Growth and risk at the industry level: The real effects of nancial liberalization

Andrei A. Levchenko
a,b,
, Romain Rancire
b,c,d
, Mathias Thoenig
c,d,e
a
University of Michigan, USA
b
IMF, USA
c
PSE, France
d
CEPR, United Kingdom
e
Universit de Genve, Switzerland
A B S T R A C T A R T I C L E I N F O
Article history:
Received 28 May 2007
Received in revised form 2 June 2008
Accepted 4 June 2008
JEL classication:
F02
F21
F36
F4
Keywords:
Financial liberalization
Growth
Volatility
Industry-level data
Difference-in-differences estimation
Propensity score matching
This paper analyzes the effects of nancial liberalization on growth and volatility at the industry level in a
large sample of countries. We estimate the impact of liberalization on production, employment, rm entry,
capital accumulation, and productivity. In order to overcome omitted variables concerns, we employ a
number of alternative difference-in-differences estimation strategies. We implement a propensity score
matching algorithm to nd a control group for each liberalizing country. In addition, we exploit variation in
industry characteristics to obtain an alternative set of difference-in-differences estimates. Financial
liberalization is found to have a positive effect on both growth and volatility of production across industries.
The positive growth effect comes from increased entry of rms, higher capital accumulation, and an
expansion in total employment. By contrast, we do not detect any effect of nancial liberalization on
measured productivity. Finally, the growth effects of liberalization appear temporary rather than permanent.
2008 Elsevier B.V. All rights reserved.
1. Introduction
Financial markets have been liberalized dramatically in many
countries over the past three decades. Fig. 1 depicts recent trends in
the indicators of nancial openness. Most de jure measures of
restrictions on domestic capital allocation or international capital ows
show a strong trend towards liberalization. Indeed, capital ows across
borders have correspondinglygrownat a higher pace thantheexpansion
of goods trade, and much faster than GDP. What are the effects of
nancial liberalization? In spite of a theoretical case that nancial
liberalization should improve the allocation of capital and increase
growth, the growth effects of nancial liberalization have not been easy
to demonstrate in cross-country data. At the same time, worries persist
that nancial liberalization may result in higher volatility.
1
This paper examines the relationship between nancial liberal-
ization, growth, and volatility using a large industry-level panel
dataset. The empirical analysis answers three sets of questions. First,
what is the impact of nancial liberalization on output growth and
volatility at the industry level? Both growth and volatility effects have
been analyzed separately in cross-country data. However, to obtain a
reliable estimate of the their relative importance it is essential to
consider these effects within a unied empirical framework. Second,
what are the channels through which nancial liberalization affects
growth? And third, are the effects of nancial liberalizationpermanent
or temporary? The answers to the last two questions shed light on the
nature of the relationship between liberalization and growth, and can
help distinguish between the different theoretical possibilities.
The main ndings can be summarized as follows. Financial
liberalization increases both growth and volatility of output. These
effects are robust to a variety of specications and estimation
strategies. The growth effect is driven by higher employment, greater
capital accumulation, and greater rm entry. By contrast, we do not
detect any impact of liberalization on TFP growth. Finally, the growth
impact is temporary rather than permanent: for output, rm entry,
and employment, the effect decreases in magnitude over time, and
Journal of Development Economics 89 (2009) 210222
Corresponding author. International Monetary Fund, 700 19th Street NW,
Washington, DC, 20431, USA.
E-mail addresses: alev@umich.edu (A.A. Levchenko), rranciere@imf.org (R. Rancire),
thoenig@ecopo.unige.ch (M. Thoenig).
1
Kose et al. (2006) provide a comprehensive exposition of basic facts about the
current wave of nancial globalization, and reviewexisting literature on its growth and
volatility effects.
0304-3878/$ see front matter 2008 Elsevier B.V. All rights reserved.
doi:10.1016/j.jdeveco.2008.06.003
Contents lists available at ScienceDirect
Journal of Development Economics
j our nal homepage: www. el sevi er. com/ l ocat e/ econbase
becomes insignicant after 6 years, while the impact on capital
accumulation is slightly more long-lasting. The only persistent effect is
on competition: the impact of nancial liberalization on the price-cost
margin a measure of markups increases progressively for the rst
few post-liberalization years, and remains signicantly negative
throughout the period we analyze. We conclude that nancial
liberalization has a permanent effect on the level of output, but no
persistent effect on output growth. While the effect of nancial
liberalization on volatility is also most pronounced on impact, we
cannot rule out the possibility of a permanent increase in the variance
of output growth.
When it comes to interpreting these results, it is useful to consider
the range of theoretical possibilities for the growth benets associated
with nancial liberalization. At one extreme, in a standard determi-
nistic neoclassical framework, capital mobility accelerates conver-
gence but has no long-run effect on growth or the level of income.
2
At
the other extreme, in an endogenous growth framework risk-
diversication and specialization in more efcient technologies can
have permanent growth-enhancing effects.
3
Our ndings of a
permanent level effect but no persistent growth effect seem to reject
either of these two polar views. However, they are consistent with the
notion that capital mobility raises production efciency by reducing
domestic distortions.
4
In particular, our empirical results can be
rationalized within a neoclassical model with imperfect competition.
In such a model, a permanent reduction in markups leads to a
temporary growth increase reecting convergence towards higher
levels of capital and income.
5
Until recently, most of the empirical literature studying nancial
liberalization used country-level data, and as a result was subject to
both conceptual and econometric problems. First, conceptually, if
nancial markets are not perfect within the country, the economy
does not behave like a representative agent. Indeed, there is strong
evidence that risk sharing between agents within a country is far from
complete even in the most advanced economies like the U.S.
(Attanasio and Davis, 1996; Hayashi et al., 1996). For developing
countries as well, there is a large amount of evidence, surveyed in
Banerjee and Duo (2005), that the representative agent assumption
is strongly violated. When that is the case, analyzing aggregate data
may in some cases lead us to miss the most important effects of
nancial liberalization, and in others produce estimates that are not
informative about welfare implications for the average individual in
the economy (Levchenko, 2005; Broner and Ventura, 2006). The use
of sector-level data therefore enables us to get a deeper understanding
of how nancial liberalization affects the typical agent. In the last
section of the paper, we demonstrate the importance of the distinction
between industry-level and aggregate effects. In particular, while the
change in aggregate growth implied by our estimates is the same as
the industry-level change, aggregate volatility increases by much less
than sector-level volatility due to diversication across sectors.
Second, existing cross-country results are most likely subject to
signicant endogeneity and omitted variables problems. The key
feature of our empirical approach is the variety of empirical strategies
we pursue in order to obtain reliable estimates. We isolate a number of
nancial liberalization episodes using the de jure liberalization indices
developed by Kaminsky and Schmukler (2008) and compare the
growth and volatility of outcomes, such as output and employment,
during the 10 years immediately before and after the liberalization
date. To address the omitted variables problem, the paper employs
two difference-in-differences strategies. The rst, more novel to this
paper, uses as the control group countries that did not liberalize in the
same period. To overcome a selection on observables problem that
could arise in such an exercise, we develop a propensity score
matching procedure to select a suitable control group for each
liberalizing country. The second approach, a more conventional one,
exploits differences in sector characteristics in the spirit of Rajan and
Zingales (1998) to identify a causal link between liberalization and
growth and volatility. As a way to assess the robustness of our results,
we also estimate the relationship between de facto measures of
nancial liberalization, such as those used by Kose et al. (2003) and
Lane and Milesi-Ferretti (2006) and growth and volatility. In spite of
important differences in the independent variables and specications,
the ndings are remarkably similar for the two types of measures.
6
This paper is related to the large literature on the growth and
volatility effects of nancial liberalization, surveyed comprehensively
by Kose et al. (2006) and Henry (2007). Here, we focus on the papers
most closely related to ours. While most existing studies in this
literature use cross-country data, Galindo et al. (2002), and Gupta and
Yuan (2006) employ industry-level data and the Rajan and Zingales
(1998) methodology to analyze the effects of nancial liberalization
on growth. Our paper differs from these two contributions in several
important respects. First, we investigate the volatility effects of
nancial liberalization, doing so within the same empirical framework
as the growth effects. This produces a more complete picture of the
effects of nancial liberalization, and enables us to evaluate its overall
impact. Second, while the RajanZingales methodology makes it
possible to identify the differential impact of nancial liberalization
across industries, it does not allowone to estimate the overall effect of
nancial liberalization. This approach is thus of limited usefulness
2
See Barro et al. (1995), and Gourinchas and Jeanne (2006).
3
See Saint-Paul (1992) and Obstfeld (1994).
4
See Tornell and Velasco (1992) and Quadrini (2005) for models in which capital
mobility reduces production inefciencies associated with imperfect property rights or
time-inconsistent scal policies. In contrast, Tressel and Verdier (2007) suggest that
nancial liberalization can increase production inefciencies by exacerbating the
misallocation of credit towards politically connected rms.
5
See Gal (1994, 1995) for a detailed analysis. Note that in this model, output growth
volatility tends to increase temporarily as an economy transitions from a steady-state
with a low level of capital and high markups to a steady-state with high level of capital
and low markups.
6
The advantage of de jure measures is that they reect policy levers, and thus
