1 Introduction

The large fiscal and external imbalances experienced by many OECD countries (among which the US and Japan) have raised the attention of the theoretical and applied literature on the determinants of current account dynamics. As a consequence, a number of empirical studies have investigated the causal relation between the public and the external deficit. Without attempting an exhaustive survey of this large body of literature, which goes back at least to the 1970s, it is fair to say that the most recent empirical analyses agree broadly on the following three points:

  1. 1)

    There is a causal relationship running from the budget to the external deficit as implied by most standard macroeconomic models.

    A direct causal nexus between the public and the current account deficits stems in a number of models, ranging from the simple “classical open economy” model (Mankiw 2002, chap. 5), to microfounded overlapping generations models (see for instance Obstfeld and Rogoff 1995, par. 3.1). Limiting ourselves to the more recent studies, this nexus is empirically proved for the US by Leachman and Francis (2002) using multicointegration tests (Engsted et al. 1997), and by Hatemi and Shukur (2002) using multivariate Granger (1969) causality tests, among others.Footnote 1

  2. 2)

    The relation between the budget and the external deficit may differ in the short- and in the long-run and depends on the long-run properties of the series involved.

    This conclusion is reached with different emphasis in a number of intertemporal explanations of the current account behavior. For instance, Normandin (1999), using a tractable version of Blanchard’s (1985) model, shows that the degree of persistence of the budget deficit affects the strength of the twin deficits relation: a persistent pattern of the budget deficit implies that the representative consumer expects current public deficits to be followed by future deficits, hence by future tax reductions; this will lead the consumer to finance a current consumption increase through a current account deficit. Kraay and Ventura (2002), in a different setting, show that the relation between national savings shocks (such as those implied by a fiscal policy change) and current account behaviour is much stronger in the long than in the short run.

  3. 3)

    However, the long-run relation appears to be weak and/or subject to structural breaks and emerges more clearly once these breaks are taken into account.

Most studies recognize the importance of structural breaks on the basis of a priori arguments, without doing formal tests. For instance, Leachman and Francis (2002) argue that the exchange rate regime plays a crucial role in the transmission mechanism between the two deficits and therefore split their sample in 1973, at the end of the Bretton Woods era; conditional on this change point they find cointegration between budget and external deficit only in the second subsample, indicating that the US budget deficit has contributed to the external deficit in recent times only. On the contrary, Mann (2002) observes that the boom in US investments caused by the “new economy” in the early 1990s has decoupled the budget and the external deficit by driving a wedge between private investment and savings, thus weakening the twin deficits relation. In fact, from 1993 to 2000 the US budget deficit has shrunk while the external deficit has moved in the opposite direction.Footnote 2

Building on this informal analysis, Fidrmuc (2003) studies the twin deficits relation in a sample of 18 OECD and transition economies by applying Johansen (1988) cointegration test to quarterly data from 1970 to 2001 conditional on a structural break in 1989, and finds significant differences in the long-run parameters between the first and the second subsample; to be more specific, Fidrmuc (2003) shows that in a large number of countries (including the US) cointegration does not hold in the second subsample. A similar result is reached, in a cross-section setting, by Obstfeld and Rogoff (1995), who regress the current account balance-to-GDP ratio on the public deficit-to-GDP ratio using a cross section of OECD countries. They construct three cross-sections using average data for the 1976–1980, 1981–1985 and 1986–1990 subsamples and find that in the most recent subsample (which overlaps the second subsample of Fidrmuc 2003) the correlation “evaporates”.

A very limited number of studies relies on econometric testing for identifying the date of the structural change: Hatemi and Shukur (2002) utilize the CUSUMSQR test of Brown et al. (1975) on a bivariate VAR representation of the US twin deficits and find a structural break in 1989. They show that when this structural change is ignored, neither variable Granger-causes the other; on the contrary, once the sample is split in 1989:4, the Granger causality test indicates an inversion of the causal nexus, which runs from the public to the external deficit in the first subsample, and from the external to the budget deficit in the second subsample.

