European Economic Review 44 (2000) 359}381
Long-term growth and short-term
economic instability
Philippe Martin!,#,*, Carol Ann Rogers"
! CERAS-ENPC, 28 rue des Saints Pe% res, 75007 Paris, France
" Department of Economics, Georgetown University, Washington, DC 20057, USA
# CEPR, London, UK
Received 1 August 1997; accepted 1 May 1998
Abstract
When learning by doing is at the origin of growth the long-run growth rate should be
negatively related to the amplitude of the business cycle if human capital accumulation is
increasing and concave in the cyclical component of production. Empirical evidence
strongly supports this "nding for industrialized countries and European regions. Using
the standard control variables, we "nd that countries and regions that have a higher
standard deviation of growth and of unemployment have lower growth rates. The result
does not come from an e!ect of instability on investment. The negative relation, however,
does not hold for non-industrialized countries, for which learning by doing may not to be
the main engine of growth. ( 2000 Elsevier Science B.V. All rights reserved.
JEL classixcation: O40; E32
Keywords: Growth; Business cycle; Learning by doing; Short-term economic instability
1. Introduction
A strong tradition among macroeconomists has been to study the business
cycle and long-term growth as two separate phenomena. Business cycle theorists
* Corresponding author. Tel.: 33 1 44 58 28 76; fax: 33 1 44 58 28 80; e-mail: martin-p@paris.
enpc.fr.
0014-2921/00/$ - see front matter ( 2000 Elsevier Science B.V. All rights reserved.
PII: S 0 0 1 4 - 2 9 2 1 ( 9 8 ) 0 0 0 7 3 - 7
360
P. Martin, C.A. Rogers / European Economic Review 44 (2000) 359}381
have considered long-term growth as an exogenous trend and growth theorists
have typically worked with models where short-term shocks have no impact on
the long-run growth rate of the economy.
A recent strand of literature has put together the two phenomena in a common theoretical framework. From a theoretical point of view, the relation
between short-term economic instability and long-run growth can be positive or
negative, depending on the mechanism at the origin of growth. As shown by
Aghion and Saint-Paul (1993), the sign of the relation depends on whether the
activity that generates growth in productivity is a complement or a substitute to
production. In the case where they are substitutes, since the opportunity cost of
productivity improving activities falls in recessions, a larger amplitude and
frequency of business-cycle #uctuations may have a positive e!ect on long-run
productivity and growth (Aghion and Saint-Paul, 1991). In the case of complementarity, they show, as does Stadler (1990), that a positive (negative) shock
will have a positive (negative) long-term impact on productivity.
Two recent papers of ours (Martin and Rogers, 1995, 1997), in which growth
is generated by learning by doing, belong to this class of models in which
production and productivity increasing activities are complements. We have
shown how stabilization policies could have a positive impact on human capital
accumulation and, through this channel, on growth. A counter cyclical policy
that smooths the impact of shocks on employment was needed for growth
maximization (Martin and Rogers, 1997) or welfare maximization (Martin and
Rogers, 1995). One natural implication from these models, that we did not
directly address, was that short-term economic instability is detrimental for
human capital accumulation and growth.
As theoretical models can predict a negative or a positive relation between the
amplitude of the business cycle and growth, the next natural step should be to
attempt to settle the question at the empirical level. One way to do this is to use
time-series methods to determine the long-run e!ect of macroeconomic shocks.
Bean (1990) and Saint-Paul (1993) have found results that are mildly supportive
of a negative long-run e!ect of demand shocks on productivity. These timeseries results may not be easy, however, to interpret in terms of the relation
between long-term growth and the business cycle. The question of whether
a positive shock has a positive or negative impact on the level of long-term
productivity is not the same as the question of whether the amplitude of the
business cycle is a determinant of the long-term growth rate. Even though the
answer to the "rst question is important from a positive point of view, it is not
certain what the policy implications are. On the other hand, if the amplitude of
the business cycle has a negative impact on long-run growth, this has important
policy implications because it gives counter cyclical stabilization policies a new
strong role.
A more direct way to test the relation is to use cross-country regressions. This
paper does so for OECD countries as well as for a sample of 90 European
P. Martin, C.A. Rogers / European Economic Review 44 (2000) 359}381
361
regions. Our empirical "ndings are the following. There is a strongly signi"cant
and negative relation between growth and the standard deviation of growth,
both for OECD countries and for European regions. This is true with di!erent
sets of control variables and this is robust to di!erent speci"cations. In particular, the relation is robust to the inclusion in the regressions of the investment
share in GDP for industrialized countries. We also "nd that European regions
and industrialized countries with high standard deviation of unemployment
have had lower growth which points to the key role of employment and
presumably learning by doing in this relation. However, we do not "nd that such
a relation holds for developing countries. This is consistent with Young's (1993)
theoretical "nding that growth will be driven by learning by doing only at
relatively high levels of development.
Ramey and Ramey (1995) have found, in a sample of 92 countries as well as in
a sample of OECD countries, that countries with higher volatility of growth
have lower growth. Their work focuses on the possible link between short-term
uncertainty and long-run growth.1 In our work (Martin and Rogers, 1995,
1997), we posit that learning by doing could generate a relation between growth
and instability. The link between #uctuations and growth then would rely
neither on uncertainty nor on an investment channel but on a labour channel so
that in our tests of this relationship in this paper, we will not di!erentiate
between predictable and unpredictable shocks.
The next section of the paper describes the relation between instability and
growth that we want to test, the methodology and the data. Section 3 reports
the empirical results for European regions and OECD countries. Section 4 does
it for developing countries. An appendix outlines the simple theoretical model
that informs our empirical tests.
