI-wide Discussion Paper series. The new series begins with number 689, reflecting the
prior publication of 688 discussion papers within the dispersed series. The earlier series are available on IFPRI’s
website at www.ifpri.org/pubs/otherpubs.htm#dp.
2
IFPRI Discussion Papers contain preliminary material and research results. They have not been subject to formal
external reviews managed by IFPRI’s Publications Review Committee, but have been reviewed by at least one
internal and/or external reviewer. They are circulated in order to stimulate discussion and critical comment

Copyright 2007 International Food Policy Research Institute. All rights reserved. Sections of this material may be reproduced for
personal and not-for-profit use without the express written permission of but with acknowledgment to IFPRI. To reproduce the
material contained herein for profit or commercial use requires express written permission. To obtain permission, contact the
Communications Division at ifpri-copyright@cgiar.org.
 CONTENTS

Acknowledgments ............................................................................................................................v
Abstract ............................................................................................................................. vii
1. Introduction ................................................................................................................................1
2. Econometric Methodology ..........................................................................................................5
3. Data ..............................................................................................................................13
4. Results ..............................................................................................................................18
4.1 Multiple Regimes and Foreign Aid ................................................................................ 18
4.1.1 Analysis of Baseline Tree Diagrams ........................................................................... 35
4.1.2 Baseline Parameter Estimates for Multiple Growth Regimes ..................................... 35
4.1.3 Results from Alternative Specifications ...................................................................... 37
4.2 Robust Relationship between Aid and Growth............................................................... 39
5. Conclusion ..............................................................................................................................46
References ..............................................................................................................................47

TABLES
1 Data Description ......................................................................................................................15
2 Summary Statistics ..................................................................................................................17
3 Model Specifications ...............................................................................................................29
4 Country Breakdowns by Growth Regimes for BTREED and TR Models ..............................30
5 BTREED Coefficient Estimates for Growth Regimes.............................................................32
6 TR Coefficient Estimates for Growth Regimes .......................................................................33

iii
 FIGURES

1 Landholding Distribution, by Rural Household Percentile .................................................7

2(a) Tree Diagram For BTREED Baseline Model ...................................................................20
2(b) Tree Diagram For BTREED Solow Model .......................................................................21
2(c) Tree Diagram For BTREED Parsimonious Model............................................................21
3(a) Tree Diagram For TR Baseline Model..............................................................................22
3(b) Tree Diagram For TR Solow Model .................................................................................23
3(c) Tree Diagram For TR Parsimonious Model......................................................................24
4(a) Hansen (2000) Confidence Intervals for Figure 3(a).........................................................25
4(b) Hansen (2000) Confidence Intervals for Figure 3(b) ........................................................26
4(c) Hansen (2000) Confidence Intervals for Figure 3(c).........................................................28
5(a) Partial Dependence Plot for Aid (Baseline/Solow Model)................................................40
5(b) Partial Dependence Plots for Solow Variables (Baseline/Solow Model)..........................41
5(c) Partial Dependence Plots for Macroeconomic Policy and Institutions (Baseline/Solow
Model) ..................................................................................................................42
5(d) Partial Dependence Plots for Other Fundamental Determinants (Baseline/Solow Model)43
6 Partial Dependence Plots (Parsimonious Model) ..............................................................44

iv
 Acknowledgments

We are deeply grateful to Rob McCulloch for his invaluable advice on the
implementation of the Bayesian tree methods used in this paper and helpful comments from
seminar participants at IFPRI. We thank the USAID/International Food Policy Research Institute
(IFPRI) for research support. Charalambos Michael and Ioanna Stylianou provided excellent
research assistance.

v
vi
 Abstract

There have been intensive debates on the role of aid in promoting economic development
in developing countries by using cross-country analyses. Cross-country regression assuming
linear relationship between aid and growth and without taking into heterogeneity of countries
would produce biased estimates. To correct this, in this paper we investigate the relationship
between foreign aid and growth using recently developed sample splitting methods that allow us
to simultaneously uncover evidence for the existence of heterogeneity and nonlinearity. We also
address model uncertainty in the context of these methods. We find some evidence that aid may
have heterogeneous effects on growth across two growth regimes defined by ethnolinguistic
fractionalization. However, when we account for model uncertainty, we find no evidence to
suggest that the relationship between aid and growth is nonlinear. In fact, our results suggest that
the partial effect of aid on growth is likely to be weakly negative. In this sense, our findings
suggest that aid is potentially counterproductive to growth with outcomes not meeting the
expectations of donors.
However due to the data comparability problem and inherent econometric problem in
cross-country regression, more detailed country case studies are needed to evaluate the impact of
aid at the country level and even project level. This can really answer the question that under
what conditions what type of aid helped growth and poverty reduction in developing countries.
The methodology developed in this paper can be used to identify typologies on other
outcome variables, such as those included in the Millennium Development Goals.

vii
viii
 1. Introduction

One of the most controversial debates in the empirical growth literature with big policy
implications is whether foreign aid is beneficial to a country's economic growth. In an influential paper,
Burnside and Dollar (2000) examine the effect of aid, as measured by the ratio of the sum of grants and
the grant equivalents of official loans in constant prices to real GDP (or, Effective Development
Assistance (EDA)), on growth. Using standard cross-country panel growth regressions that include an
interaction term of aid with a policy index, they find that aid has a positive impact on growth in
developing countries as long as these countries have sound macroeconomic policies. The policy
implication of this finding was straightforward. Policy makers at international aid agencies could now
argue that development assistance can contribute to poverty reduction in countries with good policy
environments.
On the other hand, this finding has sparked an industry of mainly empirical papers trying to
examine the sensitivity of Burnside and Dollar's results to model specification, alternative sets of
included/excluded variables, and different data series. Some of the most notable papers include
Guillaumont and Chauvet (2001), Hansen and Tarp (2001), Collier and Dehn (2001), Collier and Dollar
(2002, 2004), Collier and Hoeffler (2004), Easterly (2003), Easterly, Levine, and Roodman (2004),
Dalgaard, Hansen and Tarp (2004), Roodman (2004), and Rajan and Subramanian (2005a,b). Some of
these papers confirm the main finding of Burnside and Dollar—that aid is effective only in countries with
good policies, while others find the results fragile to the addition of particular variables and countries
covered.
One problem that the literature on aid and growth has been dealing with is the problem of how to
model heterogeneity and/or nonlinearities in growth analyses. Typically, what has been done is to treat
this issue in an ad hoc way by including squares and interaction terms for aid, policy, and other growth
variables. The unsystematic, ad hoc nature as to how specific choices are made over which
nonlinearities/heterogeneity to include and which to leave out, however, leaves much to be desired. For
instance, there is no good reason for only including an interaction term between aid and policy and not the
square of aid or even both in the model. Why not also include an interaction term between policy and
institutions? In fact, several new growth theories such as Azariadis and Drazen (1990) and Howitt and
Mayer-Foulkes (2002) suggest that the cross-country growth process is highly nonlinear.
To make things worse, as suggested by Brock and Durlauf (2001), new growth theories are
inherently open-ended. By theory open-endedness, Brock and Durlauf refer to the fact that typically a
priori statement that a particular theory of growth is relevant does not preclude other theories of growth

1
from also being relevant. Growth models typically do not provide much guidance as to the exact
specification in which growth determinants should enter the growth equation as well. Brock and Durlauf
point out that taken together, the combination of theory and specification uncertainty (what they refer to
collectively as model uncertainty) potentially renders coefficient estimates of interest to be “fragile.” The
potential fragility of coefficient estimates under model uncertainty is important because it implies that
findings on the relationship between aid and growth, which do not properly account for model
uncertainty, may be non-robust. For instance, the finding of a nonlinear relationship between aid and
growth may, in fact, be just a manifestation of some other unaccounted misspecification due to omitted
variables or even due to unaccounted heterogeneity and/or nonlinearities with respect to other growth
determinants. Our point is that strong a priori assumptions on the appropriate specification of growth
determinants and functional form of the model are hard to justify.
Nevertheless, while there is little agreement over the exact nature of nonlinearities and
heterogeneity in the growth literature, there is a growing consensus that, given that we think such
nonlinearities/heterogeneity exists, they may potentially be fruitfully modeled using empirical tools that
emphasize pattern recognition (see Durlauf (2003)). Sample splitting and threshold regression methods
and their derivatives are important constituents of such tools. For instance, Durlauf and Johnson (1995)
employed a classification and regression tree method (CART; see Breiman, Friedman, Olsen, and Stone
(1984)) to sort countries based on initial per capita income and initial literacy rates. They interpret their
findings as evidence in favor of the theory of poverty traps of Azariadis and Drazen (1992).
In this paper we employ recently developed sample splitting methods to systematically uncover
the robust relationship between aid and growth. Sample splitting methods such as threshold regression
and regression trees allow for increased flexibility in functional form and at the same time are not as
susceptible to curse of dimensionality problems as nonparametric methods. Unlike parametric models
with polynomial terms (squares, interactions, etc), sample splitting methods are parsimonious. More
importantly, these methods are structurally interpretable as they endogenously sort the data, on the basis
of some threshold determinants, into groups of countries each of which obeys the same model (i.e.,
multiple growth regimes). Other notable applications of sample splitting methods in growth include Tan
(2006) who use an improved regression tree algorithm to CART (GUIDE; see Loh (2002)) and Masanjala
and Papageorgiou (2004) who employ threshold regression (TR; see Hansen (1996, 2000) and Gonzalo
and Pitarakis (2002)).
A major problem associated with the sample splitting methods that have been employed so far in
the literature, however, is the sequential nature of the splitting process. By this we mean that choices of
threshold variables and split values made in initial sample splits are never revised as the number of splits
increases. Hence, any mistake made at the earlier stages of the process is propagated to the splits below.

2
The result is that the classification of observations into regimes can be unstable. Small changes in the
data result in large changes to the threshold or “tree” structure (see Hastie, Tibshirani, and Friedman
(2001) and also Hong, Wang, and Zhang (2005)).
To be clear, this is not an issue of statistical inference but rather it has to do with the qualitative
nature of threshold variables. It is one thing to define a 95% confidence interval for a (real-valued)
parameter as [0.3,0.8] and quite another thing to say that a 95% confidence interval for the discrete-
valued parameter associated with the choice of threshold variable includes two variables, initial per capita
income and property rights. In the former case the threshold effect is consistent with theories of poverty
traps and development while in the latter it says something about the importance of economic institutions
in posing barriers to growth.
A contribution of this paper is to employ a simultaneous sample split method, Bayesian tree
regression (BTREED; see Chipman, George, and McCulloch (1998, 2002)) to deal with this problem.
BTREED is a non-sequential regression tree procedure that generates the best tree of every size. Thus, it
is less likely to suffer from some of the consequences (e.g., tree instability issues) of sequential sample
splitting methods such as TR or CART. Nevertheless, we compare our results with TR since this method
provides formal asymptotic theory for the construction of confidence intervals for the threshold estimates.
A second key methodological contribution of this paper is to move the discussion away from
model selection towards model averaging in the context of nonlinear (and, in particular, sample split or
tree) models. As Cohen-Cole, Durlauf, and Rondina (2006) note, there has not so far been a systematic
investigation of model uncertainty and nonlinearities in the growth context. This paper can be viewed as
a first attempt towards this ambitious goal. In order to achieve this, we exploit a new statistical learning
methodology, Bayesian Additive Regression Trees (BART)1. Specifically, the idea is to generate a large
number of trees, each of which is a bad fit for the data as a whole (i.e., a “weak learner”), but gives
insight into a small part of the underlying data generation process, so that, taken together, the “sum-of-
trees” provides a good estimate of the underlying process. Also, in contrast to single-tree methods, there
is no need in BART to condition upon a particular choice of slope covariates and threshold variables.
Rather inference is obtained by averaging the sum-of-tree draws from the BART posterior distribution.
We view our methodological contribution in this paper as an extension of the standard model averaging
exercises recently applied in the empirical growth literature (see Brock and Durlauf (2001), Fernandez,
Ley, and Steel (2001), and Sala-i-Martin, Doppelhofer, and Miller (2004) among others).

