A Sensitivity Analysis of Cross-Country Growth Regressions (Levine and Renelt, 1992)

Download as pdf or txt
Download as pdf or txt
You are on page 1of 23

American Economic Association

A Sensitivity Analysis of Cross-Country Growth Regressions


Author(s): Ross Levine and David Renelt
Source: The American Economic Review, Vol. 82, No. 4 (Sep., 1992), pp. 942-963
Published by: American Economic Association
Stable URL: http://www.jstor.org/stable/2117352
Accessed: 24-04-2018 22:12 UTC

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact [email protected].

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms

American Economic Association is collaborating with JSTOR to digitize, preserve and extend
access to The American Economic Review

This content downloaded from 200.235.241.134 on Tue, 24 Apr 2018 22:12:54 UTC
All use subject to http://about.jstor.org/terms
A Sensitivity Analysis of Cross-Country Growth Regressions

By Ross LEVINE AND DAVID RENELT*

A vast literature uses cross-country regressions to search for empirical linkages


between long-run growth rates and a variety of economic policy, political, and
institutional indicators. This paper examines whether the conclusions from
existing studies are robust or fragile to small changes in the conditioning
information set. We find that almost all results are fragile. We do, however,
identify a positive, robust correlation between growth and the share of invest-
ment in GDP and between the investment share and the ratio of intemational
trade to GDP. We clarify the conditions under which there is evidence of per
capita output convergence. (JEL 047)

A vast literature uses cross-country re- over 50 variables have been found to be
gressions to search for empirical linkages significantly correlated with growth in at
between long-run average growth rates and least one regression, readers may be uncer-
a variety of economic policy, political, and tain as to the confidence they should place
institutional factors suggested by theory. in the findings of any one study.2 This paper
Most investigators consider only a small addresses the question: how much confi-
number of explanatory variables in attempt- dence should we have in the conclusions of
ing to establish a statistically significant re- cross-country growth regressions? We find
lationship between growth and a particular that only a few findings can withstand slight
variable of interest. For example, many au- alterations in the list of explanatory vari-
thors who examine the relationship between ables.
measures of fiscal policy and growth ignore As argued by Thomas F. Cooley and
the potential importance of trade policy, Stephen F. LeRoy (1981 p. 825), economic
while those authors who study the empirical theory "...ordinarily does not generate a
ties between trade and growth commonly complete specification of which variables are
ignore the role of fiscal policy.1 Given that to be held constant when statistical tests are
performed on the relation between the de-
pendent variable and the independent var-
iables of primary interest." Thus, many can-
*The World Bank, Washington, DC 20433 and Har- didate regressions have equal theoretical
vard University, Cambridge, MA 02138, respectively.
status, but the estimated coefficients on the
We received helpful comments from Robert Barro,
John Campbell, Maria Carkovic, David Dollar, Bill variables of interest in these regressions may
Easterly, Stanley Fischer, Dale Jorgenson, Lant Pritch- depend importantly on the conditioning set
ett, Dani Rodrik, Paul Romer, Larry Summers, Sara of information. We use a variant of Edward
Zervos, two anonymous referees, and seminar partici-
E. Leamer's (1983) extreme-bounds analysis
pants at Harvard University, M.I.T., the University of
Rochester, the Federal Reserve Board, the World
(EBA) to test the robustness of coefficient
Bank, and the NBER Economic Growth Conference in estimates to alterations in the conditioning
Stanford, April 1991. The findings, interpretations, and
conclusions are only those of the authors and should
not be attributed to the World Bank, its Board of
Governors, its staff, or member countries. Tragically,
we lost David Renelt in the spring of 1991 and the (1990, 1991). Gershon Feder (1983) and Sebastian Ed-
spring of his life. We will greatly miss him and all that wards (1989) study trade policy but ignore fiscal indica-
he would have taught us. tors. Roger Kormendi and Philip Meguire (1985) and
1Studies of fiscal policy that exclude trade indicators Paul M. Romer (1990a) include variables for both.
include Daniel Landau (1983), Rati Ram (1986), Kevin 2See Levine and Renelt (1991) for a review of the
Grier and Gordon Tullock (1989). and Robert J. Barro empirical growth literature.
942

This content downloaded from 200.235.241.134 on Tue, 24 Apr 2018 22:12:54 UTC
All use subject to http://about.jstor.org/terms
VOL. 82 NO. 4 LEVINE AND RENELT: CROSS-COUNTRY GROWTH REGRESSIONS 943

set of information. We study a large number conditions under which one finds evidence
of variables that have been the focus of of convergence of per capita output levels.
attention in a broad collection of growth Before detailing the methodology and the
studies, and we study the statistical relation- results, it is important to emphasize this
ship between growth and a wide array of paper's boundaries. We do not estimate a
newly constructed policy indicators. We structural model, establish causal links,
consider the relationship between growth identify growth determinants, make policy
and a particular variable of interest to be recommendations, improve the measure-
robust if it remains statistically significant ment of policy indicators, or run the full
and of the theoretically predicted sign when gamut of sensitivity analyses discussed by
the conditioning set of variables in the re- Leamer (1985) and Michael McAleer et al.
gression changes. Even though we try not to (1985). We simply examine whether partial
include variables in the conditioning set that, correlations that have drawn the attention
on a priori grounds, measure the same phe- of a large empirical literature are robust or
nomenon as the variable of interst, almost fragile to small changes in the list of right-
all identified relationships are very sensitive hand-side variables. We find that they are
to slight alterations in the conditioning set generally fragile.
of variables, and many publicized coeffi-
cients change sign with small changes in the I. Methodology and Data
conditioning set of variables.
Two themes emerge from our investiga- There does not exist a consensus theoret-
tion. First, there are many econometric ical framework to guide empirical work on
specifications in which measures of eco- growth, and existing models do not com-
nomic policy are significantly correlated with pletely specify the variables that should be
long-run per capita growth rates. The sec- held constant while conducting statistical
ond theme is that the cross-country statisti- inference on the relationship between
cal relationships between long-run average growth and the variables of primary inter-
growth rates and almost every particular est.3 This has produced a diverse and some-
policy indicator considered by the profes- times unwieldy literature, in which few stud-
sion are fragile: small alterations in the ies control for the variables analyzed by
"other" explanatory variables overturn past other researchers. To provide evidence on
results. In particular, the broad array of the sensitivity of past findings to small alter-
fiscal-expenditure variables, monetary-policy ations in the explanatory variables, we use a
indicators, and political-stability indexes variant of the EBA discussed in Leamer
considered by the profession are not ro- (1983, 1985) and Leamer and Herman
bustly correlated with growth; and a huge Leonard (1983). We first describe the EBA
assortment of new indicators that we have and then return to study the empirical
constructed to capture exchange rate, trade, growth literature.
tax, and fiscal-expenditure policies are also Based on the influential work of Kor-
not robustly correlated with growth. This mendi and Meguire (1985), a common fea-
implies that there is not a reliable, indepen- ture of most cross-country growth regres-
dent statistical relationship between a wide sions is that the explanatory variables are
variety of macroeconomic indicators and entered independently and linearly. Thus,
growth.
Our analysis also identifies some robust
relationships and clarifies some past find- 3For example Feder (1983) and Ram (1986) use an
ings. We find a positive and robust correla- augmented neoclassical production function to orga-
tion between growth and the share of in- nize their empirical studies, while Romer (1989) and

vestment in GDP, and we also find that the Barro (1990) use endogenous-growth models that high-
light a few aspects of growth. Kormendi and Meguire
ratio of trade to output is robustly, posi-
(1985) and Grier and Tullock (1989) use a variety of
tively correlated with the investment share. models to motivate an assortment of variables that they
Furthermore, this paper helps clarify the use in exploratory empirical studies.

This content downloaded from 200.235.241.134 on Tue, 24 Apr 2018 22:12:54 UTC
All use subject to http://about.jstor.org/terms
944 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 1992

our EBA uses equations of the form out that the multicollinearity problem really
reflects a weak-data problem. If one is un-
(1) Y = iI+PmM +PzZ + u able to find robust partial correlations in a
cross-section regression, this means that
where Y is either per capita GDP growth or there is not enough independent variation
the share of investment in GDP, I is a set of in that variable to explain cross-country dif-
variables always included in the regression, ferences in growth. Only when one identi-
M is the variable of interest, and Z is a fies a significant correlation while controlling
subset of variables chosen from a pool of for other relevant variables, should one have
variables identified by past studies as poten- much confidence in the correlation. How-
tially important explanatory variables of ever, finding a robust partial correlation
growth. Our EBA involves varying the sub- certainly does not imply that the variable of
set of Z-variables included in the regression interest causes growth. The crucial, though
to find the widest range of coefficient esti- nettlesome, issue of empirically identifying
mates on the variable of interest, M, that causal channels has not been adequately
standard hypothesis tests do not reject. In addressed by the cross-country growth liter-
particular, we first choose a variable that ature.
has been the focus of past empirical studies, Although we agree with Leamer that
M, and run a "base" regression that in- multicollinearity is not a procedural prob-
cludes only the I-variables and the variable lem but rather represents an inability to
of interest. Then we compute the regression identify a statistical relationship that is in-
results for all possible linear combinations sensitive to the conditioning set of informa-
of up to three Z-variables and identify the tion, our purpose is to convince as wide an
highest and lowest values for the coefficient audience as possible that certain partial cor-
on the variable of interest, 8m' that cannot relations are robust or fragile in as mean-
be rejected at the 0.05 significance level. ingful and noncontroversial a manner as
Thus, the extreme upper bound is defined possible. Consequently, we restrict the EBA
by the group of Z-variables that produces in three ways. First, to the list of variables
the maximum value of Pm plus two stan- always included in the regressions, the I-
dard deviations. The degree of confidence variables, we only allow the procedure to
that one can have in the partial correlation choose up to three Z-variables from the
between the Y and M variables can be pool of variables identified as potentially
inferred from the extreme bounds on the important for explaining cross-country
coefficient Pm. If Pm remains significant growth differentials. Consequently, we re-
and of the same sign at the extreme bounds, strict the total number of explanatory vari-
then one can maintain a fair amount of ables included in any one regression to be
confidence in that partial correlation. In eight or fewer.4 The second way we limit
such a case, we refer to the result as the EBA is that we choose a small pool of
"robust." If the coefficient does not remain variables from which the extreme-bounds
significant or if the coefficient changes sign, procedure selects Z-variables; we do not
then one might feel less confident in the search over the massive data set that we
relationship between the M and Y vari- have compiled for any variable that might
ables, because alterations in the condition- cause the variable of interest to lose its
ing information set change the statistical significance. We only search over seven in-
inferences that one draws regarding the dicators that we argue represent a reason-
M-Y relationship. In this case, we refer to able conditioning set. Thus, although we
the result as "fragile." examine the sensitivity of the relationship
One possible objection to this EBA is between growth and more variables than
that it introduces multicollinearity, inflates
the coefficient standard errors, and exagger-
ates the range on the coefficient of interest. 4This total is similar to that used by Kormendi and
Leamer (1978 pp. 170-81), however, points Meguire (1985) and Barro (1991).

