Threshold Effects in the Relationship Between

Inflation and Growth: A New Panel-Data Approach∗

David Drukker
Stata Corporation
Pere Gomis-Porqueras
Department of Economics, University of Miami
Paula Hernandez-Verme
Department of Economics, Texas A&M

February 9, 2005

Abstract
In this paper we use a new approach to throw light on the old question of the super-neutrality
of money. Recent theoretical results suggest a threshold model instead of super-neutrality.
To ascertain whether or not there is a threshold level of inflation above which the effect of
inflation on long-run growth changes, we apply new econometric methods for estimation and
inference in non-dynamic, fixed-effects, panel-data models that may contain threshold effects.
In the full sample of 138 countries over the period 1950–2000, we find that there is one
threshold that is well identified by the data; the estimated value of the threshold is 19.16%. For
the industrialized sample, our results indicate that there are two threshold points at 2.57% and
12.61%. In the full sample, if the initial inflation rate is below 19.16%, increases in inflation
do not have a statistically significant effect on growth. In contrast, when the initial inflation
is above 19.16%, further increases in inflation will decrease long-run growth.

Keywords: threshold model, nonlinear model, empirical growth, super-neutrality, panel data
JEL Codes: C23, E31, O42, O57

∗
We would like to thank Badi Baltagi, Massimo Guidolin, participants in the 11th International Conference on
Panel Data and participants of the Texas A&M brown bag seminar for their comments. All remaining errors are
our own.

1
 The recent theoretical literature has begun to ask an old question in a new way; is money super-

neutral for all possible levels of initial inflation, or is there a threshold level of inflation, above

which money is not super-neutral? Our application of new panel-data techniques to standard

datasets provides evidence that the last question should be answered affirmatively.

Many of the theoretical models in the money and growth literature analyze the impact of

inflation on long-run growth. There are four principal predictions in the literature regarding the

impact of inflation on output and growth.

1. The prediction that there is no effect of inflation on growth (money is super-neutral) was first

established by (Sidrauski 1967) when he showed that money is neutral and super-neutral in

an optimal control framework with money in the utility function.

2. (Tobin 1965) assumed that money is a substitute for capital, causing inflation to have a

positive effect on long-run growth.

3. (Stockman 1981) presented a cash in advance model in which money is complementary to

capital, causing inflation to have a negative effect on long-run growth.

4. There is a new class of models in which inflation has a negative effect on long-run growth,

but only if the level of inflation is above a threshold level. In this class of models, financial

market efficiency is affected by various informational asymmetries, as in (Huybens and Smith

1998), for example. In these models, high rates of inflation typically exacerbate financial

market frictions, interfere with the efficiency of the financial system, and thus inhibit growth.

In this paper we provide evidence that there is a threshold level of inflation above which the

effect of inflation is negative and statistically significant, and we interpret this result as supporting

prediction 4. We apply new econometric techniques due to (Hansen 1999) and (Gonzalo and

Pitarakis 2002) to estimate the number of threshold points, their values and their coefficients and

2
perform inference about whether or not these effects could be attributed to sampling error. We

overcome the limitation to balanced panels 1 by applying our own insight that we can use a cluster

bootstrap to obtain the standard errors of the parameters in the model.

We use an unbalanced panel with data for 138 countries for the period 1950–2000. As is

standard in the empirical growth literature, we use averaged data for half-decades that do not

overlap.

The remainder of the paper is organized as follows. In Section I we briefly discuss previous

empirical studies on inflation thresholds and on economic growth. In Section II we describe our

empirical specification, our procedures for estimation and inference, and the dataset that we use.

In Section III, we present our results. Finally, in Section IV we present the main conclusions of

our study and discuss future extensions of this line of research.

I The Empirical Literature

In this section, we briefly discuss previous empirical studies that focus on inflation thresholds and

on economic growth.