1
BART is closely related to so-called “ensemble” methods such as random forests (Breiman (2001)), bagging (Breiman (1996)),
and, most directly, boosting (Friedman (2001)) in the machine learning literature. Ensemble methods have been shown to have
extremely good out-of-sample prediction performance besting even those of neural networks (see, in particular, Friedman (2001)
and Hastie, Tibshirani, and Friedman (2001)). Unlike the above-mentioned machine learning methods, however, BART is not
defined purely by an algorithm, but, instead, by a statistical model within the Bayesian framework.}, developed by Chipman,
George, and McCulloch (2002).

3
 We find some evidence in the BTREED and TR results that aid may have heterogeneous effects
on growth across two growth regimes defined by ethnolinguistic fractionalization. In particular, countries
that belong to a growth regime characterized by levels of ethnolinguistic fractionalization above a
threshold value experience a negative partial relationship between aid and growth, while those in the
regime with ethnolinguistic fractionalization below the threshold experience no growth effects from aid at
all. We also find that countries in the regime with higher levels of ethnolinguistic fractionalization
experience, on average, lower growth rates than countries in the lower ethnolinguistic fractionalization
regime. Nevertheless, we do find substantial tree instability in our sample split exercises so that attempts
to characterize the typology of these growth regimes with a high degree of certainty remains elusive.
There is evidence that the typology of these regimes may be alternatively well-characterized by property
rights institutions or macroeconomic policies such as the level of inflation, and not just ethnolinguistic
fractionalization. The data simply cannot be certain.
Our BART results are therefore particularly valuable given the high degree of uncertainty
generated by tree instabiliity. Here, we find very little evidence to suggest that the relationship between
aid and growth is nonlinear for the set of developing countries who are aid recipients. Overall, our results
suggest that the partial effect of aid on growth is very likely to be negative although we cannot reject the
hypothesis that aid has no effect on growth. In this sense, our findings suggest that aid is potentially
counterproductive to growth with outcomes not meeting the expectations of donors. We are therefore
sympathetic to the positions of work such as Easterly, Levine, and Roodman (2004) and Rajan and
Subramanian (2005a) which are generally pessimistic about the potential contributions of aid to
improving economic performance.
The remainder of the paper is organized as follows: Section 2 describes our econometric
methodology, which includes Bayesian tree regression (BTREED), threshold regression (TR), and
Bayesian Additive Regression Trees (BART); Section 3 describes our data; Section 4 presents our
findings, and; Section 5 concludes.

4
 2. Econometric Methodology

We conduct our analysis of the relationship between aid and growth using a generalized sample
split model that can be defined as follows:

gi = α j + hi β j + xi/ γ j + ε i iff zi ∈ R j ({λ } ) for

b −1
s s =1 j = 1,..., b (1)

where i indexes the observations (i.e., countries) and j indexes the b growth regimes. g i is the

average growth rate of per capita income for country i across a time period. hi is the foreign aid proxy

(i.e., the variable of interest). We distinguish between two sets of growth determinants. The k-
dimensional vector x denotes the set of slope covariates while the p-dimensional vector z denotes
threshold variables.
The set of slope covariates includes the usual Solow regressors, that is, the logarithms of the
average rates of physical and human capital accumulation, the logarithm of average population growth
rate plus 0.05, and the logarithm of initial per capita income. We also include variables from a wide
range of new growth theories including macroeconomic policy, geography, ethnolinguistic
fractionalization, political institutions, and property rights institutions. Most of the covariates can also be
viewed as threshold variables. For instance, theories of development that emphasize threshold
externalities such as Azariadis and Drazen (1992) suggest that initial per capita income may act as a
threshold variable. Alternatively, theories that emphasize economic “take-off” (e.g., Galor and Weil
(2000), Galor and Moav (2002), and Galor (2005)) suggest an important role for fundamental growth
determinants (such as geography and institutions) in driving growth divergence2. To be as agnostic as
possible, a slope covariate is also a threshold variable as long as it makes sense.
To this end, we specify in our sample split exercises that, with the exception of the factors of
accumulation and population growth rates (which are period averages), all slope covariates (including aid)

(
are also threshold variables. The set of parameters is given by Ψ = b, {λs }s =1 , Θ , where
b −1
)
Θ = (α j , β j , γ j , σ 2j ) is the set of regression parameters, b is the number of regimes, and {λs }s =1 is the
b b −1
j =1

2
“The timing of the “take-off” may differ significantly across countries and regions due to historical accidents, as well as
variation in geographical, cultural, social and institutional factors, trade patterns, colonial status, and public policy that have
affected the relationship between human capital formation and technological progress [Galor, 2005, p. 80].”

5
set of threshold parameters that define the set of threshold splits. Note that {λs }s =1 , in effect, partitions
b −1

{ }
b
the support of the threshold variables Z into b mutually exclusive regions R j .
j =1

We can visualize an example of a tree or threshold regression estimation procedure using Figure
1 which is due to Hastie, Tibshirani, and Friedman (2001).
Here, the set of observations is partitioned into five regimes, R1 ,..., R5 , defined by the interaction

between variables x1 and x2 . In this example, the model in (1) is modified to be a piece-wise constant

model so that a local average is estimated within each regime. The model we use to analyze the effect of
aid on growth will be in keeping with (1); i.e., it will be a piece-wise linear model. That is, we would
replace each “step” in Figure 1 with a plane in each growth regime which slope is determined by the
coefficients to the local augmented neoclassical growth model defined by (1).

6
 Figure 1: Landholding Distribution, by Rural Household Percentile

It is worth noting the generality of (1). If we ignore the effects of z on growth; that is, if we
specify, a priori, a single growth regime, then we are back to the canonical growth regressions of
Mankiw, Romer, and Weil (1992) and Barro (1996). However, as pointed out by Brock and Durlauf
(2001), such a formulation ignores prior knowledge regarding the existence of heterogeneity across

7
country units. That is, it ignores the possibility that the effect of the right-hand side covariates on growth
may differ systematically across groups of countries. Brock and Durlauf explore a special case of (1) to
study the robust heterogeneous effects of ethnolinguistic fractionalization on growth. In their paper, the
number of regimes b is trivially fixed to two as their threshold variable is a single dummy variable for
Sub-Saharan Africa. Given the binary nature of the dummy variable there is no need to estimate a
threshold parameter and hence the classical inference is still valid3. In contrast, our methodology enables
us to have multiple regimes and multiple threshold variables. This is very important in our context given
the large number of growth determinants that can act as threshold variables. What is more, the number of
regimes b is not pre-specified, but instead is endogenously determined.
One way to estimate (1) is to use the threshold regression methodology of Hansen (2000). At
each stage of the sample splitting, we carry out Hansen's test to see whether the sample should be split. If
so, we choose the best (in the sense of minimizing sum of squared errors) threshold variable, associated
threshold value estimate, and the set of regression estimates for Θ . The same procedure is then applied
iteratively to each of the two subsequent subsamples. This “tree growing” procedure stops when either
the null of no-split fails to be rejected, or the number of observations in the (sub-) sample falls below a
pre-determined minimum value. It is worth noting that TR bears deep similarities to the classification and
regression trees (CART) method of Breiman, Friedman, Olsen, and Stone (1984). The added advantage
of using threshold regression as opposed to CART is that the statistical inference4 for both the threshold
and the regression slopes has been well developed by Hansen (2000).
Its primary weakness, however, lies in the instability of trees to small perturbations in the data as
well as in the way that variables are defined. It has been well-documented that small changes in the data
can lead to very different threshold variables, threshold values, and even number of regimes being
selected by sample splitting methods (see, Hastie, Tibshirani, and Friedman (2001) and also Hong, Wang,
and Zhang (2005)). A major reason for the instability of trees is due to the sequential nature of typical
sample splitting algorithms. That is, the tree building method does not “update” the tree as it gets bigger.
Therefore, it may be that as the tree gets bigger, the previously selected threshold variables and split
values in the “upper” parts of the tree (i.e., the initial sample splits) are no longer optimal. We should
note that Bai (1999) had suggested an alternative method for getting around the sequential nature of
traditional threshold regression models. He calls this method “repartitioning.” The idea is to revise upper

3
However, this is not true anymore when the threshold variable is not binary and we need to estimate a threshold parameter
because the threshold parameter is not identified under the null. Hansen (2000) shows that the inference is non-standard and
develops an asymptotic theory for both the threshold parameter and the regression slopes including a method to construct
asymptotic confidence intervals for the former.
4
It should be noted that Hansen (2000) only claims the validity of these results for the single threshold (i.e., two-regime) case,
even though he has shown examples of proceeding with these tests iteratively beyond this case.

8
parts of the tree once lower parts of the tree are estimated. However, we found the practical
implementation of repartitioning to be computationally expensive and quickly lost computational
tractability even when the tree size was only moderately large. This led us to consider instead Bayesian
tree regression (BTREED) developed by Chipman, George, and McCulloch (1998, 2002).
BTREED is not a sequential splitting method. Instead, what BTREED does is to search through
trees of all sizes (i.e., the (final) number of regimes) and then locate the tree with the highest evidentiary
weight for each size. Specifically, it employs MCMC to stochastically search over the posterior
distribution of trees for high posterior probability trees. We then select the final tree using BIC. Because
each of these trees (no matter the size) is generated probabilistically at every stage of tree building, we do
not have the situation, as we do with sequential splitting methods such as TR, where “upper” portions of
the tree are never revised even as we vary (increase) the size of trees5; see Chipman, George, and
McCulloch (2002).
Nevertheless, we should note that BTREED, like TR, is still ultimately a model selection
algorithm. Both sample split methods seek to present one tree as the best device for summarizing the
relationship between growth and the set of growth determinants out of the forest of possible trees. While
engaging in such model selection has advantages --- for instance, it allows us to present a structurally
interpretable typology (i.e., tree diagram) for relating aid to growth --- this strategy ignores the
evidentiary weight associated with alternative trees. Cohen-Cole, Durlauf, and Rondina (2006) have
suggested that, even in the context of nonlinear models, researchers should still attempt to report robust
estimates of relationships that take into account alternatives to the chosen or benchmark model. We
pursue this suggestion in this paper. That is, we attempt to combine the evidentiary weight on the effect
of aid on growth across a large number of tree models.
To do so, we employ a new methodology due to Chipman, George, and McCulloch (2005) known
as Bayesian Additive Regression Trees (BART). More precisely, we do not condition on a particular
choice of slope covariates and threshold variables but rather inference is performed by averaging posterior
information across a large number of tree models in order to flexibly estimate the average effect of a
variable of interest on the dependent variable.
Formally, if we define wi = ( hi , xi , zi ) , then we can write the growth model (1) as

gi = f ( wi ) + ε i (2)

5
In fact, key steps in BTREED's stochastic tree building algorithm; i.e., “swap” and “change” split decisions (see Chipman,
George, and McCulloch (1998, 2002)), are in the spirit of Bai's “repartitioning”.