This content downloaded from 200.235.241.134 on Tue, 24 Apr 2018 22:12:54 UTC
All use subject to http://about.jstor.org/terms
VOL. 82 NO. 4 LEVINE AND RENELT: CROSS-COUNTRY GROWTH REGRESSIONS 945

that considered by any other study (well income. In addition, the I-variables are con-
over 50 variables), we restrict the pool of sistent with a variety of "new" growth mod-
variables from which the procedure chooses els that rely on constant returns to repro-
Z-variables to only seven. Third, for every ducible inputs or endogenous technological
variable of interest, M, we further restrict change (e.g., Barro, 1990; Romer, 1990b).
the pool of variables from which we choose Furthermore, with these I-variables, we can
Z-variables by excluding variables that, confirm the findings of a large assortment of
a priori, might measure the same phe- empirical studies; and, in recognition of the
nomenon. For example, when we examine issues raised by McAleer et al. (1985), we
the relationship between growth and the show that changes in the I-variables do not
rate of domestic-credit creation over the alter this paper's conclusions.5
1960-1989 period, we do not allow the in- Each of these I-variables has statistical
flation rate to be a Z-variable. These re- and conceptual problems. In keeping with
strictions make it more difficult to implicatethis paper's focus on assessing the statistical
past findings as fragile. sensitivity of past findings, we discuss these
When available, the data cover the period problems only briefly. Measurement prob-
1960-1989 and the Data Appendix de- lems with RGDP60 and SEC may induce
scribes them in detail. The data set includes biased results.6 In the case of GPO, census
119 countries, but we exclude the major oil data may be very poor, and the causal links
exporters. Since detailed government ex- with GYP are ambiguous (see e.g., Gary
penditure and tax information become Becker et al., 1990). Furthermore, in the
available for a wide selection of countries case of SEC, investment in human capital
only in 1974, we conduct much of the analy- represents more than formal schooling, and
sis over the 1974-1989 period. We use two enrollment rates do not control for quality.
data sets: data obtained directly from the Nonetheless, other measures (i.e., primary-
World Bank and International Monetary school enrollment, literacy) yield similar re-
Fund (WB/IMF) and data from Barro sults.7
(1991), which is composed primarily of the There are also problems with including
Robert Summers and Alan Heston (1988) the ratio of physical-capital investment to
data set (SH). We find similar results with GDP as an I-variable. The causal relation-
the two data sets but report primarily re- ship between GYP and INV is ambiguous,
sults based on the WB/IMF data set.

II. Some First Results 5Gregory N. Mankiw et al. (1992) show that our
I-variables (except, instead of SEC, they use average
secondary-school enrollment rates over the sample pe-
We choose the I-variables based on past
riod) enter with the signs predicted by their human-
empirical- studies and economic theory. capital-augmented neoclassical growth model.
When the dependent variable is the average 6For example, if initial income is mismeasured, the
annual growth rate of GDP per capita estimated coefficient on initial income will be biased
(GYP), the I-variables consist of the invest- toward being negative. Romer (1989) shows that initial
income and the literacy rate become insignificant when
ment share of GDP (INV), the initial level one uses instrumental variables to control for measure-
of real GDP per capita in 1960 from SH ment error. Also, see Robert J. Barro and Xavier
(RGDP60), the initial secondary-school en-Sala-i-Martin (1992).
rollment rate (SEC), and the average an- In correspondence, however, Paul M. Romer (pers.
comm.) has noted that when one uses the SH measure
nual rate of population growth (GPO). Al-
of initial income but growth rates computed from
though few empirical studies include all of WB/IMF sources (as we do in this paper), there is no
these variables, most studies control for evidence that measurement error affects the coefficient
some subset. Of the 41 growth studies sur- on initial income.

veyed in Levine and Renelt (1991), 33 in- 7Secondary-school enrollment may be preferable to
primary-school enrollment and literacy rates because
clude the investment share, 29 include pop-
many countries have reached the upper bound for
ulation growth, 13 include a human-capital these other measures. The various education-attain-
measure, and 18 include a measure of initial ment measures that we tried yielded similar results.

This content downloaded from 200.235.241.134 on Tue, 24 Apr 2018 22:12:54 UTC
All use subject to http://about.jstor.org/terms
946 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 1992

and the justification for including many vari- bound, the coefficient on INV is 15.1 with a
ables in growth regressions is that they may t statistic of 4.7. This robust positive rela-
explain INV. If we include INV, the only tionship between GYP and INV is consis-
channel through which other explanatory tent with a wide assortment of growth stud-
variables can explain growth differentials is ies.
the efficiency of resource allocation. To par- A second important finding presented in
tially clarify this ambiguity, we also investi- Table 1 is the robust negative partial corre-
gate the partial correlation between INV lation between GYP and initial income over
and the macroeconomic variables of pri- the 1960-1989 period. The coefficient on
mary interest. RGDP60 is often used to test the conver-
The pool of variables from which we typi- gence hypothesis: a poor country, other
cally allow the EBA to choose Z-variables things equal, tends to grow faster than a
are the average rate of government con- rich country. Bradford J. De Long (1988)
sumption expenditures to GDP (GOV), the and Romer (1987), for example, argue that
ratio of exports to GDP (X), the average there is little empirical support for uncondi-
inflation rate (PI), the average growth rate tional convergence. In accord with Barro
of domestic credit (GDC), the standard de- (1991) and Mankiw et al. (1992), we find
viation of inflation (STDI), the standard de- evidence of conditional convergence over
viation of domestic credit growth (STDD), the 1960-1989 period (i.e., we find a robust
and an index for the number of revolutions negative correlation between GYP and
and coups (REVC). We choose these vari- RGDP60 as long as the I-variables include
ables to form the basis of the conditioning SEC).
information set because the profession has Table 1 also includes EBA tests of GPO
used these variables (or closely related vari- and SEC. As illustrated, one should not feel
ables) as fiscal, trade, monetary, uncer- very comfortable assuming that population
tainty, and political-instability indicators. growth is negatively associated with per
This pool is kept small to make the results capita growth. For some specifications, GPO
more tangible and digestible. The results do enters with a significantly negative coeffi-
not depend importantly on choosing these cient, but it enters with an insignificant co-
variables. efficient with other plausible Z-variables. In
The regression results with the I-variables fact, the coefficient on GPO is insignifi-
over the 1960-1989 period are cantly correlated with growth in the base
regression, which implies that one needs to
(2) GYP = -0.83 - 0.35 RGDP60 select a particular conditioning information
(0.85) (0.14) set to obtain a significant negative coeffi-
cient on GPO. The initial secondary-school
-0.38 GPO + 3.17 SEC enrollment rate enters with a significantly
(0.22) (1.29) positive and robust coefficient, which con-
firms the finding by Barro (1991).
+ 17.5 INV Table 2 provides the EBA of the I-vari-
(2.68) ables using the investment share as the de-
pendent variable. None of the I-variables is
(R2 = 0.46, number of observations = 101; robustly correlated with INV. In fact, the
the coefficient standard errors are in paren- coefficient on initial income is positive for
theses). The variables have the signs pre- some conditioning sets. The entry in the
dicted by a wide class of models, and all but final column indicates whether the partial
GPO are significant at the 0.05 significance correlation is robust or fragile. When the
level. The I-variables explain about half of result is fragile, the column indicates how
the cross-section variance in growth rates. many Z-variables need to be added before
Table 1 presents the EBA tests for each the variable is insignificant or of the
of the I-variables. The investment coeffi- "wrong" sign. In the case of RGDP60, the
cient is positive and robust. At the lower result is fragile. The corresponding "zero"

This content downloaded from 200.235.241.134 on Tue, 24 Apr 2018 22:12:54 UTC
All use subject to http://about.jstor.org/terms
VOL. 82 NO. 4 LEVINE AND RENELT: CROSS-COUNTRY GROWJH REGRESSIONS 947

TABLE 1-SENSITIVITY RESULTS FOR BASIC VARIABLES (DEPENDENT VARIABLE:


GROWTH RATE OF REAL PER CAPITA GDP, 1960-1989)

M-variable ,B Standard error t Countries R2 Other variables Robust/fragile

INV high: 19.07 2.87 6.66 98 0.54 STDI, REVC, GOV robust
base: 17.49 2.68 6.53 101 0.46
low: 15.13 3.21 4.72 100 0.49 X, PI, REVC

RGDP60 high: -0.34 0.13 2.53 98 0.54 STDI, PI, GOV robust
base: -0.35 0.14 2.52 101 0.46
low: -0.46 0.13 3.38 85 0.56 GDC, X, REVC

GPO high: -0.34 0.23 1.48 100 0.48 X, STDI, PI fragilea


base: -0.39 0.22 1.73 101 0.46
low: -0.49 0.20 2.42 85 0.56 X, GDC, REVC

SEC high: 3.71 1.22 3.04 84 0.55 X, GOV, GDC robust


base: 3.17 1.29 2.46 101 0.46
low: 2.50 1.15 2.17 85 0.62 X, STDD, GDC

Notes: The base /8 is the estimated coefficient from the regress


always-included variables (I-variables). The I-variables, when the
capita GDP, are INV (investment share of GDP), RGDP60 (real GDP per capita in 1960), GPO (growth in
population), and SEC (secondary-school enrollment rate in 1960). The high ,B is the estimated coefficient from th
regression with the extreme high bound (P8m +two standard deviations); the low ,B is the coefficient from the
regression with the extreme lower bound.
The "other variables" are the Z-variables included in the base regression that produce the extreme bounds. The
robust/fragile designation indicates whether the variable of interest is robust or fragile. If robust, the text provides
information about further robustness tests.
aThe coefficient is insignificant with only the I-variables included.