2. Instability and growth
In an economy in which learning by doing is at the origin of growth, business
cycle #uctuations may reduce the growth rate. Recessions are periods in which
opportunities to learn by doing are foregone, so adverse business cycle shocks
negatively a!ect human capital accumulation. As long as employment and
therefore human capital accumulation is increasing and concave in the business
cycle disturbance, the &lost' learning is not fully regained when the cycle turns
upwards again. This is precisely the channel through which business cycle
#uctuations can negatively a!ect long-term growth rates.
1 They are interested in testing theories for which the negative relation between volatility and
growth relies on uncertainty through the link of investment, see Bernanke (1983), Pindyck (1991) and
Ramey and Ramey (1991).
362
P. Martin, C.A. Rogers / European Economic Review 44 (2000) 359}381
The appendix outlines a growth model with stochastic short-term productivity
shocks. It shows that the amplitude of the business cycle a!ects negatively the
growth rate if employment is concave in the output disturbances. The source of
the disturbances (demand or supply) is not important for the conclusions, since all
that matters is that employment is increasing and concave in the disturbances. We
do not test directly this concavity condition, but note that one of the stylized facts
reported in the empirical literature on business cycles (see, for example, Danthine
and Donaldson, 1993) is that labour productivity is pro-cyclical in all industrialized countries. The positive correlation between labour productivity (output
over employment) and output itself implies that employment increases with
output but at a decreasing rate. Evidence also exists at the micro-level that, for
di!erent product lines, learning rates are initially high, declining over time as
production cumulates. Lucas (1993) insists both on &how impressive the evidence
on the productivity e!ects of learning by doing can be' and on the shape of the
learning curve.2 This evidence of the learning curve suggests that learning
increases more rapidly when production is high (as it accumulates more rapidly)
than when it is low but that this increase takes place at an decreasing rate.
We test the relation between growth and instability for three samples. The
"rst sample consists of 90 European regions for which we have data between
1979 and 1992. The second sample consists of the 24 industrialized countries for
which we have complete data between 1960 and 1988. We test the relation both
for the entire period and also for three sub-periods (1960}1969, 1970}1979 and
1980}1988) that we pool together. The third sample consists of the 72 nonindustrialized and non-oil-producing countries for which again we use data
both for the entire period and the three sub-periods already mentioned. The
data for the European regions are taken from Eurostat and the data for the 96
countries are from the World Bank and Barro and Lee (1993). As in Easterly,
1994, the growth rates are computed by running a least-squares regression of the
logarithm of per capita GDP on time.3 The annual standard deviation of growth
rates (SDGW) are taken as the measure of the amplitude of the business cycle.
Alternatively, we could have chosen the variance of the growth rate. The results
are very similar but in most cases produced a slightly better "t when using the
standard deviation. In addition, because the volatility of employment is a key
element driving the theoretical results, we also use the standard deviation of the
unemployment rate as an alternative measure of the amplitude of the business
cycle when using data for the European regions and the industrialized countries.
The list of countries and European regions and their respective growth rates and
standard deviation of the growth rates are given in Tables 1 and 2.
2 See the evidence on both of these aspects in the studies by Searle (1945), Rapping (1965) and
Bahk and Gort (1993).
3 Watson (1992) shows that the least-squares growth rate is more robust to di!erences in the serial
correlation properties of the data than the geometric rate of growth.
P. Martin, C.A. Rogers / European Economic Review 44 (2000) 359}381
363
Table 1
Per capita growth rates and standard deviations of the growth rates for 97 countries
Countries
Least-squares growth
rates (1960}1988)
SDGW
Africa
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
Algeria
Chad
Egypt
Morocco
Botswana
Cameroon
Central Africa
Congo
Gabon
Gambia (The)
Ghana
Cote d'Ivoire
Kenya
Lesotho
Liberia
Madagascar
Malawi
Mali
Mauritania
Mauritius
Mozambique
Niger
Nigeria
Rwanda
Senegal
Sierra Leone
Somalia
Sudan
Swaziland
Tanzania
Togo
Tunisia
Uganda
Zambia
Zimbabwe
4.23
!2.00
5.15
0.87
8.06
3.77
0.56
2.07
4.88
3.86
!0.77
1.75
1.11
5.60
!0.50
!1.64
16.24
1.23
0.91
0.69
!2.99
0.73
!0.60
2.90
!0.71
0.41
1.28
!0.17
0.24
2.52
2.75
0.81
0.86
!1.81
2.88
11.67
8.16
6.21
4.55
7.92
5.45
3.88
6.89
16.13
9.34
4.82
5.29
5.83
8.93
7.27
3.46
5.28
5.14
7.88
5.62
7.73
7.82
8.34
9.31
4.35
6.25
13.36
7.21
8.73
5.30
6.07
3.44
14.15
6.72
6.23
Latin America
36
37
38
39
40
41
42
43
44
Barbados
Costa Rica
Dominican Rep.
El Salvador
Guatemala
Haiti
Honduras
Jamaica
Mexico
0.98
0.53
1.53
!0.04
1.83
0.77
1.67
0.19
1.73
5.39
3.58
6.61
4.95
2.86
4.25
3.73
5.72
4.16
364
P. Martin, C.A. Rogers / European Economic Review 44 (2000) 359}381
Table 1 (Continued)
Countries
Least-squares growth
rates (1960}1988)
SDGW
!0.46
3.12
1.70
1.09
1.40
3.29
0.09
2.20
3.16
2.83
0.71
0.93
1.54
12.44
3.83
9.15
4.26
4.57
8.86
5.95
2.64
4.97
4.78
5.72
5.04
6.76
Latin America
45
46
47
48
49
50
51
52
53
54
55
56
57
Nicaragua
Panama
Trinidad
Argentina
Bolivia
Brazil
Chile
Colombia
Ecuador
Paraguay
Peru
Uruguay
Venezuela
Asia
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
Myanamar
India
Israel
Jordan
Pakistan
Syria
Korea
Malaysia
Philippines
Singapore
Taiwan
Thailand
Fiji
Cyprus
Malta
2.19
1.16
2.74
2.98
1.86
4.41
8.50
3.41
2.19
4.77
5.60
3.66
1.39
3.79
4.84
5.60
3.78
4.54
8.06
3.68
10.38
4.55
4.72
3.91
4.34
2.99
3.15
5.42
9.49
4.61
Western
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
Austria
Belgium
Denmark
Finland
France
Germany
Greece
Iceland
Ireland
Italy
Luxembourg
Netherlands
Norway
Portugal
Spain
Sweden
3.27
2.38
1.70
2.99
2.70
2.42
4.23
2.20
2.45
3.52
2.19
2.04
2.63
3.34
3.09
1.64
2.41
2.59
2.80
2.89
2.12
2.58
4.07
4.06
3.21
2.84
3.43
2.41
1.62
3.94
4.03
1.78
P. Martin, C.A. Rogers / European Economic Review 44 (2000) 359}381
365
Table 1 (Continued)
Western
89
90
91
92
93
94
95
96
Countries
Least-squares growth
rates (1960}1988)
SDGW
Switzerland
Turkey
United Kingdom
Canada
U.S.A.
