A. The System

This paper estimates the following system of equations simultaneously

where i, j indexes country pairs, ρ denotes bilateral business cycles correlation, T is bilateral trade intensity and S is a specialization index capturing how different the sectoral allocations of resources are between countries i and j. Business cycles correlations, bilateral trade and specialization are all endogenous variables, while I₁, I₂ and I₃ contain the vectors of their exogenous determinants, respectively. Identification of the system requires differences between at least I₂ and I₃. Fortunately, a substantial literature is there to provide guidance on what variables to include in the I_i vectors. I next turn to this question. To clarify matters, Figure 1 summarizes the direct and indirect channels discussed here, and relates them to the relevant literature. The reader might find it useful to refer to the Figure while going through the following review.

The dependent variable in equation (1) is one of the most topical items in the list of optimal currency area criteria.¹⁷ It is therefore of interest in its own right, and indeed its determinants have been the object of intense scrutiny. Frankel and Rose (1998) have for instance focused on α₁, reasoning that if currency unions affect trade and then trade in turn boosts cycle correlations, then currency areas can endogenously become optimal.¹⁸ Imbs (2001) focused on α₂, arguing that measured bilateral trade may partly be a manifestation of differences in the degrees of specialization between the trading countries, which could affect ρ independently.¹⁹ The estimates of α₁ can therefore be affected by inclusion of a specialization term.²⁰ Since specialization can very well be the result of trade intensity, however, accurate identification of α₁ and α₂ requires exogenous determinants for both trade and specialization, which equations (2) and (3) provide. Hence an added advantage to estimating (l)-(2)-(3) simultaneously.

The typically large estimates of α₁ pose a theoretical puzzle. While Baxter (1995) reviews the theories that imply a positive α₁, Canova and Dellas (1993), Schmitt-Grohé (1998) and Kose and Yi (2002) have used various methods to document the inability of existing models to reproduce the magnitude of standard estimates for α₁.²¹ Imbs (2001) proposed to add the potentially omitted variable S to equation (1) and corrected the estimates accordingly. But an equally plausible explanation could be that we are not using the appropriate modelling strategy when attempting to reproduce the observed effects of trade. In particular, α₁ embeds the impact of both inter- and intra-industry trade, two dimensions that the models typically do not share. The simultaneous approach makes it possible to decompose the two effects, as β₁ in equation (2) captures the extent to which bilateral trade can be accounted for by the similarities in the two countries’ economic structures, i.e. an account of intra-industry trade, denoted by an upward dotted arrow in Figure 1. Thus, the total effects of trade in the simultaneous estimation equals α₁. β₁ + α₁. β₂, where the former term captures the importance of intra-industry trade. One sector models should seek to reproduce α₁. β₂ only.²²

This leaves open the question of what additional regressors to include in (1). Since this paper is concerned with the (direct and indirect) effects of financial integration, an important variable is one capturing the extent of impediments to capital flows between each pair of countries. This is done by including a variable capturing capital account restrictions in the international dataset, and an estimate of inter-state risk-sharing in the intranational data.²³ The preferred specification for (1) therefore contends that economic structure, and integration to financial and trade markets jointly explain a sufficient fraction of ρ to rule out omitted variable bias in the estimation of the coefficients α_i.²⁴ There are, however, a few additional explanations, including most prominently convergence in policies. In Section IV, I subject the specification to some sensitivity checks, introducing among others controls for monetary policy. More importantly, I implement the same estimations using information on U.S. states, which are subjected to a unique monetary policy, and find strikingly similar estimates.²⁵,²⁶

The specification of equation (2) is more straightforward, although also an object of debate. The empirical performance of so-called gravity variables in accounting for trade flows is a result going back at least to Tinbergen (1962), which has been used extensively subsequently.²⁷ The set of gravity variables customarily included in I₂ include measures of both countries’ GDP levels, or sometimes of their populations, the geographic distance between their capitals, and binary variables capturing the presence of a common border and linguistic similarities between them. The list is usually argued to contain clearly exogenous variables with high predictive power on trade flows, thus supplying an exceptional instrument set, used for instance to identify causality between trade and growth in Frankel and Romer (1999), or to control for other determinants of trade in Rose (2000), or both in Frankel and Rose (2002). This paper uses similar insights to isolate the exogenous impact of trade on cycles synchronization and specialization.²⁸

The exogenous determinants of specialization, summarized in I₃, are empirically less established.

Two variables do, however, spring to mind. First, access to financial markets will influence specialization patterns, and how similar they are between countries. I thus include in the vector I₃ the measures of capital account restrictions and the extent of risk-sharing already used in I₁. Second, Imbs and Wacziarg (2003) show economies go through two stages of specialization as income per capita grows: they initially diversify but re-specialize once a (relatively high) level of income per capita is reached. This empirical fact suggests two additional components for I₃: GDP per capita levels in both economies, but also, because of the non-monotonicity, the gap between them.

In summary, since the intersection between I₂ and I₃ is empty, the system can be identified through a choice of instruments that is largely warranted by an existing literature. The main contribution of the present exercise is simultaneity. The main assumptions are the exogeneity of financial integration to ρ and S, and that of GDP per capita and relative GDP per capita to S.²⁹ I now turn to a detailed description of the data and measurement of the variables included in the I_i vectors.

B. Data and Measurement

Bilateral correlations in business cycles are computed on the basis of the cyclical component of quarterly GDP, isolated using the Band-Pass Filter introduced in Baxter and King (1999).³⁰ The quarterly data cover the 1980s and 1990s in 24 countries, and come from the International Financial Statistics issued by the IMF.³¹ This gives rise to a cross-section of 276 bilateral correlations.³² A concern with this cross-section is that it builds on twenty years of data, over which the U.S. for instance has experienced two, at most three, business cycles. Can one safely draw inferences on the medium to long-run characteristics of the cross-section of business cycles synchronization on such a reduced sample? This is of course a concern that pervades any such exercise, and in particular all the papers mentioned earlier.³³ Here I propose to assuage somewhat this worry using two alternative data sources. First, I also compute bilateral correlation in business cycles on the basis of yearly data taken from the version 6.1 of the Penn-World Tables, where correlations are computed over the period 1960 to 2000. Second, (annual) series on Gross State Product are taken from the Bureau of Economic Analysis, and run from 1977 to 2001. It is reassuring that similar results obtain in all these datasets.

Bilateral trade intensity is computed in two ways. The first one is standard, used in Frankel and Rose (1998) among others, and writes

where X_i,j,t denotes total merchandise exports from country i to j in year t, M_i,j,t represents imports to i from j, and Y_i denotes nominal GDP in country i. Bilateral trade data are from the IMF’s Direction of Trade Statistics. I use this standard measure for the benchmark case. Clark and van Wincoop (2001) use an alternative measure of trade intensity, independent of size, based on Deardorff (1998) model, which can be constructed as

where $Y_t^w$ is world GDP. T² differs from T¹ in that it depends only on trade barriers, and not on country size. In particular, Deardorff shows that T² equals 1 if preferences are homothetic and there are no trade barriers. I use T² in Section 4, dedicated to a sensitivity analysis.