This conflicting evidence stresses the need for a more thorough empirical analysis. In this paper we argue that these contradictions could be explained at least in part by the fact that most previous studies resort to ad hoc arguments, rather than formal testing, to locate the structural breaks in the twin deficits relation. To investigate this issue we test the twin deficits relation on an extended span of annual data (from 1960 to 2005), considering a larger number of countries (22 OECD countries for which the data were available), and using appropriate econometric techniques for detecting shifts of unknown date in both the long- and the short-run parameters of the twin deficits relation. The constancy of the long-run parameters has been investigated using the cointegration estimator by Gregory and Hansen (1996), that allows for the presence of structural breaks in the cointegrating vector; the constancy of the short-run parameters has been tested using the sup-F test by Andrews (1993). As mentioned above, both tests verify the null hypothesis of constancy against the alternative of a structural break at an unknown point in the sample.

As Gregory and Hansen (1996) show, ignoring the presence of a structural break in the long-run relation may lead a researcher to accept the null hypothesis of non-cointegration between the model variables even when a long-run relation does actually exist. Consider now for instance the study of Fidrmuc (2003): even if 1989 were a true change point date for the US (as Hatemi and Shukur 2002, confirm), there is obviously no reason for extending this finding to the other OECD countries without any further analysis. We may therefore wonder whether the finding of non-cointegration in the second subsample of Fidrmuc’s data set could actually depend on an inappropriate selection of the change point. A similar remark applies to the cross-section analysis in Obstfeld and Rogoff (1995).

The remainder of the paper falls in five sections. Section 2 derives from national accounting identities the regression model utilized for the analysis of the twin deficits relation. Section 3 illustrates the econometric methodology. Section 4 summarizes the main results. Section 5 concludes. An appendix describes the data sources.

2 Twin deficits model

The twin deficits model rests on the national account identity

$$ Y^{N} = C + I + G + NX + NFI $$
(1)

where Y N is the gross national product, C private consumption, G government consumption, NX net exports and NFI the net factors incomes from abroad. The sum of the last two items defines the current account balance \({CA = NX + NFI}\). Taking it to the left-hand side of Eq. (1) and recalling the definition of national savings we have

$$ CA = S - I $$
(2)

The current account deficit is therefore equal to the excess of national investment over savings: when a country’s investment exceeds its savings, the difference must be financed from abroad. If a country is running a current account deficit, there must be at least another country running a surplus, i.e., a country which invests abroad the excess of its savings over its investment. In principle, less developed countries that are catching up will generally feature current account deficits, while mature economies will show surpluses (Blanchard and Giavazzi 2002).

Equation (2) lies at the heart of the so-called “quantity approach” to the measurement of international financial integration (Lemmen and Eijffinger 1995). This approach originates from the study of Feldstein and Horioka (1980), who test the hypothesis β = 1 in the cross-section regression

$$ i_{h} = \alpha + \beta s_{h} + u_{h} $$
(3)

where i h is the gross investment-to-GDP ratio and s h the gross savings-to-GDP ratio in country h, and u h a well-behaved stochastic disturbance. The authors are unable to reject the hypothesis β = 1 and argue that this is inconsistent with financial integration, as it implies that each increase in domestic savings is employed mostly for financing domestic investment, instead of addressing the best investment opportunities worldwide. This puzzling evidence has promoted a large debate (for a recent review see Coakley et al. 2003). For our purposes it is of some interest to recall the contribution of Gundlach and Sinn (1992). By comparing Eqs. (3) and (2) they remark that under β = 1 (capital “immobility”) the current account-to-GDP ratio, ca t , will be generated by a stationary process with mean −α.Footnote 3 Therefore, the null hypothesis of capital mobility can be verified by performing a unit root test on ca t : non-rejection of the null implies that a country is linked to the international capital markets.

Equation (2) considers two sources of financial capital, an internal one (national savings), and an external one (the current account deficit), but only one use: domestic investment. A more detailed statement of the sources and uses of financial capital is obtained by distinguishing between private and public savings. This can be done by subtracting from both sides of (1) the net direct taxes T. Upon rearranging we get:

$$ S^{{\text{P}}} - CA = I - S^{{\text{G}}} $$
(4)

where \( S^{P} = Y^{N} - T - C \) is private savings, and \( S^{{\text{G}}} = T - G \) is government savings (i.e., the budget surplus). The left-hand side of (4) displays the two main sources of financial capital, namely private savings and the current account deficit, while the right-hand side displays its uses: private investment and public deficit.