results based on them may have clearer policy implications for reforms that a
government might consider. Their disadvantage is that they may capture quite poorly
the actual degree of nancial integration, either because the true nature of legal
restrictions is mismeasured, or because these restrictions are imperfectly enforced.
Nonetheless, we place more weight on the de jure measures, since the de facto ones
represent equilibrium outcomes, and may be more noisy reections of policy.
Fig. 1. Worldwide nancial liberalization trends. The world gross capital ows/GDP are
the sum of the gross capital ows across countries, divided by world GDP, in each year.
Source: IMF Balance of Payments Statistics and World Bank's World Development
Indicators. De jure liberalization is the average composite index of nancial liberal-
ization across countries in each year. The index ranges from 1 (least liberalized) to 3
(fully liberalized). Source: Kaminsky and Schmukler, (2008).
211 A.A. Levchenko et al. / Journal of Development Economics 89 (2009) 210222
when it comes to policy evaluation of nancial liberalization reforms.
By contrast, our paper proposes a methodology to measure the overall
effect. Third, we establish whether or not the effects of nancial
liberalization are temporary or permanent. And nally, we use both de
jure and de facto measures of nancial liberalization to assess
robustness of the results. In particular, de facto measures have not
previously been used in industry-level analysis.
7
The rest of the paper is organized as follows. Section 2 describes
the data. Section 3 lays out the empirical methodology and presents
the estimating equations. Section 4 presents the results, and discusses
the implications of our sector-level estimates for aggregate growth
and volatility. Section 5 concludes.
2. Data
Industry-level production, employment, investment, and the
number of establishments come from the 2006 UNIDO Industrial
Statistics Database (United Nations Industrial Development Organiza-
tion, 2006). This paper uses the version that reports data according to
the 3-digit ISIC Revision 2 classication for the period 19632003 in
the best cases. There are 28 manufacturing sectors, plus the
information on total manufacturing. We use data reported in current
U.S. dollars, and convert theminto constant international dollars using
the Penn World Tables (Heston et al., 2002).
8
The resulting dataset is
an unbalanced panel of 56 countries, but we ensure that for each
countryyear we have a minimum of 10 sectors, and that for each
country, there are at least 10 years of data.
The data on de jure nancial liberalization come from Kaminsky
and Schmukler (2008) (henceforth KS), who provide indices of
liberalization in the stock market, the banking system, and freedom of
international transactions for 28 countries. Along each of the three
dimensions of liberalization, KS assign a value of 1, 2, or 3 for each
country and year, with 3 indicating the most liberalized. They also
provide a composite index, which is a mean of the three subcompo-
nents. As a measure of de facto nancial liberalizationwe use the gross
capital ows as a share of GDP. The gross capital ows are the sum of
gross inows and gross outows, obtained from the IMF's Balance of
Payments Statistics. This measure, which is parallel to the aggregate
trade openness (exports plus imports), has been used by Kose, Prasad,
and Terrones (2003), as well as several subsequent papers.
9
In order to test for the differential effect of nancial liberalization
across industries, we employ the dependence on external nance
measure introduced by Rajan and Zingales (1998). The Rajan and
Zingales measure is dened as capital expenditure minus cash ow,
divided by capital expenditure, and is constructed based on U.S. rm-
level data. Intuitively, it is intended to capture the share of investment
that must be nanced with funds external to the rm.
10
We also make
use of the industry-level measure of liquidity needs compiled by
Raddatz (2006), dened as inventories as a share of sales. A sector has
a higher need for liquidity when a smaller fraction of inventory
accumulation can be nanced by ongoing cash ow. Additional
controls include nancial development private credit as a share of
GDP sourced from Beck et al. (2000), and trade openness at the
industry level constructed by di Giovanni and Levchenko (in press).
Appendix Table A1 lists the countries in the sample and the
summary statistics for growth, volatility, and gross capital ows for
each country, as well as the means and standard deviations for the
entire sample. Table A5 in the supplementary web appendix lists the
sectors used in the analysis, along with the values of external nance
dependence and liquidity needs.
3. Empirical methodology
3.1. Baseline specication
In the baseline approach to estimating the effects of nancial
liberalization, we date the liberalization events in a sample of
countries, and then compare outcomes before and after liberalization.
This strategy relies on the de jure indicators compiled by KS to identify
the liberalization episodes. Because we require precise liberalization
dates, we must set a threshold for the KS index, above which the
country is considered liberalized, and below which it is not.
11
The
resulting set of liberalization dates is listed in Appendix Table A2.
To estimate the effects of nancial liberalization on economic
outcomes, we use a conventional difference-in-differences model. For
each liberalization episode, we compute the outcome variable, as well
as the relevant controls, for the 10-year period before, and the 10-year
period after the liberalization date. Then, for each episode, we identify
a control group of countries from among those that did not liberalize
during the 20-year period around the liberalization date. Using these,
we estimate the following set of specications:
VOLATILITY
ict
=
0
POST
t
+
1
TREATED
ct
+ X
ict
+ + e
ict
1a
GROWTH
ict
=
0
POST
t
+
1
TREATED
ct
+ X
ict
+ + e
ict
: 1b
Here and throughout the paper, c indexes countries, i industries,
and t time periods. On the left-hand side is either the 10-year average
growth rate of a variable (GROWTH
ict
), or the standard deviation of
that growth rate calculated over the 10 year span (VOLATILITY
ict
).
Model (1) is the classic difference-in-differences specication. The
left-hand side variable is measured in two periods, before and after
treatment. Thus, by construction, in this model t takes on only two
values: before liberalization, and after it. The variable POST
t
takes on
the value of 0 before the liberalization episode, and 1 after. It is
common to both treated and control observations. Finally, the
coefcient of interest
1
is on the variable TREATED
ct
, which is a
binary indicator for whether a country is liberalized in a given period.
Intuitively, while the familiar RajanZingales-type model uses non-
nancially intensive sectors as a control group for the nancially
intensive sectors, this empirical strategy uses non-liberalizing
countries as a control group for the liberalizing country.
The vector of controls X
ict
contains the beginning-of-period share
of the sector in total output, as well as exports and imports as a share
of output in the sector.
12
In addition, X
ict
includes a measure of
7
A small number of studies attempt to measure the effect of nancial liberalization
by using rm-level data for several countries. Henry (2000a,b) nds that stock market
liberalizations are associated with a reduction in the cost of capital, followed by an
investment boom in a sample of listed rms in 12 emerging markets. Also using listed
rms, Mitton (2006) nds that rms with stocks that are open to foreign investors
experience higher growth, greater protability, and improved efciency. Alfaro and
Charlton (2007) use a large cross-section of both listed and non-listed rms in 1999
and 2004 to show that international nancial integration fosters the entry of new
rms, a nding in line with our industry-level results.
8
Using the variable name conventions from the Penn World Tables, this deation
procedure involves multiplying the nominal U.S. dollar value by (100/P)(RGDPL/
CGDP) for output, and (100/P)*(KI/CI)*(RGDPL/CGDP) for investment to obtain the
deated value.
9
We check the results by using instead a measure of stocks of gross foreign assets
and liabilities from Lane and Milesi-Ferretti (2006). The results are robust to this
alternative index of de facto liberalization, and we do not report them to avoid
unnecessary repetition.
10
We use the version of the variable assembled by Klingebiel, Kroszner, and Laeven
(2007), in which industries are classied according to the 3-digit ISIC Revision 2
classication.
11
Whenever the nancial liberalization index used is not binary, an important
question is how to dene a nancial liberalization event. In the baseline regressions we
classify a country as liberalized whenever all three components of the index
domestic, capital account, and stock market indicate full liberalization. This
approach emphasizes the complementarities between the different nancial liberal-
ization reforms.
12
We use beginning-of-period values rather than period averages for share to avoid
inducing a mechanical correlation with the left-hand side variable: a faster-growing
sector will tend to have higher share in the contemporaneous period.
212 A.A. Levchenko et al. / Journal of Development Economics 89 (2009) 210222
nancial development (private credit as a share of GDP), and the
interactionbetweenthe country's nancial development andthe Rajan
Zingales measure of dependence onexternal nance. Theseare meant to
control for the well-documented differential growth effects of nancial
development. Raddatz (2006) nds that volatility ina sector responds to
nancial development differentially depending on its liquidity needs. In
the volatility specications we thus control for this effect, by including
instead the interaction between nancial development and liquidity
needs. Appendix Table A3 presents the correlation matrix for the
independent variables. Both specications include a set of xed effects
. The ability to employ a variety of xed effects is a major strength of
our empirical approach, as these canpotentially control for a wide range
of omitted variables. The use of xed effects becomes especially
powerful in a three-dimensional panel, which makes it possible to use
interacted effects, such as countrysector, or sectortime.
The key questionis what countries to assign to the control group for
each liberalization episode. This paper pursues two strategies. First, for
each episode we use as the control group all of the countries that did
not liberalize around the same time as the liberalizing country. This
procedure can result in a large number of heterogeneous countries
constituting each control group. To rene this procedure one step, we
only use OECDcountries as available controls for the OECDliberalizers,
and non-OECD countries as possible controls for the non-OECD
liberalizers. The advantage of this approach is that it uses a large
amount of information for what is happening in various non-liberal-
izing countries around the time of each liberalization episode. The
disadvantage is that besides the coarse OECD/non-OECD renement,
no attempt is made to use countrycharacteristics inpicking the control