There is no data available on bilateral trade flows between U.S. states. The main virtue of the gravity model, however, is that it provides a dependable predicting tool. Just as cross-country trade flows are well accounted by the gravity variables, with R² above 0.65, so cross-state trade flows can be simulated with reasonable reliability on the basis of each state’s economic size, the distance separating each state’s main business center and whether they share a common border.³⁴ Thus, I use the estimated coefficients from a cross-country gravity estimation of $T_{i,j}^1$ regressed on distance (between capital cities, in statute miles), the product of each country’s GDP, the product of each country’s population, and binary variables capturing the presence of a common border or a common language. The estimates are used to fit a value for inter-state trade, $\widehat{T}$, which is then used in estimating jointly equations (1) and (3).³⁵

There are no standard measure of similarity in industry specialization. Krugman (1991) and Clark and van Wincoop (2001) favour a variable akin to a Herfindahl index of concentration, whereas Imbs (2001) uses a correlation coefficient between sectoral shares in aggregate output or employment. Here, I use sectoral real value added to compute

where s_n,i denotes the GDP share of industry n in country i. In words, S_i,j is the time average of the discrepancies in the economic structures of countries i and j.³⁶ Thus, S reaches its maximal value of two for two countries with no sector in common: we should therefore expect α₂< 0. For the international estimation, the sectoral shares s are computed using two alternative data sources: first the United Nations Statistical Yearbook sectoral data, which provides sectoral value added at the one-digit level, for all sectors in the economy and with (incomplete) coverage from 1960 to 1998. Second, for robustness, I also use two-digit value added data issued by the UNIDO, pertaining to manufacturing sectors only.³⁷ For the intranational estimation, I use real sectoral state value added series issued by the Bureau of Economic Analysis. These cover all economic activities, at the three-digit aggregation level.³⁸

Bilateral financial integration is notoriously difficult to measure. I therefore maximize the number of alternative measures, computed from widely different data sources, and subject the conclusions to an extensive sensitivity analysis. The measures can be separated in two distinct groups: indices capturing restrictions to capital flows, and variables reflecting effective financial flows. The two measures are fundamentally different in that the absence of any impediments to capital flows does not necessarily result in intense financial integration. Included in the first group are the IMF’s binary index of capital account restrictions, and the extension thereof reported in the IMF’s Annual Report on Exchange Arrangements and Exchange Restrictions (AREAER).³⁹ These indices are summed pairwise, and thus report the average number of countries with restrictions to financial flows, for each country pair.^40,41

Measuring effective bilateral capital flows is an even taller order. Here two proxies are proposed. For intranational data, Kalemli-Ozcan, Sorensen and Yosha (2002) computed a state-specific index of risk-sharing by estimating

where gsp denotes Gross State Product per capita and dy is state disposable income per capita. Kalemli-Ozcan, Sorensen and Yosha interpret β as an index of risk-sharing. Indeed, if inter-state risk sharing (in the guise of income insurance) is perfect, β = 1 as disposable income is unrelated with GSP per capita, and equation (4) is simply a regression of GSP on itself. Conversely, if there is no inter-state risk-sharing β = 0 since the dependent variable becomes essentially noise. A measure of cross-state financial integration is then given by pairwise sums of the state-specific estimates of β. I use data on Gross Domestic Product and Private Consumption to extend these estimates in the international data.⁴²

An alternative is possible thanks to the recent dataset constructed by Lane and Milesi-Ferretti (2001), that reports cumulated external positions for a large sample of developing and industrial economies. Pairs of countries with intense capital flows should display different (or even opposite) external positions. Two countries with massively positive (negative) net foreign assets holdings will both tend to be issuers (recipients) of capital flows, and should experience less bilateral flow than two economies where one is structurally in surplus and the other in deficit. Thus, the Lane - Milesi-Ferretti data can be used to construct $\mathit{LMF}=\left|{\left(\frac{NFA}{GDP}\right)}_i-{\left(\frac{NFA}{GDP}\right)}_j\right|$, where NFA denotes the net foreign asset position in country i. LMF will take high values for pairs of countries with diverging external positions, more likely to lend and borrow from each other than pairs of countries with similar external positions. The NFA position is computed using both accumulated current accounts, and the sum of net positions in Foreign Direct Investment, Equities and Debt.⁴³

The gravity variables in I₂ are standard, and include the following: a measure of the (log mile) distance between the countries capitals, the (log) products of each country’s GDP and binary variables indicating the presence of a common border and whether they share the same language. The vector I₃ is different from I₂, and includes: the proposed measure of financial integration, already in I₁, the (log) product of each country’s GDP per capita, and the (log) GDP disparity, defined as $\mathit{Max}\left[\frac{Y_i}{Y_j},\frac{Y_j}{Y_i}\right]$.⁴⁴

Table 1 reports summary statistics for the three endogenous variables of the system, in both datasets. Table 2 reports the corresponding unconditional correlations. As usual, and particularly in the present context of simultaneity, unconditional correlations are only informative superficially. However, several points are worth noting. First, the cross-sections of cycle synchronizations appear to be very similar regardless whether the cycles are computed on the basis of quarterly or annual data. ρ_Q and ρ_Y have similar moments and extreme values, and are highly correlated. Cycle correlations between U.S. states do not seem higher on average, although the cross-section tends to have more extreme values. Second, states are more specialized than countries, as S is on average larger than both S_YB and S_Mfg. This is a well-known result, first documented in Krugman (1991). Average specialization in manufactures is also higher than across all economic activities, and the correlation between the two cross-sections is only 0.47, which warrants some sensitivity analysis.⁴⁵ In what follows, manufacturing data is privileged in computing indices of specialization because of its large coverage; the sensitivity analysis does, however, show that the results obtain as well using a coarser specialization measure, based on all economic activities and not only manufactures. Third, both measures of trade intensity are very similar, with a correlation of 0.85. They both correlate positively with ρ, although T¹ does so more significantly. Fourth, specialization is correlated negatively with cycle synchronization, no matter the combination of measures used. The evidence for a trade-induced specialization is a priori weak, since the correlation between specialization and trade measures in only positive with T², and very weakly. Unsurprisingly, this calls for appropriate conditioning, as well as a simultaneous approach.

Table 1A:

Summary Statistics across Countries

Table 1A:

Summary Statistics across Countries

	Mean	Min	Max	Std. Dev.	N.Obs
ρQ	0.193	-0.586	0.887	0.310	276
ρy	0.201	-0.357	0.790	0.245	276
T1	0.004	7.22 x 10−5	0.069	0.007	276
T2	0.285	0.013	5.347	0.489	276
Syb	0.380	0.143	0.747	0.125	171
SMfg	0.540	0.150	0.976	0.159	276

View Table

Table 1A:

Summary Statistics across Countries

	Mean	Min	Max	Std. Dev.	N.Obs
ρQ	0.193	-0.586	0.887	0.310	276
ρy	0.201	-0.357	0.790	0.245	276
T1	0.004	7.22 x 10−5	0.069	0.007	276
T2	0.285	0.013	5.347	0.489	276
Syb	0.380	0.143	0.747	0.125	171
SMfg	0.540	0.150	0.976	0.159	276

Table 1B:

Summary Statistics across States (1275 obs)

Table 1B:

Summary Statistics across States (1275 obs)

	Mean	Min	Max	Std. Dev.
ρ	0.266	-0.896	0.968	0.387
S	0.577	0.201	1.339	0.200

View Table

Table 1B:

Summary Statistics across States (1275 obs)

	Mean	Min	Max	Std. Dev.
ρ	0.266	-0.896	0.968	0.387
S	0.577	0.201	1.339	0.200

Table 2A:

Unconditional Correlations across Countries

Table 2A:

Unconditional Correlations across Countries

	ρQ	ρY	T1	T2	Syb	SMfg
ρQ	1.000
ρy	0.743	1.000
T1	0.324	0.437	1.000
T2	0.163	0.262	0.848	1.000
Syb	-0.414	-0.469	-0.187	0.032	1.000
SMfg	-0.250	-0.410	-0.228	-0.027	0.476	1.000

View Table

Table 2A:

Unconditional Correlations across Countries

	ρQ	ρY	T1	T2	Syb	SMfg
ρQ	1.000
ρy	0.743	1.000
T1	0.324	0.437	1.000
T2	0.163	0.262	0.848	1.000
Syb	-0.414	-0.469	-0.187	0.032	1.000
SMfg	-0.250	-0.410	-0.228	-0.027	0.476	1.000

Table 2B:

Unconditional Correlations across States

Table 2B:

Unconditional Correlations across States

	ρ	S
ρ	1.000
S	-0.279	1.000

View Table

Table 2B:

Unconditional Correlations across States

	ρ	S
ρ	1.000
S	-0.279	1.000

Finally, Table 3 reports a few extreme values for the four main variables used in this study, namely cycles correlations, trade, specialization and financial integration. This is only illustrative, but has some interesting teachings. For instance, it is remarkable that the only country pair displaying both high business cycles correlation and high trade linkages should be the U.S. with Canada, precisely where Schmitt-Grohé (1998) showed the theoretical inability of trade per se to account for cycles synchronization. Thus, even between the two countries where the case for trade would be the strongest, the variable alone appears insufficient to account for observed synchronization. Similarly, low correlations between GSP tend to involve South Dakota, whereas low inter-state trade flows (admittedly, only as predicted by a gravity model) tend to involve Vermont, a rather remote and economically small state. Florida and Virginia display the most correlated GSP, yet there is no particular reason to expect strong trade linkages between the two states, at least not on the basis of a gravity model.