TABLE 2-SENSITIVITY RESULTS FOR BASIC VARIABLES (DEPENDENT VARIABLE:


INVESTMENT SHARE, 1960-1989)

M-variable 1, Standard error t Countries R2 Other variables Robust/fragile

RGDP60 high: 0.008 0.003 2.60 86 0.12 GDC, PI, STDI fragile (0)
base: 0.006 0.003 2.13 104 0.04
low: -0.002 0.003 0.52 100 0.24 PI, GOV, REVC

GPO high: -0.002 0.005 0.35 101 0.24 REVC,a GOV, STDI fragile (1)
base: - 0.013 0.005 2.47 106 0.06
low: -0.012 0.006 2.97 87 0.12 GDC, STDI, STDD

SEC high: 0.095 0.024 3.96 86 0.19 GDC, STDD, STDI fragile (1)
base: 0.080 0.023 3.45 106 0.10
low: 0.022 0.024 0.93 102 0.25 REVC,a GOV, STDI

Notes: The base ,B is the estimated coefficient from the regression with
the dependent variable is the investment share, no I-variables are incl
from the regression with the extreme high bound (Pm + two standard d
the regression with the extreme lower bound. M-variable definitions: RGDP60 = real GDP per capita in 1960;
GPO = growth in population; SEC = secondary-school enrollment.
The "other variables" are the Z-variables included in the base regression that produce the extreme bounds. The
underlined variables are the minimum additional variables that make the coefficient of interest insignificant or
change sign. The robust/fragile designation indicates whether the variable of interest is robust or fragile. If fragile,
the number in parentheses indicates how many additional variables need to be added before the variable is
insignificant or of the wrong sign. A zero indicates that the coefficient is insignificant with only the I-variables
included; if robust, the text provides information about further robustness tests.
aIf REVC is excluded from the pool of variables from which Z-variables are chosen, many other variables cause
SEC and GPO to enter insignificantly.

This content downloaded from 200.235.241.134 on Tue, 24 Apr 2018 22:12:54 UTC
All use subject to http://about.jstor.org/terms
948 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 1992

TABLE 3-CROSS-COUNTRY AVERAGES, 1960-1989

Variable Fast-growers Slow-growers t

Share of investment in GDP 0.23 0.17 5.18


Secondary-school enrollment rate in 1960 0.30 0.10 5.46
Primary-school enrollment rate in 1960 0.90 0.54 6.10
Government consumption/GDP 0.16 0.12 3.26
Inflation rate 12.34 31.13 -1.74
Black-market exchange-rate premium 13.57 57.15 - 3.79
Share of exports to GDP 0.32 0.23 2.31

Notes: Mean growth rate = 1.92. Fast-growers are countries with greater than the
mean growth rate; slow-growers are countries with less than the mean growth rate.

indicates that no additional variables need average over the 1960-1989 period tended
to be added to cause the coefficient to be to have a higher share of exports in GDP, a
insignificant. This signifies that the variable higher share of investment in GDP, larger
of interest enters with an insignificant co- primary- and secondary-school enrollment
efficient (or a coefficient of the wrong sign) rates, a lower black-market exchange-rate
in the base regression. Thus, only by selec- premium, and lower inflation rates than
tively adding right-hand-side variables can slower-growing countries. Similarly, Table 4
one find a significant coefficient of the theo- shows that the investment share, the export
retically predicted sign. Interestingly, the share, the black-market premium, and the
finding of a nonrobust relationship between index of revolutions/coups are significantly
RGDP60 and INV and the finding of a correlated with the average real per capita
conditional robust negative partial correla- growth rate. Importantly, however, none of
tion between RGDP60 and GYP suggest these variables is significantly correlated
that per capita income convergence may not with the residuals from the regression of
operate primarily through increases in do- growth on the I-variables. Thus, while many
mestic savings or international capital in- policy indicators are significantly related to
flows. growth, this relationship depends on which
factors are being held constant.
III. Macroeconomic Variables and Growth Kormendi and Meguire (1985), and Barro
(1991) present intuitively appealing results
A. Illustrative Overview for a variety of macroeconomic variables to
explain growth. Table 5 presents equations
This paper's primary aim is to evaluate based on these studies. Equation (ii) is
the degree of confidence one should have in nearly a replication of Barro's (1991) work:
the partial correlations between growth and it includes INV, GPO, RGDP60, GOV,
popular macroeconomic indicators. This measures of initial investment in human
subsection uses two comprehensive studies capital, a dummy variable for socialist eco-
of growth (Kormendi and Meguire [1985] nomic systems, indicators for revolutions
and Barro [1991]) and some simple correla- and coups, dummy variables for countries in
tions to illustrate this paper's two major Latin America and sub-Saharan Africa, and
themes: many indicators of policy, taken it is based primarily on SH data. All the
individually or in groups, are correlated with variables enter with the anticipated sign,
growth, but the relationship between growth and RGDP60, INV, primary-school enroll-
and any particular indicator or group of ment rate, GOV, revolution and coups, and
indicators is typically fragile. The following the continent dummies are significant.
subsections conduct a systematic EBA of Equation (iii) is based on Kormendi and
past findings. Meguire (1985): it includes RGDP60, INV,
Tables 3 and 4 anticipate this paper's GPO, the average annual growth rate in the
findings. Countries that grew faster than share of government consumption to GDP,

This content downloaded from 200.235.241.134 on Tue, 24 Apr 2018 22:12:54 UTC
All use subject to http://about.jstor.org/terms
VOL. 82 NO. 4 LEVINE AND RENELT: CROSS-COUNTRY GRO WTH REGRESSIONS 949

TABLE 4-CROSS-COUNTRY CORRELATIONS

Variable

Variable GYP INV RES X GOV PI GDC STDI STDD BMP REVC

GYP 1.00 0.59* 0.73* 0.32* 0.09 -0.16 -0.04 -0.14 -0.16 - 0.38* - 0.36*
INV 1.00 0.00 0.50* 0.28* -0.04 0.06 -0.01 0.14 - 0.43* - 0.40*
RES 1.00 0.09 -0.13 -0.17 -0.07 -0.16 - 0.30* -0.13 -0.16
X 1.00 0.15 -0.15 -0.07 -0.10 0.05 - 0.22* - 0.34*
GOV 1.00 -0.16 0.08 -0.14 0.17 -0.19 -0.29*
PI 1.00 0.49* 0.97* 0.35* 0.18 0.46*
GDC 1.00 0.39* 0.76* 0.14 0.21
STDI 1.00 0.32* 0.14 0.45*
STDD 1.00 0.15 0.20
BMP 1.00 0.47*
REVC 1.00

Note: The variable RES is the ordinary least-squares residual from the regression of average per capita growth
(GYP) on the I-variables: initial income (RGDP60), population growth (GPO), secondary-school enrollment rate
(SEC), and the investment share (INV).
*Significantly different from zero at the P = 0.05 significance level.

GDC, STDD, the average growth rate in More fundamentally, they illustrate that it is
the share of exports to GDP, and a measure very difficult to isolate a strong empirical
of civil liberties. As in Kormendi and relationship between any particular macroe-
Meguire (1985), this equation uses conomic-policy indicator and long-run
WB/IMF data. The coefficients have the growth.
anticipated signs, and RGDP60, INV, GPO,
GDC, and STDD are significant at the 0.05 B. Fiscal-Policy Indicators
level. Equations (ii) and (iii) explain 68 per-
cent and 61 percent, respectively, of the We first use the EBA to analyze fiscal-
cross-country variation in growth rates. policy indicators. One of the most impor-
Since both equations appear to be rea- tant and frequently studied issues in
sonable but include different independent economics is the role of fiscal policy in
variables, readers may be wary of the find- economic development. Empirical attempts
ings of each study. To highlight this to link aggregate measures of fiscal policy
quandary, we combine the two equations with average per capita growth rates in
using the union of the two sets of explana- cross-country studies have tended to use (i)
tory variables. These results are shown in measures of overall size of the government
equations (iv) and (v), using WB/IMF and in the economy; (ii) disaggregated measures
SH data, respectively. Only INV, RGDP60, of government expenditures; or (iii) mea-
and the continent dummies remain signifi- sures of the growth rate of government ex-
cant with both data sets. Since the continent penditures. In addition to examining these
dummies simply suggest the importance of fiscal indicators, we examine the role of
omitted variables, the results imply that only government deficits and disaggregated mea-
the share of investment in GDP and the sures of government taxes.
initial income level (out of the long list of Before presenting our results, it is worth
explanatory variables given in Table 5) have mentioning some problems with these
an independent, statistically significant cor- fiscal-policy measures. Governments may
relation with cross-country growth differen- provide growth-promoting public goods and
tials computed from both WB/IMF and SH design taxes to close the gap between pri-
data. These results suggest that many popu- vate and social costs. On the other hand,
lar cross-country growth findings are sensi- governments may waste funds, funnel re-
tive to the conditioning information set. sources to endeavors that do not encourage

This content downloaded from 200.235.241.134 on Tue, 24 Apr 2018 22:12:54 UTC
All use subject to http://about.jstor.org/terms
950 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 1992

TABLE 5-CROSS-COUNTRY GROWTH REGRESSIONS (DEPENDENT VARIABLE:


GROWTH RATE OF REAL PER CAPITA GDP)

Regression period [data set]

(i) (ii (iii) (iv (v)


1960-1989 1960-1985 1960-1989 1960-1985 1960-1985
Independent variable [WB/IMF] [SH] [WB/IMF] [WB/IMF] [SH]