Equation (4) can be rearranged as follows:

$$ CA = S^{{\text{G}}} + S^{P} - I $$
(5)

Equation (5) shows that if domestic investment is financed almost entirely by private savings (S P ≈ I), then the current account and government balance must move together by arithmetic (i.e., they are “twins”). Besides this extreme case, there are a number of possible intermediate cases, where the financing needs of the public sector are satisfied partly by the domestic and partly by the international financial markets. Starting from the extended relation (5), Fidrmuc (2003) assesses the occurrence and the intensity of the twin deficits phenomenon by estimating the following equation:

$$ ca_{t} = \beta _{0} + \beta _{1} s^{G}_{t} + \beta _{2} i_{t} + u_{t} $$
(6)

where as before small caps indicate ratio of the relevant variables to GDP and u t is a disturbance, and we expect β 1 > 0 and β 2 < 0. Equation (6) is the starting point of our empirical analysis.

In order to frame the discussion of the empirical results, it is useful to point out some potential sources of structural change in the parameters of Eq. (6). In the following pages we will focus on three possible sources: shifts in the average propensity to save, changes in the degree of financial integration and changes in the net foreign position of a country.

The average propensity to save is often taken as constant in the theoretical models underlying the twin deficits issue. From an empirical point of view, however, we cannot exclude that the consumption/saving behaviour of the private sector will undergo structural changes in the long run.Footnote 4 By comparing Eqs. (6) and (5) we see that such a structural change will generally appear as a shift in the intercept of Eq. (6). Moreover, using a medium-run macroeconomic model Makin (2004) shows that the response of the external deficit to a change in the fiscal deficit depends on the private propensity to save as well as on the fiscal policy mix. Therefore, depending on the fiscal policy mix implemented in a given country, a shift in the propensity to save can also affect β 1.

As for the degree of financial integration, we remark that a necessary condition for the twin deficits to emerge is that the public sector is allowed to borrow from abroad. To put it in another way, in the absence of capital mobility the deficits cannot be “twins”, and the parameters of Eq. (6) will be close to zero. This suggests that major innovations in the financial liberalization process of a country are another possible source of structural change in these parameters. In particular, we expect them to increase in absolute value as a consequence of increased financial integration.

Finally, Kraay and Ventura (2002) show that in an intertemporal model with adjustment costs the representative agent, faced with transitory income shocks, utilizes foreign assets as a buffer stock. As a consequence of this behaviour, correlation between current account and investment should be negative for (net) debtor and positive for (net) creditor countries (see Section 4 in Kraay and Ventura 2002). We expect therefore the parameter β 2 to change its sign along with the net foreign assets of a country.

3 Empirical analysis

Equation (6) was estimated using time series data on the current account-to-GDP ratio, ca t , the government balance-to-GDP ratio, \( s^{G}_{t} \), the private investment-to-GDP ratio, i t , ranging from 1960 to 2005 (positive values of \( s^{{\text{G}}}_{t} \) indicate a government surplus). The analysis was performed for the 22 OECD countries for which the data were available: Australia (AUS), Austria (AUT), Belgium (BEL), Canada (CAN), Switzerland (CHE), Germany (DEU), Denmark (DNK), Spain (ESP), Finland (FIN), France (FRA), United Kingdom (GBR), Greece (GRC), Ireland (IRE), Italy (ITA), Japan (JPN), the Netherlands (NLD), Norway (NOR), New Zealand (NZL), Portugal (PRT), Sweden (SWE), Turkey (TUR), and the United States (USA); see the Appendix for a detailed account of the data definitions and sources. The data are expressed in percentage points and represented in Figs. 1, 2 and 3.