Table 3:

Selected Minima and Maxima

Notes: Correlations are based on Band-Pass filtered quarterly data. Trade is based on T¹ for international data, and a fitted gravity model for intranational data. Specialization is based on UNYB international data, and BEA intranational data. Financial integration is K, a binary variable capturing capital account restrictions across-countries, and RS, the extent of income insurance across states.

Table 3:

Selected Minima and Maxima

Correlation		Trade
-0.586	Norway - Portugal	7.22 x 10−5	Australia - Peru
-0.522	Philippines - South Africa	7.40 x 10−5	Norway - Philippines
-0.471	Israel - U.S.	8.43 x 10−5	Norway - Peru
0.839	Switzerland - Netherlands	0.029	France - Germany
0.862	Belgium - Italy	0.031	Germany - Netherlands
0.867	U.S. - Canada	0.074	Belgium - Netherlands
-0.896	Texas - South Dakota	17.06	Vermont - Wyoming
-0.894	Connecticut - South Dakota	17.94	Vermont - South Dakota
-0.808	Rhode Island - South Dakota	18.92	Vermont - Alaska
0.964	Massachusetts - Maine	64.36	New York - Pennsylvania
0.967	Florida - Maine	66.95	California - Texas
0.968	Florida - Virginia	67.54	California - New York

	Specialization		Financial Integration
0.143	Australia - Canada	0	Mexico - U.S.
0.155	Netherlands - Canada	0	Canada - U.S.
0.157	U.K. - France	0	Mexico - Canada
0.653	Israel - Mexico	2	Italy - South Africa
0.692	Norway - Philippines	2	U.K. - South Africa
0.747	Israel - Philippines	2	Netherlands - Japan
0.201	Indiana - Ohio	0.247	Nebraska - Virginia
0.231	Arizona - Florida	0.287	Kentucky - South Dakota
0.242	Minnesota - Wisconsin	0.319	California - Rhode Island
1.295	D.C. - South Dakota	1.506	Alaska - Wyoming
1.331	D.C. - Wyoming	1.554	Alaska - Louisiana
1.339	D.C. - Delaware	1.638	Alaska - New Mexico

Notes: Correlations are based on Band-Pass filtered quarterly data. Trade is based on T¹ for international data, and a fitted gravity model for intranational data. Specialization is based on UNYB international data, and BEA intranational data. Financial integration is K, a binary variable capturing capital account restrictions across-countries, and RS, the extent of income insurance across states.

View Table

Table 3:

Selected Minima and Maxima

Correlation		Trade
-0.586	Norway - Portugal	7.22 x 10−5	Australia - Peru
-0.522	Philippines - South Africa	7.40 x 10−5	Norway - Philippines
-0.471	Israel - U.S.	8.43 x 10−5	Norway - Peru
0.839	Switzerland - Netherlands	0.029	France - Germany
0.862	Belgium - Italy	0.031	Germany - Netherlands
0.867	U.S. - Canada	0.074	Belgium - Netherlands
-0.896	Texas - South Dakota	17.06	Vermont - Wyoming
-0.894	Connecticut - South Dakota	17.94	Vermont - South Dakota
-0.808	Rhode Island - South Dakota	18.92	Vermont - Alaska
0.964	Massachusetts - Maine	64.36	New York - Pennsylvania
0.967	Florida - Maine	66.95	California - Texas
0.968	Florida - Virginia	67.54	California - New York

	Specialization		Financial Integration
0.143	Australia - Canada	0	Mexico - U.S.
0.155	Netherlands - Canada	0	Canada - U.S.
0.157	U.K. - France	0	Mexico - Canada
0.653	Israel - Mexico	2	Italy - South Africa
0.692	Norway - Philippines	2	U.K. - South Africa
0.747	Israel - Philippines	2	Netherlands - Japan
0.201	Indiana - Ohio	0.247	Nebraska - Virginia
0.231	Arizona - Florida	0.287	Kentucky - South Dakota
0.242	Minnesota - Wisconsin	0.319	California - Rhode Island
1.295	D.C. - South Dakota	1.506	Alaska - Wyoming
1.331	D.C. - Wyoming	1.554	Alaska - Louisiana
1.339	D.C. - Delaware	1.638	Alaska - New Mexico

Notes: Correlations are based on Band-Pass filtered quarterly data. Trade is based on T¹ for international data, and a fitted gravity model for intranational data. Specialization is based on UNYB international data, and BEA intranational data. Financial integration is K, a binary variable capturing capital account restrictions across-countries, and RS, the extent of income insurance across states.

To summarize, both international and intranational data suggest trade may be important, but is probably insufficient to explain all the cross-section of cycle correlations. Turning to measures of specialization, the three most similar country-pairs (on the basis of manufacturing data) are Australia - Canada, the UK - France and the Netherlands - Canada. The importance of sectoral specialization is therefore of special interest to inform the debate of Sterling’s entry into the EMU. Across states, the District of Columbia unsurprisingly displays the most idiosyncratic sectoral composition of value added. Finally, financial integration as measured by capital account restrictions is highest between the North American countries and lowest when such countries as South Africa are included. Across states, the extent of risk-sharing seems to be related with specialization, as already demonstrated in Kalemli-Ozcan, Sorensen and Yosha (2003), since pairs involving Alaska display high risk-sharing, as measured by income insurance.⁴⁶

C. Three-Stage Least Squares and Heteroscedasticity

As should be clear from the previous discussion, the proposed estimation method must combine the features of simultaneous equations procedures, and account for the possible endogeneity of the three dependent variables ρ, T, S. Three-Stage Least Squares (TSLS) does exactly that. The estimator combines insights from Instrumental Variable and Generalized Least Squares methods, achieving consistency through instrumentation and efficiency through appropriate weighting in the variance-covariance matrix. As is well-known, the procedure consists of the following steps: (i) estimate the system equation by equation using Two-Stage Least Squares, and retrieve the covariance matrix of the equations disturbances, then (ii) perform a type of Generalized Least Squares estimation on the stacked system, using the covariance matrix from the first step.⁴⁷

It is also often useful to contrast the results obtained via Three-Stage Least Squares, to those generated by the partly similar Seemingly Unrelated Regression (SUR) technique. SUR focuses on the simultaneity, but does not account for possible endogeneity of the dependent variables. In other words, the system is estimated equation-by-equation under the assumption that the variance-covariance matrix is non-diagonal. Finally, the next section also reports simple equation-by-equation Least Squares estimates, for the sake of comparison with existing estimates. Comparisons between the OLS, SUR and TSLS estimates highlight the importance of simultaneity and endogeneity, respectively.

In addition, the correlations in business cycles ρ are measured with error, in a way that is likely to create a specific type of heteroscedasticity. In particular, following Clark and van Wincoop (2001), let $\widehat{\rho }=\rho +\upsilon$ denote the estimated correlation coefficients, with υ the sampling error. It is then possible that the sampling error υ be correlated across observations in $\widehat{\rho }$, since many correlation coefficients involve the GDP series for the same country or state.⁴⁸ This will create a kind of heteroscedasticity in the residuals of (1) that standard White corrections cannot account for.