Constant -0.83 2.01 0.86 0.47 2.05


(0.85) (0.83) (0.89) (1.18) (1.12)

Initial GDP per capita (RGDP60) -0.35* - 0.69* - 0.30* - 0.40* - 0.57*
(0.14) (0.12) (0.11) (0.13) (0.12)

Investment share (INV) 17.49* 9.31* 16.77* 13.44* 10.15*


(2.68) (2.08) (2.62) (3.13) (2.43)

Population growth (GPO) -0.38 0.08 -0.53 -0.15 -0.02


(0.22) (0.18) (0.18) (0.19) (0.19)

Secondary-school enrollment (SEC) 3.17* 1.21 0.63 0.99


(1.29) (1.17) (1.26) (1.23)

Primary-school enrollment (PRI) 1.79* 0.91 1.07


(0.58) (0.73) (0.70)

Government share (GOV) - 6.37* -0.59 - 6.80*


(2.03) (3.73) (2.30)

Growth of government share (GSG) -0.08


(0.06)

Socialist economy (SOC) - 0.25 -0.21 -0.17


(0.38) (0.45) (0.43)

Revolution/coups (REVC) - 1.76* - 0.86 - 1.75*


(0.52) (0.62) (0.59)

Africa dummy (AFRICA) - 1.24* - 1.36* - 1.78*


(0.37) (0.48) (0.44)

Latin America dummy (LAAM) - 1.18* - 1.34* - 1.27*


(0.33) (0.38) (0.36)

Growth of domestic credit (GDC) 0.019* 0.013 0.008


(0.009) (0.008) (0.007)

Standard deviation of domestic - 0.009* - 0.006* -0.003


credit (STDD) (0.003) (0.003) (0.003)

Export-share growth (XSG) 0.090 0.023 -0.03


(0.052) (0.047) (0.041)

Civil liberties (CIVL) -0.22 0.01 0.15


(0.11) (0.13) (0.13)

Number of observations: 101 103 83 84 86


R2: 0.46 0.68 0.61 0.67 0.73

Notes: Regressions (i), (iii),


Summers and Heston data.
*Statistically significant at the P = 0.05 level.

This content downloaded from 200.235.241.134 on Tue, 24 Apr 2018 22:12:54 UTC
All use subject to http://about.jstor.org/terms
VOL. 82 NO. 4 LEVINE AND RENELT: CROSS-COUNTRY GROWTH REGRESSIONS 951

TABLE 6-SENSITIvITY RESULTS FOR FIscAL VARIABLES (DEPENDENT VARIABLE:


GROWTH RATE OF REAL PER CAPITA GDP)

M-variable (period) p Standard error t Countries R2 Other variables Robust/fragile

GOV (1960-1989) high: - 0.85 3.20 0.27 85 0.61 REVC, STDD, GDC fragile (0)
base: -4.17 2.96 1.41 98 0.52
low: -5.52 3.33 1.66 85 0.57 X, PI, GDC

TEX (1974-1989) high - 1.22 2.22 0.55 75 0.45 X, STDD, GDC fragile (1)
base - 5.03 2.05 2.46 85 0.36
low -5.51 2.02 2.73 86 0.41 REVC, PI, STDI

GOVX (1974-1989) high -12.95 7.81 1.66 64 0.48 X, STDD, STDI fragile (2)
base -21.96 5.64 3.90 74 0.43
low -23.73 5.64 4.21 75 0.57 REVC, PI, STDI

DEF (1974-1989) high 14.17 5.36 2.64 82 0.41 REVC, PI, STDI fragile (1)
base 15.45 4.90 3.16 82 0.40
low 6.22 5.98 1.04 72 0.47 STDD, REVC, PI

Notes: The base p is the estimated coefficient from the regression with the variable of interest (M-variable) and the
always-included variables (I-variables). The I-variables, when the dependent variable is the growth rate of real per
capita GDP, are INV (investment share of GDP), RGDPxx (initial real GDP per capita), GPO (growth in
population), and SEC or SED (initial secondary-school enrollment rate). The high p is the estimated coefficient
from the regression with the extreme high bound (Pm + two standard deviations); the low p is the coefficient from
the regression with the extreme lower bound. M-variable definitions: GOV = government consumption share;
TEX = total government expenditure; GOVX = government consumption share minus defense and educational
expenditures; DEF = central government surplus/deficit as share.
The "other variables" are the Z-variables included in the base regression that produce the extreme bounds. The
underlined variables are the minimum additional variables that make the coefficient of interest insignificant or
change sign. The robust/fragile designation indicates whether the variable of interest is robust or fragile. If fragile,
the column indicates how many additional varaibles need to be added before the variable is insignificant or of the
wrong sign. A zero indicates that the coefficient is insignificant with only the I-variables included. If robust, the text
provides information about further robustness tests.

growth, and impose taxes and regulations ernment efficiency may yield inaccurate
that distort private decisions. Aggregate measures of the actual delivery of public
measures of government size will not cap- services. While recognizing these problems,
ture the potentially important implications we focus on examining the robustness of
of how total government expenditures are past findings.
allocated. Furthermore, even if government A common measure of the role of the
funds are always spent on growth-promoting government in economic activity is the ratio
goods, there may be complex, nonlinear of government consumption expenditures to
trade-offs between the beneficial effects of GDP (GOV) (e.g., Landau, 1983; Romer,
government services and the deleterious im- 1989). Table 6 reports EBA tests of this
plications of distortionary taxes. Linear variable for the period 1960-1989. Al-
cross-country regressions will not appropri- though the estimated coefficient on GOV is
ately capture these relationships. In addi- always negative, the coefficient is not ro-
tion, disaggregated measures of government bust. In fact, the coefficient is insignificant
expenditures and tax sources are only avail- in the base regression, so that only by se-
able for a limited number of countries since lecting a very particular conditioning set can
the 1970's and are particularly prone to one identify a significant partial correlation
measurement problems. Moreover, since between GOV and GYP within the linear-
government resources may be spent effec- regression context. Similarly, the growth rate
tively or ineffectively, using simple expen- of GOV has a fragile statistical relationship
diture data without accounting for gov- with GYP.

This content downloaded from 200.235.241.134 on Tue, 24 Apr 2018 22:12:54 UTC
All use subject to http://about.jstor.org/terms
952 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 1992

Although subject to data limitations, the fects of deficits. For some specifications,
ratio of total government expenditures to SUR enters with a significantly positive co-
GDP (TEX) is a more complete proxy for efficient. The mere addition of the STDD,
the size of the government in economic however, causes SUR to enter insignifi-
activity than GOV. The partial correlation cantly. Many other specifications (e.g.,
between GYP and TEX, however, is not adding GDC) also demonstrate the fragile
robust. The sign of the coefficient remains nature of the link between GYP and SUR.
negative but becomes insignificant with the Table 7 presents EBA tests of the'fiscal
inclusion of only one additional variable. In indicators with INV. Although many theo-
Table 6, this additional variable is the ratio retical predictions of a negative relationship
of exports to GDP, but the inclusion of between the size of the government and
other macroeconomic indicators (e.g., growth are based on a negative impact of
STDD) also induces an insignificant coef- government activity on capital accumula-
ficient on TEX. tion, none of the fiscal-policy measures has
The effect of government expenditures on a robust relationship with INV. In fact, each
economic growth, however, may depend on of the fiscal indicators is either insignifi-
the allocation of those funds. Barro (1990) cantly correlated with INV or has the wrong
attempts to capture this difference empiri- sign in the base regressions.
cally by removing education and defense EBA tests of the ratio of export tax re-
expenditures from government consumption ceipts to exports, the ratio of import tax
(GOVX). In Table 6, we provide EBA re- receipts to imports, the ratio of corporate
sults for GOVX over the 1974-1989 period, tax receipts to GDP, the ratio of individual
during which data exist for a broad range of income tax receipts to GDP, and the ratio
countries. In contrast to Barro (1991), how- of social-security tax receipts to GDP did
ever, we show that the coefficient on GOVX not yield any robust correlations with either
becomes insignificant when we alter the INV or GYP. The coefficient on each of
conditioning information set (e.g., by adding these variables changes sign with different
STDD and X). Z-variables.'
Continuing to examine the effects of dis- In this subsection, we could not find a
aggregated government expenditures, we robust cross-country relationship between a
test the ratios of government capital forma- diverse collection of fiscal-policy indicators
tion, government education expenditures, and growth. Specifically, although there are
and government defense expenditures to econometric specifications that yield signif-
GDP. None of these variables is robustly icant coefficient estimates between specific
correlated with growth rates.8 fiscal-policy indicators and growth, the co-
We use the central-government surplus efficients on these same variables become
(SUR) to explore the potential negative ef- insignificant when the right-hand-side vari-
ables are slightly altered. Interestingly, stan-
dard fiscal indicators enter with the
predicted sign for many econometric spec-
8We also tested the growth rate of GOV because
ifications when the regression includes
Ram (1986) argues that this measure is positively re-
lated to growth. An obvious problem with this analysis investment, but these same indicators are
is that if government services are a normal good, one insignificantly correlated with investment (or
would expect growth in government services to parallel they enter with the wrong sign). Thus, fiscal
income growth. This measure enters with a positive
policy to the extent that it has an indepen-
coefficient, but when the average annual growth rate of
exports (studied by Feder [1983]) and the change in dent relationship with growth, appears to be
exports as a share of GDP (studied by Romer [1989]) more strongly correlated with the "ef-
are included, the coefficient on the growth rate of ficiency of resource allocation" as opposed
government consumption expenditures becomes in-
significant. The high R2 of this equation (0.98) suggests
that one only needs to include the growth rates of
enough components of GDP to explain the cross- 9See Levine (1991) for an analysis of the effects of
country variance in growth. different types of taxes on long-run growth.