A The Empirical Literature on Inflation Thresholds

The seminal work by (Fischer 1993) is among the first to examine the possibility of nonlinearities in

the relationship between economic growth and inflation in the long run. Using both cross section

and panel data for a sample that includes both developing and industrialized countries, Fischer’s

parameter estimates indicate a negative relationship between inflation and growth. Interestingly,

by using (arbitrarily chosen) break points of 15% and 40% in spline regressions, Fischer shows not

only the presence of nonlinearities in the relationship between inflation and growth, but also that
1
(Hansen 1999) uses a wild-bootstrap method for inference in balanced panels. The balanced panel assumption
is not necessary in the cluster bootstrap.

3
the strength of this relationship weakens for inflation rates above 40%. 2 One weakness of Fischer’s

study is the arbitrary nature of the sample splitting, a weakness that our approach overcomes.

In the time-series literature, (Bullard and Keating 1995) use a structural vector autoregression

to estimate the response of real output to permanent inflation shocks for each country in a sample

of 16 countries. Looking at their results, from a cross-sectional point of view, indicates the the

presence of the following nonlinear relationship. Increases in long-run inflation have a positive

(negative) effect on long-run output for sufficiently low (high) initial levels of inflation.

Finally, closer to our line of work, (Khan and Senhadji 2001) estimate the threshold level

of inflation by using a balanced panel that averages time-series data over nonoverlapping half

decades. The authors estimate the threshold level of inflation, above which inflation significantly

slows growth, to be between 0.89% and 1.11% for industrialized countries, and between 10.62%

and 11.38% for developing countries. The negative and significant relationship between inflation

and growth, for inflation rates above the threshold level, is quite robust with respect to the

estimation method, perturbations in the location of the threshold level, the exclusion of high

inflation observations, data frequency and alternative specifications. But their results are difficult

to interpret because their specification does not fit into the class of models pioneered by (Hansen

1999, Hansen 2000).3 This implies that Hansen’s methods of inference do not necessarily apply.

Moreover, they do not use the model selection methods derived by (Gonzalo and Pitarakis 2002),

which allow one to select the number of threshold points without incurring a multiple testing

problem.
2
See also (Sarel 1996), and (Ghosh and Phillips 1998). For higher frequency data see (Bruno and Easterly 1998),
and (Faria and Carneiro 2001).
3
(Khan and Senhadji 2001) use a functional form that is continuous at the threshold point, but this specification
does not fit directly into the (Hansen 1999, Hansen 2000) framework.

4
B The Empirical Literature on Economic Growth

Traditional and augmented cross-sectional analyses focus on conditional convergence and suffer

from a striking lack of attention to nonlinear alternatives to the neoclassical growth model. While

this paper does not aim to study cross-country convergence, the results therein indicate that

inflation may have a nonlinear effect on growth.

While the panel-data approach provides greater flexibility, thereby reducing potential misspec-

ifications, it also complicates the interpretation in terms of conditional convergence. This change

in interpretation is not a problem for this paper, because we do not aim to study cross-country

convergence. Rather, we simply want to know whether or not, inflation has a nonlinear effect on

growth, after removing cross-country heterogeneity,

As with all the papers in this literature that use nonoverlapping half-decade averages to

approximate long-run behavior, our choice of bandwidth is driven by conventions in the literature.

In an unpublished appendix, we report similar results that use four-year and six-year averages.

II Empirical Methodology
A The Empirical Specification and Econometric Methods

We estimate the coefficients and number of threshold points in the functional form specified in

equation (1):

j
X
yit = µi + νt + γp ditp π̃it + xit β + it , (1)
p=0

where j is the number of threshold points, y it denotes the percentage growth rate of country i at

time t, µi denotes the level of country-i’s fixed-effect, ν t is the level of time-t’s fixed-effect, γ p is

the coefficient on the semi-log transform of inflation of country i at time t (π̃ it ) in region p, ditp is

the indicator variable for region p, x it is a vector of other covariates, β is a vector of coefficients on

5
xit , and it is a zero mean, finite variance, i.i.d. disturbance that is orthogonal to the covariates

and the threshold variable.4

When j = 0, equation (1) reduces to the well-known linear fixed effects (FE) model, but, when

j ≥ 0, this specification is nonlinear in the semi-log of inflation. To see these points, note that

the threshold-region indicator variable is


 1 if π̃p∗ < π̃it ≤ π̃p+1
∗


ditp = .