9
 where ε i | wi N ( 0, σ 2 ) and f ( wi ) = E ( gi | wi ) . Then BART provides a way to estimate (1)

μ ( m | w)
by combining information across tree models drawn from the posterior distribution, ,

M
fˆ ( wi ) = ∑ fˆm ( wi , Tm , Θ m )
m =1 (3)

Here, the j-th regime for each of the M trees Tm , m = 1,..., M , is associated with a real parameter

θ j . Hence, any wi is associated with one of the θ j within each tree. Letting Θ = (θ1 ,θ 2 ,...,θb ) where

b is the number of regimes in T, a single tree model may be denoted by the pair (T , Θ ) . Let

fˆ ( wi , Tm , Θ m ) denote the θ j associated with wi in the m-th tree. The posterior distribution for tree

models, μ ( m | w ) , is given by Bayes rule,

μ ( m | w) ∝ μ ( w | m ) μ ( m )
(4)

so that each weight is the product of the likelihood of the data given a model, μ ( w | m ) , and the prior

probability for a model, μ ( m ) . The latter is implicitly given by,

μ (1,..., M ) = μ ( (T1 , Θ1 ) , (T2 , Θ2 ) ,..., (TM , Θ M ) , σ )

= μ (T1 , T2 ,..., TM ) μ ( Θ1 , Θ 2 ,..., Θ M | T1 , T2 ,..., TM ) μ (σ )
(5)

For computational reasons we follow Chipman et al and assume independence so that,

M
μ (1,..., M ) = μ (σ ) ∏ μ (T j )μ ( Θ j | T j )
j =1
(6)

BART samples from the above posterior distribution using a Markov Chain Monte Carlo
(MCMC) algorithm. The construction of each tree Tm for m = 1,..., M employs precisely the tree-

building algorithm of BTREED. However, each tree is constrained to be small by appropriately setting

10
the tree priors. The choice of parameter priors are also essentially similar to those of BTREED.
Specifically, they are the normal-inverse gamma conjugate priors for the special case where the growth
model is constrained to just estimating a constant term θ j . We refer the reader to Chipman, George, and

McCulloch (2005) for details.

For better approximations, we would want to set M to be relatively large. In our exercises, we
follow Chipman et. al. and set M = 200. Notice that BART is greatly more flexible than (1). To see this
consider first the case of M = 1, then f1 ( wi , T1 , Θ1 ) is the conditional mean of g given w . However,

when M > 1, the terminal node parameters are merely components of the conditional mean of g given
w . Furthermore, these terminal node parameters will represent direct and indirect effects (interaction
terms) depending on sizes of the trees. In the special case where every terminal node assignment depends
on just a single component of w , the sum-of-trees model reduces to a simple additive function of splits
on the individual components of w .
To assess the effect of each of the determinants on growth we use Friedman's (2001) partial
dependence plot. To do so, first rewrite f ( w ) as f ( h, hc ) where hc is the complement of h in the set

w . To estimate the (partial) effect of $h$ on growth, Friedman suggests that we average out the effect of
hc on growth; i.e.,

E ( g | h ) = Ehc ⎡⎣ E ( g | w ) ⎤⎦
= Ehc ⎡⎣ f ( h, hc ) ⎤⎦
= ∫ f ( h, hc ) p ( hc ) dhc
(7)

However, when the data is i.i.d., then, we can approximate (7) with,

N
1
fˆh ( h ) =
N
∑ f ( h, h )
c ,i
i =1 (8)

where each hc for i = 1,..., n is an observation in the data. The above is the prediction by BART

of the partial dependence of growth rates on h at each level in its support. The pointwise posterior 95%

confidence intervals for fˆh ( h ) can also be easily obtain from its posterior distribution using the 2.5-th

and 97.5-th percentiles of the MCMC draws.

11
 One weakness that applies to all sample splitting methods is that there are almost no results on
dealing with endogeneity. A notable exception is Caner and Hansen (2004) who develop a two-stage
instrumental variable estimation procedure for threshold regression when the slope variables are
endogenous but the threshold variable is exogenous. A more attractive estimator would consider the
endogeneity of both slope and threshold variables. However, such an estimator would require an
alternative estimation approach due to the nonlinear nature of the endogeneity. Unfortunately, an
estimator for the case when both the slope and threshold variables are endogenous does not currently
exist.
This lack of results for how to deal with endogeneity is a particularly important weakness in our
context because we introduce aid as both a slope covariate as well as a potential threshold variable in this
paper. The reason we do so is to capture possible differential impacts of aid on growth for countries
above or below a threshold level of aid. Such a specification potentially describes the position of aid
proponents who argue that aid needs to be at a high enough level before it has a positive impact on
growth. However, by doing so endogeneity becomes a potential problem since aid is not randomly
assigned to recipient countries. Typically, more aid is given to those who are less developed leading to
potential bias in the estimation of regression coefficients because of problems with reverse causality as
well as correlation with unobserved heterogeneity.
Another aspect of the problem is that the use of instrumental variables in the context of growth
may be invalid; see Brock and Durlauf (2001). For an instrumental variable of aid to be valid, one has to
assume that it is uncorrelated with all the omitted growth determinants. However, the inherent open-
ended nature of growth theories makes this event unlikely. To explain their position, Brock and Durlauf
considered the paper by Frankel and Romer (1999) on trade and growth. Using similar reasoning let us
consider here the paper of Rajan and Subramanian (2005a). Rajan and Subramanian argue that since aid
is clearly endogenous, then it is necessary to use instrumental variables to obtain consistent estimates for
the effect of aid on growth. Their instrumental variables include dummies for colonial relationships
involving Britain, France, Spain and Portugal as well as dummies that indicate whether the donor and
recipient are common members of, or signatories to, an Entente or Alliance. However, notice that the
recent growth literature provides many theories of how colonial status may affect institutions. Rajan and
Subramanian do not account for any of these theories and hence it is plausible that the colonial dummies
are correlated with the omitted growth determinants (regression error) rendering their instrumental
variable method invalid.
In sum, there is no doubt that aid is potentially endogenous. Nevertheless, controlling for
endogeneity in the context of aid and growth is a difficult task mainly for two reasons. First, current
estimation procedures of sample splitting models are unable to deal with endogeneity in general. Second,

12
the open-ended nature of growth theories present unique difficulties for researchers arguing for the
validity of instruments. For these reasons, this paper should be viewed as a step towards getting the facts
straight rather than making strong structural claims.

3. Data

We use an unbalanced panel dataset (see Table 4) over two periods 1965-79 (42 countries) and
1979-94 (56 countries) based on a broad set of cross-country growth variables. As discussed in the
previous section, the dependent variable in (1) is the average growth rate of real per capita GDP
corresponding to the two periods. The set of explanatory variables includes a time dummy for the time
period 1979-94 and the canonical Solow variables; i.e., the logarithm of the sum of average population
growth plus 0.05 for net depreciation, the logarithm of the average proportion of real investments
(including government) to real GDP, the logarithm of years of male secondary and higher school
attainment, and the logarithm of real per capita GDP for the initial year of the time period. The national
accounting data used to construct these data series are obtained form Penn World Table 6.1 (see, Heston,
Summers, and Aten (2002)), while schooling data comes from Barro and Lee (2001).
To proxy foreign aid we use data on Effective Development Assistance (EDA) as a share of real
GDP constructed by Easterly, Levine, and Roodman (2004) and revised by Roodman (2004). Other
studies have also measured aid flows using OECD data for net Overseas Development Assistance (net
ODA). Net ODA is defined as transfers --- essentially, any assistance, save military aid, with a grant
element of at least 25% --- from a donor minus any repayment during a given period. We chose to use
EDA data (by Chang, Fernandez-Arias, and Serven (1999)) instead of ODA data in this paper for the
same reason that Easterly et al. (2004), Roodman (2004), and others do so. As pointed out by Chang, et.
al. the net ODA data potentially overstates the level of assistance to recipient countries. Instead, they
propose to exclude technical assistance, which tend to go primarily to consultants instead of governments,
and to account for different degrees of concessionality in loans.
The EDA data we employ is the most current version of the panel data used in much of the aid-
growth literature (see, for instance, Burnside and Dollar (2000), Hansen and Tarp (2001), Dalgaard,
Hansen and Tarp (2004)). This panel data set is available in 5-year periods from 1970-1999. Previously,
aid data were only available every 4 years. We use the 5-year panel data set to construct average
measures of EDA for the two sample periods 1965-79 and 1980-94.
Following the literature on growth and aid we include four macroeconomic policy variables. We
include the logarithm of inflation rate plus one, the ratio of budget surplus to GDP, money supply (M2),
and the Sachs-Warner (1995) variable measuring openness to trade. It is worth noting that we deviate

13
from Burnside and Dollar who include a single measure of economic policies. Burnside and Dollar first
estimate a growth regression without aid but with all the covariates and three indicators of
macroeconomic policy --- log (1+inflation), budget balance to GDP, and the Sachs-Warner (1995)
variable. Then, they construct their policy measure by forming a linear combination of the three using the
coefficients as weights. We believe that the inclusion of generated regressors in the analysis will result in
unnecessary biases so we include all four variables, instead (see also, Lubotsky and Wittenberg (2006)).
Additionally, we expand the Solow space with fundamental determinants of economic growth
that include proxies for geography, ethnolinguistic fractionalization, political institutions, and property
rights institutions. Following Rodrik, Subramanian, and Trebbi (2004) and Sachs (2003) we proxy
geography using a climate variable that measures the percentage of a country's land area classified as
tropical and subtropical via the Koeppen-Geiger system (KG Tropics) and a variable that measures the
percentage of a country's land area that lies within the geographic tropics (TROPICAR; Gallup, Sachs,
and Mellinger (1999)). We also include a variable of geographic isolation or the degree of landlocked-
ness that measures the percentage of a country's land area within 100km of an ice-free coast
(LCR100KM). To proxy the effect of ethnolinguistic fractionalization we use two measures due to
Alesina et al (2003). We include a variable of racial and linguistic characteristics (ethnic fractionalization)
and a measure of linguistic fractionalization (language). To capture tensions between social groups, we
also include a measure of ethnic tensions from the International Country Risk Guide. Furthermore, we
proxy political institutions using the average of Freedom House index of political rights (see Barro
(1991)) while for property rights we use the ratio of assassinations to GDP (see Banks (2002)), a measure
of the risk of expropriation of private investments (see Acemoglu, Johnson, and Robinson (2001),
executive constraints (Polity IV), and a composite governance index (KKZ96; see Kaufmann, Kraay and
Mastruzzi (2005)). Finally, we include time dummies and regional dummies for East Asia, Sub-Saharan
Africa, and Latin America and the Caribbean to account for time and regional heterogeneity, respectively.
Please refer to Table 1 for a detailed description of variables. Table 2 provides some summary statistics.