Fig. 1
figure 1

The time series plot of the current account, CA, the government balance, SG, and private investment, I (right-hand scale), expressed as a ratio to GDP, for Australia, Austria, Belgium, Canada, Switzerland, Germany, Denmark and Spain

Full size image
Fig. 2
figure 2

The time series plot of the current account, CA, the government balance, SG, and private investment, I (right-hand scale), expressed as a ratio to GDP, for Finland, France, the United Kingdom, Greece, Ireland, Italy, Japan and the Netherlands

Full size image
Fig. 3
figure 3

The time series plot of the current account, CA, the government balance, SG, and private investment, I (right-hand scale), expressed as a ratio to GDP, for Norway, New Zealand, Portugal, Sweden, Turkey and the United States

Full size image

The time series of ca t , \( s^{{\text{G}}}_{t} \) and i t were first tested for unit roots using the ADF test of Dickey and Fuller (1979). The order of lags in the ADF regression was selected by a model reduction procedure as suggested among others by Enders (2004), while the structure of the deterministic component of the underlying process was specified following the general-to-specific approach suggested by Dolado et al. (1990).

The existence of a long-run relation between these variables was then tested by the usual Engle and Granger (1987) cointegrating residual ADF (CRADF) test in the long-run Eq. (6). The order of the CRADF test was selected looking at the Q statistics of the auxiliary regression residuals.Footnote 5 When the ordinary cointegration test failed to reject the null of non-cointegration, we hypothesized that the non-rejection could depend on the presence of a structural break in the long-run parameters and re-estimated Eq. (6) using the cointegration estimator proposed by Gregory and Hansen (1996), which tests the null of non-cointegration against the alternative of cointegration in the presence of a structural break of unknown date. The breaks are modelled using the dummy variable \( \varphi _{{\tau t}} = I{\left( {t > {\left[ {T\tau } \right]}} \right)} \), where I is the indicator function, T is the sample size (T = 46), τ the relative timing of the change point, and [.] the integer part function. The null of non-cointegration has been tested against two kinds of breaks: the first one is an intercept or “level” shift

$$ ca_{t} = \beta _{0} + \mu _{0} \varphi _{{\tau t}} + \beta _{1} s^{G}_{t} + \beta _{2} i_{t} + u_{t} $$
(7)

while the second one is a “regime” shift

$$ ca_{t} = \beta _{0} + \mu _{0} \varphi _{{\tau t}} + \beta _{1} s^{{\text{G}}}_{t} + \mu _{1} \varphi _{{\tau t}} s^{{\text{G}}}_{t} + \beta _{2} i_{t} + \mu _{2} \varphi _{{\tau t}} i_{t} + u_{t} $$
(8)

where the β i (i = 0, 1, 2) indicate the values taken in the first subsample, \( \varphi _{{\tau t}} \) is the shift dummy variable defined before, the μ i (i = 0, 1, 2) are the parameter shifts, so that the parameters in the second subsample can be defined as \( \beta _{{i2}} = \beta _{i} + \mu _{i} \), and u t is the cointegrating residual. When μ i  = 0 (i = 0, 1, 2) Eqs. (7) and (8) reduce to Eq. (6). In the “level shift” model (7) only the intercept undergoes a structural break shifting by an amount μ 0 starting in t > [Tτ], while in the “regime shift” model (8) all the parameters of the long-run equation are allowed to change.

The test statistic is evaluated as \(ADF_{r} ^{*} = {\mathop {\inf }\limits_\tau }ADF_{r} {\left( \tau \right)}\), where ADF r (τ) is the cointegrating ADF statistic calculated using the OLS residuals in model r (r = L, R, where “L” indicates the “level shift” and “R” the “regime shift” model).Footnote 6 In other words, ADF r * is the smallest among all the ADF statistics that can be evaluated in model r across all possible dates of structural breaks. As we generally had no a priori information on the shape of the relevant alternative, we calculated the ADF r * statistics for each of the two models L, and R. Where the null of non-cointegration was rejected in favour of more than one alternative, we chose either the model corresponding to the more significant statistic, or that with the more meaningful parameters from the point of view of economic theory. The break date T 1 = [τT] reported refers to the last year of the first regime (i.e., the change occurs between T 1 and T 1 + 1).