 0 otherwise


where π̃0∗ = −∞, π̃j+1

∗ = ∞ are the endpoints, and π̃p∗ , p ∈ {1, 2, . . . , j}, are the j threshold

points.5

The fact that some observations have negative inflation rates prohibits using the log of infla-

tion. For this reason, we use the semi-log transform of the inflation rate π it given by



 πit − 1 if πit ≤ 1

π̃it = .

 ln(πit )

if πit > 1

A specification with j threshold points defines j+1 distinct regions over the inflation rate space

and requires the estimation of and inference about the j + 1 coefficients γ p , and the j threshold

points π̃q∗ .6 Treating the fixed effects as nuisance parameters implies that the parameters of

interest are the number of threshold points j, the j + 1 coefficients γ p , the j thresholds π̃q∗ , and

the vector β.

If the true number of threshold points and their values were known, equation (1) would be a

linear panel-data model with two-way fixed effects, and all the standard methods would apply. 7
4
The orthogonality condition (ii) of Assumption A2 in (Gonzalo and Pitarakis 2002) is satisfied when each
covariate in xit and the generated covariates ditp π̃it are strictly exogenous.
5
When j = 0, there are no threshold points, only the endpoints remain.
6
The linear FE model has zero thresholds and one region; a specification with one threshold point has two
distinct regions, and so on.
7
See (Baltagi 2001) and (Wooldridge 2001).

6
The lack of knowledge of the number of threshold points and their values complicates estimation

and inference.

We apply the results of (Gonzalo and Pitarakis 2002) to estimate the number of threshold

points (j). As in the time-series literature 8 , their method selects the model that minimizes the

Bayesian information criteria (BIC). 9,10

The parameters of the models are estimated using the sequential estimation procedure derived

by (Hansen 1999, Hansen 2000, Gonzalo and Pitarakis 2002). 11 (Drukker and Hernandez-Verme

2005) discuss the details of the estimation procedure and the computation methods.

Our empirical specification is discontinuous in inflation at the threshold point(s). This discon-

tinuity implies that small changes in inflation, in a neighborhood of the threshold point, may have

different effects depending on whether initial inflation is above or below the threshold. Intuitively,

in the spirit of (Huybens and Smith 1998) and the related literature, non-convexities in the econ-

omy may create a situation in which an increase in inflation causes a discontinuous drop in per

capita growth when initial inflation is just below the threshold point. Similarly, reducing inflation

in a country with initial inflation just above the threshold value may cause a discontinuous jump

in per capita growth.

As is standard in the empirical, panel-data, growth literature, our specification is based on

five-year averages of the dependent variable and of the covariates of interest. This practice is

conventional because it allows researchers to analyze the long-run relationships without having
8
For time-series examples, see (Hannan 1979, Lütkepohl 1993)
9
(Hansen 1999, Hansen 2000) developed a likelihood-ratio (LR) testing procedure. We use the (Gonzalo and
Pitarakis 2002) method because it is applicable to unbalanced panels and it avoids the multiple testing problem
that arises when choosing among zero, one and two thresholds using Hansen’s LR test procedure.
10
In a related paper, (Khan and Senhadji 2001) include the threshold point as part of their semi-log transform
so that the model for growth is continuous at the threshold point. Their specification does not directly fit into the
(Hansen 1999, Hansen 2000, Gonzalo and Pitarakis 2002) framework for estimation and inference.
11
These authors were building on advances made in the time-series literature by (Bai and Perron 1998, Bai and
Perron 2003).