groups. Potentially, this can result in the control group countries
having very different characteristics from the treated ones for each
episode. Note that the large size of the control groups should help in
this respect, since the country heterogeneity would be averaged out
among the large number of control countries. Also, many of the
obvious differences, such as the overall level of development, which
can arise between a treated country and its control, would be
accounted for by the country xed effects included in the estimation.
Nonetheless, potential selection concerns remain. In order to
overcome them, we also employ a propensity score matching
procedure (henceforth PSM) to nd a suitable control group. The
supplementary web appendix to this paper describes it in detail. The
PSM procedure seeks to use information on observable characteristics
of subjects to estimate a probability model for being treated. Then, for
each instance of a treated observation, it uses the information on the
observables to identify a non-treated observation closest to the
treated one. That non-treated observation then becomes the control
group for the treated one. The rst economic applications of the
propensity score techniques are due to Dehejia and Wahba (1999,
2002), while in international economics they were rst used by
Persson (2001) and Glick, Guo and Hutchinson (2006). Though it has
been applied widely in various empirical analyses, it must be kept in
mind that the PSM method corrects only for selection on observables,
not unobservables. Furthermore, it can be sensitive to the set of
conditioning variables used to predict propensity scores (see Smith
and Todd, 2005).
3.2. Alternative estimation strategies
Because the nancial liberalization variable varies at the coun-
trytime level, in the baseline empirical model we cannot include
countrytime effects that would capture any other time-varying
country characteristics not picked up by the controls. An alternative
approach is to exploit sector-level characteristics in the spirit of Rajan
and Zingales (1998) to identify a causal relationship between nancial
liberalization and outcomes. We rely on the variation in the
dependence on external nance introduced by Rajan and Zingales
(1998), as well as the liquidity needs measure fromRaddatz (2006). In
particular, we estimate the following specications on the sample of
liberalizing countries:
VOLATILITY
ict
= CHAR
i
TREATED
ct
+ X
ict
+
ct
+
i
+ e
ict
2a
GROWTH
ict
= CHAR
i
TREATED
ct
+ X
ict
+
ct
+
i
+ e
ict
; 2b
where c indexes countries, i industries, and t time periods. Same as
above, GROWTH
ict
and VOLATILITY
ict
are the average growth rates
over the 10-year period, and the standard deviation of the growth rate
over the same period, respectively. TREATED
ct
is dened identically to
the above specication: it is zero except in the post-liberalization
period for the country that liberalized. CHAR
i
refers to the industry
characteristic used in estimation. This characteristic is either the Rajan
and Zingales measure of dependence on external nance, or the
Raddatz measure of liquidity needs. X
ict
is a vector of controls. All of
the specications include a full set of countrytime effects
ct
, as well
as sector effects
i
. Thus, in this model we identify the effect of
nancial liberalization purely from the differential effects across
industries within a country. The Rajan and Zingales-type approach is a
common one in the literature, indeed we are not the rst to analyze
the growth effects of nancial liberalizationwith this strategy (though
we are the rst, to our knowledge, to address the issue of volatility).
It is important toemphasize the pros andcons of model (1) compared
to (2). The disadvantage of the former is that it may suffer from an
omitted variables problem, because of our inability to include coun-
trytime effects. Its main advantage is that it allows us to estimate the
direct effect of nancial liberalization on the average growth and
volatility across sectors within a country. By contrast, the omitted
variables problem is overcome in the RajanZingales-type model.
However, its key shortcoming is that because it relies solely on the
within-country cross-industry variation, it does not allowthe researcher
to identify the magnitude of the overall effect. That is, the growtheffect of
nancial liberalizationtheobject of muchstudyusingthecross-country
regression approach is subsumed in the countrytime xed effect.
To further check robustness of the results to alternative measures
of nancial liberalization, we estimate an empirical model based on de
facto indices rather than de jure ones:
VOLATILITY
ict
= FINOPEN
ct
+ X
ict
+ + e
ict
3a
GROWTH
ict
= FINOPEN
ct
+ X
ict
+ + e
ict
: 3b
The sample is a non-overlapping panel of 10-year averages, 1970
79, 198089, 199099, thus the subscript t refers to decades. The
variable of interest, FINOPEN
ct
, is the gross capital ows as a share of
GDP (see Kose et al. 2003). Finally, we also consider a RajanZingales-
type difference-in-differences panel specication in which the de
facto measure of nancial integration, FINOPEN
ct
, is interacted with
industry characteristics.
4. Results
4.1. Volatility and growth
We now discuss the results of estimating the baseline model (1).
Table 1 reports the estimates of the relationship between nancial
liberalization and volatility of output. The rst four columns use the
full control group, while the last four use the PSMgroup. As we cannot
use countrytime effects, we experiment with various congurations
of xed effects to control for omitted variables. Column 1 presents
estimation results with country xed effects, while column 2 uses
countrysector xed effects. Column 3 uses country and grouptime
xed effects, where we dene a group to be a single liberalizing
country plus all its control countries. The grouptime effects control
for the time variation in the variables affecting both the treated and
the control countries, such as the changes in the global conditions.
213 A.A. Levchenko et al. / Journal of Development Economics 89 (2009) 210222
Finally, column 4 uses the country and groupsector xed effects. The
latter is the same as using sector xed effects, but within each
individual group (as, for example, the sector effects may change over
time). Because nancial liberalization occurs at countrytime level,
we cluster the standard errors at countrytime level as well, in order
to avoid biasing the standard errors downwards.
Financial liberalization appears to increase volatility, as the
coefcients of interest with both the full and the PSM control groups
are positive and signicant in all but one case. The coefcient is stable
across the control groups and xed effects congurations. It implies
that a nancial liberalization event is associated with a rise in the
standard deviation of sectoral growth of 1.52.4 percentage points, or
about 0.130.21 standard deviations of volatility found in the sample.
Table 2 reports the results of estimating Eq. (1b), with the average
growthrate of output over the 10-year period as the dependent variable.
Once again, the rst four columns use all available countries as control
groups, while the last four report the results withthe PSMcontrol group.
The columns differ in their use of xed effects, identically to the
estimates of the volatility effect of nancial liberalization in Table 1.
We can see that nancial liberalization has a robust positive effect on
growth of output across sectors. This effect is present across all cong-
urations of xed effects except one. Using the PSM control group, the
coefcient is signicant at 1% in all cases. The magnitude of the effect is
large. A nancial liberalization, captured by moving the TREATED
variable from 0 to 1, is associated with a sector-level growth rate that is
between1.5(full control group) and3.5(PSMcontrol group) percentage
points higher. This is equivalent to 0.17 and 0.40 of a standard deviation
of the 10-year average sector-level growth rate observed in the sample.
Among the other controls, the most signicant one is the initial share
in total output, which has a negative sign in both the volatility and
growth specications. We interpret this as a standard convergence
effect: sectors that are already large and established experience less
growth in the subsequent period. Trade openness and nancial
development on its own do not appear to be robustly signicant. The
RajanZingales term private credit interacted with external nance
dependence has a signicant impact on growth in most specications,
as has been extensively documented. By contrast, the interaction
between nancial development and the Raddatz measure of liquidity
needs does not appear tohave arobust impact onvolatilityinour sample.
We next discuss the results of the two alternative estimation
strategies laid out above. First, we employ an alternative difference-in-
differences model based on sector characteristics, models (2a) and
(2b). Results are presented in Table 3. We can see that within
countries, sectors that rely more on external nance tend to grow
faster. In addition, sectors that rely on external nance, as well as
sectors with greater liquidity needs tend to become more volatile as a
result of nancial liberalization.
The secondalternative strategy, model (3), uses de facto measures of
nancial liberalization instead of de jure ones. Table 4 reports the results
of estimating Eq. (3a), in which the dependent variable is the standard
deviation of the growth rate of output over the 10-year period, while
Table 5 reports the results of estimating the impact of nancial
liberalization on growth, Eq. (3b). Unless otherwise indicated, we use
the same specications, controls, and congurations of xed effects
throughout for maximum comparability. The independent variable of
interest, FINOPEN, is the average gross capital ows over the same 10-
year period. Because FINOPEN is measured at countrytime level, we
cluster the standard errors at the countrytime level as well. The rst
four columns add progressively more xed effects.
13
FINOPEN has a positive effect on volatility for all congurations of
xed effects, though the level of signicance is at 10% in most
specications. The magnitude of the impact of FINOPEN on volatility is
Table 1
Difference-in-differences results based on control countries, volatility.
(1) (2) (3) (4) (5) (6) (7) (8)
Dep. var.: standard deviation of the growth rate of output
Treated 0.022 0.021 0.015 0.022 0.022 0.022 0.024 0.023
[0.009] [0.014] [0.006] [0.009] [0.011] [0.011] [0.006] [0.013]
Post 0.002 0.001 0.013 0.001 0.005 0.004 0.034 0.003
[0.007] [0.011] [0.013] [0.007] [0.016] [0.017] [0.012] [0.020]
Exports/output 0.030 0.016 0.030 0.016 0.021 0.016 0.021 0.008
[0.005] [0.015] [0.005] [0.005] [0.008] [0.008] [0.008] [0.007]
Imports/output 0.003 0.000 0.002 0.001 0.003 0.002 0.003 0.002
[0.002] [0.002] [0.002] [0.001] [0.002] [0.001] [0.002] [0.002]
Initial share 0.281 0.888 0.282 0.250 0.234 0.445 0.235 0.103
[0.026] [0.216] [0.026] [0.059] [0.047] [0.259] [0.047] [0.092]
Private credit 0.016 0.124 0.073 0.069 0 0.027 0.017 0.017
[0.037] [0.062] [0.037] [0.041] [0.092] [0.109] [0.060] [0.118]
Private credit Liq. needs 0.089 0.617 0.088 0.258 0.067 0.099 0.062 0.03
[0.057] [0.397] [0.057] [0.131] [0.072] [0.383] [0.072] [0.245]
Country FE Yes No Yes Yes Yes No Yes Yes