At this stage, in all but a limited number of cases we obtained a cointegrating residual coming from the estimation of Eq. (6), or (7), or (8). This residual was then utilized in the short-run error correction model:

$$ \Delta ca_{t} = \delta _{0} + \delta _{1} \Delta s^{G}_{t} + \delta _{2} \Delta i_{t} + \delta _{3} \widehat{u}_{{t - 1}} + \varepsilon _{t} $$
(9)

Equation (9) was tested for non autocorrelation, homoskedasticity and normality of the residuals using the standard Lagrange multiplier test statistics (Engle 1984), as well as for constancy of the short-run coefficients using the sup-F test by Andrews (1993). The sup-F test verifies the null hypothesis of parameter constancy against the alternative of a structural break at an unknown point of the sample. Its statistic is the maximal F statistic of the Chow (1960) tests performed over any possible sample split of the kind 1,..., T 1*; T 1* + 1,..., T, with T 1* = m,..., T − m, where m is a “trimming” value.Footnote 7 The critical values are provided by Andrews (1993, Table I) and depend on the number of the regressors and on the trimming parameter; for a model with four regressors and a trimming value equal to 10% of the sample the 5% critical value is equal to 16.98. The parameters of Eq. (9) have been tested for significance using their t-statistics and the model has been reduced according to the response of the significance tests.

We have therefore five possible outcomes for our specification procedure: non-cointegration, cointegration with no structural change, cointegration with shifts only in the long-run parameters, or only in the short-run parameters, or in both.

4 Results

4.1 Unit root tests

Table 1 reports the results of the unit root tests on the levels of ca t , \( s^{{\text{G}}}_{t} \), and i t . The ADF carried out on the first differences of the series (not reported) always strongly rejected the null of unit root. Therefore, the order of integration of our time series does not exceed one. Some time series however appear to be generated by I(0) processes. In fact, the unit root null is rejected in a limited number of cases. The analysis of these cases allows us to draw some conclusions on the existence of “twin deficit” behaviour before estimating Eq. (6). To this purpose, recall first that while a linear combination of I(1) variables can be I(0) (if they are cointegrated), a linear combination of I(0) variables cannot be I(1); second, that for a regression model to be balanced, the left-and the right-hand side must be of the same order of integration (Maddala and Kim 1998).

Table 1 Results of the unit roots tests on the model variables
Full size table

Consider first the case of Switzerland, where the current account is I(1), but the regressors are both I(0). Equation (6) is therefore unbalanced; this means that the trending pattern of ca t (apparent in Fig. 1) cannot be explained by either regressors or a combination of them. This result rules out the existence of a twin deficits phenomenon in the long-run.

In three other cases (Germany, Denmark and Sweden) the government balance appears to be I(0), while the current account deficit and investment are still I(1). In these cases the long-run pattern of the current account balance appears to be dominated by the stochastic trend in the investment series: this is most apparent for Sweden, where a long-run current account improvement is associated with a relatively steady decline in the investment ratio (see Fig. 3). In this as well as in the other two countries we can therefore rule out the existence of long-run “twin deficits” behaviour.

In three more cases (Italy, Spain and Turkey) the dependent variable ca t appears to be I(0), while both the regressors are I(1). In these cases Eq. (6) may still be balanced, provided that the regressors are cointegrated, resulting in a stationary linear combination. Irrespective of the statistical properties of the equation, from an economic point of view, this outcome indicates either long-run crowding-out or current account targeting, depending on the direction of causality between the two series: in both cases, this rules out the existence of a long-run relation between the government and the external deficit. This is especially evident in the case of Italy (see Fig. 2), where the long-run correlation between the investment ratio and the government balance is very strong at least until 1992.

On the contrary, in another case (Canada) the investment series turns out to be generated by an I(0) process, while the government and current account balances are I(1). This implies that the stochastic trend in the current account balance must be generated by the unit root in the government balance: we have therefore strong evidence of twin deficits.

Finally, in the other 14 countries both the dependent variable and the regressors appear to be I(1); therefore, the analysis of the orders of integration provides us with no prior information on the existence and strength of the twin deficits behaviour.

Summing up the results, in seven out of 22 cases the unit root tests allow us to rule out the existence of a “twin deficits” phenomenon. This occurs, for different reasons, in Denmark, Germany, Italy, Spain, Sweden, Switzerland and Turkey. In one of the remaining 15 countries (Canada) the unit roots tests imply the presence of twin deficits (although the degree of the phenomenon is still an open question), while in the other 14 countries we cannot reach a conclusion before estimating Eq. (6).