7
to specify the short-run dynamics. 12 Thus, in our particular case, the sub-index t refers to the

half-decade t or, equivalently, the length of each period t is five years.

B Covariates and predicted signs of coefficients

The dependent variable yit is the five-year arithmetic average of the annual percentage growth

rate of the GDP per capita in country i during the half-decade t. We chose our covariates so that

they would be similar to those used by other researchers. 13 We define our vector of covariates as

xit = (igdpit , dpopit , dtotit , tot sdit , openit , open sdit ).

In the cross-sectional growth literature, some of these variables are treated as endogenous and

instrumental variables (IV) estimators are used. The methods used in this paper have not yet

been extended to the case of instrumental variables. This paper assumes that any endogenous

components are perfectly correlated with the fixed effects, and therefore controlled for by our

fixed-effect estimation procedure. 14

Recall that the subscript i indicates the country and the subscript t indicates the half-decade.

Henceforth, to simplify the notation, we drop the subscripts from the names of the covariates.

igdp is the arithmetic average of the annual percentage of GDP dedicated to investment. dpop is

the arithmetic average of the percentage change in population. dtot is the arithmetic average of

the annual percentage change in the terms-of-trade. As is standard in the literature, igdp, dpop

and dtot each have a linear effect on the growth rate of per capita GDP.

tot sd is the semi-log transform of the standard deviation of the terms-of-trade during the half-

decade.15 open is the arithmetic average of the natural log of openness. Openness is measured as
12
Below we discuss the results based on five-year averages. In an unpublished appendix, we discuss the results
based on four-year and six-year averages. Changing the window size for the average does not change any of the
fundamental results.
13
For the purpose of comparability, our choice of covariates is similar to the one used by (Khan and Senhadji
2001) for the open economy case, and (Islam 1995) for the closed economy case.
14
While we plan to use IV methods in future research, the regression results reported here are an important
baseline against which future IV results can be compared.
15
As discussed in (Heston, Summers and Aten 2004), the terms-of-trade index is relative, with the base of 100

8
the share of exports plus imports in the GDP. Finally, open sd is the natural log of the standard

deviation of openness during the half-decade. tot sd, open and open sd each have a multiplicative

effect on the growth rate of per capita GDP.

Although the threshold(s) and the semi-log transform of inflation complicate the computation

of the marginal effect of inflation on long-run growth, the monetary-theory literature generates at

least two interesting hypotheses about the γ p coefficients. First, the null hypothesis that money

is super-neutral can be investigated by testing whether all the γ p are zero. Second, suppose that

there is a single threshold point.16 In this case, evidence that γ0 ≥ 0 > γ1 , with γ1 statistically

significant would support the hypothesis of a nonlinear effect of inflation on long-run per capita

growth in accordance with the theory in (Huybens and Smith 1998). The latter case would also

provide evidence against the super-neutrality of money.

Defining the vector β as β 0 = (βigdp , βdpop , βdtot , βtot sd , βopen , βopen sd ) simplifies the nota-

tion. Thus, βigdp is the coefficient on igdp, βdpop is the coefficient on dpop, βdtot is the coefficient

on dtot, βtot sd is the coefficient on tot sd, βopen is the coefficient on open and βopen sd is the

coefficient on open sd. Below we describe the interpretation of each one of these coefficients as

well as their predicted signs.

The coefficient βigdp indicates that an increase (decrease) of one point in the percentage of GDP

dedicated to investment would result in an increase (decrease) of β igdp points in the percentage

growth of GDP per capita. The standard growth model would predict that β igdp > 0. From a

theoretical standpoint, there is no obvious reason to expect different coefficients when comparing

industrial versus developing countries.

for the United States (US). This implies that the standard deviation of the terms-of trade for the US during a
half-decade is always zero. But the univariate distribution of the five-year standard deviation of the terms-of-trade
series indicated that a log specification was preferable to a level specification. To avoid dropping the US from the
sample the same semi-log transform used for inflation was used to create tot sd.
16
There are more elaborate versions of this hypothesis for models with multiple threshold points, but the issue is
much clearer in the case of a single threshold point.