14
 Table 1: Data description
VARIABLE DESCRIPTION PANEL SOURCE
1965-79,
Growth GDP growth rates (using rgdpch) PWT61
1980-94
1965-79,
Population Growth logarithm of population growth + 0.05 PWT61
1980-94
Investment logarithm of average investments/gdp 1965-79, 1980-94 PWT61
Schooling logarithm of average years of male secondary and higher school attainment 1965, 1980 Barro-Lee(2000)
Initial Income log of initial per capita income 1965, 1980 PWT61
Percentage of land area classified as tropical and subtropical via the
KG Tropics CID, Harvard University
Koeppen-Geiger system.
Landlocked
Percentage of a country’s land area within 100km of an ice- free coast. CID, Harvard University
(LCR100KM)
Tropical Area
Fraction of land area in geographic tropics. Gallup and Sachs, 1999
(TROPICAR)
Alesina, A., A.
Measure of linguistic fractionalization based on data describing Devleeschauwer, W. Easterly,
Language
shares of languages spoken as “mother tongues”. S. Kurlat, and R. Wacziarg
(2003)
Alesina, A., A.
Ethnic Devleeschauwer, W. Easterly,
Measure of ethnic fractionalization based on racial and linguistic characteristics.
Fractionalization S. Kurlat, and R. Wacziarg
(2003)
This variable, which has been transformed to lie between 0 to 1, “measures the degree of
tension within a country attributable to racial, nationality, or language divisions.” Higher International Country Risk
Ethnic Tension 1982-1994
values correspond to lower degrees of ethnic tension. We take the average value of Ethnic Guide
Tension for the available data (1982-94) and repeat in each period.
Political Rights. The variable was tranformed using ( 7-x)/6 so that lower ratings (closer to
Political Rights zero) are given to countries with poor political rights and higher ratings (closer to one) are 1972-79, 1980-94 Freedom House 2005
given to countries with better political rights.

15
 Table 1 (cont.): Data description
VARIABLE DESCRIPTION PANEL SOURCE
Assassinations Assassinations per capita 1965-79, 1980-94 Banks (2002)
Rescaled, from 0 to 1, with a higher score indicating more constraint:
Executive
0 indicates unlimited authority; score of 1 indicates executive parity or subordination. We 1965-79, 1980-94 Polity IV dataset
Constraints
calculated the average for each period.
Risk of “outright confiscation and forced nationalization" of property.
Expropriation Risk 1982-94, 1982-94 IRIS
Rescaled, from 0 to 1, with a higher score indicating higher less risk of expropriation.
Composite Governance index. It is calculated as the average of six variables: voice and
Governance Kaufmann, Kraay and
accountability, political stability and absence of violence, government effectiveness, 1996, 1996
(KKZ96) Mastruzzi (2005)
regulatory quality, rule of law, and control of corruption.
Aid Effective Development Assistance/ real GDP 1970-79, 1980-94 Roodman (2004)
Budget Surplus Budget surplus 1965-79, 1980-94 Roodman (2004)
Global Development Network
Inflation ln(1+ inflation rate) 1965-79, 1980-94
Growth Database.
Global Development Network
M2 Average Ratio of M2 to GDP 1965-79, 1980-94
Growth Database.
Sachs and Warner, 1995;
Openness Average openness measure proposed by Sachs and Warner 1965-79, 1980-94 Easterly et al., 2004; Wacziarg
and Welch, 2002
East Asia A dummy variable for East Asia
Latin America &
A dummy variable for Latin America
Caribbean
Sub-Saharan
A dummy variable for Sub-Saharan Africa
Africa

16
 Table 2: Summary statistics
Min. Max. Median Mean Std. Dev.
KG Tropics 0.000 1.000 0.656 0.547 0.400
Landlocked (LCR100KM) 0.000 1.000 0.363 0.433 0.349
Tropical Area (TROPICAR) 0.000 1.000 1.000 0.689 0.432
Language 0.003 0.923 0.427 0.418 0.320
Ethnic Fractionalization 0.039 0.930 0.540 0.507 0.235
Ethnic Tension 0.131 1.000 0.587 0.570 0.241
Political Rights 0.000 1.000 0.417 0.484 0.278
Assassinations 0.000 4.000 0.067 0.373 0.722
Executive Constraints 0.000 1.000 0.389 0.471 0.328
Expropriation Risk 0.346 0.883 0.613 0.614 0.120
Governance (KKZ96) -1.869 1.159 -0.270 -0.195 0.589
East Asia 0.000 1.000 0.000 0.092 0.290
Sub-Saharan Africa 0.000 1.000 0.000 0.214 0.412
Latin America & Caribbean 0.000 1.000 0.000 0.245 0.432
M2 0.073 1.001 0.225 0.278 0.153
Budget Surplus -0.206 0.092 -0.033 -0.039 0.040
Inflation 0.031 3.127 0.119 0.320 0.587
Openness 0.000 1.000 0.067 0.255 0.339
Aid -0.328 9.482 0.495 1.316 1.823
Population Growth -3.059 -2.365 -2.580 -2.602 0.111
Investment 0.698 3.563 2.568 2.501 0.520
Schooling -4.017 1.226 -0.298 -0.489 0.936
Initial Income 6.094 9.344 7.906 7.843 0.742
Growth -0.053 0.081 0.014 0.014 0.024

17
 4. Results

4.1 Multiple Regimes and Foreign Aid

We first turn to our sample splitting (TR and BTREED) results. These methods require us to pre-
specify which growth variables should be treated as slope covariates, which as potential threshold
variables, and finally which as both. We carry out exercises for many alternative specifications. Our aim
in carrying out these different exercises is to observe two forms of robustness. Firstly, we want to see if
the trees obtained by TR and BTREED are stable. That is, we investigate whether the uncovered tree
structures do not vary dramatically across specifications when we (1) vary the set of covariates, (2) given
a set of covariates, vary the choices on which variables should be threshold variables, split variables, or
both, and (3) vary the number of observations in the data due to the inclusion or exclusion of countries
because of variations in missing values across specifications. And secondly, we want to see the extent to
which the results obtained by these different sample splitting methods --- TR (sequential) and BTREED
(non-sequential) --- are in agreement.
We report results for three specifications --- Baseline, Solow, and Parsimonious --- that turned out
to be most interesting. The Baseline specification is meant to reflect closely the cross-country growth
equation in the aid literature (see Burnside and Dollar (2000)). The set of slope covariates includes the
Solow variables (i.e., population growth, investment, schooling, and initial income), aid (EDA),
macroeconomic policy variables (i.e., openness, inflation, budget surplus, M2), geography (i.e.,
TROPICAR), linguistic fractionalization (language), regional dummies, political institutions (political
rights), and property rights (assassinations, expropriation risk, executive constraints, governance
(KKZ96)). The set of threshold variables includes most of the slope variables. We do not include in this
set the rates of human and physical capital accumulation and population growth rates because these are
period averages and not initial conditions.
The Solow specification differs from the Baseline specification in that the set of covariates only
includes the Solow and the Aid variables. Following Durlauf, Kourtellos, and Minkin (2001), the idea
behind the Solow specification is to examine local generalizations of the Solow model in the sense that a
Solow model applies to each country within a growth regime, but the model's parameters vary across
regimes.
Finally, the Parsimonious specification aims to maximize the number of observations by
excluding the macroeconomic policy variables. Specifically, the set of slope covariates includes
population growth, investment, schooling, initial income, aid, TROPICAR, language, political rights,
governance (KKZ96), and the three regional dummies. The set of threshold variables for the

18
Parsimonious specification comprises TROPICAR, language, political rights, governance (KKZ96), aid,
and initial income.
Figures 2(a)-(c) show the tree diagrams for BTREED for, respectively, the Baseline, Solow, and
Parsimonious specifications, while Figures 3(a)-(c) show the corresponding diagrams for TR. We also
report Hansen's 95% confidence bounds in Figures 4(a)-(c); these correspond to the TR threshold value
estimates in Figures 3(a)-(c). While the tree structures generated by TR in Figures 3(a)-(c) offer us an
interpretable relationship between various growth determinants and economic growth, the confidence
bounds provide us with a measure of the uncertainty over the classification of particular countries into
each growth regime. The classification of countries into regimes is given, for both BTREED and TR and
for all three specifications, in Table 4. Where applicable (i.e., in the TR cases), a superscript “c” denotes
countries within Hansen's 95% confidence bounds for the first threshold split as given in Figures 4(a)-(c).
Finally, the coefficient estimates and standard errors for each of the BTREED growth regimes are given
in Table 5. The corresponding numbers for the TR growth regimes are given in Table 6.

19
 Figure 2(a)*: Tree diagram for BTREED baseline model

N = 98

Language Language
≤ 0.411 > 0.411

N = 49 N = 49
g = 0.0162 g = 0.0115
Y = 8.0672 Y = 7.3331
(1) (2)

*Note that N gives the number of observations in each node, g stands for the average growth rate in real
per capita GDP from 1965-94 for countries in the terminal node, and Y is the corresponding average
level of log initial real per capita GDP in 1965.

20
Figure 2(b)*: Tree diagram for BTREED solow model

Figure 2(c)*: Tree diagram for BTREED parsimonious model

21
Figure 3(a)*: Tree diagram for TR baseline model

22
Figure 3(b)*: Tree diagram for TR solow model

23
Figure 3(c)*: Tree diagram for TR parsimonious model

24
 Figure 4(a): Hansen (2000) confidence intervals for Figure 3(a)

First Split:

Threshold Variable Language

Threshold Estimate 0.44720000
95% Confidence Interval: [0.411000, 0.458600]

25
 Figure 4(b): Hansen (2000) confidence intervals for Figure 3(b)

First Split

Threshold Variable Expropriation Risk

Threshold Estimate 0.45538462
95% Confidence Interval: [0.455384, 0.455384]

Second Split

Threshold Variable Expropriation Risk

Threshold Estimate 0.66692308
95% Confidence Interval: [0.649230, 0.677692]

26
 Figure 4(b) (cont.): Hansen (2000) confidence intervals for Figure 3(b)

Third Split

Threshold Variable Landlocked (LCR100KM)

Threshold Estimate 0.38887620
95% Confidence Interval: [0.270992, 0.582102]

Fourth Split

Threshold Variable Tropical Area (TROPICAR)

Threshold Estimate 0.56140000
95% Confidence Interval: [0.561400, 0.561400]

27
 Figure 4(c): Hansen (2000) confidence intervals for Figure 3(c)

First Split:

Threshold Variable Language

Threshold Estimate 0.44720000
95% Confidence Interval: [0.335800, 0.597500]

28
 Table 3◊: Model specifications
Baseline Exercise Solow Exercise Parsimonious Exercise
Slope Threshold Slope Threshold Slope Threshold
1 KG Tropics X X - -
2 Landlocked (LCR100KM) X X - -
3 Tropical Area (TROPICAR) X X X X X
4 Language X X X X X
5 Ethnic Fractionalization X X - -
6 Ethnic Tension X X - -
7 Political Rights X X X X
8 Assassinations X X X - -
9 Executive Constraints X X - -
10 Expropriation Risk X X X - -
11 Governance (KKZ96) X X X X
12 East Asia X X X
13 Sub-Saharan Africa X X X
14 Latin America & Caribbean X X X
15 M2 X X X - -
16 Budget Surplus X X X - -
17 Inflation X X X - -
18 Openness X X X - -
19 Aid X X X X X X
20 Population Growth X X X
21 Investment X X X
22 Schooling X X X
23 Initial Income X X X X X X
24 Dummy 1980-94 X X X
Number of obs. 98 98 123

◊
This Table describes the set of variables in the model space for each of the three specifications – Baseline, Solow, and
Parsimonious. An “X” means that a variable was designated either to be a potential threshold variable, or a slope covariate (or, as
the case may be, both). An “-” means that that variable was dropped from the model space.