4.2 Long-run estimates

We proceed therefore to the estimation of Eq. (6) for the 15 countries in which we cannot rule out the existence of twin deficits.Footnote 8 Table 2 reports the results of the cointegration tests along with the estimates of the long-run coefficients. The first column reports the CRADF test. When this statistic rejects the null of non-cointegration there is a stable long-run relation between the three variables throughout the sample. This applies in particular to Australia, Austria, and the United States.

Table 2 Estimates of the long-run equation
Full size table

The evidence for Austria seems to imply the absence of twin deficits, as β 1 is close to zero and statistically insignificant. However, the Gregory and Hansen test results in a 10% significant statistic equal to −4.85 for a level shift in 1966. This accounts for a large shift in private savings that explains the current account improvement taking place between those two decades.Footnote 9 Conditional on this shift, we obtain an estimate of β 1 equal to 0.22 (t-statistic 2.0) and of β 2 equal to −0.61 (t-statistic 3.9).

The estimates for the United States are especially interesting because, as mentioned in the Introduction, a number of previous studies have either found or imposed a structural break in the US twin deficits relation. The absence of structural breaks in our sample depends basically on the fact that we analyzed a large enough sample of data. As Fig. 3 shows, the increase in the investment ratio determined by the “new economy” bubble has actually decoupled the two deficits during the Nineties; however, after the bubble burst, and the investment ratio declined, the public and external deficits started to move together again (Roubini and Setser 2004). As of 2005, the decoupling of the two deficits determined by the increase in investment during the Nineties can be interpreted as a transitory phenomenon, an unusually large swing on a general downward common trend. Imagine however looking at the (possibly quarterly) data from 1980 to 2000: then it would be much harder not to consider 1989 as a persistent change point from a subsample of positive to a subsample of negative correlation between the two deficits. This could explain why Leachmann and Francis (2002) obtain a weak empirical relation between budget and current account deficit in their second subsample (1974–1992).

The second and third column report the Gregory and Hansen (1996) cointegration tests, respectively for a level shift (Eq. (7)) and for a regime change (Eq. (8)), along with the date of the structural break (i.e., the last year of the first regime). When neither statistic is significant we conclude that there is no significant long-run relation between the variables involved in Eq. (6). This happens in two countries only: the United Kingdom and Portugal.

4.3 Interpreting the structural changes

In the remaining ten countries (Belgium, Canada, Finland, France, Greece, Ireland, Japan, the Netherlands, Norway and New Zealand) the conventional CRADF statistics do not reject the null, while applying the Gregory and Hansen estimator we find a strong long-run twin deficits relation in at least one subsample, the only possible exceptions being the Netherlands, where the β 1 coefficient is close to zero and insignificant throughout the sample, and Japan, where β 1 has the wrong sign in the first subsample.

In a number of cases the endogenous determination of the change point reverses or qualifies previous empirical results. For instance, by imposing an arbitrary change point in 1989 Fidrmuc (2003) finds non-cointegration for Canada and Finland,Footnote 10 and obtains a wrong (negative) sign in the coefficients of public savings for Australia, Canada and Finland, a wrong (positive) sign in the coefficients of private investments for Australia and Canada, and an implausible 16.7 long-run coefficient for public savings in France. The estimates reported in Table 2 fall in a more plausible range and display the correct signs.

As stated in Section 2, recent economic theory points out a number of possible sources of structural breaks in Eq. (6). We mentioned, among others, the presence of structural changes in private consumption behaviour (Makin 2004), changes in the net foreign position (Kraay and Ventura 2002), and more generally changes in the degree of financial integration. It should be stressed that these events are themselves intrinsically difficult to date: for instance, the graphs reported by Milesi-Ferretti and Lane (1999) show that the switch in the sign of the net foreign assets of a country may be located at quite distant dates, depending on the method used to construct the data. Moreover, different sources of structural change may coexist at different dates in the sample, leading to multiple structural breaks. The estimator utilized allows for only one structural break and may lead to biased estimates of both the change point dates and the parameters whenever the underlying data generating process undergoes multiple structural breaks. However, the recent testing procedures for multiple structural breaks do not allow for I(1) regressors (Bai and Perron 1998) and their application would be unwarranted in this context.