Japan
Australia
New Zealand
1.43
3.45
1.85
1.36
1.11
4.99
1.58
0.37
2.93
3.72
2.17
2.98
2.54
3.30
2.59
3.39
The other variables are the usual control variables that are used in growth
regressions (see Barro, 1991; Levine and Renelt, 1992; Barro and Sala-i-Martin,
1995). For all samples, the regression contains the initial GDP of the period
(GDPI). The reason is twofold. First, the literature on conditional convergence
has shown that initial income is a determinant of growth. Second, if transitional
dynamics exist and resemble those of the Solow model they would give rise to
a convergence result4 in the sense of a decline over time in per capita growth
rates. This creates a bias in the relation between growth rates and the standard
deviation of growth rates. Suppose we look at two countries identical in all
dimensions (in particular, in the underlying temporary shocks that hit the
economy) except in their position with respect to the (common) steady state
(de"ned as a situation where its growth rate is constant). The country further
away from the steady state will have a higher growth rate because of the
transitional dynamics. Its growth rate will also be decreasing faster than the
country close to the steady state so that mechanically the variance of its growth
rate will also be higher. This implies that the presence of transitional dynamics
in itself creates a positive relation between growth rates and the standard
deviation of the growth rates. It will therefore be especially important in our
growth regressions to account for the transitional dynamics so as to avert
a positive bias on the coe$cient of the standard deviation of growth.
In the cross-country regressions, we also have the average investment share in
GDP (INV), the average share of government expenditures in GDP (SGOV),
and the primary schooling (PRIM) from Barro (1991) which refers to the initial
human capital. For European regions, secondary education (SEC) is also
4 Even though we need to take into account of transitional dynamics in the empirical tests
performed in the next section, we do not conduct &convergence' tests so that our tests are not marred
by Galton's fallacy as described by Quah (1993).
16
17
18
19
20
21
14
15
13
1
2
3
4
5
6
7
8
9
10
11
12
Deutschland
Denmark
Belgium
Countries
Baden Wuerttemberg
Bayern
Berlin
Bremen
Hamburg
Hessen
Sjvlland-Lolland
Falster-Bornholm
Fyn
Jylland
Vlaams Gewest
Region Vallone
Bruxelles
Regions
Antwerpen
Brabant
Hainaut
Liege
Limburg
Luxembourg
Namur
Oost-Vlaandern
West-Vlaanderen
Subregions
Table 2
Per capita growth rates and standard deviations of the growth rates for 90 European regions
4.08
4.58
2.03
4.00
4.10
4.70
0.61
1.32
1.22
1.67
1.05
1.64
1.37
1.32
0.93
1.06
2.30
1.86
1.19
1.91
1.54
Least-square
SDGWgrowth rates
(1978}1992)(%)
2.24
2.44
6.79
2.29
4.33
2.84
3.73
3.77
3.18
2.58
2.41
3.16
3.08
2.66
2.56
2.61
3.40
2.15
2.75
2.96
2.86
SDGW
(%)
366
P. Martin, C.A. Rogers / European Economic Review 44 (2000) 359}381
46
47
44
45
41
42
43
38
39
40
35
36
37
28
29
30
31
32
33
34
27
22
23
24
25
26
France
Mediterranee
Centre Est
Sud Ouest
Ouest
Nord Pas de Calais
Est
Ile de France
Bassin Parisien
Niedersachsen
Nordrhein Westfalen
Rheinland Pfalz
Saarland
Schleswig Holstein
Languedoc Roussillon
Provence Alpes Cote d'Azur
Rhones Alpes
Auvergne
Aquitaine
Midi Pyrenees
Limousin
Pays de la Loire
Bretagne
Poitou Charentes
Lorraine
Alsace
Franche Comte
Champage Ardenne
Picardie
Haute Normandie
Centre
Basse Normandie
Bourgogne
0.18
!0.32
0.20
0.20
0.15
0.70
0.34
0.16
0.18
0.19
!0.52
0.24
!0.08
0.00
!0.65
!0.29
0.23
0.21
0.19
!0.32
0.73
3.96
3.49
3.62
1.68
3.71
3.79
3.52
3.13
2.96
3.75
3.45
3.35
3.20
3.29
3.18
3.88
3.37
2.56
3.59
3.35
4.61
3.39
2.93
3.17
3.39
2.68
2.26
2.14
2.74
1.81
2.48
P. Martin, C.A. Rogers / European Economic Review 44 (2000) 359}381
367
64
65
66
62
63
53
54
55
56
57
58
59
60
61
49
50
51
52
48
Italia
Ireland
Countries
Table 2 (Continued)
Sud
Lazio
Campania
Abruzzi Molise
Emilia Romagna
Lombardia
Nord Est
Nord Ovest
Regions
Puglia
Basilicata
Calabria
Abruzzi
Molise
Toscana
Umbria
Marche
Trentino alto adige
Veneto
Friuli Venezia Giulia
Piemonte
Valle d'Aosta
Liguria
Subregions
0.51
!0.65
0.11
0.79
0.83
0.52
0.90
0.75
!0.09
!0.14
!0.15
0.32
0.86
0.27
!0.15
!0.44
!0.22
0.12
1.06
Least-square
SDGWgrowth rates
(1978}1992)(%)
5.61
6.10
6.69
5.84
6.08
5.93
6.15
4.93
5.87
4.76
5.92
7.24
4.82
4.76
3.90
4.04
3.75
4.62
5.09
SDGW
(%)
368
P. Martin, C.A. Rogers / European Economic Review 44 (2000) 359}381
80
81
82
83
84
85
86
87
88
89
90
78
79
74
75
76
77
70
71
72
73
69
67
68
United Kingdom
Nederland
Luxembourg
North
Yorkshire
East Midlands
East Anglia
South East
South West
West Midlands
North West
Wales
Scotland