CountrySector FE No Yes No No No Yes No No
GroupTime FE No No Yes No No No Yes No
GroupSector FE No No No Yes No No No Yes
Control group All All All All PSM PSM PSM PSM
Observations 3789 3789 3789 3789 1738 1738 1738 1738
R-squared 0.28 0.71 0.29 0.48 0.3 0.72 0.33 0.57
Notes: Robust standard errors clustered at countrytime level in brackets; signicant at 10%; signicant at 5%; signicant at 1%. The dependent variable is the standard
deviation of the growth rate of output during the 10 years immediately before or immediately after an episode of nancial liberalization. Treated takes on the value of 1 if a
liberalization event took place in a country, and zero otherwise. Post takes on the value of zero before the liberalization event, and 1 after, for all countries irrespective of whether they
liberalized. Initial share is the beginning-of-period share of output in a sector in total manufacturing output. Exports/output and Imports/output are the exports and the imports in the
sector divided by the total output in the sector. Private credit is the private credit by banks and other nancial institutions as a share of GDP. Liq. needs is the sector-level measure of
liquidity needs. In the rst 4 columns the control group consists of all countries (within the group of OECD/non-OECD) that did not liberalize within the 20-year period. In the last
four columns the control group is the country selected by the propensity score matching procedure (PSM). All specications are estimated using OLS, and including the xed effects
specied in the table. Variable denitions and sources are described in detail in the text.
13
Column 1 includes country, sector, and time effects separately. Column 2 uses
instead country and sectortime xed effects. Column 3 adds countrysector and
time effects. Note that in this column, identication comes purely from the time series
variation in the variables of interest. Column 4 includes countrysector and
sectortime xed effects. This is the most stringent possible array of xed effects
(in terms of remaining degrees of freedom) that can be included in this specication.
The only difference between the volatility and the growth specications is that Eq. (3a)
interacts private credit with the Raddatz measure of liquidity needs, while Eq. (3b)
interacts private credit with the Rajan and Zingales measure of external nance
dependence instead. This difference in the control variables does not affect the results.
214 A.A. Levchenko et al. / Journal of Development Economics 89 (2009) 210222
economically signicant. Aone standarddeviationchange inFINOPENis
associated with a rise in the standard deviation of sector-level growth
rate of 1.6 percentage points, equivalent to a movement of 0.13 standard
deviations of the sectoral volatility in the sample. From Table 5, it is
evident that the nancial openness variable also has a positive effect on
the growth rate of total output. The magnitude of the coefcient of
interest is economically signicant. A one standard deviation change in
de facto nancial openness is associated with a 1.3 percentage points
increase in the output growth rate, a change of 0.16 standard deviations.
Finally, columns 5 and6 of Tables 4 and5 interact FINOPENwiththe
Rajan and Zingales measure of dependence on external nance andthe
Raddatz measure of liquidity needs. We include countrytime xed
effects, controlling for other changes such as reforms that occur at
country level and differ across time. Note that this makes it impossible
toestimate the effect of FINOPENonvolatilityor growth, but enables us
to make a statement about its differential impact across sectors. Higher
levels of FINOPENincrease volatility more in sectors that depend more
on external nance, or with higher liquidity needs. When it comes to
growth effects, it does appear to be the case that more nancially
dependent sectors grow faster as a result of liberalization than less
nancially dependent sectors. We do not nd a signicant differential
growth effect for sectors with higher liquidity needs.
In sum, the estimates using the de jure and de facto measures
nancial integration yield strikingly similar results. Both reveal that
nancial liberalization increases both growth and volatility. Both
effects are magnied in sectors that are more dependent of external
nance, suggesting that those sectors are growing faster inpart thanks
to higher leverage in the post-liberalization period.
14
When it comes
to magnitudes, the impact of a de jure liberalization on both growth
and volatility is larger than that of increasing de facto capital ows by
one standard deviation. The two measures of nancial liberalization
Table 3
Difference-in-differences results based on industry characters.
(1) (2) (3) (4)
Output
Growth Volatility
Extern. nTreated 0.016 0.024
[0.010] [0.015]
Liq. needsTreated 0.100 0.158
[0.067] [0.097]
Exports/output 0.004 0.005 0.007 0.008
[0.005] [0.005] [0.011] [0.011]
Imports/output 0.004 0.004 0.001 0.001
[0.002] [0.002] [0.001] [0.001]
Initial share 0.109 0.113 0.332 0.329
[0.066] [0.064] [0.105] [0.103]
Private credit Extern. n 0.005 0.031
[0.020] [0.029]
Private credit Liq. needs 0.237 0.316
[0.152] [0.202]
CountryTime FE Yes Yes Yes Yes
Sector FE yes yes yes yes
Observations 852 852 851 851
R-squared 0.55 0.55 0.46 0.46
Notes: Robust standard errors in brackets; signicant at 10%; signicant at 5%;
signicant at 1%. The dependent variable is the average growth rate, or the standard
deviation of the growth rate of output during the 10 years immediately before or
immediately after an episode of nancial liberalization. Treated takes on the value of 1 if
a liberalization event took place, and zero otherwise. Private credit is the private credit
by banks and other nancial institutions as a share of GDP. Extern. n. is the sector-level
measure of reliance on external nance. Liq. needs is the sector-level measure of
liquidity needs. Initial share is the beginning-of-period share of output in a sector in
total manufacturing output. Exports/output and Imports/output are the exports and the
imports in the sector divided by the total output in the sector. All specications are
estimated using OLS, and including countrytime and sector xed effects. Variable
denitions and sources are described in detail in the text.
Table 2
Difference-in-difference results based on control countries, growth.
(1) (2) (3) (4) (5) (6) (7) (8)
Dep. var.: growth rate of output
Treated 0.015 0.151 0.015 0.016 0.034 0.035 0.025 0.035
[0.008] [0.013] [0.006] [0.009] [0.008] [0.013] [0.006] [0.0010]
Post 0.012 0.011 0.003 0.013 0.021 0.021 0.007 0.021
[0.004] [0.006] [0.002] [0.004] [0.007] [0.011] [0.004] [0.009]
Exports/output 0.004 0.013 0.004 0.004 0.003 0.004 0.003 0.004
[0.004] [0.006] [0.004] [0.003] [0.005] [0.013] [0.005] 0.006]
Imports/output 0 0.001 0 0 0.004 0.003 0.005 0.005
[0.000] [0.002] [0.001] [0.001] [0..002] [0.003] [0.002] [0.033]
Initial share 0.053 1.553 0.054 0.218 0.036 1.598 0.038 0.191
[0.019] [0.146] [0.019] [0.041] [0.030] [0.205] [0.030] [0.072]
Private credit 0.026 0.006 0.028 0.039 0.051 0.091 0.011 0.055
[0.018] [0.029] [0.017] [0.020] [0.039] [0.058] [0.034] [0.049]
Private credit Extern. n 0.053 0.176 0.052 0.007 0.059 0.182 0.059 0.053
[0.005] [0.016] [0.005] [0.016] [0.007] [0.027] [0.007] [0.026]
Country FE Yes No Yes Yes Yes No Yes Yes
Countrysector FE No Yes No No No Yes No No
Grouptime FE No No Yes No No No Yes No
Groupsector FE No No No Yes No No No Yes
Control group All All All All PSM PSM PSM PSM
Observations 3799 3799 3799 3799 1738 1738 1738 1738
R-squared 0.35 0.75 0.39 0.5 0.43 0.79 0.47 0.63
Notes: Robust standard errors clustered at countrytime level in brackets; Signicant at 10%; signicant at 5%; signicant at 1%. The dependent variable is the average growth
rate of output during the 10 years immediately before or immediately after an episode of nancial liberalization. Treated takes on the value of 1 if a liberalization event took place in a
country, and zero otherwise. Post takes on the value of zero before the liberalization event, and 1 after, for all countries irrespective of whether they liberalized. Initial share is the
beginning-of-period share of output in a sector in total manufacturing output. Exports/output and Imports/output are the exports and the imports in the sector divided by the total
output in the sector. Private credit is the private credit by banks and other nancial institutions as a share of GDP. Extern. n. is the sector-level measure of reliance on external nance.
In the rst 4 columns the control group consists of all countries (within the group of OECD/non-OECD) that did not liberalize within the 20-year period. In the last four columns the
control group is the country selected by the propensity score matching procedure (PSM). All specications are estimated using OLS, and including the xed effects specied in the
table. Variable denitions and sources are described in detail in the text.
14
How do our volatility results compare to the existing estimates? The literature
using cross-country data has focused on the volatility of aggregate consumption rather
than output. Even for aggregate consumption, the results are inconclusive: while Kose
et al. (2003) nd, paradoxically, that nancial integration increases consumption
volatility, Bekaert et al. (2006) nd the opposite. Glick et al. (2006) demonstrate that
nancial integration reduces the likelihood of currency crises. However, these results
are not directly comparable to ours, as currency crises are a different object than the
year-on-year volatility studied here.
215 A.A. Levchenko et al. / Journal of Development Economics 89 (2009) 210222
are not directly comparable, however. It could be, for instance, that a
typical de jure episode we analyze is equivalent to a more than one
standard deviation change in de facto openness.
Our difference-in-differences approach that uses the de jure
indicators requires a precise dating of nancial liberalization reforms.
Because of this, we transform the KS measure into a binary indicator
and thus overlook the gradual nature of the nancial liberalization
process. As a robustness check, we used the original KS index in place
of FINOPEN in the panel specication model (3)).
15
Table A6 in the
supplementary web appendix shows that our results are robust to this
alternative estimation strategy.
4.2. Factor accumulation vs. total factor productivity growth
We next investigate the channels through which nancial liberal-
ization increases the growth rate of output. We would like to know
whether it is associated with greater entry (the number of rms).
Furthermore, as in a standard growth accounting framework, growth
in total production can come from increased employment, capital
accumulation, and growth in total factor productivity (TFP). We use
the standard techniques to construct the capital stock and a TFP series
for each country and sector (see, for example, Hall and Jones, 1999).
The capital stock in each year t is given by K
ict
=(1)K
ict 1
+I
ict
,
where I
ict
denotes investment. We take a depreciation rate =0.08,
and adopt the standard assumption that the initial level of capital
stock is equal to I
ic0
/. We then follow Jorgenson and Stiroh (2000) to
compute total factor productivity at the industry level. Log of TFP in
year t is equal to lnTFP
ict
=lnY
ict
(1
ic
)lnL
ict