9
 An increase (decrease) of one point in the percentage growth rate of the population reduces

(increases) the percentage growth rate of GDP per capita in β dpop points. The standard growth

model would predict that βdpop < 0. However, some studies17 have observed the potential problem

of endogeneity in this covariate, which could lead to evidence of β dpop ≤ 0, after controlling for

fertility rates. However, the latter studies are cross-sectional in nature. We do not explicitly

control for fertility in this paper because this series is not available as a time series with the

frequency that we need.18

dtot measures long-run changes in the real exchange rate, and standard trade theory would

predict βdtot ≤ 0. tot sd measures volatility in the terms-of-trade or the real exchange rate, and

one would predict that βtot sd < 0.

With respect to the level of openness, while many theoretical models predict β open > 0, many

empirical studies have found it to be negative. Finally, in so far as open sd measures volatility in

a country’s commitment to a level of openness, standard theory would predict β open sd < 0.19

C Data

As discussed above, we use data from two different sources: the Penn World Tables 6.1 20 and

the International Financial Statistics. From the Penn World Tables, we start with an unbalanced

panel of 174 countries. This panel contained information for GDP per capita and the covariates,

except inflation, for the period 1950 to 2000. 21 For the same time span, we obtained data for the

inflation rate for a sample of 149 countries from the International Financial Statistics. 22 After
17
See (Barro and Lee 1994) for an example.
18
The country level FE control for fertility rates to the extent that they vary over countries but not time.
19
Frequent changes in rules may reduce the expected returns to foreign investment and to domestic investment
in exports. The fall in returns reduces the optimal level of foreign investment and diverts domestic investment to
less productive sectors, thereby lowering growth.
20
The official citation of this dataset is (Heston, Summers and Aten October 2002).
21
The series used for GDP per capita is the one reported in constant prices, Laspeyres’ index in the Penn World
Tables 6.1. Similarly, we used investment as a share of the GDP in current prices, population, openness in current
prices and the price level of GDP.
22
The official citation for this dataset is (International Financial Statistics 2004).

10
first-differencing variables and matching the information from the aforementioned sources, we

were left with an unbalanced panel of 138 countries. 23

As discussed above, in order to focus on long-run effects, we followed a procedure frequently

used in the literature, which begins by partitioning each time series into nonoverlapping half-

decade intervals. By including 2000 into the last of half-decades, we have 10 potential half-decade

intervals.24 After forming the partitions, we average the available data inside these half-decades

into a single observation, creating an unbalanced panel of 916 observations from 138 countries. 25

As is standard in the empirical growth literature, we were interested in checking whether our

results are similar for the industrialized and nonindustrialized countries in our sample. We use

the IFS definition of industrialized countries. 26

III Results

Table 1 presents the model selection results for the three different samples. Models with one

threshold point were selected for the full and nonindustrialized samples, while a two threshold-

point model was selected for the industrialized sample. The fact that the linear FE model was

not selected in any of the three samples is strong evidence in favor of the presence of at least one

threshold.

Table 2 presents the estimates of the thresholds and the γ p coefficients for all the different