This content downloaded from 200.235.241.134 on Tue, 24 Apr 2018 22:12:54 UTC
All use subject to http://about.jstor.org/terms
VOL. 82 NO. 4 LEVINE AND RENELT: CROSS-COUNTRY GRO WT7H REGRESSIONS 953

TABLE 7-SENSITIvITY RESULTS FOR FISCAL VARIABLES (DEPENDENT VARIABLE: INVESTMENT SHARE)

M-variable (period) ,B Standard error t Countries R2 Other variables Robust/fragile

GOV (1960-1989) high: 0.244 0.13 1.90 85 0.07 GDC, STDD, STDI fragile (0)
base: 0.310 0.11 2.92 102 0.08
low: 0.097 0.11 0.87 85 0.31 X, GDC, GOV

GOVX (1974-1989) high: -0.018 0.15 0.12 74 0.13 STDI, X, PI fragile (0)
base: -0.011 0.16 0.07 76 0.01
low: -0.444 0.20 2.26 65 0.13 GDC, PI, REVC

TEX (1974-1989) high: 0.110 0.05 2.19 76 0.08 GDC, STDD, STDI fragile (0)
base: 0.120 0.05 2.65 87 0.08
low: 0.060 0.05 1.17 75 0.23 GDC, X, REVC

DEF (1974-1989) high: -0.004 0.19 0.02 72 0.04 PI, GDC, STDI fragile (0)
base: -0.009 0.14 0.06 83 0.01
low: -0.158 0.15 1.05 71 0.21 X, STDD, REVC

Notes: The base ,B is the estimated coefficient from the regression with th
the dependent variable is the investment share, no I-variables are included
from the regression with the extreme high bound (P8m + two standard de
the regression with the extreme lower bound. M-variable definitions: GOV = government consumption share;
GOVX = government consumption share minus defense and educational expenditures; TEX = total government
expenditure share; DEF = central-government surplus/deficit as share.
The "other" variables are the Z-variables included in the base regression that produce the extreme bounds. The
robust/fragile designation indicates whether the variable of interest is robust or fragile. If fragile, the column
indicates how many additional variables need to be added before the variable is insignificant or of the wrong sign. A
zero indicates that the coefficient is insignificant with only the I-variables included. If robust, the text provides
information about further robustness tests.

to the accumulation of physical capital per in past studies. In addition, we examine the
se. These results suggest that the interac- relationship between growth and import in-
tions among fiscal policy, investment, and dicators, total-trade indicators, and more
growth may be more complicated than can direct estimates of trade policy and the dis-
be captured in simple linear models using tortion between domestic and international
fairly aggregate measures of fiscal activity. prices.
The EBA analysis yields three important
C. International Trade and Price Distortions results. First, if one substitutes imports or
total trade for exports in cross-country
Over 200 years ago Adam Smith argued growth or investment regressions one ob-
that openness to international markets could tains essentially the same coefficient esti-
enhance productivity by encouraging spe- mate and coefficient standard error.10 Thus,
cialization that would be unprofitable in researchers who identify a significant corre-
smaller markets. Recently, this argument lation using an export performance measure
and other theoretical ties between trade should not associate this result with exports
and growth have been formalized by Louis per se, because it could be obtained using a
Rivera-Batiz and Romer (1991), Gene M. corresponding measure of imports or total
Grossman and Elhanan Helpman (1990), trade. Second, the share of trade in GDP is
and Romer (1986, 1990b). Although theo- robustly positively correlated with the share
retical discussions frequently focus on the
relationship between international trade and
growth, empirical examinations have typi-
10Although this result may not be surprising, it seems
cally examined the relationship between ex- to be frequently overlooked. Many authors interpret
ports and growth. Consequently, we exam- their results as establishing an exclusive relationship
ine the robustness of export indicators used between exports and growth.

This content downloaded from 200.235.241.134 on Tue, 24 Apr 2018 22:12:54 UTC
All use subject to http://about.jstor.org/terms
954 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 1992

TABLE 8-SENSITIVITY RESULTS FOR TRADE VARIABLES


(DEPENDENT VARIABLE: GROWTH RATE OF REAL PER CAPITA GDP)

M-variable (period) , Standard error t Countries R2 Other variables Robust/fragile

X (1960-1989) high: 0.99 0.81 1.23 98 0.55 GOV, PI, STDI fragile (0)
base: 0.88 0.84 1.05 100 0.47
low: 0.14 0.91 0.16 86 0.57 GDC, PI, STDI

IMP (1960-1989) high: 1.27 0.94 1.35 97 0.52 GOV, PI, STDI fragile (0)
base: 0.56 0.89 0.63 99 0.44
low: - 1.11 1.02 1.09 85 0.55 GDC, PI, STDI

LEAM1 (1974-1989) high: -0.08 1.78 0.04 50 0.45 GOV, PI, REVC fragile (0)
base: 1.11 1.79 0.62 50 0.36
low: -2.03 1.84 1.10 41 0.51 DEF, PI, REVC

LEAM2 (1974-1989) high: -0.33 2.11 0.15 50 0.46 REVC, PI, STDI fragile (0)
base: -0.63 2.16 0.29 50 0.36
low: -4.61 2.33 1.98 41 0.51 REVC, GOV, DEF

BMP (1960-1989) high: -0.002 0.003 0.76 90 0.56 REVC, GOV, PI fragile (0)
base: -0.003 0.003 1.02 92 0.55
low: -0.005 0.003 1.53 79 0.57 REVC, GOV, GDC

RERDB (1974-1989) high: -0.011 0.006 1.78 59 0.57 PI, GOV, GDC fragile (3)
base: -0.019 0.006 3.08 63 0.53
low: -0.019 0.006 3.11 63 0.58 PI, REVC, STDI

Notes: The base 83 is the estimated coefficient from the regression with the
always-included variables (I-variables). The I-variables, when the dependent variable is the growth rate of real per
capita GDP, are INV (investment share of GDP), RGDPxx (initial real GDP per capita), GPO (growth in
population), and SEC or SED (initial secondary-school enrollment rate). The high ,3 is the estimated coefficient
from the regression with the extreme high bound (I3m + two standard deviations); the low 8 is the coefficient from
the regression with the extreme lower bound. M-variable definitions: X = exports as percentage of GDP; IMP =
imports as percentage of GDP; LEAM1 = Leamer's (1988) openness measure based on factor-adjusted trade;
LEAM2 = Leamer's (1988) trade-distortion measure based on Heckscher-Ohlin deviations; BMP = black-market
exchange-rate premium; RERDB = Dollar's (1992) real exchange-rate distortion for SH benchmark countries.
The "other variables" are the Z-variables included in the base regression that produce the extreme bounds. The
underlined variables are the -minimum additional variables that make the coefficient of interest insignificant or
change sign. The robust/fragile designation indicates whether the variable of interest is robust or fragile. If fragile,
the column indicates how many additional variables need to be added before the variable is insignificant or of the
wrong s~ign. A zeTo indicates Vnal Vne coefficient is insignilicant witih only the l-variables included. It robust, the text
provides information about further robustness tests.

of investment in GDP. Finally, when con- fact, one needs to search beyond the seven
trolling for the share of investment in GDP, variables considered as potential Z-vari-
we could not find a robust independent ables by the EBA to find a regression in
relationship between any trade or interna- which X enters positively and significantly.
tional price-distortion indicator and growth. However, as in Romer (1990a), we find a
These three results indicate that the rela- positive and robust link between X and
tionship between trade and growth may be INV. When we substituted the ratio of total
based on enhanced resource accumulation trade to GDP or the ratio of imports
and not necessarily on the improved alloca- to GDP for X, the results are almost identi-
tion of resources. cal.1" X was also found to be robust in the
The major results are in Tables 8 and 9.
The ratio of exports to GDP (X) is not
robustly correlated with growth when in- itWhen we dropped countries with X greater than
vestment is included as an I-variable. In 0.75, the results did not change.

This content downloaded from 200.235.241.134 on Tue, 24 Apr 2018 22:12:54 UTC
All use subject to http://about.jstor.org/terms
VOL. 82 NO. 4 LE VINE AND RENELT: CROSS-COUNTRY GROWTH REGRESSIONS 955

TABLE 9-SENSITIVITY RESULTS FOR TRADE VARIABLES


(DEPENDENT VARIABLE: INVESTMENT SHARE)

M-variable (period) ,3 Standard error t Countries R2 Other variables Robust/fragile

X (1960-1989) high: 0.16 0.030 5.31 87 0.26 GDC, STDI robust


base: 0.14 0.024 5.90 106 0.25
low: 0.09 0.024 3.90 101 0.35 GOV, REVC, STDI

LEAM1 (1974-1989) high: 0.15 0.055 2.68 40 0.20 DEF, STDD, GDC robust
base: 0.15 0.043 3.40 50 0.19
low: 0.10 0.050 2.08 48 0.24 REVC, STDD

LEAM2 (1974-1989) high: 0.24 0.044 5.32 48 0.39 GOV, STDD robust
base: 0.22 0.039 5.55 50 0.39
low: 0.18 0.041 4.30 52 0.46 REVC, PI, GOV

BMP (1960-1989) high: - 0.0002 0.0001 1.58 79 0.19 GDC, GOV, REVC fragile (3)
base: -0.0004 0.0001 4.54 95 0.18
low: -0.0004 0.0001 3.78 81 0.18 PI, STDD, GDC

RERDB (1974-1989) high: - 0.0002 0.0002 0.96 52 0.07 DEF, REVC fragile (0)
base: - 0.0002 0.0002 1.12 63 0.02
low: -0.0003 0.0002 1.46 59 0.15 STDD, GDC

Notes: The base 83 is the estimated coefficient from the regression with the variable of interest (M-variable). When
the dependent variable is the investment share, no I-variables are included. The high 83 is the estimated coefficient
from the regression with the extreme high bound (Pm + two standard deviations); the low 83 is the coefficient from
the regression with the extreme lower bound. M-variable definitions: X = exports as percentage of GDP; LEAM1
= Leamer's (1988) openness measure based on factor-adjusted trade; LEAM2 = Leamer's (1988) trade-distortion
measure based on Heckscher-Ohlin deviations; BMP = black-market exchange-rate premium; RERDB = Dollar's
(1992) real exchange-rate distortion for SH benchmark countries.
The "other variables" are the Z-variables included in the base regression that produce the extreme bounds. The
underlined variables are the minimum additional variables that make the coefficient of interest insignificant or
change sign. The robust/fragile designation indicates whether the variable of interest is robust or fragile. If fragile,
the column indicates how many additional variables need to be added before the variable is insignificant or of the
wrong sign. A zero indicates that the coefficient is insignificant with only the I-variables included. If robust, the text
provides information about further robustness tests.