Ignoring this heteroscedasticity is likely to result in understated standard errors for the estimates in (1), and since the estimation is simultaneous, in the rest of the system as well. It is, however, possible to account for this potential issue, at the cost of relatively mild assumptions. If the true vector of bilateral correlations ρ is assumed to be deterministic, (1) rewrites

The variance-covariance matrix of equation (5) involves ${\widehat{{\Sigma }}}_{\upsilon }=E(\upsilon \upsilon )$, which requires using a GMM estimator. As part of the robustness checks discussed in Section 4, I present results applying a GMM estimator to the system (l)-(2)-(3), that also accounts for both endogeneity and simultaneity.

A. Equation-by-Equation Estimates

Table 4 reports the results of equation-by-equation estimations for the system (l)-(2)-(3), using the benchmark set of control variables described in Section 2.⁴⁹ The purpose of Table 4 is to confirm that the well-known results pertaining to each equation in the estimated system, taken in isolation, are present in the dataset used here.⁵⁰ Column (i) confirms the large and significant effect of trade in accounting for ρ in equation (1). The point estimate in column (i) means that doubling trade results in a correlation higher by 0.053, which is close to the estimate in Clark and van Wincoop (2001), and exactly equal to the revised estimates in Kose and Yi (2002). As in Frankel and Romer (1999), instrumenting trade with exogenous gravity variables results in an even higher point estimates. This probably happens because the endogeneity bias is negative, as non-synchronized economies tend to trade more. The point estimate for α₁ in column (ii) means doubling trade now results in a correlation coefficient higher by 0.08, comparable in magnitude with the estimates in Frankel and Rose (1998). We conclude that the dataset used here is perfectly standard from the point of view of the relationship between ρ and T.

Table 4:

Equation-by-equation Estimations

Notes: Intercepts are not reported. All variables are in logs, excepted binary ones and correlation coefficients. The instruments in the IV estimations are the gravity variables, K and the exogenous determinants of specialization, which is based on manufacturing sectors only, t-statistics between parentheses.

Table 4:

Equation-by-equation Estimations

	(i) OLS	(ii) IV	(iii) OLS	(iv) IV
(1) Correlation ρ
T	0.077 (6.00)	0.115 (7.55)	0.064 (4.25)	0.083 (3.73)
S			-0.123 (1.92)	-0.217 (1.92)
K			-0.019 (0.57)	0.026 (0.77)
R2	0.116	0.089	0.124	0.100
(2) Trade T
Distance	-0.975 (19.34)		-0.970 (17.43)	-0.922 (15.36)
Border	0.038 (0.17)		0.119 (0.48)	0.041 (0.16)
Language	0.325 (2.23)		0.347 (2.02)	0.248 (1.36)
GDP Product	0.261 (9.68)		0.231 (7.26)	0.166 (4.15)
S			-0.401 (2.22)	-1.197 (3.62)
R2	0.701		0.697	0.674
(3) Specialization S
GDP per capita Product	-0.139 (6.02)	-0.119 (4.79)	-0.185 (7.74)	-0.163 (6.34)
GDP Gap	0.613 (2.21)	0.524 (1.86)	0.783 (2.81)	0.671 (2.36)
T	-0.051 (3.64)	-0.074 (4.17)	-0.047 (3.37)	-0.073 (4.14)
K			-0.102 (3.39)	-0.106 (3.51)
R2	0.269	0.267	0.347	0.338

Notes: Intercepts are not reported. All variables are in logs, excepted binary ones and correlation coefficients. The instruments in the IV estimations are the gravity variables, K and the exogenous determinants of specialization, which is based on manufacturing sectors only, t-statistics between parentheses.

View Table

Table 4:

Equation-by-equation Estimations

	(i) OLS	(ii) IV	(iii) OLS	(iv) IV
(1) Correlation ρ
T	0.077 (6.00)	0.115 (7.55)	0.064 (4.25)	0.083 (3.73)
S			-0.123 (1.92)	-0.217 (1.92)
K			-0.019 (0.57)	0.026 (0.77)
R2	0.116	0.089	0.124	0.100
(2) Trade T
Distance	-0.975 (19.34)		-0.970 (17.43)	-0.922 (15.36)
Border	0.038 (0.17)		0.119 (0.48)	0.041 (0.16)
Language	0.325 (2.23)		0.347 (2.02)	0.248 (1.36)
GDP Product	0.261 (9.68)		0.231 (7.26)	0.166 (4.15)
S			-0.401 (2.22)	-1.197 (3.62)
R2	0.701		0.697	0.674
(3) Specialization S
GDP per capita Product	-0.139 (6.02)	-0.119 (4.79)	-0.185 (7.74)	-0.163 (6.34)
GDP Gap	0.613 (2.21)	0.524 (1.86)	0.783 (2.81)	0.671 (2.36)
T	-0.051 (3.64)	-0.074 (4.17)	-0.047 (3.37)	-0.073 (4.14)
K			-0.102 (3.39)	-0.106 (3.51)
R2	0.269	0.267	0.347	0.338

Notes: Intercepts are not reported. All variables are in logs, excepted binary ones and correlation coefficients. The instruments in the IV estimations are the gravity variables, K and the exogenous determinants of specialization, which is based on manufacturing sectors only, t-statistics between parentheses.

Equation (2) in specification (i) is unsurprising: the gravity variables all have the expected signs, as is now completely standard in any empirical work concerned with explaining the geography of trade. The same is true of equation (3), with estimates of the γs in line with theory. In particular, pairs of rich countries tend to have significantly lower value of S, i.e. their economic structures are more similar, as would happen if growth resulted in diversification. But on the other hand, pairs of countries at different stages of development, as measured by the gap between their GDPs, have significantly higher S, i.e., tend to display different economic structures. Finally, the effect of trade on S is significant, but with the “wrong” sign, i.e. more bilateral trade results in lower S, or in more similar economies. This last result is even stronger when trade is instrumented in column (ii), a puzzling fact, though it is later shown to be non robust.

Columns (iii) and (iv) add the two variables specific to the simultaneous approach, namely specialization S and financial integration K. Three results are of particular interest. (a) Financial integration has the predicted specialization effect, since a low value of K is significantly associated with high S, that is, financially integrated economies have different specialization patterns. (b) Country pairs with low S have significantly higher ρ: similarities in economic structure result in correlated business cycles. (c) Estimates for α₁ are now much smaller, with a point estimate such that a doubling of trade would result in a correlation higher by 0.044.⁵¹ Column (iv) implements the corresponding instrument variables estimations, with the same instrument set as in (i) and (ii). This tends to magnify all three results (a) to (c), and adds a fourth interesting one: (d) β₁ becomes very significantly negative once specialization is instrumented. This makes sense as the endogeneity of S to T would if anything tend to bias β₁ upwards, as trading partners specialize and thus have high S. A negative β₁ can be interpreted as meaning that countries with similar economic structures trade more, a quantification of the extent of intra-industry trade. Estimates of β₁ are therefore key to disentangling the impacts of inter- versus intra-industry trade in simultaneous estimations.

Although they are largely consistent with existing work, all these results may alter once both endogeneity and simultaneity are controlled for. For instance in equation (1), the three regressors are theoretically related, and a proper account of the covariance term may very well result in different estimates throughout the system. For instance, without controlling for simultaneity, the effect of K on ρ in equation (1) is but a “reduced form” measure, which does not discriminate between indirect specialization effects (which equation (3) suggests are important) and putative direct effects. I next turn to the simultaneous estimations.

B. Simultaneity and Channels

Table 5 is symmetric to Table 4, and presents SUR and TSLS estimates, where the former only account for simultaneity while the latter also instruments the three endogenous variables ρ, S and T. Table 5 can help answer two questions: 1) are the estimates sensitive to simultaneity? 2) can we quantify the direct and indirect channels relating trade, finance, specialization and business cycles synchronization?

Table 5:

Simultaneous Estimations

Notes: Intercepts are not reported. All variables are in logs, excepted binary ones and correlation coefficients. The instruments in the IV estimations are the gravity variables, K and the exogenous determinants of specialization, which is based on manufacturing sectors only. t-statistics between parentheses.