Northern Ireland
Zuid
Oost
West
Sicilia
Sardegna
Noord Brabant
Limburg
Utrecht
Noord Holland
Zuid Holland
Zeeland
Groningen
Friesland
Drenthe
0.24
0.45
0.64
1.15
0.80
0.91
0.43
0.11
0.69
0.52
0.60
3.54
3.86
3.06
2.90
2.77
3.68
!0.88
3.08
2.22
2.86
3.48
0.03
0.07
9.48
8.47
9.42
9.34
9.47
9.48
8.16
8.94
8.78
9.09
8.78
3.12
2.66
4.49
2.34
3.30
4.09
10.75
3.76
4.91
2.75
6.11
4.86
4.19
P. Martin, C.A. Rogers / European Economic Review 44 (2000) 359}381
369
370
P. Martin, C.A. Rogers / European Economic Review 44 (2000) 359}381
included. For the developing countries, we include variables for political instability. These are the average number of revolutions and coups (REVC) over the
period and the number of political assassinations per million inhabitants
(ASSP). The shocks we consider in our theoretical model could be thought of, as
in Barro and Sala-i-Martin, as supply shocks and equivalent to temporary
declines in the security of property rights. We include political variables in the
regressions because we want to distinguish between economic and political
instability.5
3. Empirical evidence for European regions and industrialized countries
We now report results of the tests of the link between the amplitude of the
business cycle and the growth rate. Table 3 shows the results for the regressions
for European regions between 1979 and 1992.6 We also include country dummies. The "rst and second column report regressions without and with sectoral
shares: the share of agriculture in production (AGRI) and the share of industry
in production (IND). In both cases, SDGW has the right sign and is very
signi"cant. We have checked that these results are not due to the presence of
outliers.
Table 4 reports the results for the sample of 24 industrialized countries. The
"rst two columns have the results for pooled data set whereas the next two
columns report the growth rate between 1960 and 1988 as the dependent
variable. In all speci"cations, the coe$cient on SDGW is negative and signi"cant at the 5% level or less. We have checked that the coe$cient on SDGW
remains negative and signi"cant even when, in the pooled data set, we control
for "xed time e!ects. The coe$cient becomes more negative and signi"cant
when the investment ratio (INV) is included in the regression7 (see columns
2 and 4). This shows that the impact of short-term instability does not go
through an e!ect on investment which could be a natural alternative explanation to our empirical "ndings. This con"rms the results of Ramey and Ramey
(1995). In their cross-country regressions, they "nd no relation between the
investment share in GDP and the standard deviation of growth rates.
A further implication of our theoretical framework is that employment
instability has a negative impact on long-term growth. To test this, we used, as
5 On the impact of political instability on growth see Alesina et al. (1992).
6 The standard errors in all the regressions reported in the paper are computed using White's
heteroskedacity-robust procedure.
7 Note that the coe$cient on the investment ratio has the wrong sign and is not signi"cant. In the
di!erent regressions we have performed the coe$cient on the investment ratio in the industrialized
countries is signi"cant and positive only when SGOV and PRIM are excluded from the regression.
P. Martin, C.A. Rogers / European Economic Review 44 (2000) 359}381
371
an alternative measure of the amplitude of the business cycle, the standard
deviation of the unemployment rate (SDUN). The two columns M3N and M4N in
Table 3 report for European regions the same regressions as in M1N and M2N
except that SDGW is replaced by SDUN. Because of lack of data on unemployment, the growth rate for these regressions is computed on the period
1983}1992. The coe$cient on SDUN is negative and signi"cant at the 5% level.
We also tried as a measure of the e!ect of the business cycle on employment the
standard deviation of the unemployment rate of workers under 25 (SDUN25).
We thought that young workers should have the steepest learning curve so that
the e!ect of employment instability on growth should be stronger in this case.
The results are given in columns M5N and M6N. The coe$cient has the right sign in
both regressions but is not very signi"cant whether sectoral shares are included
or not in the regression.
We also tested this implication of the model for OECD countries. In columns
M5N}M8N, the measure of the amplitude of the business cycle, the standard deviation
of growth, SDGW, is replaced by the standard deviation of the unemployment
rate, SDUN. Again, as for the European regions the coe$cient has the expected
sign but is signi"cant at the 5% level only in the pooled data set.