models and samples. The number in parentheses under the estimates of the γ p coefficients, is the
23
The countries lost are Antigua, Azerbaijan, Bahamas, Bahrain, Belarus, Bermuda, Bhutan, China, Comoros,
Cuba, Djibouti, Equatorial Guinea, Eritrea, Guinea, Iraq, Kuwait, Laos, Lebanon, Liberia, Mongolia, Oman,
Puerto Rico, Qatar, Russia, Sao Tome and Principe, Saudi Arabia, Somalia, Sudan, Suriname, Taiwan, Tajikistan,
Turkmenistan, Ukraine, United Arab Emirates, Uzbekistan and Yugoslavia
24
The intervals are 1950-1954, 1955-1959, 1960-1964, 1965-1969, 1970-1974, 1975-1979, 1980-1984, 1985-1989,
1990-1994, and 1995-2000.
25
In the estimations discussed below, we do not weight the observations to account for the unbalanced nature of
the averages.
26
By this definition, the industrialized countries, for the years in our sample, are Australia, Austria, Belgium,
Canada, Denmark, Finland, France, Germany, Greece, Iceland, Ireland, Italy, Japan, Luxembourg, Netherlands,
New Zealand, Norway, Portugal, Spain, Sweden, Switzerland, the United Kingdom, and the United States.

11
 Table 1: Model selection for the original Specification
Sample No. of SSR Parameters BIC
thresholds
Full 0 5063.7 154 9.676456
Full 1 4972.2 155 9.665663
Full 2 4935.6 156 9.665708
Full 3 4917.6 157 9.669506
Industrialized 0 341.5 39 6.768816
Industrialized 1 324.7 40 6.742298
Industrialized 2 311.2 41 6.723658
Industrialized 3 303.9 42 6.723983
Nonindustrialized 0 4625.3 131 9.680327
Nonindustrialized 1 4544.9 132 9.672248
Nonindustrialized 2 4508.9 133 9.673775
Nonindustrialized 3 4466.0 134 9.673703

number of observations in that threshold region. 27 Note that the estimates of the threshold points

for the industrialized sample are smaller than their counterparts for the full and nonindustrialized

samples. Inverting the semi-log transform produces the level estimates of 19.16% for the single

threshold-point model in the full and nonindustrialized samples. The level estimates for the

industrialized sample are 12.61% for the one threshold-point model and 2.57% and 12.61% for the

two threshold-point model. All of the threshold regions contain strictly more observations than

the minimum required by the estimation procedure.

Table 3 presents the point estimates and standard errors for the coefficients and threshold

points. For the full and nonindustrialized samples, we report results for single threshold-point
27
The estimation procedure requires that there is a minimum number of observations between threshold points.

The results presented here require a minimum number of 15 observations between threshold points. In an un-

published appendix, we report estimates that enforce minimums of 10 and 20 observations. As is to be expected,

using these two alternative minimums, changes some of the point estimates by small amounts, but all the model

selection statistics remain the same, except that in the case of at least 10 observations, a two threshold-point model

is selected for the full and nonindustrialized samples. Note that the linear FE model is not selected for any sample

over all three cases of minimum observations.

12
 Table 2: Threshold estimates

Sample No. of π̃1∗ π̃2∗ π̃3∗ γ0 γ1 γ2 γ3

thresholds
Full 0 -.20281
(916)
Full 1 2.952739 .106364 -.296385
(758) (158)
Full 2 2.952739 3.061248 .078371 -.876014 -.275562
(758) (27) (131)
Full 3 2.952739 3.061248 4.099721 .132461 -.814541 -.125527 -.330627
(758) (17) (86) (55)
Industrialized 0 -.123496
(226)
Industrialized 1 2.534730 .071569 -.419821
(208) (18)
Industrialized 2 0.945216 2.534730 1.14071 .110641 -.392363
(49) (159) (18)
Industrialized 3 0.472216 0.945216 2.534730 -.236838 1.6281 .243785 -.283182
(20) (29) (159) (18)
Nonindustrialized 0 -.198475
(690)
Nonindustrialized 1 2.952739 .135708 -.271194
(538) (17) (135)
Nonindustrialized 2 2.952739 3.061248 .10056 -.848798 -.250674
(410) (21) (107) (152)
Nonindustrialized 3 2.413063 2.482869 2.952739 .012823 .92544 .15022 -.260373

Note: The number of observations in each threshold region is given in parentheses under the corresponding
coefficient estimate. The total number of observations for sample is provided under the coefficient of the
zero threshold-point model.