Table 5:

Simultaneous Estimations

	(i) SUR	(ii) TSLS	(iii) SUR	(iv) TSLS
(1) Correlation ρ
T	0.102 (7.97)	0.122 (8.07)	0.083 (5.66)	0.074 (3.32)
S			-0.132 (2.09)	-0.298 (2.67)
K			0.020 (0.63)	0.011 (0.35)
R2	0.105	0.078	0.116	0.083
(2) Trade T
Distance	-0.968 (19.67)	-0.946 (19.39)	-0.908 (16.93)	-0.815 (14.49)
Border	0.053 (0.25)	0.062 (0.30)	0.147 (0.62)	0.152 (0.69)
Language	0.359 (2.54)	0.367 (2.68)	0.394 (2.42)	0.401 (2.58)
GDP Product	0.275 (10.51)	0.278 (10.88)	0.232 (7.57)	0.216 (5.66)
S			-0.820 (4.68)	-1.470 (4.54)
R2	0.701	0.700	0.688	0.640
(3) Specialization S
GDP/capita Product	-0.131 (5.75)	-0.122 (4.98)	-0.167 (7.24)	-0.144 (6.11)
GDP Gap	0.666 (2.42)	0.691 (2.49)	0.946 (3.51)	1.098 (4.26)
T	-0.061 (4.37)	-0.074 (4.24)	-0.071 (5.26)	-0.079 (4.68)
K			-0.096 (3.28)	-0.074 (2.62)
R2	0.273	0.266	0.336	0.322

Notes: Intercepts are not reported. All variables are in logs, excepted binary ones and correlation coefficients. The instruments in the IV estimations are the gravity variables, K and the exogenous determinants of specialization, which is based on manufacturing sectors only. t-statistics between parentheses.

View Table

Table 5:

Simultaneous Estimations

	(i) SUR	(ii) TSLS	(iii) SUR	(iv) TSLS
(1) Correlation ρ
T	0.102 (7.97)	0.122 (8.07)	0.083 (5.66)	0.074 (3.32)
S			-0.132 (2.09)	-0.298 (2.67)
K			0.020 (0.63)	0.011 (0.35)
R2	0.105	0.078	0.116	0.083
(2) Trade T
Distance	-0.968 (19.67)	-0.946 (19.39)	-0.908 (16.93)	-0.815 (14.49)
Border	0.053 (0.25)	0.062 (0.30)	0.147 (0.62)	0.152 (0.69)
Language	0.359 (2.54)	0.367 (2.68)	0.394 (2.42)	0.401 (2.58)
GDP Product	0.275 (10.51)	0.278 (10.88)	0.232 (7.57)	0.216 (5.66)
S			-0.820 (4.68)	-1.470 (4.54)
R2	0.701	0.700	0.688	0.640
(3) Specialization S
GDP/capita Product	-0.131 (5.75)	-0.122 (4.98)	-0.167 (7.24)	-0.144 (6.11)
GDP Gap	0.666 (2.42)	0.691 (2.49)	0.946 (3.51)	1.098 (4.26)
T	-0.061 (4.37)	-0.074 (4.24)	-0.071 (5.26)	-0.079 (4.68)
K			-0.096 (3.28)	-0.074 (2.62)
R2	0.273	0.266	0.336	0.322

Notes: Intercepts are not reported. All variables are in logs, excepted binary ones and correlation coefficients. The instruments in the IV estimations are the gravity variables, K and the exogenous determinants of specialization, which is based on manufacturing sectors only. t-statistics between parentheses.

Table 6:

Alternative Measures of Financial Integration

Notes: Intercepts are not reported. All estimations are run using TSLS, with the gravity variables, K and the exogenous determinants of specialization as instruments. Restrictions denotes the composite index for capital account restrictions from AREAER. Risk-Sharing denotes the index proposed by Kalemli-Ozcan et al (2002). Cum. CA denotes the net cumulated current - account, whereas NFA denotes the net foreign position in FDI, equity and debt, both from Lane and Milesi-Ferretti (2001), and measured in 1980. t-statistics between parentheses.

Table 6:

Alternative Measures of Financial Integration

	(i) Restrictions	(ii) Risk Sharing	(iii) Cum.CA	(iv) NFA
(1) Correlation ρ
T	0.073 (3.33)	0.090 (3.52)	0.070 (3.13)	0.074 (3.27)
S	-0.321 (2.93)	-0.352 (2.24)	-0.388 (2.75)	-0.294 (2.31)
K	-0.037 (0.92)	0.087 (2.38)	0.117 (0.75)	0.224 (3.51)
R2	0.085	0.103	0.054	0.106
(2) Trade T
Distance	-0.828 (16.02)	-0.731 (10.55)	-0.855 (15.96)	-0.872 (16.30)
Border	0.061 (0.30)	0.037 (0.18)	0.037 (0.18)	0.041 (0.20)
Language	0.317 (2.42)	0.214 (1.59)	0.330 (2.45)	0.327 (2.44)
GDP Product	0.218 (5.85)	0.197 (3.29)	0.213 (5.15)	0.224 (5.44)
S	-1.239 (3.89)	-2.269 (3.68)	-1.079 (2.83)	-0.898 (2.34)
R2	0.674	0.578	0.687	0.695
(3) Specialization S
GDP/capita Product	-0.166 (6.74)	-0.090 (3.70)	-0.127 (5.77)	-0.144 (6.31)
GDP Gap	0.971 (3.90)	0.765 (2.84)	0.687 (2.70)	1.088 (4.14)
T	-0.076 (4.72)	-0.085 (4.74)	-0.068 (4.29)	-0.074 (4.57)
K	-0.168 (4.64)	-0.020 (0.77)	0.491 (4.54)	0.213 (4.23)
R2	0.331	0.256	0.340	0.335

Notes: Intercepts are not reported. All estimations are run using TSLS, with the gravity variables, K and the exogenous determinants of specialization as instruments. Restrictions denotes the composite index for capital account restrictions from AREAER. Risk-Sharing denotes the index proposed by Kalemli-Ozcan et al (2002). Cum. CA denotes the net cumulated current - account, whereas NFA denotes the net foreign position in FDI, equity and debt, both from Lane and Milesi-Ferretti (2001), and measured in 1980. t-statistics between parentheses.

View Table

Table 6:

Alternative Measures of Financial Integration

	(i) Restrictions	(ii) Risk Sharing	(iii) Cum.CA	(iv) NFA
(1) Correlation ρ
T	0.073 (3.33)	0.090 (3.52)	0.070 (3.13)	0.074 (3.27)
S	-0.321 (2.93)	-0.352 (2.24)	-0.388 (2.75)	-0.294 (2.31)
K	-0.037 (0.92)	0.087 (2.38)	0.117 (0.75)	0.224 (3.51)
R2	0.085	0.103	0.054	0.106
(2) Trade T
Distance	-0.828 (16.02)	-0.731 (10.55)	-0.855 (15.96)	-0.872 (16.30)
Border	0.061 (0.30)	0.037 (0.18)	0.037 (0.18)	0.041 (0.20)
Language	0.317 (2.42)	0.214 (1.59)	0.330 (2.45)	0.327 (2.44)
GDP Product	0.218 (5.85)	0.197 (3.29)	0.213 (5.15)	0.224 (5.44)
S	-1.239 (3.89)	-2.269 (3.68)	-1.079 (2.83)	-0.898 (2.34)
R2	0.674	0.578	0.687	0.695
(3) Specialization S
GDP/capita Product	-0.166 (6.74)	-0.090 (3.70)	-0.127 (5.77)	-0.144 (6.31)
GDP Gap	0.971 (3.90)	0.765 (2.84)	0.687 (2.70)	1.088 (4.14)
T	-0.076 (4.72)	-0.085 (4.74)	-0.068 (4.29)	-0.074 (4.57)
K	-0.168 (4.64)	-0.020 (0.77)	0.491 (4.54)	0.213 (4.23)
R2	0.331	0.256	0.340	0.335

Notes: Intercepts are not reported. All estimations are run using TSLS, with the gravity variables, K and the exogenous determinants of specialization as instruments. Restrictions denotes the composite index for capital account restrictions from AREAER. Risk-Sharing denotes the index proposed by Kalemli-Ozcan et al (2002). Cum. CA denotes the net cumulated current - account, whereas NFA denotes the net foreign position in FDI, equity and debt, both from Lane and Milesi-Ferretti (2001), and measured in 1980. t-statistics between parentheses.