Another implication of the channel we identify in the theoretical model is that
average e!ective employment should have a positive impact on human capital
accumulation and growth. Average unemployment rates are imperfect measures
of this e!ect but we nonetheless include them in our growth regressions to see if
they have a negative impact on average growth. In all samples, the coe$cient
(not reported) for the average unemployment rate is negative and very signi"cant. There is, however, an obvious problem of reverse causality as average
growth rates are also a determinant of unemployment. To tackle this problem,
we used two-stage least-squares regressions.8 We report these regressions in
columns M7N and M8N of Table 3 and columns M9N and M10N of Table 4. In
regression M7N of Table 3, where the instability measure is not included, the
average unemployment rate (AVUN) has a strong negative and signi"cant
impact on growth in European regions. For industrialized countries (see regression M9N in Table 4) this is also the case. When we include both the instability
measure and the average unemployment rate in our regressions, the later
becomes insigni"cant in the European regions sample and even though the
coe$cient on instability remains verysigni"cant, it decreases in both samples.
This suggests that short-term instability and unemployment a!ect long-term
growth in similar ways.
8 For European regions, the instruments for the unemployment are a constant, the share of
industry in regional output, the initial GDP and SDGW. For industrialized countries (pooled data),
the instruments are a constant, the initial unemployment rate of the period, a dummy for the "rst
decade and the initial GDP. These instruments are unlikely to be caused by average growth rates.
SDUN25
SDUN
SDGW
IND
AGRI
SEC
PRIM
GDPI
Constant
No of obs.
!0.4255
[0.00]
90
0.0278
[0.31]
!0.00023
[0.70]
0.65315
[0.97]
!0.0199
[0.14]
M1N
87
0.0479
[0.04]
!0.00019
[0.78]
!0.0115
[0.58]
!0.0259
[0.07]
!0.0197
[0.41]
!0.0117
[0.11]
!0.4169
[0.00]
M2N
!0.00218
[0.33]
90
0.00493
[0.84]
!0.991
[0.205]
0.0158
[0.43]
!0.00218
[0.51]
M3N
!0.00212
[0.04]
87
0.0292
[0.22]
!1.02]10~6
[0.22]
0.00279
[0.90]
!0.0182
[0.17]
!0.0321
[0.25]
!0.0152
[0.07]
M4N
M6N
M7N
M8N
!0.000635
[0.19]
!0.000821
[0.075]
TSLS
TSLS
87
90
90
90
0.023
0.000358
0.0312
0.0203
[0.34]
[0.99]
[0.10]
[0.20]
!9.94]10~7 !9.28]10~7 !2.13]10~6 !9.17]10~7
[0.25]
[0.25]
[0.00]
[0.07]
0.00448
0.0136
0.025
0.0184
[0.85]
[0.49]
[0.20]
[0.25]
!0.0151
!0.00377
!0.0189
!0.0162
[0.27]
[0.82]
[0.10]
[0.12]
!0.005
[0.13]
!0.0135
[0.09]
!0.274
[0.005]
M5N
Table 3
European regions dependent variable: growth rate 1979}1972 (regression M1N and M2N), growth rate 1983}1992 (regressions M3N to M8N)
372
P. Martin, C.A. Rogers / European Economic Review 44 (2000) 359}381
0.84
!0.0208
[0.005]
0.0401
[0.00]
0.00355
[0.62]
0.00628
[0.37]
0.0335
[0.00]
0.0191
[0.01]
0.85
0.0206
[0.00]
0.0410
[0.00]
0.00374
[0.57]
0.00701
[0.31]
0.0325
[0.00]
0.0192
[0.00]
Note: The signi"cance level is in brackets.
R2
Uki
Ned
Ita
Fra
Ger
Bedlux
AVUN
0.81
0.0155
[0.00]
0.0405
[0.00]
0.00131
[0.78]
0.00342
[0.45]
0.0293
[0.00]
0.0065
[0.15]
0.81
0.0152
[0.00]
0.04
[0.00]
0.00191
[0.66]
0.00444
[0.31]
0.0287
[0.00]
0.00706
[0.07]
0.81
0.0153
[0.00]
0.042
[0.00]
0.00359
[0.49]
0.00615
[0.23]
0.0293
[0.00]
0.00735
[0.12]
0.80
0.0161
[0.00]
0.0413
[0.00]
0.00278
[0.62]
0.00449
[0.39]
0.0293
[0.00]
0.00673
[0.20]
0.84
!0.230
[0.00]
0.0137
[0.00]
0.0399
[0.00]
0.00189
[0.30]
0.0061
[0.09]
0.0308
[0.00]
0.0082
[0.04]
0.86
!0.084
[0.20]
0.0165
[0.00]
0.0393
[0.00]
0.00220
[0.30]
0.0057
[0.09]
0.0319
[0.00]
0.0136
[0.00]
P. Martin, C.A. Rogers / European Economic Review 44 (2000) 359}381
373
0.53
}
0.296
[0.00]
!0.0282
[0.00]
!0.074
[0.041]
0.0002
[0.37]
!0.3768
[0.00]
}
70
0.60
!0.0266
[0.25]
0.307
[0.00]
!0.0287
[0.00]
0.0898
[0.02]
0.0003
[0.20]
!0.4630
[0.00]
}
68
M2N
0.67
}
0.210
[0.00]
!0.019
[0.00]
!0.0404
[0.22]
0.00015
[0.36]
!0.5334
[0.03]
}
24
M3N
Note: The signi"cance level is in brackets.
R2
AVUN
INV
SDUN
SDGW
PRIM
SGOV
GDPI
No.
of obs.
Const.