13
models. For the industrialized sample, we report results for a one threshold-point model and a

two threshold-point model.28 We used a cluster bootstrap with 2000 replications to estimate the

standard errors.29

That the full-sample estimate of γ1 has the predicted signs and is statistically different from

zero provides evidence in favor of the (Huybens and Smith 1998) model. More evidence in favor

of this class of models is provided by the pattern in the estimates of the γ p coefficients. In all the

selected models, the estimates of γ p are decreasing in p.

The signs of the coefficients on the covariates are in line with the theoretical predictions. It

is worth highlighting that the estimates of the coefficients on both open and open sd have the

predicted signs and are statistically significant in the full and nonindustrialized samples.

The coefficient on igdp is positive and statistically significant for the nonindustrialized coun-

tries, as predicted by the neoclassical growth model. However, it is negative and insignificant for

the industrialized countries.

IV Conclusions

In order to throw new light on the question of the super-neutrality of money, this paper applies new

econometric techniques for estimation and inference in panel-data models with threshold effects

to standard datasets. We find strong evidence that inflation has a nonlinear effect on growth.

In particular, our results support the newer theoretical work 30 in which inflation has a negative
28
We report results for both models to enable a more direct comparison of the coefficient estimates across samples.
29
In the cluster bootstrap, each country is a cluster, and each bootstrap sample draws uniformly over the clusters

and selects whole clusters to be part of a bootstrap sample. Aside from the standard regularity conditions, treated

in (Hansen 1999), this bootstrap method only requires independence over the clusters and is robust to within-cluster

correlation.
30
For example, see (Huybens and Smith 1998).

14
 Table 3: Coefficient Estimates
Coefficient Full Sample Industrialized 1 Industrialized 2 Nonindustrialized H0
igdp .07869884∗ -.0008136 -.00786869 .08546722 βigdp > 0
(.03298677) (.06538358) (.0612653) (.03601169)
dpop .04004897 .56315732 .49889868 .01821566 βdpop > 0
(.44308504) (.54702438) (.52669043) (.51305887)
dtot -.04637448∗ .01724174 .01175966 -.05022043∗ βdtot > 0
(.02388336) (.03698134) (.04037888) (.02644399)
tot sd -.11291561 -.46695158∗ -.40730724∗ -.05769045 βtot sd > 0
(.13827403) (.18026063) (.18754344) (.1885233)
open av 1.341518∗ .08480795 .39049208 1.5217344∗ βopen > 0
(.59568356) (1.2400415) (1.1075688) (.66425252)
open sd -.59122938∗ -.47738841 -.50857913 -.64259756∗ βopen sd > 0
(.18920561) (.22706086) (.23575455) (.2317477)
γ0 .10636439 .07156886 1.1407106 .13570771
(.40851964) (1.0493123) (1.1774337) (.41349783)
γ1 -.29638508∗ -.41982096 .11064136 -.27119407
(.14238313) (.42617385) (.74547128) (.17292488)
γ2 -.39236268
(.37224348)
π̃1 2.952739∗ 2.5347304 ∗
.9452157∗ 2.952739∗
(.75821987) (.70680222) (.26390206) (.76054499)
π̃2 2.5347304∗
(.57202002)

15
effect on growth, but only after inflation reaches a threshold level. That the sign of the coefficient

on investment differs between the industrialized and nonindustrialized subsamples, and that the

coefficient on openness has a positive sign are interesting results. We are especially intrigued by

the latter, since it is a standard theoretical result that has often eluded empirical researchers. In

future research, we plan to investigate whether applying threshold techniques causes the estimate

of the coefficient on openness to agree with standard theory in cases where it previously has not.