ic
lnK
ict
, where Y
ict
is the total output, and L
ict
is the total employment in the sector. Each
sector in each country has its own labor share
ic
, computed as the
average of the total wage bill divided by value added.
16
Table 6 investigates the impact of nancial liberalization on each of
these components of overall growth. We estimate Eq. (1b) with the
growth rate in the number of establishments, employment, capital
accumulation, and TFP as dependent variables. For each of these, we
report results with both the full and PSM control groups. All of the
specications are presented only with country and grouptime xed
effects, though the results are robust across the various xed effects
congurations. Columns 1 and 2 use the growth rate in the number of
establishments as the dependent variable. The evidence here is mixed.
While the full control group sample produces zero effect, when we
select the control group with the PSM procedure, it turns out that the
effect of nancial liberalization on entry is strongly positive. Columns
3 and 4 show that the growth rate of sector-level employment
increases signicantly with nancial liberalization. Columns 5 and 6
investigate the impact of nancial liberalization on capital accumula-
tion. The effect is positive and robustly signicant. Finally, there does
not appear to be a robust positive effect of nancial liberalization on
TFP. In one of the specications it is not signicant, while in the other
there is a positive and marginally signicant coefcient.
We estimated the impact of nancial liberalization on the channels
for growth in two additional ways. First, we used the RajanZingales
identication strategy in model (2). Second, we used the de facto
Table 4
De facto nancial liberalization and volatility, 10-year panel estimates.
(1) (2) (3) (4) (5) (6)
Dep. var.: standard deviation of the growth rate of output
FINOPEN 0.266 0.271 0.269 0.277
[0.108] [0.109] [0.147] [0.149]
Extern. nFINOPEN 0.164
[0.081]
Liq. needsFINOPEN 1.836
[0.598]
Log(Output/worker) 0.021 0.021 0.021 0.023 0.020 0.020
[0.005] [0.005] [0.013] [0.014] [0.005] [0.005]
Initial share 0.283 0.282 0.289 0.335 0.288 0.280
[0.041] [0.042] [0.147] [0.166] [0.042] [0.042]
Exports/output 0.005 0.018 0.145 0.211 0.043 0.039
[0.124] [0.120] [0.210] [0.211] [0.128] [0.129]
Imports/output 0.005 0.000 0.005 0.016 0.000 0.001
[0.025] [0.024] [0.031] [0.031] [0.025] [0.025]
Private credit 0.007 0.014 0.067 0.011
[0.048] [0.047] [0.080] [0.093]
Private credit Liq. needs 0.08 0.128 0.38 0.1 0.330
[0.160] [0.155] [0.410] [0.525] [0.148]
Private credit Extern. n 0.006
[0.020]
Country FE Yes Yes No No No No
Sector FE Yes No No No No No
Time FE Yes No Yes No No No
Countrysector FE No No Yes Yes No No
SectorTime FE No Yes No Yes Yes Yes
CountryTime FE No No No No Yes Yes
Observations 3761 3761 3761 3761 3761 3761
R-squared 0.39 0.41 0.65 0.66 0.48 0.48
Notes: Robust standard errors in brackets; standard errors are clustered at country-time level in columns (1)(4); signicant at 10%; signicant at 5%; signicant at 1%. The
sample is a panel of three decades, 197079, 198089 and 199099; all of the variables are 10-year averages unless otherwise indicated. The dependent variable is the standard
deviation of the growth rate of output over the 10-year period. FINOPENis gross capital ows, dened as the absolute value of total inows plus the absolute value of total outows, as
a share of GDP. Log(Initial output/worker) is the log of output per worker in a sector. Initial share is the beginning-of-period share of output in a sector in total manufacturing output.
Exports/output and Imports/output are the exports and the imports in the sector divided by the total output in the sector. Private credit is the private credit by banks and other
nancial institutions as a share of GDP. Extern. n. is the sector-level measure of reliance on external nance. Liq. needs is the sector-level measure of liquidity needs. All specications
are estimated using OLS, and including the xed effects specied in the table. Variable denitions and sources are described in detail in the text.
15
The original KS index varies from 1 (no liberalization in any dimension) to 3 (full
liberalization in all dimensions).
16
Alternatively, we applied to all countries the labor share in sector i in the U.S., or
the average labor share in sector i across all countries in the sample. We also used labor
productivity (value added per worker) instead of TFP. The results were unchanged.
216 A.A. Levchenko et al. / Journal of Development Economics 89 (2009) 210222
nancial openness indices and estimated model (3) for each
subcomponent of growth. In addition, we combined the Rajan
Zingales strategy with de facto indices. The results are reported in
Tables A7 and A8 of the supplementary web appendix to this paper.
They conrm our conclusions regarding the channels for the growth
impact of nancial liberalization: there is a robust effect on employ-
ment and capital accumulation, and suggestive evidence of increased
entry. However, by and large there is no robust impact of nancial
liberalization on TFP growth.
17
4.3. Temporary vs. permanent effects
This paper uses a variety of empirical strategies to document the
effect of nancial liberalization on growth, volatility, and the various
subcomponents of output at a 10-year horizon. Going much beyond
10 years would be impractical, as there aren't many liberalization
episodes in the sample that occurred more than 10 years before our
data ends. However, we can still investigate whether the magnitude of
the effect of nancial liberalization changes over time. This will allow
us to establish whether the impact of liberalization on various
outcomes is short-lived, or has a chance to be long-lasting.
In this section, we break the post-liberalization periods into 3-year
intervals: 02, 35, 68, and 911 years, and use the difference-in-
differences model (1) with the PSM control group to estimate the
treatment effect (
1
) for each 3-year period after liberalization.
Examining these coefcients will tell us at which lag the effect of
nancial liberalization is at its strongest. Fig. 2 presents the results. It
plots
1
over time, along with the 90% condence intervals.
The left panel of Fig. 2 presents the timing of the growth effects. It
is clear that the positive effect of nancial liberalization occurs early in
the sample: the rst 6 years. At longer lags, the effect of nancial
liberalization on growth becomes muted and not statistically
signicant. The time pattern also indicates that the growth effect in
the post-liberalization period is highly non-stationary: growth rises
on impact, accelerates further 3 to 5 years after liberalization, and then
decelerates to reach zero at the end of the 12 year period. An
interesting question is how much of the increase in growth volatility
within 10 years found in Section 4.1 is due to the non-stationarity of
the growth transition. To measure this, we compute the increase in
growth volatility implied by the time evolution of the growth effects.
We nd that it amounts to 1.8%, a gure only slightly lower than the
average post-liberalization effect for volatility in the 10 year window
presented in Table 1 (2.22%).
The right panel of Fig. 2 presents the timing of the volatility effects.
Note that we measure the impact of nancial liberalization on short-
run growth volatility within each 3-year interval, abstracting from the
impact of change in growth between intervals discussed above. We
nd that the growth volatility experiences a sharp increase in the
immediate aftermath of nancial liberalization. This effect is reduced
over time but remains positive. In the last interval 911 years the
effect on volatility is equal to 1.7%, though it is not statistically
different from zero (P-value of 15.6%). Therefore, we cannot
17
We do not report here the decomposition of the volatility results into channels as
we did with the growth results. While in growth accounting the growth rates of each
component of the production function add up to the total, the volatilities of the
subcomponents do not add up to the volatility of the total because of the covariances
among the subcomponents. Thus, it is not as informative to report the effect of
nancial liberalization on each subcomponent, and may be misleading as to what is
responsible for the overall effect if the covariances are also changing. Results are
nevertheless available upon request.
Table 5
De facto nancial liberalization and growth, 10-year panel estimates.
(1) (2) (3) (4) (5) (6)
Dep. var.: growth rate of output
FINOPEN 0.260 0.259 0.282 0.286
[0.066] [0.066] [0.092] [0.091]
FINOPENExtern. n. 0.187
[0.061]
FINOPENLiq. needs 0.445
[0.411]
Log(Initial output/worker) 0.015 0.014 0.036 0.036 0.013 0.013
[0.004] [0.004] [0.008] [0.008] [0.004] [0.004]
Initial share 0.100 0.098 0.641 0.738 0.100 0.092
[0.028] [0.029] [0.118] [0.128] [0.029] [0.029]
Exports/output 0.001 0.002 0.142 0.160 0.020 0.036
[0.140] [0.137] [0.258] [0.266] [0.124] [0.124]
Imports/output 0.021 0.019 0.004 0.009 0.027 0.029
[0.024] [0.024] [0.035] [0.035] [0.021] [0.021]
Private credit 0.008 0.01 0.003 0.005
[0.029] [0.029] [0.042] [0.044]
Private Credit Extern. n 0.036 0.038 0.096 0.112 0.017
[0.017] [0.017] [0.062] [0.076] [0.015]
Private credit Liq. needs 0.078
[0.111]
Country FE Yes Yes No No No No
Sector FE Yes No No No No No
Time FE Yes No Yes No No No
Countrysector FE No No Yes Yes No No
Sectortime FE No Yes No Yes Yes Yes
Countrytime FE No No No No Yes Yes
Observations 3777 3777 3777 3777 3777 3777
R-squared 0.31 0.33 0.57 0.59 0.41 0.41
Notes: Robust standard errors in brackets; standard errors are clustered at countrytime level in columns (1)(4); signicant at 10%; signicant at 5%; signicant at 1%. The
sample is a panel of three decades, 197079, 198089 and 199099; all of the variables are 10-year averages, unless otherwise indicated. The dependent variable is the growth rate of
output. FINOPEN is gross capital ows, dened as the absolute value of total inows plus the absolute value of total outows as a share of GDP. Log(Initial output/worker) is the log of
beginning-of-period output per worker in a sector. Initial share is the beginning-of-period share of output in a sector in total manufacturing output. Exports/output and Imports/
output are the exports and the imports in the sector divided by the total output in the sector. Private credit is the private credit by banks and other nancial institutions as a share of
GDP. Extern. n. is the sector-level measure of reliance on external nance. Liq. needs is the sector-level measure of liquidity needs. All specications are estimated using OLS, and
including the xed effects specied in the table. Variable denitions and sources are described in detail in the text.
217 A.A. Levchenko et al. / Journal of Development Economics 89 (2009) 210222
denitively rule out the possibility of a permanent increase in short-
run volatility on top on the temporary increase in medium-run
volatility resulting from the growth transition.
Howdoes nancial liberalization affect the subcomponents of total
output analyzed in this paper? Fig. 3 presents the timing of the effects
for each channel affecting growth.
18
Panel 1 presents the treatment
effect on the growth of the number of establishments. There is a
positive effect in the short-run, same as for the total output.
19
Panel 2
presents the results on employment growth. These mirror the overall
output results: a positive and signicant short-run effect, becoming
muted at longer lags. The results for capital accumulation growth are
presented in Panel 3. What is interesting here is that the effect of
nancial liberalization is both longer-lasting, and increasing over time,
until the 9th year or so after liberalization. Thus, the capital
accumulation effect is more persistent than the other outcomes:
since capital apparently adjusts slowly, it takes longer to attain the full
impact. Unlike the output and employment effects, the effect of
nancial liberalization on capital accumulation is still positive at the
longest lag, but it is not signicant due to substantially widened error
bands. Panel 4 presents the TFP chart. Consistent with the regression
results from almost all of our specications, there is no persistent
effect of nancial liberalization onTFP growth. It is only in the rst two
periods that TFP growth increases signicantly. To see whether there
is an effect on the level of TFP at 10 year horizon, we compound the
point estimates for each subsequent three-year interval. We nd a
cumulative level effect on TFP close to zero.
Finally, panel 5 considers another outcome, the level of the price-
cost margin. It is dened as follows:
PCM =
value of sales wages cost of inputs
value of sales
;
andis meant tocapturethesizeof markups, andthus thecompetitiveness
of the industry (see Braun and Raddatz, 2008).
20
The effect of nancial
liberalization on the price-cost margin is negative and signicant, quite
pronounced, and appears persistent. We call this reduction in markups
the pro-competitive effect of nancial liberalization.
The nding of a reduction in the markups can, in part, explain why
we nd a permanent effect on the level output without any detectable
effects on TFP. Since the presence of markups introduces a wedge
between the marginal product of capital and the rental rate of capital,
their reduction can lead to a higher steady-state level of capital and
output, as shown by Gal (1994, 1995). Such a permanent effect on the
level of output is also likely to result in much larger welfare gains from
nancial liberalization than the ones implied by the standard
neoclassical model (Gourinchas and Jeanne, 2006).
A general feature of our results is the apparent lack of signicant
effects of nancial liberalization on total factor productivity growth.
These results should be interpreted with caution, as the construction of
TFP may be subject to several measurement biases. First, we do not have
direct information on the use of intermediate inputs in sectoral
production. The direction of the resulting bias is hard to assess since it
depends on the change in the use of intermediate inputs relative to the
other factors of production.
21
Second, as shown by Hall (1988), a change
Table 6
Difference-in-differences results based on control countries, channels.
(1) (2) (3) (4) (5) (6) (7) (8)
Number of establishments Employment Capital accumulation Total factor productivity
Treated 0.006 0.028 0.011 0.023 0.018 0.040 0.010 0.000
[0.009] [0.007] [0.005] [0.005] [0.011] [0.011] [0.006] [0.006]
Post 0.084 0.018 0.054 0.050 0.023 0.061 0.008 0.002
[0.027] [0.007] [0.007] [0.006] [0.011] [0.009] [0.006] [0.008]
Exports/output 0.003 0.011 0.007 0.004 0.003 0.002 0.003 0.000
[0.003] [0.006] [0.004] [0.004] [0.003] [0.005] [0.003] [0.003]
Imports/output 0.000 0.003 0.001 0.003 0.000 0.002 0.000 0.005
[0.000] [0.001] [0.000] [0.002] [0.000] [0.002] [0.000] [0.001]
Initial share 0.023 0.044 0.049 0.031 0.048 0.049 0.056 0.060
[0.014] [0.026] [0.018] [0.019] [0.014] [0.018] [0.013] [0.023]
Private credit 0.124 0.092 0.019 0.018 0.033 0.057 0.068 0.065
[0.078] [0.056] [0.015] [0.033] [0.027] [0.068] [0.017] [0.039]
Private credit Extern. n 0.045 0.049 0.048 0.055 0.048 0.060 0.007 0.002
[0.005] [0.006] [0.004] [0.005] [0.004] [0.005] [0.002] [0.003]
Country FE Yes Yes Yes Yes Yes Yes Yes Yes
GroupTime FE Yes Yes Yes Yes Yes Yes Yes Yes
Control group All PSM All PSM All PSM All PSM
Observations 2870 1510 3839 1764 3287 1539 3267 1536
R-squared 0.42 0.47 0.4 0.48 0.58 0.65 0.24 0.2
Notes: Robust standard errors clustered at countrytime level in brackets; signicant at 10%; signicant at 5%; signicant at 1%. The dependent variable is the average growth
rate of the number of establishments, total employment, capital stock, and TFP during the 10 years immediately before or immediately after an episode of nancial liberalization.
Treated takes on the value of 1 if a liberalization event took place in a country, and zero otherwise. Post takes on the value of zero before the liberalization event, and 1 after, for all
countries irrespective of whether they liberalized. Initial share is the beginning-of-period share of output in a sector in total manufacturing output. Exports/output and Imports/
output are the exports and the imports in the sector divided by the total output in the sector. Private credit is the private credit by banks and other nancial institutions as a share of
GDP. Extern. n. is the sector-level measure of reliance on external nance. In columns (1), (3), (5), and (7) the control group consists of all countries (within the group of OECD/non-
OECD) that did not liberalize within the 20-year period. In columns (2), (4), (6) and (8) the control group is the country selected by the propensity score matching procedure (PSM).
All specications are estimated using OLS, and including the xed effects specied in the table. Variable denitions and sources are described in detail in the text.
18
For the reasons mentioned in footnote 17, we do not report the time evolution of
the volatility of components of output. This gure is available upon request.
19
The results for the number of rms are not presented for the last period (911
years), as the coverage for the number of rms is more sparse than for other variables,
and thus there are not enough observations to obtain a reliable last period estimate.
20
The PCM is essentially a measure of protability, or the ow accrued to owners of
capital. Though imperfect as a measure of markups, it has the advantage of simplicity, and
has been widely used in the literature. It is also highly correlated to other indicators of
competitiveness, such as industry concentration ratios (see, e.g., Domowitz et al. 1986).
Furthermore, note that our empirical strategy relies on the time variation in this index.
Thus, tothe extent that mismeasurement occurs mainly inthe cross-sectionof countries or
industries rather than differentially over time, the results are still informative.
21
In particular, the fact that a large number of industrial sectors produce
intermediate inputs and have experienced higher growth following liberalization
possibly suggesting a higher demand for intermediates from other industries is not
directly informative of the direction of the bias in measured TFP growth. An alternative
method is to derive total factor productivity growth from value added instead of
output. This approach has the advantage of controlling for the role intermediate inputs
but it requires separability between the value added production function and
intermediate inputs, a condition generally not met in industry-level data (see
Jorgenson et al., 1987). We nevertheless computed an alternative measure of TFP
based on value added and the results were unchanged.
218 A.A. Levchenko et al. / Journal of Development Economics 89 (2009) 210222
in the Solow residual under imperfect competition can reect both a
change in total factor productivity and a change in markups. Note that a
reduction in markups suggested by the observed reduction in the
price-cost margin following liberalization would if anything bias our
results in favor of nding a positive TFP effect.
22,23
Finally, beyond
measurement issues, our results are consistent with the recent ndings
of Hale and Long (2007) on the lack of productivity spillovers on
domestic rms stemming from foreign direct investment ows.
4.4. Aggregation
Armed with point estimates of hownancial liberalization changes
sector-level growth and volatility, we can now calculate what these
imply for the aggregate economy. In a country comprised of sectors
i =i =1,, I, denote the level of an aggregate variable by upper case
Y
A
, and its growth rate by lower case y
A
. The aggregate growth rate can
be written as:
y
A
=