29
 Table 4∇: Country breakdowns by growth regimes for BTREED and TR models
BTREED TR
Country Baseline Solow Pars. Baseline Solow Pars.

Africa

- -
Benin - 1 - 2
Botswana 1 1 1 1c 4 1c
Cameroon 2 1 1 2 3 2
Central African Rep. - - 1 - - 2
Congo, Rep. 2 1 1 2 1 2
Gambia 2 1 1 2 5 2
Ghana 2 1 1 2 3 2
Kenya 2 1 1 2 3 2
Lesotho - - 1 - - 1
Malawi 2 1 1 2 3 2
Mali 2 1 1 2 1 2
Mauritius - - 1 - - 2c
Mozambique - - 1 - - 2
Niger 2 1 1 2 3 2
Senegal 2 1 1 2 3 2
Sierra Leone 2 1 1 2 3 2
South Africa 2 1 1 2 4 2
Togo 2 1 1 2 3 2
Uganda 2 2 1 2 1 2
Congo, Dem. Rep. 2 2 1 2 1 2
Zambia 2 2 1 2 3 2
Zimbabwe 2 1 1 1c 3 2c

Asia

Bangladesh 1 1 1 1 2 1
China 1 1 1 1 4 1
India 2 1 1 2 4 2
Indonesia 2 1/2 1 2 5 2
Korea, Rep. of 1 1 1 1 5 1
Malaysia 2 1 1 2 5 2c
Nepal - - 1 - - 2
Pakistan 2 1 1 2 2 2
Papua New Guinea 1 1 1 1 5 1c
Philippines 2 1 1 2 3 2
Singapore - - 1 - - 1c
Sri Lanka 2 1 1 2 3 2c
Thailand 2 1 1 2c 4 2

∇
A superscript “c” denotes countries within Hansen’s 95% CI bound for the first threshold split as given in Figures 4(a)-(c).
“1/2” indicates that a country was in one regime in one time period and another in the other.

30
 Table 4∇ (cont.): Country breakdowns by growth regimes for BTREED and TR models
BTREED TR
Country Baseline Solow Pars. Baseline Solow Pars.

Latin America & the Caribbean

Argentina 1 2 1 1 2 1
Bolivia 1 1/2 1 1 3 1
Brazil 1 2 1 1 4 1
Chile 1 2 1 1 5 1
Colombia 1 1 1 1 4 1
Costa Rica 1 1 1 1 3 1
Dominican Republic 1 1 1 1 3 1
Ecuador 1 1 1 1 3 1
Guatemala 2 1 1 2c 1 2c
Honduras 1 1 1 1 3 1
Jamaica 1 1 1 1 3 1
Mexico 1 1 1 1 4 1
Nicaragua 1 2 1 1 1c 1
Panama - - 1 - - 1c
Paraguay 2 1 1 2 2 2c
Peru 1 2 1 1 3 1c
Trinidad & Tobago 1 1 1 1 5 1
Uruguay 1 1 1 1 2 1
Venezuela 1 1 1 1 3 1

Europe, North Africa, & Middle

East

Algeria 2 1 1 1c 2 1c
Egypt, Arab Rep. 1 1 1 1 2 1
Hungary 1 1 1 1 5 1
Iran 2 1 1 2 1 2
Israel 2 2 1 2 5 2c
Jordan 1 1 1 1 2 1
Poland 1 2 1 1 5 1
Syrian Arab Rep. 1 1 1 1 2 1
Tunisia 1 1 1 1 2 1
Turkey 1 1 1 1 4 1

∇
A superscript “c” denotes countries within Hansen’s 95% CI bound for the first threshold split as given in Figures 4(a)-(c).
“1/2” indicates that a country was in one regime in one time period and another in the other.

31
 Table 5+: BTREED coefficient estimates for growth regimes

Baseline Solow Parsimonious

(1) (2) (1) (2) (1)
Constant 0.1798*** -0.0160 -0.0027 -0.0660 0.1329**
(0.0559) (0.0848) (0.0499) (0.4004) (0.0555)
Dummy 1980-94 -0.0117* -0.0022 -0.0134 -0.0338 -0.0089**
(0.0061) (0.0046) (0.0039) (0.0249) (0.0039)
Tropical Area -0.0151*** -0.0155** - - -0.0084**
(TROPICAR) (0.0051) (0.0064) - - (0.0040)
Language 0.0004 -0.0003 - - -0.0010
(0.0184) (0.0123) - - (0.0067)
Political Rights - - - - 0.0084
- - - - (0.0065)
Assassinations 0.0019 -0.0011 - - -
(0.0026) (0.0024) - - -
Expropriation Risk 0.0735*** 0.0271* - - -
(0.0238) (0.0155) - - -
Governance - - - - 0.0128***
(KKZ96) - - - - (0.0033)
East Asia -0.0062 0.0129** - - 0.0115*
(0.0081) (0.0055) - - (0.0059)
Sub-Saharan Africa 0.0078 0.0155*** - - 0.0039
(0.0059) (0.0051) - - (0.0044)
Latin America & -0.0020 0.0487*** - - 0.0068
Caribbean (0.0057) (0.0070) - - (0.0047)
M2 0.0042 -0.0072 - - -
(0.0192) (0.0205) - - -
Budget Surplus 0.1509** -0.0887** - - -
(0.0621) (0.0424) - - -
Inflation 0.0007 -0.0118*** - - -
(0.0042) (0.0035) - - -
Openness 0.0069 0.0227*** - - -
(0.0077) (0.0068) - - -
Aid -0.0022 -0.0045*** -0.0009 -0.0035 -0.0027**
(0.0017) (0.0013) (0.0011) (0.0088) (0.0011)
Population Growth -0.0008 -0.0486* -0.0392** -0.0672 -0.0019
(0.0199) (0.0280) (0.0177) (0.1438) (0.0184)
Investment 0.0072 0.0150*** 0.0260*** -0.0172 0.0173***
(0.0083) (0.0036) (0.0037) (0.0405) (0.0030)
Schooling -0.0018 0.0078*** 0.0082*** 0.0117 0.0061***
(0.0047) (0.0025) (0.0024) (0.0253) (0.0022)
Initial Income -0.0260*** -0.0181*** -0.0174*** -0.0045 -0.0198***
(0.0039) (0.0032) (0.0030) (0.0344) (0.0029)

Number of obs. 49 49 85 13 123

+
Dependent variable is the growth rate of real GDP per capita across, respectively, the periods 1965-79 and 1980-94. Standard
errors are in parentheses. Model specifications are described in Table 3. “***” indicates significance at the 1% level while “**”
indicates significance at the 5% level and “*” at the 10% level.

32
 Table 6+: TR coefficient estimates for growth regimes
Baseline Solow
(1) (2) (1) (2) (3) (4) (5)
Constant 0.1858*** -0.0421 -2.2178*** 0.1827*** 0.1292* -0.0084 0.1559***
(0.0451) (0.0572) (0.1590) (0.0247) (0.0720) (0.1940) (0.0500)
Dummy -0.0109** -0.0038 -0.0533*** -0.0157*** -0.0079* -0.0157*** -0.0049
1980-94 (0.0052) (0.0030) (0.0069) (0.0033) (0.0043) (0.0039) (0.0047)
Tropical Area -0.0151*** -0.0169*** - - - - -
(TROPICAR) (0.0041) (0.0049) - - - - -
Language -0.0052 0.0060 - - - - -
(0.0099) (0.0122) - - - - -
Political - - - - - - -
Rights - - - - - - -
Assassinations 0.0015 -0.0004 - - - - -
(0.0013) (0.0014) - - - - -
Expropriation 0.0759*** 0.0238** - - - - -
Risk (0.0184) (0.0093) - - - - -
Governance - - - - - - -
(KKZ96) - - - - - - -
East Asia -0.0064 0.0147*** - - - - -
(0.0068) (0.0048) - - - - -
Sub-Saharan 0.0079 0.0149*** - - - - -
Africa (0.0049) (0.0041) - - - - -
Latin America -0.0012 0.0494*** - - - - -
& Caribbean (0.0050) (0.0062) - - - - -
M2 -0.0035 0.0032 - - - - -
(0.0154) (0.0164) - - - - -
Budget 0.1494*** -0.0877*** - - - - -
Surplus (0.0445) (0.0285) - - - - -
Inflation 0.0001 -0.0117*** - - - - -
(0.0030) (0.0021) - - - - -
Openness 0.0061 0.0250*** - - - - -
(0.0048) (0.0061) - - - - -
Aid -0.0016 -0.0040*** -0.0067*** -0.0015** -0.0072*** 0.0155*** -0.0068***
(0.0013) (0.0009) (0.0016) (0.0006) (0.0017) (0.0036) (0.0008)
Population -0.0003 -0.0581*** -1.0114*** 0.0090 -0.0120 -0.0400 0.0280
Growth (0.0157) (0.0187) (0.0888) (0.0090) (0.0262) (0.0676) (0.0201)
Investment 0.0069 0.0128*** 0.0132 -0.0253*** 0.0171*** 0.0347*** 0.0346***
(0.0071) (0.0027) (0.0086) (0.0058) (0.0030) (0.0082) (0.0040)
Schooling -0.0006 0.0075*** 0.0280*** 0.0037 0.0042** 0.0198*** -0.0036
(0.0030) (0.0020) (0.0071) (0.0045) (0.0019) (0.0060) (0.0039)
Initial Income -0.0264*** -0.0179*** -0.0422*** -0.0079*** -0.0233*** -0.0192*** -0.0166***
(0.0027) (0.0021) (0.0096) (0.0025) (0.0029) (0.0044) (0.0050)

Number of
52 46 11 16 38 15 18
obs.

+
Dependent variable is the growth rate of real GDP per capita across, respectively, the periods 1965-79 and 1980-94. Standard
errors are in parentheses. Model specifications are described in Table 3. “***” indicates significance at the 1% level while “**”
indicates significance at the 5% level and “*” at the 10% level.

33
 Table 6+ (cont.): TR coefficient estimates for growth regimes
Parsimonious
(1) (2)
Constant 0.1811*** -0.0307
(0.0652) (0.0662)
Dummy 1980-94 -0.0099** -0.0069
(0.0047) (0.0046)
Tropical Area -0.0101** -0.0036
(TROPICAR) (0.0042) (0.0048)
Language 0.0140 0.0192
(0.0117) (0.0163)
Political Rights 0.0097 0.0052
(0.0081) (0.0085)
Assassinations - -
- -
Expropriation - -
Risk - -
Governance 0.0158*** 0.0098**
(KKZ96) (0.0032) (0.0045)
East Asia 0.0010 0.0089*
(0.0046) (0.0051)
Sub-Saharan -0.0014 0.0028
Africa (0.0068) (0.0049)
Latin America & - -
Caribbean - -
M2 - -
- -
Budget Surplus - -
- -
Inflation - -
- -
Openness - -
- -
Aid -0.0023 -0.0040***
(0.0014) (0.0013)
Population 0.0089 -0.0666***
Growth (0.0187) (0.0212)
Investment 0.0263*** 0.0128***
(0.0074) (0.0034)
Schooling 0.0033 0.0073***
(0.0032) (0.0024)
Initial Income -0.0250*** -0.0171***
(0.0043) (0.0035)

Number of obs. 60 63

+
Dependent variable is the growth rate of real GDP per capita across, respectively, the periods 1965-79 and 1980-94. White
standard errors are in parentheses. Model specifications are described in Table 3. “***” indicates significance at the 1% level
while “**” indicates significance at the 5% level and “*” at the 10% level.