M1N
0.73
!0.0322
[0.31]
0.244
[0.00]
!0.023
[0.00]
!0.0587
[0.10]
0.0003
[0.01]
!0.7231
[0.00]
}
23
M4N
0.59
!0.0069
[0.00]
}
0.265
[0.00]
!0.026
[0.00]
!0.0121
[0.76]
0.00031
[0.28]
}
60
M5N
0.6
!0.0071
[0.00]
!0.0405
[0.24]
0.284
[0.00]
!0.0264
[0.00]
!0.039
[0.45]
0.00019
[0.64]
}
57
M6N
0.74
!0.0008
[0.12]
}
0.1845
[0.00]
!0.0192
[0.00]
!0.0437
[0.11]
0.00028
[0.00]
}
24
M7N
0.73
!0.00055
[0.29]
!0.001
[0.97]
0.1826
[0.00]
!0.0199
[0.00]
!0.0415
[0.13]
0.0004
[0.00]
}
23
M8N
Table 4
Industrialized countries dependent variable: growth rates 1960}1969; 1970}1979; 1980}1988 and growth rate 1960}1988
!0.004
[0.10]
!0.001
[0.02]
0.58
}
68
TSLS
0.278
[0.00]
!0.026
[0.00]
!0.0518
[0.10]
0.0002
[0.25]
}
M9N
!0.036
[0.15]
!0.001
[0.025]
0.62
68
TSLS
0.310
[0.00]
!0.028
[0.00]
!0.0781
[0.02]
0.0002
[0.25]
!0.413
[0.01]
}
M10N
374
P. Martin, C.A. Rogers / European Economic Review 44 (2000) 359}381
P. Martin, C.A. Rogers / European Economic Review 44 (2000) 359}381
375
In both the European regions sample and the industrialized countries sample,
the impact of short-term instability is quantitatively important. Lowering the
instability measure (SDGW) across European regions by one standard deviation is associated with an increase in the average growth rate of around half
a percentage point of annual per capita growth. For industrialized countries this
number is around 0.4 percentage point. Reducing the instability of the unemployment rate by one standard deviation has even a larger impact as it increases
the average growth rate by 0.8}0.9 percentage point of annual per capita growth
in European regions and around 0.6 percentage point for industrialized
countries.
4. Empirical evidence for developing countries
Table 5 reports results for non-industrialized countries. The "rst two columns
report the regression results for the three decades pooled together without and
with the investment ratio. The next two columns report the same regressions for
the period 1960}1988. Except in one regression, the coe$cient on the instability
measure is positive and insigni"cant. We "rst note that these results contradict
those of Ramey and Ramey whereas our results for the developed countries do
not.9 We conclude that the negative relation between short-term instability and
growth is robust only for the developed countries.10
There are several possible reasons for the di!erence in the relation between
growth and the standard deviation of growth in developed countries (European
regions and industrialized countries) and in developing countries.
(i) The lack of relation between the two variables in the developing countries
could be due to measurement error. To see whether our results are due to
measurement errors, we reestimated our regression for the non-industrialized
countries sample using instrumental variables. The instruments for the standard
deviation of the growth rate are the standard deviation of the growth rate of the
preceding decade, the initial in#ation rate of the decade, the initial GDP per
capita level and the number of revolutions and coups. This implied that we
could not use the observations of the "rst decade (1950}1960) when we used
these instruments in the regression. The results did not change much. For this
9 This may be due to di!erent factors: our growth rates are calculated over the period 1960}1988
rather than 1962}1985, using least square growth rates rather than geometric growth rates and with
a slightly larger set of developing countries (72 in our sample, 68 in their sample, the di!erence being
mostly in African countries). A natural candidate explanation would also be that political instability
variables are included in our regressions but not theirs. However, we have checked that their
exclusion does not alter the results.
10 We have also performed the same type of regressions with all countries, both developed and
developing. The coe$cient on the instability measure is positive and insigni"cant.
376
P. Martin, C.A. Rogers / European Economic Review 44 (2000) 359}381
Table 5
Developing countires dependent variable: growth rates 1960}1969; 1970}1979; 1980}1988 and
growth rate 1960}1988
No of obs.
Const.
GDPI
SGOV
PRIM
REVC
ASSP
SDGW
M1N
M2N
M3N
M4N
M5N
224
0.1643
[0.0000]
!0.0154
[0.0001]
!0.0763
[0.0022]
0.0001
[0.2822]
!0.0117
[0.0543]
!4.7523
[0.6585]
0.097
[0.2112]
205
0.1706
[0.00000]
!0.0204
[0.0000]
!0.0683
[0.0043]
0.0001
[0.681]
!0.0012
[0.8142]
!6.835
[0.53078]
0.1516
[0.0494]
0.148
[0.0000]
72
0.01763
[0.0012]
!0.0198
[0.0012]
!0.0268
[0.325]
0.0004
[0.0109]
!0.021
[0.0436]
7.2319
[0.07061]
0.0967
[0.2902]
67
0.168
[0.0003]
!0.0218
[0.00058]
!0.0331
[0.2168]
0.0003
[0.0268]
!0.0134
[0.1952]
8.4874
[0.6392]
0.03962
[0.7021]
0.1207
[0.0018]
70
0.12
[0.006]
!0.0156
[0.0123]
!0.0507
[0.1774]
0.0001
[0.2489]
!0.0031
[0.5989]
!14.3062
[0.5987]
0.0807
[0.4115]
0.1268
[0.0016]
INV
DAGL
SUBAFRICA
LAAM
R2
!0.0244
[0.0000]
!0.0134
[0.0101]
!0.0253
[0.0000]
!0.0044
[0.4116]
!0.0186
[0.0263]
!0.0143
[0.0143]
!0.01315
[0.1366]
!0.0069
[0.2855]
!0.0017
[0.0034]
!0.0184
[0.0209]
!0.0086
[0.1797]
0.17
0.26
0.29
0.30
0.28
Note: The signi"cance level is in brackets.
limited sample, the coe$cient became negative but only signi"cant at the 30%
level so that this does not make a very convincing case that measurement error
is at the origin of our results for these countries.
(ii) Transitional dynamics should create a positive mechanical bias between
growth and the standard deviation of growth. We have checked that there is
indeed a negative correlation between the initial level of GDP and the standard
deviation of growth rates. This bias will be more important the more important
the role of transitional dynamics in explaining growth. It should therefore be
stronger for developing countries. We have accounted for transitional dynamics
through the insertion in the regression of the initial level of GDP per capita and
for proxies for the steady-state levels of growth rates. Our results may re#ect
that not all the e!ect of transitional dynamics has been eliminated.