Table 7:

US States

Notes: Intercepts are not reported. All variables are in logs, excepted binary ones and correlation coefficients. Trade between states is a fitted value implied by a gravity model estimated on cross-country data; in particular, $\widehat{T}$ = −1.355* ln(Distance) + 1.057* ln(GDP Product) − 0.635* ln(Population Product) − 29.834. t-statistics between parentheses. Risk Sharing measures income insurance between US states, and reports the pairwise sum of β from the regression lnGSP_t− lny_t = α + β lnGSP_t, where the cyclical component of all variables is isolated using the Baxter-King filter (except in (iv) and (v)). Size is measured by the pairwise discrepancy in GSP.

Table 7:

US States

	(i) TSLS	(ii) TSLS	(iii) TSLS	(iv) HP	(v) GMM
(1) Correlation ρ
$\widehat{T}$	0.031 (13.74)	0.032 (14.09)	0.031 (13.84)	0.029 (12.65)	0.028 (7.44)
S		-0.351 (8.49)	-0.326 (7.64)	-0.571 (13.11)	-0.571 (8.52)
Risk Sharing		0.063 (2.28)	0.051 (1.86)	0.140 (4.95)	0.142 (4.60)
Size			-0.049 (3.51)	-0.056 (3.94)	-0.042 (2.47)
R2	0.176	0.267	0.277	0.338	0.341
(3) Specialization S
GDP per capita Product	-0.185 (17.22)	-0.206 (18.21)	-0.206 (18.23)	-0.205 (18.33)	-0.460 (8.03)
GDP Gap	0.400 (9.55)	0.405 (9.09)	0.404 (9.08)	0.406 (9.23)	-0.131 (0.99)
$\widehat{T}$	-0.033 (17.66)	-0.035 (17.96)	-0.035 (17.97)	-0.035 (17.98)	-0.103 (6.46)
Risk Sharing		0.050 (2.81)	0.050 (2.81)	0.050 (2.81)	0.014 (0.44)
R2	0.628	0.624	0.624	0.624	0.471

Notes: Intercepts are not reported. All variables are in logs, excepted binary ones and correlation coefficients. Trade between states is a fitted value implied by a gravity model estimated on cross-country data; in particular, $\widehat{T}$ = −1.355* ln(Distance) + 1.057* ln(GDP Product) − 0.635* ln(Population Product) − 29.834. t-statistics between parentheses. Risk Sharing measures income insurance between US states, and reports the pairwise sum of β from the regression lnGSP_t− lny_t = α + β lnGSP_t, where the cyclical component of all variables is isolated using the Baxter-King filter (except in (iv) and (v)). Size is measured by the pairwise discrepancy in GSP.

View Table

Table 7:

US States

	(i) TSLS	(ii) TSLS	(iii) TSLS	(iv) HP	(v) GMM
(1) Correlation ρ
$\widehat{T}$	0.031 (13.74)	0.032 (14.09)	0.031 (13.84)	0.029 (12.65)	0.028 (7.44)
S		-0.351 (8.49)	-0.326 (7.64)	-0.571 (13.11)	-0.571 (8.52)
Risk Sharing		0.063 (2.28)	0.051 (1.86)	0.140 (4.95)	0.142 (4.60)
Size			-0.049 (3.51)	-0.056 (3.94)	-0.042 (2.47)
R2	0.176	0.267	0.277	0.338	0.341
(3) Specialization S
GDP per capita Product	-0.185 (17.22)	-0.206 (18.21)	-0.206 (18.23)	-0.205 (18.33)	-0.460 (8.03)
GDP Gap	0.400 (9.55)	0.405 (9.09)	0.404 (9.08)	0.406 (9.23)	-0.131 (0.99)
$\widehat{T}$	-0.033 (17.66)	-0.035 (17.96)	-0.035 (17.97)	-0.035 (17.98)	-0.103 (6.46)
Risk Sharing		0.050 (2.81)	0.050 (2.81)	0.050 (2.81)	0.014 (0.44)
R2	0.628	0.624	0.624	0.624	0.471

Notes: Intercepts are not reported. All variables are in logs, excepted binary ones and correlation coefficients. Trade between states is a fitted value implied by a gravity model estimated on cross-country data; in particular, $\widehat{T}$ = −1.355* ln(Distance) + 1.057* ln(GDP Product) − 0.635* ln(Population Product) − 29.834. t-statistics between parentheses. Risk Sharing measures income insurance between US states, and reports the pairwise sum of β from the regression lnGSP_t− lny_t = α + β lnGSP_t, where the cyclical component of all variables is isolated using the Baxter-King filter (except in (iv) and (v)). Size is measured by the pairwise discrepancy in GSP.

The estimates in Table 5 are not dissimilar to the equation-by-equation approach, but some coefficients do change substantially in magnitude. First, estimates of α₁ increase slightly in specifications (i) and (ii). Since T in equation (2) is only explained by the (exogenous) gravity variables, this may happen because the systems (i) and (ii) focus de facto on the geographic component of trade, and thus the attenuating effect of trade’s endogeneity is neutralized. This, however, disappears once S in included in equations (1) and (2) in specifications (iii) and (iv). Then, estimates of α₁ revert to relatively low levels, roughly twice lower than the Frankel-Rose estimates. Of course, part of the overall effect of trade may now go through specialization S, as γ₁ captures in equation (3). Furthermore, trade itself is decomposed into two components: intra-industry trade, as captured by β₁ in equation (2), and the classic “geographic” component, as captured by β₂.

One word of comment is in order here. A number of authors have puzzled over the inability of standard one-sector models to reproduce the large effect trade has on cycle correlations.⁵² Recently, Kose and Yi (2002) have calibrated and simulated an international business cycles model with transport costs and technology shocks. Depending on the parametrization, their model yield simulated values for α₁ ranging from 0.009 to 0.069. It is interesting to note that the simultaneous estimates in Table 5 are well within this range: thus, once focused on the link between inter-industry trade and business cycles correlations, the data is not inconsistent with elasticities predicted by one-sector models. The estimates for α₁ in the last two columns of Table 5 are low for two reasons. First, the presence of an adequate control for the influence of common sectors, where the sectoral pattern of production is in turn allowed to respond endogenously to trade intensity and other exogenous determinants. Second, an explicit account of the possibility that (at least part of) the high estimate for α₁ originates in the prevalence of intra-industry trade, something not typically part of international business cycles models.^53,54

Second, estimates of the effect of S on ρ become larger in Table 5, now that S is allowed to depend on both trade and financial integration. The R² suggest S is a quantitatively important determinant of business cycles correlation. Furthermore, the estimates from equation (3) suggest its determinants are largely beyond the reach of policy makers. While financial integration does affect specialization patterns as predicted by theory, the bulk of the cross-section in S depends on the stages of diversification reached by both economies, largely a low frequency, long-run phenomenon. Finally, financial integration does still result in specialization, although no direct effect on cycle observation is apparent. Importantly, alternative measures of financial integration, and in particular the variables capturing effective rather than potential capital flows turn out to have substantial direct synchronization effects as well. I next turn to a quantification of these direct and indirect channels.