I
i =1
s
i
y
i
4
where s
i
is the share of sector i in the overall level of Y
A
in the country,
and y
i
is the growth rate of this variable in sector i. This paper
estimates the change in sector-level growth rate, y, that comes as a
result of nancial liberalization. The change in the aggregate growth
rate could be obtained from Eq. (4) in a straightforward manner:
y
A
=

I
i =1
s
i
y = y: 5
That is, if in each sector the growth rate of a variable increases by
y, then the aggregate growth of that variable will rise by the same
amount.
Note that this expression applies to all variables we analyzed:
output, employment, capital stock, and TFP. To compute the change in
aggregate TFP in the same way as we compute changes in output,
employment, and capital requires two additional simplifying assump-
tions. First, we must assume there is a CobbDouglas aggregate
production function Y
t
=A
t
K
t

L
t
1
. Second, we must assume a time-
invariant share of each industry in total output, total capital and total
employment. This assumption rules out composition effects. In the
presence of composition effects, aggregate TFP can increase purely
from expansion of high-TFP sectors and contraction of low-TFP
sectors, without any change in TFP at individual sector level. In
order to assess the empirical relevance of this mechanism, we re-
estimated the baseline model while allowing nancial liberalization to
affect growth in high-TFP sectors differently from growth in low-TFP
sectors. We found no evidence of composition effects due to nancial
liberalization: sectors with higher than average initial TFP do not
appear to grow systematically faster than sectors with low initial TFP
after liberalization (estimation results are available upon request).
24
The empirical analysis above produces the point estimate y

of the
change in sector-level average growth rate in the 10 years following
nancial liberalization. From Eq. (5) we immediately get y
A