34
4.1.1 Analysis of Baseline Tree Diagrams

Our Baseline results for BTREED and TR are essentially in agreement. In terms of the tree
structures, comparing Figure 2(a) with Figure 3(a), we find that both BTREED and TR identify two
growth regimes defined by ethnolinguistic heterogeneity (language). The size of the regimes are roughly
equal. We also note that the regime with ethnolinguistic heterogeneity falling below the threshold value
(regime (1)) is initially richer and has a faster rate of per capita income growth on average than the regime
where ethnolinguisitc heterogeneity falls above the threshold value (regime (2)). If we look at the country
breakdowns for the regimes (columns 1 and 4 of Table 4), we find that the breakdowns are also very
similar for both BTREED and TR. Those countries for which the two are not in agreement --- i.e.,
Algeria and Zimbabwe --- fall within Hansen's 95% confidence bounds.
The countries in the high ethnolinguistic fractionalization growth regime are predominantly Sub-
Saharan African countries (with the key exception of Botswana which is classified as belonging to the
other regime). On the other hand, the low ethnolinguistic fractionalization growth regime is composed
mostly of Latin American and Caribbean countries (with the exception of Paraguay and possibly
Guatemala). The countries in Asia, Europe, North Africa, and the Middle East have more heterogeneous
predicted growth experiences. While most countries in Asia appear to fall in the worse performing (high
ethnolinguistic fractionalization) regime, some such as Bangladesh, China, South Korea, and Papua New
Guinea are predicted to fall in the better performing (low ethnolinguistic fractionalization) group.
Similarly, while most countries in the set that we label for convenience as Europe, North Africa, and the
Middle East are classified as belonging to the better performing (low ethnolinguistic fractionalization)
regime, there are notable exceptions such as Iran and Israel that get placed into the worse performing
(high ethnolinguistic fractionalization) regime.
The finding that ethnolinguistic fractionalization is an important driver of heterogeneity in growth
is consistent with work by Easterly and Levine (1997) and Alesina et. al. (2003). Easterly and Levine, in
particular, argue that ethnolinguistic fractionalization is critically important in accounting for Sub-
Saharan Africa's underdevelopment. Given that the set of countries in this study are necessarily confined
to the set of developing countries (aid recipients), the fact that almost all Sub-Saharan African countries
(with the lone and well-documented exception of Botswana (see, for instance, Acemoglu, Johnson, and
Robinson (2003)) are separated out in this way and classified under the worse performing regime would
appear to provide especially strong support for Easterly and Levine's hypothesis.

4.1.2 Baseline Parameter Estimates for Multiple Growth Regimes

The evidence on the nature of the growth regimes has important implications for the recent
debates over the effect of aid on growth. In contrast to the current literature, our Baseline results suggest

35
that the effect of aid on growth (if any) does not depend on policy variables but rather depends on the
fundamental determinant, ethnolinguistic fractionalization. Specifically, columns 1 and 2 of Table 5 (for
BTREED) and Table 6 (for TR) provide the results for the two growth regimes for the respective sample
split methods. We find that aid has no significant effect for countries in the regime with low
ethnolinguistic fractionalization, but, has a negative and highly significant (at the 1% level) effect for
countries in the regime with high ethnolinguistic fractionalization. Since the countries in the latter regime
are, on average, initially poorer to begin with, our results suggest that aid is in fact potentially counter-
productive for this set of countries. Our results therefore are consistent with Easterly, Levine, and
Roodman (2004) and Roodman (2004).
In terms of the coefficient estimates and standard errors for growth determinants, the results in
Tables 5 and 6 are revealing. For both BTREED and TR, we find that the coefficients to initial per capita
income for countries in both the high and low ethnolinguistic fractionalization growth regimes are highly
significant at the 1% level and negative. A negative coefficient on log initial income per capita is
typically taken as evidence in the literature that poorer countries within the regime are catching up with
richer countries in the same regime after controlling for other growth factors. Our findings are therefore
consistent with the interpretation in the literature of “conditional convergence” within each of the two
growth regimes. In this sense, the findings appear to suggest the existence of two convergence clubs
defined by ethnolinguistic fractionalization, where countries within each club are converging to a
different steady state.
Both BTREED and TR find that climate (TROPICAR) has a significant negative effect on growth
for countries in both regimes, while property rights institutions (expropriation risk) exhibit a significant
positive relationship for both regimes. Macroeconomic policies also appear to be important for countries
in the worse performing (high ethnolinguistic fractionalization) regime. For instance, conditional on the
other growth determinants, countries with higher rates of inflation experience significantly lower growth
rates in this regime. Finally, the Solow variables; i.e., population growth, investment, and schooling, are
all significant and have the correct signs; that is, negative, positive, and positive, respectively, for
countries in the worse performing (high ethnolinguistic fractionalization) regime, although they are
insignificant for countries in the low ethnolinguistic fractionalization regime.
In sum, the findings from the Baseline specification, which is meant to reflect the literature at
large, would appear so far to be stable --- in the sense that both BTREED and TR are in agreement --- and
reflect the consensus of the recent work on the relationship between aid and growth. Nevertheless, we
would like to go a step further in order to investigate whether the results we obtained for the Baseline
specification holds when we perturb the exercises a bit. We turn now, therefore, to the results for the
Solow and Parsimonious specifications.

36
4.1.3 Results from Alternative Specifications

Figures 2(b) and 3(b) show the tree diagrams for BTREED and TR, respectively, for the Solow
specification. Recall that the only difference between the Solow and Baseline specifications is that,
except for aid and the canonical Solow variables, all other variables that were pre-assigned to be slope
covariates in the Baseline setup are now assigned to be solely potential threshold variables. As can be
seen, the tree diagrams for the Solow specification are dramatically different from those obtained for the
Baseline specification. The BTREED tree for the Solow specification is split into two regimes. But,
now, the threshold variable selected is no longer ethnolinguistic fractionalization, but inflation.
Furthermore, the set of countries within each regime also differs dramatically from what we obtained
before. There are now 85 observations in one regime (the low inflation regime) and 13 in the other (the
high inflation regime) as opposed to 49 for both under the Baseline specification. Also, as far as the
breakdown of countries into regimes is concerned (see column 2 of Table 4), there does not appear to be
such a strong separation according to geographic regions as we obtained before. Essentially, a few
countries from each regional grouping with particularly high levels of inflation are picked out to form the
high inflation regime. Nevertheless, if we look at the estimates for the relationship between aid and
growth (see columns 4 and 5 of Table 5), we see that it is (negative but) insignificant from zero for both
regimes. These results, therefore, should not be taken as evidence to support the position that aid may be
beneficial to those developing countries that are made to implement desirable macroeconomic policies as
precondition to receiving aid (policy conditionality).
The situation for TR is worse. As can be seen from Figure 3(b), TR now splits the set of
countries into five growth regimes according to institutions and geography. These are the low-quality
institutions regime (regime (1)), the medium-quality institutions/less tropical regime (regime (2)),
medium-quality institutions/more tropical regime (regime (3)), the high-quality institutions/less
geographically accessible regime (regime (4)), and the high-quality institutions/more geographically
accessible regime ((regime (5)). The classification of countries into regimes is therefore not at all similar
to what was achieved before under the Baseline specification. However, if we consider the classification
of observations for the Solow specification according to just the first split; i.e., according to whether or
not expropriation risk for countries are above or below the threshold value of 0.455, then, the sample
splits obtained under TR are somewhat similar to those obtained under BTREED. For instance, if we
compare the country breakdown for the first regime in TR with the second regime in BTREED (i.e.,
compare columns 2 and 4 of Table 4), we see that these are largely similar. Therefore, at least at some
level, we find that BTREED and TR do agree on the classification of countries into regimes. However,
even if we are willing to concede that, we cannot escape from the fact that BTREED and TR do not agree

37
on the exact source of heterogeneity. Given the same choices for possible threshold variables, BTREED
chooses macroeconomic policies (i.e., inflation) while TR chooses institutions (expropriation risk). It is
therefore very difficult to assign a consistent structural interpretation to these findings.
The tree diagrams for the Parsimonious specification bear somewhat better news. Recall that the
difference between the Parsimonious specification and the Baseline and Solow specifications is that for
the Parsimonious specification, we drop the set of policy variables (except for aid). The reason we did so
was to attempt to maximize the number of observations in the sample. If we compare the TR tree for the
Parsimonious specification (Figure 3(c)) with that for the Baseline specification (Figure 3(a)), we see that
they are identical. However, when we carry out the analogous comparison for BTREED (i.e., cf. Figure
2(c) with Figure 2(a)), we find that BTREED has selected a single regime (no heterogeneity) model for
the Parsimonious specification. Hence, yet again, there is no clear message from our tree diagrams.
In other unreported exercises where we consider alternative choices for designating variables as
threshold, slope, or both for these three specifications, we find very little evidence of tree stability. As
represented by the trees in Figures 2(a)-(c) and 3(a)-(c), we find that the trees we obtain tend to (1) vary
in size, (2) classify countries quite differently, and (3) choose different threshold variables; occasionally
by fundamental determinants (such as geography, institutions, or ethnolinguistic fractionalization) and
other times by policy variables (such as aid, inflation, or government budget surplus). The instability of
the trees obtained under both BTREED and TR renders attempts to interpret them structurally to be,
unfortunately, precarious. We are forced to conclude that there is very little evidence of a robust/reliable
typology that would relate aid to growth. Another way of putting this is that we are severely limited in
our ability to engage in tree (model) selection in any sensible way.
Nevertheless, there are some strong regularities in the results across specifications (please refer to
Tables 5 and 6). We find that the relationship between aid and growth tends to be negative with most
cases being significant. The exception is to be found in the high-quality institutions/less geographically
accessible regime (regime (4)) for the Solow specification where the relationship between aid and growth
appears to be positive and highly significant. Also, consistent with the larger debate in the growth
literature over the importance of institutions versus geography to economic performance, we find that, at
least for the set of developing countries in our sample, both these fundamental determinants are important
to growth. Climate (TROPICAR) has a significant negative effect on growth for countries across
specifications and regimes with the sole exception of the high ethnolinguistic fractionalization regime
(regime (2)) for the Parsimonious specification for which its effect is also negative but insignificant.
Similarly, property rights institutions (as measured by expropriation risk and governance (KKZ96)) have
a significant positive effect on growth for countries in all regimes and for all specifications. We also find
that conditional convergence holds strongly in the growth regression. For almost all regimes across all

38
specifications (the exception being regime (2) of the Solow specification), we find the coefficient to initial
per capita income to be negative and highly significant.

4.2 Robust Relationship between Aid and Growth

These regularities are encouraging because they suggest that even though the instability of the
trees we obtained implies that finding one that would be robust enough to tell a structurally interpretable
story about the relationship between aid and growth may be difficult, there may be a way for us
nevertheless to give policymakers some sense of a “robust” relationship between growth determinants of
interest, such as aid, and growth. As described in the Econometric Methodology section above, we
attempt to uncover such robust relationships using partial dependency plots generated using the BART
algorithm.
Figure 5(a) shows the partial (i.e., conditioning upon heterogeneity in terms of the other
covariates) dependency plot of growth on aid for the Baseline/Solow set of variables. Similarly, the top
left-hand graph in Figure 6 shows the partial dependency plot of growth on aid for the Parsimonious set
of variables. We also show the corresponding MCMC posterior 95% confidence bounds around the point
estimates in both figures. We find that the (partial) relationship between growth and international aid is
probably not nonlinear, and very likely negative. Nevertheless, the posterior 95% confidence bounds do
not allow us to reject the possibility that the relationship is flat.