growth equation when we dropped invest- value represents more openness. As Table 8
ment from the list of I-variables. These re- indicates, neither the intervention nor the
sults suggest an important two-link chain openness index is robustly correlated with
between trade and growth through invest- GYP. Both of Leamer's indexes, however,
ment. Interestingly, however, the theoretical are robustly, positively correlated with INV,
ties between growth and trade typically seem as seen in Table 9. On the one hand this is
to run through improved resource allocation not surprising, because both of Leamer's
and not through a higher physical invest- indexes are highly and significantly corre-
ment share. lated with X (e.g., r = 0.70, P < 0.01), which
We also examine more direct measures of we found to be significantly correlated with
trade policy. Leamer (1988) uses the INV. On the other hand, these results are
Heckscher-Ohlin-Vanek trade model to difficult to interpret because the interven-
construct measures of "openness" and "in- tion and openness indexes are positively and
tervention." The intervention index repre- significantly correlated with each other (e.g.,
sents the deviation between the actual and r = 0.63, P < 0.01).12
predicted pattern of trade. The openness
index represents the difference between the
actual and predicted level of trade (as op-
posed to the pattern of trade). Leamer con- 12After carefully examining the relationship among
structs the openness index so that a higher different measures of trade policy, Lant Pritchett (1991

This content downloaded from 200.235.241.134 on Tue, 24 Apr 2018 22:12:54 UTC
All use subject to http://about.jstor.org/terms
956 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 1992

TABLE 10-SENSITIVITY RESULTS FOR MONETARY AND POLITICAL VARIABLES


(DEPENDENT VARIABLE: GROWTH OF REAL PER CAPITA GDP 1960-1989)

M-variable 63 Standard error t Countries R2 Other variables Robust/fragile

PI high: - 0.0022 0.0028 0.80 101 0.49 REVC, X fragile (0)


base: - 0.0039 0.0023 1.67 102 0.48
low: -0.0041 0.0026 1.57 99 0.54 REVC, GOV

STDI high: -0.005 0.0007 0.79 101 0.49 REVC, X fragile (0)
base: -0.0010 0.0006 1.63 102 0.48
low: -0.0010 0.0006 1.52 99 0.54 REVC, GOV

GDC high: 0.026 0.009 2.79 86 0.64 X, STDI, STDD fragile (0)
base: -0.004 0.006 0.59 86 0.56
low: - 0.004 0.006 0.56 86 0.56 X

STDD high: -0.004 0.002 1.93 87 0.59 REVC, GOV, PI fragile (3)
base: - 0.005 0.002 2.90 88 0.60
low: -0.010 0.003 3.92 86 0.64 X, PI, GDC

REVC high: 0.217 0.758 0.29 86 0.57 GDC, STDI, X fragile (0)
base: -1.178 0.647 1.82 102 0.48
low: -1.096 0.659 1.66 101 0.48 X

Notes: The base ,B is the estimated coefficient from the reg


always-included variables (I-variables). The I-variables, whe
capita GDP, are INV (investment share of GDP), RGDP60 (real GDP per capita in 1960), GPO (growth in
population), and SEC (secondary-school enrollment rate in 1960). The high ,3 is the estimated coefficient from the
regression with the extreme high bound (,Pm +two standard deviations); the low /3 is the coefficient from the
regression with the extreme lower bound. M-variable definitions: PI = average rate of inflation; STDI = standard
deviation of the rate of inflation; GDC = average growth rate of domestic credit; STDD = standard deviation of
domestic credit growth; REVC = revolutions and coups.
The "other variables" are the Z-variables included in the base regression that produce the extreme bounds. The
underlined variables are the minimum additional variables that make the coefficient of interest insignificant or
change sign. The robust/fragile designation indicates whether the variable of interest is robust or fragile. If fragile,
the column indicates how many additional variables need to be added before the variable is insignificant or of the
wrong sign. A zero indicates that the coefficient is insignificant with only the I-variables included. If robust, the text
provides information about further robustness tests.

We also examine the average black- and international prices. This "real ex-
market exchange-rate premium (BMP). change-rate distortion" index is significantly
Since this variable represents the interac- positively correlated with BMP, but it is
tions of many policies, we find it difficult tonegatively correlated with X. These correla-
interpret this variable as an indicator of any tions plus the analysis by Pritchett (1991)
one policy. BMP is not robustly correlated suggest that one may want to interpret Dol-
with GYP or INV."3 lar's index as a general measure of interna-
Finally, we examine David Dollar's (1992) tional distortions and not as a narrow mea-
measure of the distortion between domestic sure of trade policy. For the benchmark
countries that have actual as opposed to
interpolated data, Table 8 shows that Dol-
lar's index is negatively though not robustly
rarP 1a to Allz t arllz h14
p. 29) concludes that ". . . alternative objective summary
measures of policy outward orientation produce en-
tirely different country rankings." This assessment has
obviously dour implications for attempts to quantify
the relationship between trade policy and growth. 14We also examined measures of import penetration
13Similar results were found when we excluded (e.g., MP in the Appendix) and indexes of outward
OECD countries. orientation (e.g., SCOUT in the Appendix). Neither

This content downloaded from 200.235.241.134 on Tue, 24 Apr 2018 22:12:54 UTC
All use subject to http://about.jstor.org/terms
VOL. 82 NO. 4 LEVINE AND RENELT: CROSS-COUNTRY GROWTH REGRESSIONS 957

TABLE 11-SENSITIVITY RESULTS FOR MONETARY AND POLITICAL VARIABLES


(DEPENDENT VARIABLE: INVESTMENT SHARE 1960-1989)

M-variable ,p Standard error t Countries R2 Other variables Robust/fragile

PI high: -0.0001 0.0003 0.16 101 0.27 X, GOV, STDI fragile (0)
base: - 0.0001 0.0001 0.46 106 0.01
low: -0.0005 0.0004 1.25 90 0.04 STDD, STDI

STDI high: - 0.00001 0.00002 0.24 102 0.08 GOV fragile (0)
base: -0.00000 0.00002 0.15 106 0.00
low: -0.00005 0.00002 2.28 102 0.24 REVC, GOV

GDC high: 0.0003 0.0003 1.26 85 0.16 REVC, GOV fragile (0)
base: 0.0001 0.0003 0.58 85 0.01
low: 0.0001 0.0003 0.46 85 0.06 STDI, GOV

STDD high: 0.0002 0.00007 2.14 89 0.17 REVC fragile (0)


base: 0.0001 0.00008 1.29 90 0.02
low: 0.0001 0.00007 0.74 88 0.27 X, GOV, PI

REVC high: -0.045 0.022 2.03 88 0.30 X, GOV, STDD robust


base: -0.088 0.020 4.47 106 0.16
low: -0.106 0.025 4.24 86 0.22 GDC, PI, STDI

Notes: The base /3 is the estimated coefficient from the regre


the dependent variable is the investment share, no I-variables are included. The high 83 is the estimated coefficient
from the regression with the extreme high bound (,lm + two standard deviations); the low ,3 is the coefficient from
the regression with the extreme lower bound. M-variable definitions: PI = average rate of inflation; STDI = standard
deviation of the rate of inflation; GDC = average growth rate of domestic credit; STDD = standard deviation of
domestic credit growth; REVC = revolutions and coups.
The "other variables" are the Z-variables included in the base regression that produce the extreme bounds. The
robust/fragile designation indicates whether the variable of interest is robust or fragile. If fragile, the column
indicates how many additional variables need to be added before the variable is insignificant or of the wrong sign. A
zero indicates that the coefficient is insignificant with only the I-variables included. If robust, the text provides
information about further robustness tests.

D. Monetary and Political Indicators ship between measures of monetary policy


and growth. Kormendi and Meguire (1975)
This section examines the empirical rela- find that the average growth rate of the
tionship between growth and measures of money supply, the mean growth in the rate
monetary policy and indicators of the politi- of inflation, and the standard deviation of
cal climate. Based on work by Robert E. money-supply shocks are negatively related
Lucas (1973), Barro (1976), Stanley Fischer to growth, while Grier and Tullock (1989),
(1979), and Alan C. Stockman (1981), previ- using a pooled cross-section, time-series
ous researchers have explored the relation- analysis find that the standard deviation of
inflation is negatively related to growth. We
examined these and other indicators of
monetary policy and report the results of
four indicators in Tables 10 and 11: PI, the
was robustly correlated with GYP when INV was in-
STDI, the GDC, and the STDD. Each indi-
cluded in the regression. Furthermore, in light of stud-
ies by Feder (1983) and Kormendi and Meguire (1985),
cator has conceptual and statistical prob-
we studied export growth and export growth times the lems. For example, PI probably represents
share of exports in GDP. Neither of these variables is less of a direct indicator of monetary policy
robustly correlated with growth when the regression and more of a conglomerate index of the
includes corresponding fiscal indicators (e.g., growth of
result of many policies and shocks, while
GOV). Given the national-accounts identity, even if we
found a robust relationship, it is not clear what worth- the endogeneity and identification issues as-
while inferences could be drawn. sociated with GDC may be particularly