Panel A of Table 8 reports all direct and indirect channels in the system under study, illustrated in Figure 1, as a function of the estimated parameters. Of interest is the decomposition of the coefficients αs in equation (1) into overall contributions from trade, specialization and financial integration. The direct effects of trade are either a reflection of intra-industry flows (α₁ β₁) or of standard Ricardian trade (α₁ β₂); the indirect effect of trade comes from the possibility that economies open to goods are also specialized, and may have a higher value of S as a result (α₂ γ₁). The direct effects of specialization can, again, originate in trade (α₂ γ₁), or in financial integration (α₂ γ₂), or in exogenous stages of development reached by both economies (α₂ γ₃); the indirect effect of specialization may, once again, consist in the manifestation of intra-industry trade (α₁ β₁). The direct effect of financial integration is captured by α₃, and its indirect effect is going via specialization (α₂ γ₂).

Table 8A:

Channels to Business Cycles Synchronization

Table 8A:

Channels to Business Cycles Synchronization

	Direct	Indirect
(1) Trade
Intra-Industry Trade	α1β1
Geographic Trade	α1β2
via Specialization		α2 γ1
(2) Specialization
Trade Induced	α2 γ1
Finance Induced	α2 γ2
Stages of Diversification	α2 γ3
via Trade		α1β1
(3) Finance
Financial Integration	α3
via Specialization		α2 γ2

View Table

Table 8A:

Channels to Business Cycles Synchronization

	Direct	Indirect
(1) Trade
Intra-Industry Trade	α1β1
Geographic Trade	α1β2
via Specialization		α2 γ1
(2) Specialization
Trade Induced	α2 γ1
Finance Induced	α2 γ2
Stages of Diversification	α2 γ3
via Trade		α1β1
(3) Finance
Financial Integration	α3
via Specialization		α2 γ2

Table 8B: Channels to Business Cycles Synchronization - International Data

Table 8B: Channels to Business Cycles Synchronization - International Data

	Direct	Indirect
(1) Trade
Intra-Industry Trade	-0.108 (0.008)
Geographic Trade	0.105 (0.000)
via Specialization		0.024 (0.024)
(2) Specialization
Trade Induced	0.024 (0.024)
Finance Induced (CAPA)	0.022 (0.060)
Stages of Diversification	-0.343 (2.72)
via Trade		-0.108 (0.008)
(3) Finance
Capital Account	0.011 (0.727)
via Specialization (CAPA)		0.022 (0.060)

View Table

Table 8B: Channels to Business Cycles Synchronization - International Data

	Direct	Indirect
(1) Trade
Intra-Industry Trade	-0.108 (0.008)
Geographic Trade	0.105 (0.000)
via Specialization		0.024 (0.024)
(2) Specialization
Trade Induced	0.024 (0.024)
Finance Induced (CAPA)	0.022 (0.060)
Stages of Diversification	-0.343 (2.72)
via Trade		-0.108 (0.008)
(3) Finance
Capital Account	0.011 (0.727)
via Specialization (CAPA)		0.022 (0.060)

Table 8C: Channels to Business Cycles Synchronization - Intranational Data

Notes: P-values between parentheses.

Table 8C: Channels to Business Cycles Synchronization - Intranational Data

	Direct	Indirect
(1) Trade
Intra-Industry Trade	N/A
Geographic Trade	N/A
via Specialization		0.020 (0.00)
(2) Specialization
Trade Induced	0.020 (0.00)
Finance Induced	-0.029 (0.01)
Stages of Diversification	-0.694 (0.00)
via Trade		N/A
(3) Finance
Financial Integration	0.140 (0.00)
via Specialization		-0.029 (0.01)

Notes: P-values between parentheses.

View Table

Table 8C: Channels to Business Cycles Synchronization - Intranational Data

	Direct	Indirect
(1) Trade
Intra-Industry Trade	N/A
Geographic Trade	N/A
via Specialization		0.020 (0.00)
(2) Specialization
Trade Induced	0.020 (0.00)
Finance Induced	-0.029 (0.01)
Stages of Diversification	-0.694 (0.00)
via Trade		N/A
(3) Finance
Financial Integration	0.140 (0.00)
via Specialization		-0.029 (0.01)

Notes: P-values between parentheses.

Panel B of Table 8 reports the values for these channels as implied by the Three-Stage Least Squares estimates in Table 5.⁵⁵ A number of results are of interest. First, trade affects business cycles synchronization equally because of inter- and intra-industry deals.⁵⁶ The impact of intra-industry trade is quantified by the result that economies sharing similar economic structure trade substantially more, and are more synchronized as a result. We conclude that the puzzlingly high estimates for α₁ in standard single equation estimations are largely driven by intra-industry trade, an ingredient models seeking to reproduce these estimates should include. Models with inter-industry trade should only seek to reproduce the much smaller value of α₁ = 0.074.

The indirect effect of trade, in turn, is estimated to increase business cycles correlations, if anything. This is a somewhat surprising result, as one would expect trading partners to specialize according to comparative advantage, and have higher S as a result. Here there is evidence for the opposite result. This may find an explanation in the fact that the sample includes a vast majority of developed economies. Perhaps specialization patterns amongst these countries may indeed follow standard trade theories, but the variable S does capture it only imperfectly. This result is also not robust: applying an alternative filter to the data, or measuring trade differently results in a coefficient with the opposite sign. The absence of much of an indirect effect of trade on ρ is consistent with existing evidence, for instance Frankel and Rose (1998).

Most of the direct effect of S on ρ works through the exogenous determinants of specialization, i.e. the degree of economic advancement of countries i and j. As mentioned above, trade does appear to result in less specialization (and higher ρ), while financial integration, in turn, does result in higher S, i.e. countries specialized in different range of activities. In other words, the bulk of the determinants of S is largely beyond the reach of short-term policymaking. In as much as the international correlation of business cycles is an important constraint to policy, e.g. through its relation with monetary union, this puts into perspective the significantly positive estimates for α₁ arising from single equation estimations. While it may be possible to manipulate T through trade policy, there is no immediate equivalent for S. Finally, most of the effect of financial integration is indirect, as countries with liberalized capital account tend to be more specialized, and less correlated as a result. I next investigate how these conclusions are affected by alternate measures of financial integration, or by focusing on international as opposed to intranational data.

C. Measures of Effective Financial Integration

Results in the previous section were based on a particular measure of financial integration, which in effect merely quantifies the potential capital flows between economies. The results in Kalemli-Ozcan et al (2002) were confirmed, in that financially integrated economies were shown to be specialized, and less synchronized as a result. There was, however, no evidence on a direct link between business cycles synchronization and financial integration. In what follows, I use alternative measures to investigate the robustness of these conclusions.

Table 6 reports three-stage least squares estimates for the system (l)-(2)-(3), where the composite index of capital account restrictions is used in column (i), the index of risk-sharing introduced in Kalemli-Ozcan et al (2002) is used in column (ii), and two measures of LMF are used in the last two columns. Table 6 confirms overall that financial integration results in specialization, but provides further evidence suggesting that financial integration also affects cycle synchronization directly, and positively. This direct effect seems to become significant when measures of effective financial integration are used, rather than mere indices of restrictions.

Specification (i) uses the composite index of capital account restrictions issued by AREAER, similar to the index in Table 5. Results are almost identical: finance has an indirect, but no direct effect on cycles. The other channels are unchanged. When financial integration is measured by an index of (effective) risk-sharing between countries in specification (ii), however, the results are reversed: financial integration results in more synchronized cycles, but there is no substantial indirect specialization effect. Again, other results remain unchanged. The last two columns report results corresponding to LMF, which tends to be high between economies with differing net foreign positions, presumably more likely to exchange capital than two debitors’ or creditors’ countries. In specification (iii), LMF is computed on the basis of accumulated current account whereas it uses the accumulated net stock of direct investment, equity and debt in specification (iv). Finance-induced specialization is significant in both cases, while the last column provides further support for a positive direct channel. In specification (iv), both the direct and indirect effects of finance are also large economically. Other channels are unchanged relative to Table 5.