39
 Figure 5(a): Partial dependence plot for aid (Baseline/Solowϒ Model)

ϒ
The BART partial dependence diagrams are the same for the Baseline and Solow Models (see Table 3 for model specifications)
since both model specifications have the same set of variables.

40
 Figure 5(b): Partial dependence plots for solow variables (Baseline/Solowϒ Model)

ϒ
The BART partial dependence diagrams are the same for the Baseline and Solow Models (see Table 3 for model specifications) since both model specifications have the same set
of variables.

41
 Figure 5(c): Partial dependence plots for macroeconomic policy and institutions (Baseline/Solowϒ Model)

ϒ
The BART partial dependence diagrams are the same for the Baseline and Solow Models (see Table 3 for model specifications) since both model specifications have the same set
of variables.

42
 Figure 5(d): Partial dependence plots for other fundamental determinants (Baseline/Solowϒ Model)

ϒ
The BART partial dependence diagrams are the same for the Baseline and Solow Models (see Table 3 for model specifications) since both model specifications have the same set
of variables.

43
 Figure 6: Partial dependence plots (Parsimoniousϒ Model)

ϒ
Please see Table 3 for model specification.

44
 The rest of Figures 5 and 6 show the partial dependence plots for the other growth variables and
growth for the respective sets of variables (i.e., Baseline/Solow and Parsimonious). While some of these
partial dependence plots --- notably those for ethnolinguistic fractionalization (language) --- are
suggestive of possible nonlinear relationships, the large posterior 95% confidence bounds make it difficult
for us to find conclusively in favor of this outcome. Taken together with the sample split (i.e., TR and
BTREED) results, the evidence for a nonlinear relationship between ethnolinguistic fractionalization and
growth appears to be the strongest amongst the set of regressors.
Like Figures 2(a) and 3(a) for the Baseline model, the partial dependence plots for ethnolinguistic
fractionalization suggest that there exists a positive relationship between growth and ethnolinguistic
fractionalization when the degree of fractionalization is low (below approximately 0.45), and a negative
relationship when the degree of fractionalization is high (above 0.45).
The plots also show the correct relationships, as suggested by the neoclassical growth model,
between the Solow variables and growth; i.e., negative for population growth, positive for investment and
schooling, and negative for initial per capita income. They confirm the regularities from the TR and
BTREED findings that property rights institutions (expropriation risk and governance (KKZ96)) have
strong positive relationships with growth while climate (TROPICAR) has a strong negative relationship.
Finally, policies such as trade openness and inflation also appear to have (positive and negative,
respectively) consequences for growth.

45
 5. Conclusion

In this paper, we attempt to characterize the relationship between aid and growth using recently
developed sample splitting methods such as Bayesian tree regression (BTREED) and threshold regression
(TR). Our aim is to uncover the factors that cause divergent effects, if any, of aid on growth for particular
subsets of countries. We also sought evidence of a nonlinear relationship between aid and growth. While
our results are suggestive of an interaction effect between ethnolinguistic fractionalization and aid --- so
that countries with levels of ethnolinguistic fractionalization above a threshold value experience a
negative relationship between aid and growth, while those with ethnolinguistic fractionalization below the
threshold experience no growth effects --- our efforts are severely complicated by the high degree of tree
instability, and therefore model uncertainty, associated with these sample splitting methods.
A key methodological contribution of our paper therefore is to implement in the growth context a
strategy for obtaining robust characterizations of the aid/growth nexus using model averaging methods
such as Bayesian Additive Regression Trees (BART). When we do so, we find no evidence of a
nonlinear relationship between aid and growth. The relationship between aid and growth is, in fact, likely
to be negative. Our findings therefore leave us skeptical as to any potential positive contributions to
growth from increasing foreign aid to developing countries. Nevertheless, the evidence from the data is
noisy (as seen from the large posterior 95% confidence bounds we obtained), and we therefore expect the
debate over the role of foreign aid in promoting growth to continue.
An additional caveat to the interpretation of our findings is the problem of the endogeneity of aid.
The problem of endogeneity is one that is endemic to the growth literature. The standard method for
getting around endogeneity is to use instrumental variables. However, instrumental variables estimation
procedures for sample splitting methods are scarce and non-existent for the case of interest here (i.e.,
where we have endogeneity in the threshold variables). Further, as argued by Brock and Durlauf (2001),
the open-ended nature of growth theories present difficulties for researchers arguing for the validity of
instruments on the grounds of predetermination. We therefore suggest that our findings, while strongly
consistent with recent findings in the literature, be interpreted with appropriate caution.
Although we do not find a nonlinear relationship between aid and growth, the methodology
developed in this paper can be used to identify typologies on other outcome variables, such as those
included in the Millennium Development Goals.

46
 References

Acemoglu, D., J. A. Robinson, and S. Johnson (2001), “The Colonial Origins of Comparative Development: An Empirical
Investigation,” American Economic Review, 91, 1369-1401.
Acemoglu, D., J. A. Robinson, and S. Johnson (2003), “An African Success Story: Botswana,” in In Search of Prosperity:
Analytical Narrative on Economic Growth, edited by Dani Rodrik, Princeton University Press.
Alesina, A., A. Devleeschauwer, W. Easterly, S. Kurlat, and R. Wacziarg (2003), “Fractionalization,” Journal of Economic
Growth, 8(2), 155-94.
Azariadis, C. and A. Drazen (1990), “Threshold Externalities in Economic Development,” Quarterly Journal of Economics, 105,
501-526.
Bai, J., (1997), “ Estimating Multiple Breaks One at a Time,” Econometric Theory, 13, 315-352.
Banks, A. (2002), Cross-National Time-Series Data Archive. Bronx, NY: Databanks International.
Barro, R. J. (1991), “Economic Growth in a Cross-Section of Countries,” Quarterly Journal of Economics, 106(2), 407-43.
Barro, R. J. and J-W Lee (2001), “International Data on Educational Attainment: Updates and Implications,” Oxford Economic
Papers, 53(3), 541-63.
Breiman, L. (1996), “Bagging Predictors,” Machine Learning, 26, 123-140.
Breiman, L. (2001), “Random Forests,” Machine Learning, 45, 5-32.
Breiman, L., J. H. Friedman, R. A. Olsen, and C. J. Stone (1984), Classification and Regression Trees, Wadsworth, Belmont.
Brock, W. and S. Durlauf, (2001), “Growth Empirics and Reality,” World Bank Economic Review, 15, 2, 229-272.
Burnside, C. and D. Dollar (2000), “Aid, Policies, and Growth,” American Economic Review, 90(4), 847-868.
Chang, C. C., E. Fernandez-Arias, and L. Serven (1999), “Measuring Aid Flows: A New Approach,” Washington: World Bank,
Development Economics Research Group.
Chipman, H., E. George, and R. McCullough (1998), “Bayesian CART Model Search,” Journal of the American Statistical
Society, 93, 935-960.
Chipman, H., E. George, and R. McCullough (2002), “Bayesian Treed Models,” Machine Learning, 48, 299-320.
Chipman, H., E. George, and R. McCullough (2005), “BART: Bayesian Additive Regression Trees,” University of Chicago GSB
mimeo.
Collier, P. and J. Dehn (2001), “Aid, Shocks, and Growth,” World Bank Policy Research Working Paper No. 2688.
Cohen-Cole, E., S. N. Durlauf, and G. Rondina (2005), “Nonlinearities in Growth: From Evidence to Policy; Foreign Aid and
Macroeconomic Policy,” University of Wisconsin SSRI Working Paper 2005-09.
Collier, P. and D. Dollar (2002), “Aid Allocation and Poverty Reduction,” European Economic Review, 46 (8), 1475-1500.
Collier, P. and D. Dollar (2004), ``Development Effectiveness: What Have We Learnt?” Economic Journal, 114 (496), F244-
F271.
Collier, P. and A. Hoeffler (2004), “Aid, Policy and Growth in Post-Conflict Countries,” European Economic Review, 48, 1125-
1145.
Dalgaard, C-J, H. Hansen, and F. Tarp (2004), “On The Empirics of Foreign Aid and Growth,” Economic Journal, 114(496),
F191-F216.
Durlauf, S. N. and P. Johnson, (1995), “Multiple Regimes and Cross-Country Growth Behavior,” Journal of Applied
Econometrics, 10, 4, 363-384.
Durlauf, S. N., A. Kourtellos, and A. Minkin (2001), “The Local Solow Growth Model,” European Economic Review, 45(4-6),
928-940.

47
Durlauf, S. N. (2003), “The Convergence Hypothesis After 10 Years,” University of Wisconsin SSRI Working Paper #2003-06.
Easterly, W. and R. Levine (1997), “Africa's Growth Tragedy: Policies and Ethnic Divisions,” Quarterly Journal of Economics,
112 (4), p.1203-50.
Easterly, W. (2003), “Can Foreign Aid Buy Growth?” Journal of Economic Perspectives, 17(3), 23-48.
Easterly, W., R. Levine, and D. Roodman (2004), “New Data, New Doubts: A Comment on Burnside and Dollar's “Aid, Policies,
and Growth”,” American Economic Review, 94(3), 774-780.
Fernandez, C, E. Ley, and M.F.J. Steel (2001), “Model Uncertainty in Cross-Country Growth Regressions,” Journal of Applied
Econometrics, 16(5), 563-576.
Friedman, J. H. (2001), “Greedy Function Approximation: A Gradient Boosting Machine,” Annals of Statistics, 29, 1189-1232.
Gallup, J., J. D. Sachs, and A. Mellinger (1999), “Geography and Economic Development,” International Regional Science
Review, 22(2), 179-232.
Galor, O. and D. N. Weil (2000), “Population, Technology and Growth: From the Malthusian Regime to the Demographic
Transition,” American Economic Review, 90, 806-828.
Galor, O. and O. Moav (2002), “Natural Selection and the Origin of Economic Growth,” Quarterly Journal of Economics, 117,
1133-1192.
Galor, O. (2005), “Unified Growth Theory: From Stagnation to Growth,” in Aghion, P. and S. N. Durlauf, eds., Handbook of
Economic Growth, Elsevier: Amsterdam.
Guillaumont, P. and L. Chauvet (2001), “Aid and Performance: A Reassessment,” Journal of Development Studies, 37(6), 66-92.
Gonzalo, J and J. Pitarakis (2002), “Estimation and Model Selection based Inference in Single and Multiple Threshold Models,”
Journal of Econometrics, 110, 319-352.
Hansen, B. E. (1996), “Inference When a Nuisance Parameter Is Not Identified under the Null Hypothesis,” Econometrica, 64(2),
413-30.
Hansen, B. E. (2000), “Sample Splitting and Threshold Estimation,” Econometrica, 68(3), 575-604.
Hansen, H. and F. Tarp (2001), “Aid and Growth Regressions,” Journal of Development Economics, 64(2), 547-570.
Hastie, T., R. Tibshirani, and J. H. Friedman (2001), The Elements of Statistical Learning, Springer-Verlag.
Heston A., R. Summers and B. Aten (2002), Penn World Table Version 6.1, Center for International Comparisons at the
University of Pennsylvania (CICUP).
Hong, Y., D. Wang, and X. Zhang (2005), “Identifying Threshold Effects and Typologies in Economic Growth: A Panel
Approach,” IFPRI Working Paper.
Howitt, P., and D. Mayer-Foulkes (2002), “R&D, Implementation and Stagnation: A Schumpeterian Theory of Convergence
Clubs,” NBER Working Paper No. 9104.
Kaufmann, D., A. Kraay and M. Mastruzzi (2005), “Governance Matters IV: Governance Indicators for 1996-2004,” The World
Bank, mimeo.
Loh, W.-Y. (2002), ''Regression Trees with Unbiased Variable Selection and Interaction Detection,” Statistica Sinica, 12, p. 361-
386.
Lubotsky, D. and W. Wittenberg (2006), “Interpretation of Regressions with Multiple Proxies,” Review of Economic Studies,
88(3), 549-562.
Mankiw, N. G., D. Romer, and D. Weil (1992), “A Contribution to the Empirics of Economic Growth,” Quarterly Journal of
Economics, 107, 407-437.
Masanjala, W. and C. Papageorgiou (2004), “The Solow Model with CES Technology: Nonlinearities and Parameter
Heterogeneity,” Journal of Applied Econometrics, 19, 2, 171-202.