This content downloaded from 200.235.241.134 on Tue, 24 Apr 2018 22:12:54 UTC
All use subject to http://about.jstor.org/terms
958 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 1992

acute. Nonetheless, the wide assortment of set, the determination of whether variables
indicators that we test produce similar re- have robust or fragile partial correlations
sults: none of the indicators is robustly cor- with GYP or INV did not depend on the
related with GYP or INV. particular variables chosen for the condi-
The profession has also used a variety of tioning set.
political indicators in searching for explana- To provide some evidence concerning the
tions of long-run growth. Kormendi and reasons underlying our findings, we also ex-
Meguire (1985) find that greater civil liber- amined the importance of maximizing the
ties are positively related to growth, while differences in the Pm's rather than the dif-
Barro (1991) finds a negative relationship ferences in B-bounds (,Bm's plus two coef-
between growth and an index of wars and ficient standard errors). We found that this
revolutions. We find that indexes of revolu- alteration in the EBA did not alter the
tions and coups (REVC) and civil liberties results. This suggests, as does the fact that
(CIVL) are not robustly correlated with coefficient standard errors are generally
GYP. REVC, however, is robustly, nega- similar between upper and lower bounds,
tively correlated with INV. Thus, not sur- that alterations in the Z-variables change
prisingly, countries that experience a high the estimated 1's more than the standard
number of revolutions and coups tend to be errors.
countries that invest less of their resources We gauged the sensitivity of our results to
domestically than countries with stable po- data quality and comparability. Wherever
litical environments. possible, we did the analysis using both the
SH and WB/IMF data sets, and the results
IV. Sensitivity of the Sensitivity Analysis did not importantly change."6 Also, the SH
and Variable Groupings data set ranks the quality of each country's
data from A to D with A being the best-
We selected the I-variables based on the- quality data. To test for the importance of
oretical grounds, past empirical findings, and data quality, we did the analysis (i) eliminat-
the ability to replicate past finding with this ing all quality-D data and (ii) using weighted
set of included variables. Nonetheless, we least squares with A-D as the weights.
examined the robustness of our findings to Again, these specifications did not alter the
alterations in the I-variables. We conducted results.
the entire EBA with two alternative sets of The restrictions we impose on the EBA,
I-variables. The first set is the original I- such as limiting the pool of variables from
variables plus the sub-Saharan African and which we choose Z-variables and limiting
Latin American dummy variables. We added the number of Z-variables to three, make it
these dummies because a number of previ- easier to classify a finding as robust. Thus,
ous researchers have found significant ef- we conducted additional sensitivity analyses
fects for the continent variables (see Romer, of the robust correlations. We briefly dis-
1989, 1990a; Grier and Tullock, 1989; Barro, cuss two findings. First, the partial correla-
1991). The second alternative set of I-vari- tion between GYP and INV remains signif-
ables includes only INV. The alternative icantly positive even when we allow the EBA
choices of the I-variables did not signifi- to choose five Z-variables, drastically ex-
cantly alter the results.15 pand the pool of variables from which the
In addition, we experimented with dif- EBA chooses Z-variables, and examine dif-
ferent variable pools from which the EBA ferent subperiods and subgroups of coun-
chooses Z-variables. As long as we included tries. Second, the conditional-convergence
a diverse set of variables in the conditioning result is not robust over the 1974-1989 pe-
riod or when we exclude OECD countries.

15SEC is not robustly correlated with growth when


the regression includes a dummy variable for sub- 16The Z-variables chosen by the EBA are some-
Saharan African countries. times different for the two data sets.

This content downloaded from 200.235.241.134 on Tue, 24 Apr 2018 22:12:54 UTC
All use subject to http://about.jstor.org/terms
VOL. 82 NO. 4 LEVINE AND RENELT: CROSS-COUNTRY GROWTH REGRESSIONS 959

Finally, we made some attempts to mea- tors than any previous study. We find that
sure the notion that policies should be in- very few economic variables are robustly
terpreted more broadly than any particular correlated with cross-country growth rates
measure of fiscal, trade, or monetary perfor- or the ratio of investment expenditures to
mance can capture. For example, BMP is GDP. We do, however, identify some corre-
related to exchange-rate policy, monetary lations that, with some qualifications, are
policy, trade policy, and political uncer- robust to slight alterations in the list of
tainty; thus, it may be "unfair" to include independent variables. We hope that this
other policy indicators while examining the will provide useful information for future
partial correlation between BMP and GYP. theoretical and empirical work.
Of course, if a significant coefficient is then We briefly summarize our findings as fol-
found when other policy indicators are ex- lows.
cluded, the significance should not be inter-
preted as representing a correlation be- (i) We found a positive and robust corre-
tween GYP and BMP per se, but between lation between average growth rates
growth and a general indicator of "distor- and the average share of investment in
tions." Consequently, we used factor analy- GDP.
sis to construct aggregate policy indicators (ii) We found a positive and robust corre-
from groups of individual policy indicators. lation between the share of investment
For example, we tested the robustness of in GDP and the average share of trade
various "international" distortion indexes, in GDP.
"domestic" distortion indexes, and "uncer- (iii) We found that all findings using the
tainty" indexes constructed from up to four share of exports in GDP could be ob-
individual indicators. None was robustly tained almost identically using the to-
correlated with growth. This again indicates tal trade or import share. Thus, studies
the difficulty of isolating the independent that use export indicators should not
importance of any single policy. be interpreted as studying the relation-
ship between growth and exports per
V. Conclusion se but rather as studying the relation-
ship between growth and trade defined
In many respects, this paper is a natural more broadly.
extension of the types of exploratory cross- (iv) We found that a large variety of trade
country empirical investigations of growth policy measures were not robustly cor-
pioneered by Kormendi and Meguire (1985) related with growth when the equation
and recently advanced by Barro (1990, 1991). included the investment share.
Representative of the large empirical (v) We found qualified support for the
cross-country growth literature, each of conditional-convergence hypothesis:
these studies uses an assortment of theoret- we find a robust, negative correlation
ical papers to motivate a variety of eco- between the initial level income and
nomic variables that are then used in cross- growth over the 1960-1989 period
country growth regressions. Although each when the equation includes a measure
study presents intuitively appealing results, of the initial level of investment in
they use different explanatory variables. In human capital; but this result does not
addition to showing for the specific cases of hold over the 1974-1989 period.
Kormendi and Meguire (1985) and Barro (vi) We found that none of the broad array
(1991) that a union of the two sets of ex- of fiscal indicators that we studied is
planatory variables leaves none of the eco- robustly correlated with growth or the
nomic policy indicators significantly corre- investment share.
lated with growth, this paper systematically (vii) We found that a large assortment of
evaluates the robustness of the partial cor- other economic and political indicators
relation between per capita growth rates are not robustly correlated with growth
and a wider assortment of economic indica- or the investment share.

This content downloaded from 200.235.241.134 on Tue, 24 Apr 2018 22:12:54 UTC
All use subject to http://about.jstor.org/terms
960 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 1992

We have tried to distinguish partial domestic product (Source: World Bank


National Accounts)
growth correlations that seem robust from
GOVXa Government consumption less defense and
those that are fragile. We find that, al-
education share of GDP (Source: GOV -
though there are many econometric speci- DEE-EDE)
fications in which macroeconomic policy in- GPO Growth of population (Source: World Bank
dicators-taken individually or in groups Social Indicators)
GR Growth of real per capita GDP (Source:
are significantly correlated with growth, the
Summers-Heston data set)
cross-country statistical relationship be- GSG Growth of the share of government con-
tween long-run average growth rates and sumption (GOV) (Source: World Bank Na-
almost every particular macroeconomic in- tional Accounts)
dicator is fragile. National policies appear GX Growth of exports (Source: World Bank
National Accounts)
to be a complex package, and future re-
GYP Growth of real per capita gross domestic
searchers may wish to focus on macroeco- product (Source: World Bank National Ac-
nomic policy regimes and interactions counts)
among policies as opposed to the indepen- HSGVX Share of real government consumption ex-
penditures minus defense and education
dent influence of any particular policy.
expenditures (Source: Barro, 1991)
IMP Import share of GDP (Source: World Bank
DATA APPENDIX National Accounts)
INV Investment share of gross domestic prod-
Variables and Sources uct (Source: World Bank National Ac-
counts)
Variable Definition and source ITXa Share of central-government individual in-
come tax revenue to GDP (Source: IMF's
AFRICA Dummy variable for sub-Saharan African Government Finance Statistics Yearbook)
countries LAAM Dummy variable for Latin American
AREA Land area (in thousands of square kilome- countries
ters) (Source: World Bank Social Indica- LEAM1 Measure of overall trade openness
tors) (Source: Leamer, 1988)
BMP Black-market exchange-rate premium LEAM2 Measure of overall trade intervention
(Source: Picks Currency Yearbook [World (Source: Leamer, 1988)
Bank Updates]) LIT Literacy rate in 1960 (Source: World Bank
BMS Standard deviation of BMP Social Indicators)
CGCa Central-government gross capital forma- MIX Dummy variable for mixed government
tion (Source: IMF's Government Finance (Source: Barro, 1991)
Statistics Yearbook) MSG Growth of import share (Source: World
CIVL Index of civil liberties (Source: Barro, Bank National Accounts)
1991) MTXa Ratio of import taxes to imports (Source:
CTXa Ratio of central-government corporate- IMF's International Financial Statistics
income-tax revenue to GDP (Source: and Government Finance Statistics Year-
IMF's Government Finance Statistics Year- book)
book) MP Measure of openness based on import
DEEa Ratio of central-government defense ex- penetration (Source: residuals of regres-
penditure to GDP (Source: IMF's Gov- sion of IMP on RGDP60, RGDP602,
emnment Finance Statistics Yearbook) AREA, and POP)
DEFa Ratio of central-government deficit to OECD Dummy for OECD countries (members of
GDP (Source: IMF's Government Finance the Organization for Economic Coopera-
Statistics Yearbook) tion and Development)
EDEa Ratio of government educational expendi- OIL Dummy for OPEC countries (members of
tures to GDP (Source: IMF's Government the Organization of Petroleum Exporting
Finance Statistics Yearbook) Countries)
GDC Growth rate of domestic credit (Source: PI Average inflation of GDP deflator (Source:
IMF's International Financial Statistics) World Bank National Accounts)
GGCFD Real government capital formation POP70 Population in 1970 (Source: Summers-
(Source: Barro, 1991) Heston data set)
GG Growth of government consumption ex- PRI Primary-school enrollment rate in 1960
penditures (Source: World Bank National (Source: Barro, 1991)
Accounts) PRJ Primary-school enrollment rate in 1970
GM Growth of imports (Source: World Bank (Source: Barro, 1991)
National Accounts) RERD Real exchange-rate distortion (Source:
GOV Government consumption share of gross Dollar, 1991)