In order to verify the robustness of our results with respect to the possible occurrence of multiple breaks we performed a sensitivity analysis by reestimating Eq. (6) on a subsample of 31 observations starting in 1975. Because of sample “trimming”, the search for structural breaks takes place actually in the 1980–2000 range, skipping the two oil-price shocks, widely seen as a major source of structural changes in world economies. This should help to single out any possible structural change located in the second half of the sample.

The results of the sensitivity analysis, reported in Table 3, generally confirm the robustness of the full sample estimates. In five out of ten structural break regressions the date of the change point is confirmed exactly (this happens in Belgium, Canada, Finland, Greece and Norway) and in two other countries it changes only slightly (Japan and New Zealand).Footnote 11 In the case of Ireland, once the first structural break (occurring in 1973) is dropped from the sample, the CRADF confirms the presence of cointegration without a structural break in the second subsample (thus supporting the hypothesis of a single structural break).Footnote 12 The results for the United Kingdom point out the possible presence of multiple structural breaks: while in the full sample the null of non-cointegration cannot be rejected, in the reduced sample there is evidence of cointegration with level or regime shift in 1996. The associated estimate of β 1 however has the wrong sign and is only marginally significant.

Table 3 Sensitivity analysis on the long-run structural breaks
Full size table

Keeping in mind these qualifications, we conclude that the timing of structural changes is generally robust and broadly consistent with a priori information, in that most structural changes are associated with either a switch in the external position (as in Belgium, France, Japan), or with a large shift in the average propensity to consume (as in Canada, Finland, Norway).

4.4 Short-run estimates

Table 4 reports the estimates of the short-run Eq. (9). Two features emerge from the results. First, the impact coefficient of the budget deficit is often insignificant, thus showing that the current account adjusts more slowly to shocks originated in the public sector. Second, the sup-F test by Andrews (1993) never rejects the hypothesis of constancy of the short-run parameters.

Table 4 Estimates of the short-run equation
Full size table

5 Conclusions

Using time series data over the last four decades we analyze the impact of structural breaks in the twin deficit relation in 22 OECD countries. To this end we estimate the long- and short-run relation between the current account, government budget and investment, and we assess the presence of structural changes of unknown date in both relations. Summing up the results, in seven countries (Denmark, Germany, Italy, Spain, Sweden, Switzerland and Turkey) the existence of twin deficits can be ruled out by analyzing the orders of integration of the relevant time series; in another country (Portugal) we are unable to find any evidence of cointegration between the model variables; in Austria and the Netherlands the impact of the budget on the external deficit is statistically insignificant. In the other twelve countries we are able to identify a long-run twin deficits relation, and in ten out of these twelve countries the relation becomes statistically significant only once the presence of structural breaks is taken into account. These results confirm that ignoring the structural breaks, or imposing them on the basis of a priori arguments, hides in general the existence of a long-run relation or leads to biased estimates of its parameters.

Since the techniques adopted in this paper account for only one change point, we tested the robustness of the empirical results by performing the tests on a subsample excluding the two oil-price shocks. The estimated change points and the associated parameters are generally robust to this change, and there is only very limited evidence of multiple breaks.

The average long-run coefficient of government budget is slightly below 0.4, i.e., the link between the government and the external deficit is not one for one, as implied by the standard “classical open economy” model, but falls generally between zero and one, as in the medium-run open economy model of Makin (2004). The coefficients of investment are generally larger, around 0.5 on average, and become generally even larger after the structural break. This happens in particular in some countries belonging to the Euro area (Greece, Ireland), thus supporting the view expressed in previous empirical research that European integration has significantly raised the degree of financial openness of the participating countries, and in particular of the “periphery” economies.Footnote 13 By considering the impact of structural breaks, we are able to show a weakening of the Feldstein–Horioka result also in other leading non-EU economies (among which the United States and Japan).Footnote 14

The results obtained appear to be consistent with both economic theory and intuition. From a methodological point of view, they confirm the opinion expressed among others by Hatemi and Shukur (2002) that empirical research still pays very little attention to the possible structural breaks in the economic process. From an economic point of view they confirm that the government deficit is a significant determinant of the external deficit and that domestic investment in OECD countries is increasingly financed by resorting to foreign savings. This increased degree of financial openness is shown to hold not only in EU countries, but also in other leading G7 economies: a result that, while broadly in accordance with direct evidence on “globalization”, did not always feature in applied economic research.