Choice of exchange rate regime and currency zones

International Review of Economics and Finance 17 (2008) 436 – 456 www.elsevier.com/locate/iref

Choice of exchange rate regime and currency zones ☆ Isamu Kato a,1 , Merih Uctum a,b,⁎ a

Program in Economics, The Graduate Center, City University of New York, 365 Fifth Avenue, New York, NY 10016-4309, United States b Economics Department, Brooklyn College of the CUNY, 2900 Bedford Avenue, Brooklyn, NY, 11210, United States Received 15 April 2006; received in revised form 13 November 2006; accepted 8 January 2007 Available online 21 February 2007

Abstract We investigate the choice of exchange rate regimes in different currency zones (the US dollar, Euro and the CFA zones), and geographic regions (Latin America and Caribbean, East Asia and Pacific, Europe, core Europe, and the CFA countries). We control for country and regional heterogeneity, time dummies, endogeneity and perform various robustness checks. Results from regional analysis substantially differ from the aggregate analysis despite controlling for random effects. Even at the regional level controlling for currency zones affects our findings. Regional results are generally robust to regime measurement, and sample changes (number of observations). © 2007 Elsevier Inc. All rights reserved. JEL classification: F31; F33; F21 Keywords: Exchange rate regimes; Discrete choice model; Optimum currency area; Currency crises

1. Introduction After the dollar crisis that led to the collapse of the Bretton Woods system in the early 1970's, several industrial countries abandoned their fixed exchange rate regimes and shifted to floating rates. Since then, the choice of exchange rate regimes has been the subject of a lively debate in international finance. To this day there is still no consensus over issues such as the optimal choice of regimes, their determinants, and whether regimes are sustainable or not. One reason for this inconclusiveness is, of course, lack of consensus on a specific model to be used in the empirical work. Another reason is the diversity in the cross-sectional and time series samples used by various studies. In the model specification, few studies control for unobservable country specific effects and none controls for regional currency effects explicitly, leading to model misspecifications. It is a well known fact that ignoring countries' different characteristics leads to misleading inference. Likewise, economies also belong to different groups and exhibit

☆ The authors would like to thank Mike Wickens and Denis Bolduc for many helpful suggestions, and the participants in the Winter 2004 Meeting of the Econometric Society, San Diego, January 3–5 and the Seminar in Applied Economics at the Graduate Center of CUNY for useful comments on an earlier version of this paper and two anonymous referees whose comments improved the paper considerably. ⁎ Corresponding author. Department of Economics, Brooklyn College and the Graduate Center of the City University of New York, 2900 Bedford Avenue, Brooklyn, New York 11210, USA. Tel.: +1 732 549 5252. E-mail addresses: [email protected] (I. Kato), [email protected], [email protected] (M. Uctum). 1 Tel.: +1 718 349 7893.

1059-0560/$ - see front matter © 2007 Elsevier Inc. All rights reserved. doi:10.1016/j.iref.2007.01.004