Overall, the picture painted by Table 6 is that of a substantial role for financial integration in affecting business cycles synchronization. While indices of capital account restrictions imply an important finance-induced specialization effect, measures of effective financial integration provide evidence of a direct positive channel from K to ρ. This is particularly true between economies with large differences in their net positions in FDI, equities and debt. The measures of LMF in specifications (iii) and (iv) are both warranted in theory. Nevertheless, there is anecdotal support for the view that the real effects of reversals in financial flows tend to go through asset holdings.⁵⁷ In particular, (unreported) results suggest that these real effects work through FDI flows.⁵⁸ In the next section, I investigate whether the importance of the financial channel is confirmed across U.S. states.

D. The Channels between U.S. States

Table 7 reports three-stage least squares estimates for U.S. states data. Since no data are available on direct trade flows between U.S. states, equation (2) is subsumed in implementing a gravity model to predict inter-state trade. This does not come at zero cost. First of all, all results depend on the reliability of a gravity model for inter-state trade. Luckily, the descriptive statistics in Table 3 do not seem out of line, as indeed the route between California and New York is probably amongst the most heavily traded within the U.S. The gravity model has reached almost universal validity, and is probably particularly appropriate for intranational data given the absence of any tariffs: it is hard to think of any impediments to commerce between U.S. states that are not related to geography. Second, not having any estimates for the coefficients β in equation (2) prevents a separate assessment of the effects of inter- and intra-industry trade on ρ.

On the upside, however, a dataset where trade treatment to all third parties can be taken as equal in the cross-section is a precious gift, since it helps in quantifying the extent of a potentially important bias. Anderson and van Wincoop (2002) have shown that bilateral trade flows are heavily influenced by the trade treatment each party is imposing on the rest of the world. Correcting for this “multilateral resistance” effect is crucial when investigating the determinants, as well as the impact of bilateral trade. The previous estimates of α₁ in equation (1) may have suffered from this bias, to an extent that estimates based on intranational data can help evaluate. Finally, inter-state data provides an important robustness check. Indeed: the data is coming from completely different sources, the sectoral information is more disaggregated, the sample universe is more complete, as there is data on the whole of an economic entity, and the measure of financial integration is conceptually and practically different from an index of capital account restrictions.

Yet, as Table 7 reports, the results are largely unchanged. Perhaps the largest alteration in the results pertains to the estimates for α₁, which are still significantly positive but much smaller in magnitude. Recall that, given the manner in which inter-state trade was simulated, there is no way of discriminating between inter- and intra-industry trade. Estimates for α₁ are, however, almost half those implied by international data. Doubling inter-state trade (as predicted by a gravity model) results in GSP correlations being higher by 0.021, which is well into the range implied by standard (real) international business cycles models with plausible parameters. This suggests the bias demonstrated by Anderson and van Wincoop (2002) is prevalent in the international dataset. But trade does affect significantly cycles synchronization even after it is corrected for.

The other coefficients, however, remain largely unaffected by the use of an alternate dataset. In particular: (i) states with similar specialization patterns do display significantly more synchronized GSP, and quite remarkably the estimates for α₂ in Table 7 are not significantly different from their equivalent in Tables 5 or 6. Specialization patterns continue to be an economically and statistically important determinant of business cycles synchronization. (ii) States with high risk-sharing, as measured by an index of income insurance, tend also to be more synchronized even though they also become more specialized. Both these channels remain significant in all specifications, in a way that is reminiscent of Table 6 when a measure of effective financial integration was used. (iii) Inter-state trade seems to result in less specialized states, a somewhat surprising outcome that already obtains in the international data. It is not robust, however, as the next section shows.

The channels between trade, finance, specialization and synchronization are stronger between states than between countries, although not qualitatively different. Panel C in Table 8 presents the estimated direct and indirect linkages between the four variables of interest, although no channel involving an estimate of the βs can be obtained given the simulated trade data. The most remarkable difference with panel B has to do with the significant channel from risk-sharing to ρ: financial integration results in specialization, and thus less correlated cycles. The significance of this coefficient summarizes the results in Kalemli-Ozcan, Sorensen and Yosha (2001 and 2003). But the approach here also shows the presence of a direct link between financial integration to cycles synchronization, which in the intranational data is positive at the one percent confidence level. Indeed the estimates in panel C suggest that financial integration has an overall positive effect on business cycles synchronization.

Finally, the lower panel of Table 7 presents some robustness analysis for inter-state data. In particular, three checks are performed. First, the possibility that the economic size of each state be independently and significantly affecting ρ is investigated, but without any sizable changes. Second, the Hodrick-Prescott filter is used to isolate the cyclical component of GSP when computing ρ (and the indices of risk-sharing), again without any sizable effects. Third, the GMM estimator described in section II is implemented instead of TSLS. This affects the estimates of financial integration’s specialization effect, γ₂, which then becomes non-significantly different from zero. The next section extends this sensitivity analysis to the results pertaining to international data.

A. GMM Estimates

The first specification in Table A9 implements the exact same specification as in Section III.B, but using a GMM estimator instead of Three-Stage Least Squares. The results are very similar, usually stronger both economically and statistically, with the notable exception of α₁, which is not significantly different from zero. This does not appear to be a robust fact, though, as the next sub-section argues.

B. Alternative Measure of Trade

The second specification in Table A9 contains the alternative measure of trade T² introduced in Deardorff (1998). T² is not scale dependent, which manifests itself in the sudden reversal of the size variable in I₂ (namely GDP product). This measure of trade does once again affect directly cycles correlations, with an estimate for α₁ in line with earlier results. Results pertaining to other channels are unchanged: they are, if anything, once again stronger than in the TSLS case with T¹. Note also that T² does not appear to correlate at all with specialization patterns, confirming the sensitivity of the earlier negative point estimates.

C. Hodrick-Prescott Filter

Specification (iii) in Table A9 uses the Hodrick-Prescott filter to isolate the cyclical component of GDP fluctuations. None of the results change significantly, except for the correlation between trade and specialization (β₁), which becomes insignificant here as well.

D. All Sectors

The last column in Table A9 Panel a reports estimates when S is computed using sectoral value added at the one-digit level, but for all sectors in the economy, as issued by the UN. The main difference has to do with the (much) larger magnitude of α₂, the effect of specialization on cycles correlation. The coefficient on K in equation (1) does also become significantly negative, suggesting the absence of a direct effect from K to ρ in Table 5 might have been due to focusing on manufactures only.

E. Yearly Data

The second panel of Table A9 reproduces the previous four checks, but using yearly data since 1960 to compute cycles correlations. The estimates are virtually identical to those in the first panel.

F. Monetary Policy

The first two specifications in Table A9 Panel C report estimates where convergence in monetary policies is proxied and controlled for in equations (1) and (2). Identifying monetary policy is the object of an enormous literature, whose purpose is to track the effect of exogenous monetary policy decisions over time. There is fortunately no purpose in being that ambitious in the present context. Rather, this sub-section purports to ensure that the channels identified in the previous section are not merely a manifestation of similar monetary policies. Note that this is unlikely, for, while converging monetary policies, manifested by a stable exchange rate for instance, are known to result in more trade and more synchronized business cycles, there is no obvious theoretical link with countries’ specialization patterns.⁶⁰ In other words, these controls will most likely affect the estimates of α₁ only. Furthermore, the fact that intranational results are almost identical to international ones suggests monetary policy is not driving the results in this paper.

Specifications (ix) and (x) in Table A9 control for the volatility in the (growth rate of the) nominal exchange rate, and for five-year averages of inflation differentials, respectively. Once again, the results are unchanged. It is interesting to notice that the effects of a stable exchange rate and/or small inflation differentials seem to work through trade, as trade is estimated to significantly increase in the face of stable and similar monetary policies.

G. Size

Specifications (xi) and (xii) in Table A9 report estimates once the relative size of the two economies is controlled for in equation (1). The idea behind this control is to check whether country size continues to have a direct impact on business cycles correlation, beyond the indirect channels via trade or specialization, and to verify that the results carry through. Once again, the results are unchanged: large countries are indeed more synchronized, but it is happening indirectly because they tend to trade more.⁶¹