P. Martin, C.A. Rogers / European Economic Review 44 (2000) 359}381
377
A related argument is that some of the high growth countries in our sample of
developing countries may have been hit by important permanent shocks. This
will be the case for countries where industrialization has played an important
role in growth. This would induce a positive bias between the standard deviation
of growth and the growth rate. We have tested this hypothesis by adding in the
regression the di!erence in the initial and "nal shares of agriculture in GDP
(DAGL) in column M5N of Table 5. Unfortunately, we lose a lot of observations
because we do not have this data for all the countries of our sample. Introducing
this variable (which non-surprisingly is very signi"cant and negative) reduces the
positive coe$cient on the SDGW and also its signi"cance compared to regression M2N. This may constitute weak evidence that developing countries have been
hit by important permanent shocks which obscure the relation between growth
and the standard deviation of growth.
(iii) Our theoretical prediction should only hold in countries for which
growth is driven by learning by doing. In developing countries where the
economy is dominated by traditional economic activities such as agriculture, the
learning curve can be thought as almost #at. In this case, our model predicts no
relation between growth and the standard deviation of growth. Our empirical
results are also consistent with Young's (1993) theoretical "nding that growth
will be driven by learning by doing only at relatively high levels of development
that is when the market size is large relative to the cost of invention.
(iv) If growth is driven by learning by doing then the level of employment is
key to our results. In particular, it is important that employment is procyclical.
It is quite likely that in developing countries employment responds di!erently to
shocks than in developed countries. In particular, contrarily to industrialized
countries, we have no stylized facts about the procyclical nature of employment
in developing countries. More generally, our results may simply re#ect the fact
that the business cycle is an industrialized countries phenomenon.
It is di$cult at this stage to discriminate between these di!erent explanations
which may all play a role. Our theoretical model coupled with Young's (1993)
"nding that the learning by doing model should apply only at high development
stages would predict that (iii) is enough to explain the empirical di!erences
between developed and developing countries.
5. Conclusion
We have studied the impact of learning by doing on the relation between
growth and short-term instability at the aggregate level. For developed countries, our empirical results show a signi"cant and quantitatively important
negative relation between growth and the amplitude of the business cycle
whether measured by the standard deviation of growth or the standard deviation of unemployment. We have seen that this relation does not work through
378
P. Martin, C.A. Rogers / European Economic Review 44 (2000) 359}381
an impact of short-term instability on the level of investment in industrialized
countries which could be a natural explanation of the empirical results. Furthermore, if investment played an important role in this relation it would be di$cult
to explain the di!erence in results between developed and developing countries.
If short-term instability is detrimental to investment it should be so in both sets
of countries. Instead, our results are consistent with a model where human
capital accumulation is increasing and concave in production and Young's
(1993) "nding that growth is driven by learning by doing only at high levels of
development.
Our conclusions have interesting policy implications. They give a clear and
novel rationale in favour of short-term stabilization policies, be they monetary
or "scal policies. Two recent papers of ours (Martin and Rogers, 1995, 1997)
study in similar models under which conditions a "scal counter-cyclical policy
can improve growth prospects. An interesting characteristic of this policy
implication is that it does not come out of a Keynesian-type model. In particular, markets clear, and the origin of the shocks, supply or demand, does not
matter for the results.
Acknowledgements
We gratefully acknowledge the "nancial support of the Swiss Fonds National
de la Recherche Scienti"que. We thank an anonymous referee, Gilles Dowek,
Hans Genberg, Pierre-Yves Geo!ard, Claire Lefevbre, Danny Quah and Pierre
Villa as well as seminar participants at CEPII for helpful comments and Marco
Fugazza for excellent research assistance.
Appendix: A model of growth with learning by doing with stochastic shocks
A representative household chooses consumption c and labour l over an
t
t
in"nite horizon to maximize the expected utility function:
A B
t ac1~1@p#(1!a)[h (1!l )]1~1@p
=
1
t
t
t
E ;"E +
.
(A.1)
0
0
1!1/p
1#o
t/0
There is only one factor of production, labour, and goods and factor markets are
perfectly competitive so that the income of the household is given by the level of
production and the household faces the budget constraints:
w l "c , t"0, 12,
(A.2)
tt
t
where w is the wage rate. There is no saving. The maximum amount of labour
t
the consumer can supply in any date t equals unity: 04 l 41.
t
P. Martin, C.A. Rogers / European Economic Review 44 (2000) 359}381
379
Output is produced using e!ective labour, which equals hours worked (l)
times human capital (h): y "/ l h . The parameter / is an exogenous stochast
t t t
t
tic productivity disturbance. The wage rate per unit of time worked is: w "/ h .
t
t t
Human capital accumulates via learning by doing:
h
"(1!d#bl ) h ,
(A.3)
t`1
t t
where d is the rate of depreciation of human capital, b is a parameter that tells
how much is learned through experience. According to Eq. (A.3), returns to
learning are not bounded and the rate of learning depends upon the #ow of
e!ective labour. All bene"ts of human capital accumulation are &external' that is,
that individual workers do not internalize the fact that experience a!ects future
wages in the economy.
The business cycle is characterized by a two-state stationary Markov process.
In good states the productivity level, / , takes the value G. In bad states, it takes
t
the value B(G. We thus assume that all disturbances are transitory. The
two-state Markov chain is de"ned by the following probabilities:
PrM/ "G; / "GN"P , PrM/ "B; / "BN"P ,
t`1
t
G
t`1
t
B
PrM/ "G; / "BN"1!P , PrM/ "B; / "GN"1!P .
t`1
t
B
t`1
t
G
The long-term expected value of productivity is therefore:
G(1!P )#B(1!P )
B
G,
E/ "
t
1!j
(A.4)
where j"P #P !1. We also assume that agents observe the state of the
G
B
economy at the beginning of the period.