References

Bai, J. and P. Perron, “Testing for and Estimation of Multiple Structural Changes,” Econo-

metrica, 1998, 66, 47–79.

and , “Computation Analysis of Testing Multiple Structural Changes,” Journal of

Applied Econometrics, 2003, 18, 1–22.

Baltagi, Badi H., Econometric Analysis of Panel Data, second ed., New York: John Wiley &

Sons, 2001.

Barro, Robert and J.-W. Lee, “Sources of Economic Growth,” Carnegie–Rochester Conference

Series, 1994, 40, 1–57.

Bruno, M. and W. Easterly, “Inflation Crises and Long-Run Growth,” Journal of Monetary

Economics, 1998, 41, 3–26.

Bullard, James and John W. Keating, “The long-run relationship between inflation and

output in postwar economies,” Journal of Monetary Economics, 1995, 36, 477–496.

Drukker, David and Paula Hernandez-Verme, “Estimation and model selection in fixed-

effects panel-data models with threshold effects,” Technical Report 2005.

16
Faria, J. and F. Carneiro, “Does High Inflation Affect Growth in the Long and Short Run?,”

Journal of Applied Economics, 2001, (1), 89–105.

Fischer, Stanley, “The Role of Macroeconomic Factors in Growth,” Journal of Monetary Eco-

nomics, 1993, (32), 485–512.

Ghosh, A. and S. Phillips, “Inflation May Be Harmful to Your Growth,” Technical Report,

IMF Staff Papers 1998.

Gonzalo, Jesús and Jean-Yves Pitarakis, “Estimation and Model Selection Based on In-

ference in Single and Multiple Threshold Methods,” Journal of Econometrics, 110 2002,

pp. 319–352.

Hannan, Quinn B., “The determination of the order of an autoregression,” Journal of the Royal

Statistical Society, Series B, 1979, 41, 190–195.

Hansen, Bruce, “Threshold Effects in Non-dynamic Panels: Estimation, Testing and Inference,”

Journal of Econometrics, 1999, 81, 594–607.

, “Sample Splitting and Threshold Estimation,” Econometrica, 2000, 68, 575–603.

Heston, Alan, Robert Summers, and Bettina Aten, “Data Appendix for a Space-Time

System of National Accounts: Penn World Table 6.1,” Technical Report, Center for Interna-

tional Comparisons at the University of Pennsylvania (CICUP) 2004.

, , and , “Penn World Table Version 6.1,” Technical Report, Center for Interna-

tional Comparisons at the University of Pennsylvania (CICUP) October 2002.

Huybens, Elizabeth and Bruce D. Smith, “Financial Market Frictions, Monetary Policy,

and Capital Accumulation in a Small Open Economy,” Journal of Economic Theory, 1998,

81, 353–400.

17
International Financial Statistics, Technical Report, IMF 2004.

Islam, Nazrul, “Growth Empirics: A Panel data Approach,” Quarterly Journal of Economics,

1995, 110, 1127–1170.

Khan, Mohsin and Abdelhak Senhadji, “Threshold Effects in the Relationship Between

Inflation and Growth,” Technical Report, IMF Staff Papers 2001.

Lütkepohl, Helmut, Introduction to Multiple Time Series Analysis, second ed., New York:

Springer-Velag, 1993.

Sarel, M., “Nonlinear Effects of Inflation on Economic Growth,” Technical Report, IMF Staff

Papers 1996.

Sidrauski, Miguel, “Rational Choice and Patterns of Growth in a Monetary Economy,” Amer-

ican Economic Review, 1967, 57, 534–544.

Stockman, Alan C., “Anticipated Inflation and the Capital Stock in a Cash-in-Advance Econ-

omy,” Journal of Monetary Economics, 1981, 8, 387–393.

Tobin, James, “Money and Economic Growth,” Econometrica, 1965, 33, 671–684.

Wooldridge, Jeffrey, Econometric Analysis of Cross Section and Panel Data, Cambridge, Mas-

sachusetts: MIT Press, 2001.

18