48
Rajan, R. G. and A. Subramanian (2005a), “Aid and Growth: What Does the Cross-Country Evidence Really Show?” NBER
Working Paper No. 11513.
Rajan, R. G. and A. Subramanian (2005b), “What Undermines Aid's Impact on Growth?” NBER Working Paper No. 11657.
Rodrik, D., A. Subramanian, and F. Trebbi (2004), “Institutions Rule: The Primacy of Institutions Over Geography and
Integration in Economic Development,” Journal of Economic Growth, 9(2), 131-165.
Roodman, D. (2004), “The Anarchy of Numbers: Aid, Development, and Cross-country Empirics,” Center for Global
Development Working Paper No. 32.
Sachs, J. D. and A. Warner (1995), “Economic Reform and the Process of Global Integration, Brookings Papers on Economic
Activity, 1-118.
Sachs, J. D. (2003), “Institutions Don't Rule: Direct Effects of Geography on Per Capita Income,” NBER Working Paper No.
9490.
Sala-i-Martin, X., G. Doppelhofer, and R. Miller (2004), “Determinants of Long-term Growth: a Bayesian Averaging of Classical
Estimates (BACE) Approach,” American Economic Review, 94, 4, 813-835.
Tan, C. M., (2005), “No One True Path: Uncovering the Interplay between Geography, Institutions, and Fractionalization in
Economic Development,” Tufts University, Dept. of Economics Working Paper No. 2005-12.

49
50
 Related Publications

Agricultural Growth and Investment Options for Poverty Reduction in Rwanda by Xinshen Diao, Shenggen Fan,
Sam Kanyarukiga, and Bingxin Yu, January 2007
Comparing the Evolution of Spatial Inequality in China and India: A Fifty-Year Perspective by Kiran Gajwani,
Ravi Kanbur and Xiaobo Zhang, December 2006
Development Domains for Ethiopia: Capturing the Geographical Context of Smallholder Development Options
by Jordan Chamberlin, John Pender and Bingxin Yu, November 2006
Exploring Growth Linkages and Market Opportunities for Agriculture in Southern Africa by Alejandro Nin Pratt
and Xinshen Diao, September 2006
A Multilevel Analysis of Public Spending, Growth and Poverty Reduction in Egypt by Shenggen Fan, Perrihan Al-
Riffai, Moataz El-Said, Bingxin Yu, and Ahmed Kamaly, September 2006
Assessing Potential Impact of Avian Influenza on Poultry in West Africa – A Spatial Equilibrium Model Analysis
by Liangzhi You and Xinshen Diao, September 2006
Agricultural Trade Liberalization Under Doha: The Risks Facing African Countries by Ousmane Badiane,
September 2006
Shocks, Livestock Asset Dynamics and Social Capital in Ethiopia by Tewodaj Mogues, August 2006
From “Best Practice” to “Best Fit”: A Framework for Analyzing Pluralistic Agricultural Advisory Services
Worldwide by Regina Birner, Kristin Davis, John Pender, Ephraim Nkonya, Ponniah Anandajayasekeram,
Javier Ekboir, Adiel Mbabu, David J. Spielman, Daniela Horna, Samuel Benin, and Marc Cohen, August
2006
Has Trade Liberalization in South Africa Affected Men and Women Differently? by James Thurlow, July 2006
Public Investment to Reverse Dutch Disease: The Case of Chad by Stephanie Levy, June 2006
Moving Up and Down: A New Way of Examining Country Growth Dynamics by Marc* Rockmore and Xiaobo
Zhang, June 2006
Trade Liberalization under CAFTA: An Analysis of the Agreement with Special Reference to Agriculture and
Smallholders in Central America by Sam Morley, May 2006
Shocks, Sensitivity and Resilience: Tracking the Economic Impacts of Environmental Disaster on Assets in
Ethiopia and Honduras by Michael R. Carter, Peter D. Little, Tewodaj Mogues, and Workneh Negatu,
April 2006
Village Inequality in Western China: Implications for Development Strategy in Lagging Regions by Li Xing,
Shenggen Fan, Xiaopeng Luo, and Xiaobo Zhang, February 2006
Does Good Governance Contribute to Pro-poor Growth?: A Review of the Evidence from Cross-Country Studies
by Danielle Resnick and Regina Birner (February 2006)
The Role of Agriculture in Development: Implications for Sub-Saharan Africa by Xinshen Diao, Peter Hazell,
Danielle Resnick, and James Thurlow, February 2006
Asymmetric Property Rights in China’s Economic Growth by Xiaobo Zhang, January 2006

51
Determinants of Change in Household-Level Consumption and Poverty in Uganda, 1992/93-1999/00 by Sam
Benin and Samuel Mugarura, January 2006
Geographic Space, Assets, Livelihoods and Well-being in Rural Central America: Empirical Evidence from
Guatemala, Honduras and Nicaragua by Jeffrey Alwang, Hans G.P. Jansen, Paul B. Siegel and Francisco
Pichon, November 2005
Social Capital and the Reproduction of Economic Inequality in Polarized Societies by Tewodaj Mogues and
Michael R. Carter, November 2005
Rural Nonfarm Development in China and India: The Role of Policies and Institutions by Anit Mukherjee and
Xiaobo Zhang, September 2005
Rural and Urban Dynamics and Poverty: Evidence from China and India by Shenggen Fan, Connie Chan-Kang
and Anit Mukherjee, August 2005
The Dragon and the Elephant: Agricultural and Rural Reforms in China and India by Ashok Gulati, Shenggen
Fan and Sara Dalafi, August 2005
Fiscal Decentralization and Political Centralization in China: Implications for Regional Inequality by Xiaobo
Zhang, July 2005
Growth Options and Poverty Reduction in Ethiopia: A Spatial, Economywide Model Analysis for 2004-15 by
Xinshen Diao and Alejandro Nin Pratt with Madhur Gautam, James Keough, Jordan Chamberlin, Liangzhi
You, Detlev Puetz, Danille Resnick and Bingxi Yu, May 2005
Identifying the Drivers of Sustainable Rural Growth and Poverty Reduction in Honduras by Hans G.P. Jansen,
Paul B. Siegel and Francisco Pichón, April 2005
Public Investment and Poverty Reduction in Tanzania: Evidence from Household Survey Data by Shenggen Fan,
David Nyange and Neetha Rao, April 2005
Achieving Regional Growth Dynamics in African Agriculture by Awudu Abdulai, Xinshen Diao and Michael
Johnson, January 2005
The Road to Pro-poor Growth in Zambia: Past Lessons and Future Challenges by James Thurlow and Peter
Wobst, December 2004
Institutions and Economic Policies for Pro-poor Agricultural Growth by Andrew Dorward, Shenggen Fan,
Jonathan Kydd, Hans Lofgren, Jamie Morrison, Colin Poulton, Neetha Rao, Laurence Smith, Hardwick
Tchale, Sukhadeo Thorat, Ian Urey, and Peter Wobst, November 2004
Strategic Analysis and Knowledge Support Systems for Rural Development Strategies in Sub-Saharan Africa by
Michael Johnson and Danielle Resnick, with Simon Bolwig, Jordan Chamberlin, Liangzhi You, Stanley
Wood, and Peter Hazell, October 2004
Blunt to Sharpened Razor: Incremental Reform and Distortions in the Product and Capital Markets in China by
Xiaobo Zhang and Kong-Yam Tan, August 2004
Road Development, Economic Growth, and Poverty Reduction in China by Shenggen Fan and Connie Chan-
Kang, August 2004

52
Prospects for Growth and Poverty Reduction in Zambia, 2001-2015 by Hans Lofgren, James Thurlow, and
Sherman Robinson, August 2004
Bridging Research, Policy, and Practice in African Agriculture by Steven Were Omamo, July 2004
Smallholder African Agriculture: Progress and Problems in Confronting Hunger and Poverty by Danielle
Resnick, July 2004
Cross-Country Typologies and Development Strategies to End Hunger in Africa by Xiaobo Zhang, Michael
Johnson, Danielle Resnick, and Sherman Robinson, June 2004
The Importance of Public Investment for Reducing Rural Poverty in Middle-income Countries: The Case of
Thailand by Shenggen Fan, Somchai Jitsuchon, and Nuntaporn Methakunnavut, June 2004
Security Is Like Oxygen: Evidence from Uganda by Xiaobo Zhang, May 2004
Food Aid for Market Development in Sub-Saharan Africa by Awudu Abdulai, Christopher B. Barrett, and Peter
Hazell, April 2004
Public Expenditure, Growth, and Poverty Reduction in Rural Uganda by Shenggen Fan, Xiaobo Zhang, and
Neetha Rao, March 2004
The Effect of WTO and FTAA on Agriculture and the Rural Sector in Latin America by Samuel Morley and
Valeria Piñeiro, February 2004
Exploring Regional Dynamics in Sub-Saharan African Agriculture by Xinshen Diao and Yukitsugu Yanoma,
October 2003
Market Opportunities for African Agriculture: An Examination of Demand-Side Constraints on Agricultural
Growth by Xinshen Diao, Paul Dorosh, and Shaikh Mahfuzur Rahman with Siet Meijer, Mark Rosegrant,
Yukitsugu Yanoma, and Weibo Li, September 2003

53
INTERNATIONAL FOOD POLICY RESEARCH INSTITUTE
www.ifpri.org
IFPRI HEADQUARTERS
2033 K Street, NW
Washington, DC 20006-1002 USA
Tel.: +1-202-862-5600
Fax: +1-202-467-4439
Email: ifpri@cgiar.org
IFPRI ADDIS ABABA
P. O. Box 5689
Addis Ababa, Ethiopia
Tel.: +251 11 6463215
Fax: +251 11 6462927
Email: ifpri-addisababa@cgiar.org
IFPRI NEW DELHI
CG Block, NASC Complex, PUSA
New Delhi 110-012 India
Tel.: 91 11 2584-6565
Fax: 91 11 2584-8008 / 2584-6572
Email: ifpri-newdelhi@cgiar.org