The optimal private choice of labour supply is then derived from the "rstorder conditions of the maximization problem of the consumer:
1
l"
,
(A.5)
t 1#k/1~p
t
where k"[(1!a)/a]p.11 The labour supply is increasing in the cyclical component of the wage rate } the productivity level } and will be procyclical if p is
more than 1. There are two levels of labour supply: l and l .
G
B
The expected growth rate of output between date 0 and date ¹ is E (y /y ).
0 T 0
The average annual growth rate between those dates then equals
C A BD
y
T
E
0 y
0
1@T
.
11 For utility to be "nite, a su$cient condition is that l ((o#d)/b for all t.
t
(A.6)
380
P. Martin, C.A. Rogers / European Economic Review 44 (2000) 359}381
If the economy is the same states in dates 0 and ¹, then as ¹PR, the
long-run expected growth rate can be rewritten as
C A BD
1@T
C A
BD
C A BD
1@T
1@T
/ l h
h
TT T
T
" E
" E
0 / l h
0 h
00 0
0
(A.7)
"(1!d#bl )(1~PB)@(1~j) (1!d#bl )(1~PG)@(1~j).
B
G
In Eq. (A.7), (1!P )/(1!j) is the expected long-run proportion of good
B
states (when the growth rate equals (1!d#b1 )) and (1!P )/(1!j) is the
B
G
expected long-run proportion of bad states (when the growth rate equals
(1!d#bl )). To determine the e!ect of the amplitude of the business cycle on
G
the expected long-run growth rate, we di!erentiate Eq. (A.7) with respect to the
productivity levels B and G. For the exercise to be meaningful, the change in the
amplitude of the business cycle must leave the long-run expected level of
productivity / unchanged. This requires that the changes in B and G satisfy the
t
condition
y
T
E
0 y
0
1!P
B dG.
dB"!
1!P
G
In this case, we obtain
C A BD
C A BD
d E
y
T
0 y
0
(A.8)
1@T
C
D
1@T 1!P Ll
y
1
1
Ll
T
G
B
! B
.
(A.9)
"b E
0 y
1!j LG 1!d#bl
LB 1!d#bl
G
B
0
As long as the labour supply is increasing and concave in the productivity level,
this expression is negative so that the expected long-run growth rate decreases
with the amplitude of the business cycle.
References
Aghion, P., Saint-Paul, G., 1991. On the virtues of bad times. Working Paper no. 578. CEPR,
London.
Aghion, P., Saint-Paul, G., 1993. Uncovering some causal relationships between productivity
growth and the structure of economic #uctuations: A tentative survey. Working Paper d 4603.
NBER, Cambridge, MA.
Alesina, A., Sule, O., Nouriel, R., Philip, S. 1992. Political instability and economic growth. Working
Paper 4173, September, NBER, Cambridge, MA.
Bahk, B.H., Gort, M., 1993. Decomposing learning by doing in new plants. Journal of Political
Economy 101, 4.
Barro, R., 1991. Economic growth in a cross-section of countries. Quarterly Journal of Economics
106 (2), 407}444.
P. Martin, C.A. Rogers / European Economic Review 44 (2000) 359}381
381
Barro, R., Lee J.-W., 1993. International comparisons of educational attainment. Journal of Monetary Economics, December 32 (3), 363}394.
Barro, R., Sala-I-Martin, X., 1995. Economic Growth. McGraw-Hill, New York.
Bean, C., 1990. Endogenous growth and the pro-cyclical behaviour of productivity. European
Economic Review 34, 355}363.
Bernanke, B.S., 1983. Irreversibility, uncertainty, and cyclical investment. Quarterly Journal of
Economics 98 (1), 85}106.
Danthine, J.P., Donaldson, J., 1993. Methodological and empirical issues in real business cycle
theory. European Economic Review 33, 31}50.
Easterly, W., Rebelo, S., 1994. Fiscal policy and economic growth: An empirical investigation,
Discussion Paper Series No. 885. CEPR, London.
Levine, R., Renelt, D., 1992. A sensitivity analysis of cross-country growth regressions. American
Economic Review 82 (4), 942}963.
Lucas, R., 1993. Making a miracle, Econometrica 61, 2.
Martin, P., Rogers, C.A., 1997. Stabilization policy, learning by doing, and economic growth. Oxford
Economic Papers 49 (2), 152}166.
Martin, P., Rogers, C.A., 1995. Optimal stabilization policy, in the presence of learning by doing,
Discussion Paper No. 1129. CEPR, London.
Pindyck, R.S., 1991. Irreversibility, uncertainty, and investment. Journal of Economic Literature 29
(3), 1110}1148.
Quah, D., 1993. Galton's fallacy and tests of the convergence hypothesis. The Scandinavian Journal
of Economics 95 (4), 427}443.
Ramey, G., Ramey, V.A., 1991. Technology commitment and the cost of economic #uctuations,
Working Paper No. 3755, June. NBER, Cambridge, MA.
Ramey, G., Ramey, V.A., 1995. Cross-country evidence on the link between volatility and growth.
American Economic Review 85 (5), 1138}1159.
Rapping, L., 1965. Learning and World War II production functions. Review of Economics and
Statistics 47, 81}86.
Saint-Paul, G., 1993. Productivity growth and the structure of the business cycle. European
Economic Review 37, 861}890.
Searle, A., 1945. Productivity changes in selected wartime shipbuilding programs, Monthly Labor
Review 61.
Stadler, G.W., 1990. Business cycle models with endogenous technology. American Economic
Review 80 (4), 763}778.
Watson, M., 1992, A note on estimating deterministic trends in the presence of serially correlated
errors, Working paper, Northwestern University, Evanston, IL.
Young, A., 1993. Invention and bounded learning by doing. Journal of Political Economy 101 (3),
443}472.