This content downloaded from 200.235.241.134 on Tue, 24 Apr 2018 22:12:54 UTC
All use subject to http://about.jstor.org/terms
VOL. 82 NO. 4 LEVINE AND RENELT: CROSS-COUNTRY GROWTH REGRESSIONS 961

RERDB RERD for Summers-Heston benchmark STDI Standard deviation of PI (inflation)


countries (Source: World Bank National Accounts)
REVC Number of revolutions and coups per year TAXa Ratio of central-government tax revenue
(Source: Barro, 1991) to GDP (Source: IMF's Government Fi-
RGDPxx Real GDP per capita in 19xx (Source: nance Statistics Yearbook)
Summers-Heston data set) TEXa Ratio of total government expenditure to
SCOUT Dummy for outward orientation (Source: GDP (Source: IMF's Government Finance
Moshe Syrquin and Hollis Chenery, 1988) Statistics Yearbook)
SEC Secondary-school enrollment rate in 1960 TRD Ratio of total trade (exports+imports) to
(Source: Barro, 1991) GDP (Source: World Bank National Ac-
SED Secondary-school enrollment rate in 1970 counts)
(Source: Barro, 1991) XSG Growth of export share of GDP (Source:
SGOV Real government consumption share of World Bank National Accounts)
GDP (Source: Summers-Heston data set) XTXa Ratio of central-government export-tax
SINV Real investment share of GDP (Source: revenue to exports (Source: IMF's Gov-
Summers-Heston data set) ernment Finance Statistics Yearbook)
SOC Dummy for socialist economy (Source: X Export share of GDP (Source: World Bank
Barro, 1991) National Accounts)
SSTa Ratio of social-security tax revenue to
GDP (Source: IMF's Government Finance aOnly available for 1
Statistics Yearbook)
STDD Standard deviation of GDC (growth of
domestic credit) (Source: IMF's Interna-
tional Financial Statistics)

Country List, 119-Country Sample

Afghanistan El Salvador Lesotho Senegal


Algeria Ethiopia Liberia Sierra Leone
Angola Fiji Luxembourg Singapore
Argentina Finland Madagascar Somalia
Australia France Malawi South Africa
Austria Gabon Malaysia Spain
Bangladesh Gambia Mali Sri Lanka
Barbados Germany Malta Sudan
Belgium Ghana Mauritania Suriname
Benin Greece Mauritius Swaziland
Bolivia Guatemala Mexico Sweden
Botswana Guinea-Bissau Morocco Switzerland
Brazil Guyana Mozambique Syria
Burkina Faso Haiti Nepal Taiwan
Burma Honduras Netherlands Tanzania
Burundi Hong Kong New Zealand Thailand
Cameroon Iceland Nicaragua Togo
Canada India Niger Trinidad and Tobago
Central African Republic Indonesia Nigeria Tunisia
Chad Iran Norway Turkey
Chile Iraq Oman Uganda
Colombia Ireland Pakistan United Kingdom
Congo Israel Panama United States
Costa Rica Italy Papua New Guinea Uruguay
Cote d'Ivoire Jamaica Paraguay Venezuela
Cyprus Japan Peru Yemen
Denmark Jordan Philippines Zaire
Dominican Republic Kenya Portugal Zambia
Ecuador Korea Rwanda Zimbabwe
Egypt Kuwait Saudi Arabia

This content downloaded from 200.235.241.134 on Tue, 24 Apr 2018 22:12:54 UTC
All use subject to http://about.jstor.org/terms
962 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 1992

REFERENCES ican Economic Review, May 1990 (Papers


and Proceedings), 80, 86-91.
Kormendi, Roger and Meguire, Philip, "Macro-
Barro, Robert J., "Rational Expectations and economic Determinants of Growth:
the Role of Monetary Policy," Journal of Cross-Country Evidence," Journal of
Monetary Economics, January 1976, 2, Monetary Economics, September 1985, 16,
1-32. 141-63.
,_ "Government Spending in a Simple Landau, Daniel, "Government Expenditure
Model of Endogenous Growth,." Journal and Economic Growth: A Cross-Country
of Political Economy, October 1990, part Study," Southern Economic Journal, Jan-
2, 98, 103-25. uary 1983, 49, 783-92.
,_ "Economic Growth in a Cross Sec- Leamer, Edward E., Specification Searches:
tion of Countries," Quarterly Journal of Ad Hoc Inference from Non-Experimental
Economics, May 1991, 106, 407-44. Data, New York: Wiley, 1978.
and Sala-i-Martin, Xavier, "Conver- ,_ "Let's Take the Con Out of Econo-
gence," Journal of Political Economy, metrics," American Economic Review,
April 1992, 100, 223-51. March 1983, 73, 31-43.
Becker, Gary, Murphy, Kevin and Tamura, ,"Sensitivity Analyses Would Help,"
Robert, "Economic Growth, Human Capi- American Economic Review, June 1985,
tal and Population Growth," Journal of 75, 308-13.
Political Economy, October 1990, part 2, , "Measures of Openness," in R.
98, 12-37. Baldwin, ed., Trade Policy Issues and Em-
Cooley, Thomas F. and LeRoy, Stephen F., pirical Analysis, National Bureau of Eco-
"Identification and Estimation of Money nomic Research Conference Report Se-
Demand," American Economic Review, ries, Chicago: University of Chicago Press,
December 1981, 71, 825-44. 1988, pp. 147-200.
De Long, J. Bradford, "Productivity Growth, and Leonard, Herman, "Reporting the
Convergence, and Welfare: Comment," Fragility of Regression Estimates," Re-
American Economic Review, December view of Economics and Statistics, May
1988, 78, 1138-54. 1983, 65, 306-17.
Dollar, David, "Outward-Oriented Develop- Levine, Ross, "Stock Markets, Growth, and
ing Economies Really Do Grow More Tax Policy," Journal of Finance, Septem-
Rapidly: Evidence from 95 LDCs, ber 1991, 46, 1445-65.
1976-1985," Economic Development and and Renelt, David, "Cross Country
Cultural Change, April 1992, 40, 523-44. Studies of Growth and Policy: Some
Edwards, Sebastian, "Openness, Outward Methodological, Conceptual, and Statisti-
Orientation, Trade Liberalization and cal Problems," World Bank Working Pa-
Economic Performance in Developing per Series No. 608, 1991.
Countries," World Bank Working Paper Lucas, Robert E., Jr., "Some International
Series No. 191, 1989. Evidence on Output-Inflation Tradeoffs,"
Feder, Gershon, "On Exports and Economic American Economic Review, June 1973,
Growth," Journal of Development Eco- 63, 326-34.
nomics, February/April 1983, 12, 59-74. Mankiw, N. Gregory, Romer, David and Weil,
Fischer, Stanley, "Anticipations and the Non- David, "A Contribution to the Empirics of
neutrality of Money," Journal of Political Economic Growth," Quarterly Journal of
Economy, April 1979, 87, 225-52. Economics, May 1992, 152, 407-37.
Grier, Kevin and Tullock, Gordon, "An Empir- McAleer, Michael, Pagan, Adrian and Volker,
ical Analysis of Cross-National Economic Paul A., "What Will Take the Con Out of
Growth, 1951-1980," Journal of Monetary Econometrics?" American Economic Re-
Economics, September 1989, 24, 259-76. view, June 1985, 75, 293-307.
Grossman, Gene M. and Helpman, Elhanan, Pritchett, Lant, "Measuring Outward-Orien-
"Trade, Innovation, and Growth," Amer- tation in LDCs: Can It Be Done?" World

This content downloaded from 200.235.241.134 on Tue, 24 Apr 2018 22:12:54 UTC
All use subject to http://about.jstor.org/terms
VOL. 82 NO. 4 LEVINE AND RENELT: CROSS-COUNTRY GROWTH REGRESSIONS 963

Bank Working Paper Series No. 566, 1991. cal Change," Journal of Political Econ-
Ram, Rati, "Government Size and Economic omy, October 1990, part 2, 98, 71-102.
Growth: A New Framework and Some Stockman, Alan C., "Anticipated Inflation and
Evidence from Cross-Section and Time the Capital Stock in a Cash-in-Advance
Series Data," American Economic Re- Economy," Journal of Monetary Eco-
view, March 1986, 76, 191-203. nomics, November 1981, 8, 387-93.
Rivera-Batiz, Louis and Romer, Paul M., "Eco- Summers, Robert and Heston, Alan, "A New
nomic Integration and Endogenous Set of International Comparisons of Real
Growth," Quarterly Journal of Economics, Product and Price Levels Estimates for
May 1991, 106, 531-56. 130 Countries, 1950-1985," Review of In-
Romer, Paul M., "Increasing Returns and come and Wealth, March 1988, 34, 1-25.
Long-Run Growth," Journal of Political Syrquin, Moshe and Chenery, Hollis, "Patterns
Economy, October 1986, 94, 1002-37. of Development, 1950-1983," World
"Crazy Explanations for the Pro- Bank Working Paper Series No. 41, 1989.
ductivity Slowdown," in Stanley Fischer, Government Finance Statistics Yearbook, Wash-
ed., NBER Macroeconomics Annual, ington, DC: International Monetary Fund,
Cambridge, MA: MIT Press, 1987, pp. various years.
163-202. International Financial Statistics, Washington,
, "Human Capital and Growth: The- DC: International Monetary Fund, vari-
ory and Evidence," National Bureau of ous years.
Economic Research (Cambridge, MA), Pick's Currency Yearbook, New York: Pick Pub-
Working Paper No. 3173, 1989. lishing Corp., various years.
, (1990a) "Capital, Labor, and Pro- World Bank National Accounts, Washington, DC:
ductivity," Brookings Papers on Economic World Bank, various years.
Activity, special issue, 1990, 337-420. World Bank Social Indicators, Washington, DC:
_ (1990b), "Endogenous Technologi- World Bank, various years.

This content downloaded from 200.235.241.134 on Tue, 24 Apr 2018 22:12:54 UTC
All use subject to http://about.jstor.org/terms

You might also like