=y

. By
the same argument, the standard errors of the aggregate growth
effects are simply equal to the standard errors of the sectoral growth
estimates for each variable of interest. Panel A of Table 7 reports the
estimated impact of nancial liberalization on aggregate growth rates
of the variables used in the analysis. Not surprisingly, they correspond
to the values reported in Tables 2 and 6 above.
Next, we evaluate the long-run impact of nancial liberalization.
Section 4.3 establishes that there is no long-run impact on growth.
However, the temporary growth impact over the rst 10 years following
liberalization compounds to yield a long-run level effect. Since the
growth effects of liberalization do not persist after 10 years, the
permanent level effect results from compounding each sectoral growth
22
Using a fully specied model, Jaimovich (2007) shows that true TFP growth (z) is
related to the change in the markups () and the Solow residual (SR) as follows:
z = SR + :
Since our measure of TFP is the Solow residual, a reduction in markups negative
implies that the true change in TFP is actually lower than our estimates, not higher.
23
See Hsieh and Klenow (2007) for a comprehensive analysis of the effect of
distortions on sectoral TFP in China and India.
Fig. 2. The time evolution of the growth and volatility effects of nancial liberalization. This gure depicts the treatment effect of nancial liberalization for the outcome variables
over time. The solid line is the coefcient on the TREATED dummy variable in the years 02, 35, 68, and 911 after the liberalization episode. Dashed lines represent the 10%
signicance bands.
24
The rst assumption yields the level of aggregate TFP as a function of sector-level
TFPs:
A
t
=
Y
t
K

t
L
1
t
=

i
A
it
K
it
K
t
_ _

L
it
L
t
_ _
1
:
The second assumption leads to the change in the growth rate of aggregate TFP equal
to:
a
A
t
=
dA
t
A
t
=

i
dA
it
A
t
K
it
K
t
_ _

L
it
L
t
_ _
1
=

i
dA
it
A
it
Y
it
Y
t
=

i
s
i
a
t
= a
t
;
same as Eq. (5).
219 A.A. Levchenko et al. / Journal of Development Economics 89 (2009) 210222
effect over 10 years: 1 + Y
A

= 1 + y
10
. Hence we set the
estimated impact to be

Y
A

= 1 +

y
_ _
10
1. Panel C of Table 7
reports the results. Given the point estimates, we compute the standard
errors using the delta method.
25
In addition, in order to assess whether
this long-run impact is statistically different from zero, the t-test is not
sufcient due to the non-linear transformation of the regression
estimates. Therefore, the statistical signicance of the long-run level
effect reported in Table 7 comes from a Wald test of this non-linear
relationship. We can see that while the growth effect is conned to the
rst decade after liberalization, its estimated long-run level effect is still
substantial. The level of aggregate output increases by 28%, behinda rise
of 25.5% in employment, and of 48% in the capital stock.
26
Fig. 3. The time evolution of the effect of nancial liberalization: channels. Notes: This gure depicts the treatment effect of nancial liberalization for the outcome variables over
time. The solid line is the coefcient on the TREATED dummy variable in the years 02, 35, 68, and 911 after the liberalization episode. Dashed lines represent the 10% signicance
bands. All variables are in growth rates with the exception of the price-cost margin which is in level.
26
We choose to report the long-run level effect based on the estimates from the
baseline 10-year regressions in Section 4.1, rather than the 3-year interval regressions
in Section 4.3. This choice is dictated primarily by the substantial difculty in
computing the standard errors of the level effect estimates based on the 3-year interval
regressions. The two methods produce very similar point estimates, however, that are
well within the condence interval for the level effects reported in Table 7.
25
Assuming

y is close enough to its true value, we use the following rst order
Taylor approximation:

Y
A

Y
A

g10 1 + y
9

y y
_ _
:
Therefore, E

Y
A

_ _
gY
A

and

Var

Y
A

_ _
_
g10 1 +

y
_ _
9

e
2
_
where is the esti-
mated standard error of y.
220 A.A. Levchenko et al. / Journal of Development Economics 89 (2009) 210222
When it comes to volatility, Eq. (4) implies that the change in the
standard deviation of aggregate output is equal to:

A
=

I
i =1
s
2
i

2
=

h
p
;

_
where is the impact of nancial liberalization on sector-level
volatility, and hu

I
i = 1
s
2
i
is the Herndahl of production shares in
the economy.
27
In contrast to the growth increase, the change in
sector-level volatility is moderated by the Herndahl index of
production shares in the economy. Thus, for any given change in
sector-level volatility, the increase in aggregate volatility is much
lower. For instance, the median value of h in our sample is 0.087,
leading to the change in the aggregate volatility equal to about one
third of the magnitude of the change in sector-level volatility:

A
=0.29. Panel B of Table 7 reports the change in aggregate
volatility implied by the sector-level volatility estimates in this paper.
While the standard deviation of volatility at sector level is predicted to
rise by 2.2% in the decade following liberalization, the median
economy's aggregate volatility rises by only 0.7% due to diversication
across sectors. This effect is quite minor: the average standard
deviation of aggregate output among the countries in our sample is
8.3%. In contrast to the growth effect above, the volatility impact will
vary depending on a country's level of diversication across sectors. To
get a sense of the variation in the volatility impact across countries,
Table 7 also reports the change in aggregate volatility predicted for
countries in the 25th and the 75th percentiles of the diversication
distribution (countries with lower Herndahls are more diversied).
We can see that the impact does not differ too much, ranging from
0.6% for the more diversied country to 0.9% for the less diversied
one. The contrast between the aggregation of the growth and the
volatility estimates yields an interesting conclusion. While at the
sector level the growth and volatility effects appear similar in
magnitude, the aggregate growth effect is on average three times
larger than the aggregate volatility effect.
5. Conclusion
It is often argued, both theoretically and empirically, that nancial
liberalization should affect economic growth. At the same time, claims
that nancial liberalization increases volatility are made just as often.
This paper uses a large panel of industry-level data to analyze both
growth and volatility effects within the same empirical framework. A
key strength of our approach is the number of alternative strategies
we use to estimate these relationships. We employ a variety of
difference-in-differences estimates, and use both de jure and de facto
measures of liberalization. We exploit sector characteristics, use non-
liberalizing countries as controls, develop a propensity score matching
procedure to overcome selection on observables, and use a variety of
xed effects throughout to control for omitted variables. What is
remarkable is that the conclusions are virtually the same across all
empirical strategies.
There is strong evidence that nancial liberalization increases both
growth and volatility of output. Those effects are not long-lasting:
they typically vanish after 6 years. When it comes to channels, we nd
that nancial liberalization is accompanied by an increase in the
growth of employment and capital formation. Furthermore, liberal-
ization exerts pro-competitive pressures on the product market: there
is a transitory increase in the entry of rms and a permanent drop in
the price to cost margin. By contrast, the growth rate of TFP does not
appear to be affected by liberalization.
Thus, both growth and volatility increase as a result of nancial
liberalization, though admittedly the signicance of the volatility
results is uniformly lower. Can we say something about the net
welfare impact? While a complete treatment of the welfare question
would require a fully specied growth model and is therefore outside
the scope of this paper, it is relatively easy to pin down the direction of
the net effect. Lucas (1987) shows that the welfare benets of
removing all of the U.S. business cycle volatility are minuscule about
0.05% of consumption. By the same logic, the adverse welfare impact
of higher volatility due to liberalization is quite small. In fact, the 3
percentage point reduction in consumption volatility Lucas consid-
ered is actually higher than the 1.52.5 percentage point increase in
output volatility implied by the estimates in this paper. Even for this
small adverse effect there are several mitigating factors. First, the
estimates in this paper are for the volatility of output, not consump-
tion. If agents can self-insure by smoothing intertemporally, the
implied welfare effect is lower. Second, the estimates are at sector
level. Thus, if there is any amount of risk sharing across sectors, the
adverse welfare impact would be reduced further. On the other hand,
though the growth effect of nancial liberalization is estimated to be
temporary, it still translates into a permanent level effect. The results
presented in this paper therefore imply that the welfare impact of
nancial liberalization is positive.
Acknowledgments
We are grateful to Mathieu Taschereau-Dumouchel for the superb
research assistance, and to the editors, two anonymous referees, Jean
Imbs, Olivier Jeanne, Raphael Lam, Gilles Saint-Paul, Aaron Tornell,
Thierry Tressel, Thijs van Rens, and workshop participants at the IMF,
Universit de Genve, HEC-Lausanne, UCLA, PSE, INSEAD, HEC-Paris,
UC Santa Cruz, IADB, ESSIM (Izmir), and CREI/CEPR Conference on
Finance, Growth and The Structure of the Economy (Barcelona) for
helpful comments. The views expressed in this paper are those of the
authors and should not be attributed to the International Monetary
Fund, its Executive Board, or its management.
Table 7
The impact on aggregate growth and aggregate volatility.
Short-run impact Long-run impact
Panel A: aggregate growth rate Panel B: aggregate volatility of output Panel C: aggregate level
Output Capital Employment TFP Herf-25 Herf-50 Herf-75 Output Capital Employment TFP
0.024 0.040 0.023 0.000 0.006 0.007 0.009 0.280 0.477 0.255 0.000
[0.006] [0.011] [0.005] [0.006] [0.0005] [0.0005] [0.0006] [0.072] [0.151] [0.057] [0.062]
Notes: Standard errors in brackets ( signicant at 1%). The short-run impact corresponds to the estimated changes in the aggregate growth rate/volatility over the 10 years
following an episode of nancial liberalization. These changes are computed based on estimates of Eqs. (1a) and (1b) and they use the industry-level estimates with the PSM control
group and Country and GroupTime xed effects (Tables 1, 2, and 6). The aggregate volatility impact is reported for three different values for the Herndahl index: the 25th, 50th, and
75th percentile. The long-run impact corresponds to the permanent effect of nancial liberalization on the level of aggregate output, employment, capital and TFP. Details of the
computation of point estimates and standard errors are described in detail in the text.
27
This assumes that liberalization does not have a signicant effect on the
covariances between the sectors in the economy, which appears to be the case in
our data.
221 A.A. Levchenko et al. / Journal of Development Economics 89 (2009) 210222
Appendix A. Supplementary data
Supplementary data associated with this article can be found, in
the online version, at doi:10.1016/j.jdeveco.2008.06.003.
References
Alfaro, Laura, and Charlton, Andrew, 2007. International Financial Integration and
Entrepreneurship, NBER Working Paper 13118.
Attanasio, Orazio, Davis, Steven, 1996. Relative wage movements and the distribution of
consumption. Journal of Political Economy 104 (6), 12271262.
Banerjee, Abhijit, Duo, Esther, 2005. Growth theory through the lens of development
economics. In: Durlauf, Steven, Aghion, Philippe (Eds.), Handbook of Economic
Growth, vol. 1A. Elsevier Science Ltd., North Holland.
Barro, Robert J, Mankiw, N. Gregory, Sala-i-Martin, Xavier, 1995. Capital mobility in
neoclassical models of growth. American Economic Review 85, 103115.
Beck, Thorsten, Demirg-Kunt, Asl, Levine, Ross, 2000. A new database on nancial
development and structure. World Bank Economic Review 14, 597605.
Bekaert, Geert, Harvey, Campbell, Lundblad, Christian, 2006. Growth Volatility and
nancial liberalization. Journal of International Money and Finance 25, 370403.
Braun, Matias, Raddatz, Claudio, 2008. The politics of nancial development: evidence
from trade liberalization. Journal of Finance 63, 14691508.
Broner, Fernando, and Ventura, Jaume, 2006. Globalization and Risk Sharing, mimeo,
CREI-UPF.
Dehejia, Rajeev, Wahba, Sadek, 1999. Causal effects in non-experimental studies: re-
evaluating the evaluation of training programs. Journal of the American Statistical
Association 94, 10531062.
Dehejia, Rajeev, Sadek, Wahba, 2002. Propensity score matching methods for non-
experimental causal studies. Review of Economics and Statistics 84, 151161.
di Giovanni, Julian, and Levchenko, Andrei A, in press. Trade Openness and Volatility,
Review of Economics and Statistics.
Domowitz, Ian, Hubbard, R. Glenn, Petersen, Bruce, 1986. Business cycles and the
relationship between concentration and price-cost margins. Rand Journal of
Economics 17, 117.
Gal, Jordi, 1994. Monopolistic competition, endogenous markups, and growth.
European Economic Review 38, 748756.
Gal, Jordi, 1995. Product diversity, endogenous markups, and development traps.
Journal of Monetary Economics 36, 3963.
Galindo, Arturo, Micco, Alejandro, Ordoez, Guillermo, 2002. Financial liberalization:
does it pay to join the party? Economia 3, 231252.
Glick, Reuven, Guo, Xueyan, Hutchinson, Michael, 2006. Currency crises, capital account
liberalization, and selection bias. Review of Economics and Statistics 88, 698714.
Gourinchas, Pierre-Olivier, Jeanne, Olivier, 2006. The elusive gains from international
nancial integration. Review of Economic Studies 73, 715741.
Gupta, Nandini, and Yuan, Kathy, 2006. On the Growth Effects of Stock Market
Liberalizations, mimeo, Indiana University and University of Michigan.
Hale, Galina, and Long, Cheryl, 2007. Is There Evidence of FDI Spillovers on Chinese
Firms' Productivity and Innovation? mimeo, Federal Reserve Bank of San Francisco.
Hall, Robert E., 1988. The Relation between Price and Marginal Cost in U.S. Industry.
Journal of Political Economy 96, 921947.
Hall, Robert, Jones, Chad, 1999. Why do some countries produce so much more output
per worker then others. Quarterly Journal of Economics 114, 83116.
Hayashi, Fumio, Altonji, Joseph, Kotlikoff, Laurence, 1996. Risk-sharing between and
within families. Econometrica 64 (2), 261294.
Henry, Peter Blair, 2000a. Stock market liberalization, economic reform, and emerging
market equity prices. Journal of Finance 55 (2), 529564.
Henry, Peter Blair, 2000b. Do stock market liberalizations cause investment booms?
Journal of Financial Economics 58 (12), 301334.
Henry, Peter Blair, 2007. Capital account liberalization: theory, evidence, and
speculation. Journal of Economic Literature 45, 887935.
Heston, Alan, Summers, Robert, Aten, Bettina, 2002. Penn World Table Version 6.1.
Center for International Comparisons at the University of Pennsylvania (CICUP).
Hsieh, Chang-Tai, and Klenow, Peter, 2007. Misallocation and Manufacturing TFP in
China and India, mimeo, UC Berkeley.
Jaimovich, Nir, 2007. Firm dynamics and markup variations: implications for multiple
equilibria and endogenous economic uctuations. Journal of Economic Theory 137,
300325.
Jorgenson, Dale W., Gollop, Frank M., Fraumeni, Barbara M., 1987. Productivity and U.S.
economic growth. Harvard University Press, Cambridge, MA.
Jorgenson, Dale W., Stiroh, Kevin J., 2000. U.S. economic growth at the industry level.
American Economic Review 90 (2), 161167.
Kaminsky, Graciela, Schmukler, Sergio, 2008. Short-run pain, long-run gain: the effects
of nancial liberalization. Review of Finance 12, 253292.
Klingebiel, Daniela, Kroszner, Randall, Laeven, Luc, 2007. Financial crises, nancial
dependence, and industry growth. Journal of Financial Economics 84, 187228.
Kose, M. Ayhan, Prasad, Eswar, Terrones, Marco, 2003. Financial integration and
macroeconomic volatility. IMF Staff Papers 50, 119142 Special Issue.
Kose, M. Ayhan, Prasad, Eswar, Rogoff, Kenneth, and Wei, Shang-Jin, 2006. Financial
Globalization, A Reappraisal, IMF Working Paper 06/189.
Lane, Philip, and Milesi-Ferretti, Gian Maria, 2006. The External Wealth of Nations Mark
II: Revised and Extended Estimates of Foreign Assets and Liabilities, 19702004,
IMF Working Paper, 06/69.
Levchenko, Andrei A., 2005. Financial liberalization and consumption volatility in
developing countries. IMF Staff Papers 52 (2), 237259.
Lucas Jr., Robert E., 1987. Models of Business Cycles. Basil Blackwell, New York.
Mitton, Todd, 2006. Stock market liberalization and operating performance at the rm
level. Journal of Financial Economics 81 (3), 625647.
Obstfeld, Maurice, 1994. Risk-taking, global diversication, and growth. American
Economic Review 84 (5), 13101329.
Persson, Torsten, 2001. Currency unions and trade: how large is the treatment effect?
Economic Policy 435448 (October).
Quadrini, Vincenzo, 2005. Policy commitment and the welfare gains from capital
market liberalization. European Economic Review 49 (8), 19271951.
Raddatz, Claudio, 2006. Liquidity needs and vulnerability to nancial underdevelop-
ment. Journal of Financial Economics 80, 677722.
Rajan, Raghuram, Zingales, Luigi, 1998. Financial dependence and growth. American
Economic Review 88 (3), 559586.
Saint-Paul, Gilles, 1992. Technological choice, nancial markets and economic
development. European Economic Review 36, 763781.
Smith, Jeffrey, Todd, Petra, 2005. Does matching overcome LaLonde's critique of
nonexperimental estimators? Journal of Econometrics 125 (12), 305353.
Tornell, Aaron, Velasco, Andres, 1992. The tragedy of the commons and economic
growth, why does capital ow from poor to rich countries? Journal of Political
Economy 100, 12081231.
Tressel, Thierry, and Verdier, Thierry, 2007. Does Financial Globalization Discipline
Politically Connected Firms, mimeo, IMF.
United Nations Industrial Development Organization, 2006. Industrial Statistics
Database, CD-ROM.
222 A.A. Levchenko et al. / Journal of Development Economics 89 (2009) 210222