I. Kato, M. Uctum / International Review of Economics and Finance 17 (2008) 436–456

437

group characteristics. More specifically, a number of currency zones exist in the world, and many currencies are linked to currencies other than the US dollar. Shocks that affect the euro area for example, are not necessarily the same as the ones hitting the dollar area because the anchor currency country is likely to have a greater impact on member countries of the currency zone than the United States. The existing studies on exchange rate regime determination lump all foreign variables into a single “foreign country”, which consists of either the United States or an average of the OECD countries. This methodology implicitly assumes the US economy or the OECD is the key external factor for all currencies. Not recognizing this currency-zone effect is likely to bias the results by introducing cross-sectional dependence into the analysis. This paper attempts at filling this gap by investigating whether these two types of heterogeneity play a role in the determination of the exchange rate regimes. The first objective of this paper is, therefore, to decompose the choice of exchange rate regime into components attributable to the effects of observed variables and unobserved heterogeneity. We find that in most cases, controlling for country heterogeneity affects estimates. The second and more important objective goes one step further and aims to explore the differences attributable to different currency zones. This represents a major departure from the existing studies because our analysis covers three distinct currency zones. The currency zones we consider consist of countries that tie their currencies to the US dollar, those that anchor to the ECU/ euro, and the countries belonging to the franc zone of the African financial community, the CFA franc zone, which linked their currencies to the French franc and now to the Euro. For each zone, we compute the foreign variables with respect to the anchor currency country's variables.2 Our findings show that results controlling currency zone/regional analysis can substantially differ from the aggregate analysis even when country-specific effects are accounted for. The second reason that contributes to inconclusiveness in the literature is the variety of period averages used in the existing studies. These analyses are based on period averages, and treat the choice of the exchange rate regime as permanent throughout the sample period. Although this methodology is believed to help avoiding business cycle effects and biases due to autocorrelation and simultaneity problems, it is undesirable for several reasons. First, it throws away valuable information and eliminates the dynamic interaction of different factors. Over a period of 15–20 years, very few countries stick to their choice, and a majority changes it more than once to respond to domestic and foreign shocks to the economy. As a result of this fact, empirical results also change depending on which period average is used and where the cutoff point is set. Second, it is not clear that even the business cycle effects are eliminated since the cutoff point for time averaging is arbitrary and cycles may last through any two time intervals. To capture the change in countries' choice of exchange rate regime and control for factors affecting the choice, we favored using a larger data span than typical in these studies, and keep its time dimension. This allows us to take into account both the time series characteristics of the data as well as their cross-section aspects. We perform several robustness tests and check the sensitivity of the results to factors such as regime categorization (de facto versus de jure), sample specification, and also time series characteristics such as regime persistence, multicollinearity, and endogeneity. Our findings indicate significant differences between pooled data, which does not differentiate between currency zones, and individual currency zones. Country size is in general robust across the groups and results show that as countries grow, they tend to choose more flexible rates. Exchange rate volatility is a significant determinant in Europe, Latin America and Caribbean (LAC), and in the CFA countries where it increases the probability of choosing relatively flexible regimes. It is significant in East Asia and Pacific (EAP) countries only if the Asian crisis is controlled for. Openness and inflation unambiguously affect the regime choice in LAC, CFA, the EAP regions and inflation in CFA and the LAC regions. This is in contrast to pooled sample results where they are not robust to sample change. Finally, although we find a general move towards relatively flexible rates, in both Latin America and Europe, the end of the 1990s are marked by the return of relatively fixed regimes. Our findings thus indicate that full sample results can mask regional differences and should be dealt with cautiously, in particular by controlling for geographical/regional idiosyncrasies.

2 Recently some studies correct for cross-regional effects to explain economies' responses to global shocks. However, this methodology presents two inconveniences. Conceptually, it may not capture the monetary and financial constraints imposed on a country for belonging to a currency zone or a geographical area as does our approach. In practice, unlike for continuous time data, the unavailability of an established methodology in discrete-time data makes it difficult to address this problem econometrically. We explain below an alternative methodology we resort to capture any additional regional dependence left in our analysis.

438

I. Kato, M. Uctum / International Review of Economics and Finance 17 (2008) 436–456

2. Conceptual framework Although there is no well defined model used in the empirical literature on exchange rate regimes, we consider the most common variables that are consistent with the OCA model, which emphasizes the role of economic characteristics of a country in the determination of the choice of the regimes: trade openness, financial openness (size of capital transactions), the economy's size, patterns of international trade, inflation differential.3,4 Following the most recent studies, we also check whether inclusion of institutional variables affect the results. We can summarize the model as follows: Yit⁎ ¼ b Vxit þ hVwit þ ui þ vit

for i ¼ 1; 2; N ; N ; t ¼ 0; 1; 2; N ; T ;

ð1Þ

where Yit⁎ is the exchange rate regime, xit is a vector of macroeconomic and financial variables relevant for the OCA model, wit is a vector of institutional variables, β, θ are vectors of coefficients for the traditional and institutional variables, respectively, and ui is a country specific random component that is constant across time with ui ∼ N(0,σu), vit is a normally distributed error term, vi ∼ N(0,σv) and σv = 1. 2.1. Dependent variable Following the earlier literature, which drew on various methods such as discriminant analysis (Heller, 1978), flexibility index (Holden, Holden, & Suss, 1979), the consensus now is to use discrete variables. The latter consist of the following categories: two regimes with fixed and flexible rates (Bosco, 1987; Dreyer, 1978; Savvides, 1990), three regimes with fixed, intermediate, and float (Bosco, 1987; Rizzo, 1998; Savvides, 1990; Poirson, 2001), and four or more regimes with single-currency peg, basket peg, crawling peg and float (Juhn & Mauro, 2002; Melvin, 1985). The IMF exchange rate classification (1983–1998) broadly divides the exchange rate regimes into four categories: fixed, flexibility limited (crawling peg), managed float (dirty float), and independent float. For our dependent variable, we consider three regimes following Masson's (2000) categorization, and define the two middle categories as an “intermediate” regime.5 We index the fixed/pegged regime, the intermediate regime and the floating regimes respectively as 1, 2, and 3. Recently, several studies relied on de facto exchange rate regime categorization instead of de jure (official)6. In general, each de facto approach classifies the exchange rate regimes according to a measure of exchange rates (official/unofficial), and/or fundamental variables, such as reserves. This classification presents several limitations, such as narrow country coverage, and endogeneity of the quantitative measures, which are affected by other economic and political factors. In this paper we use de jure classification from the IMF's Exchange Rate Arrangements and Annual Restrictions annual report for several reasons. First, so far no particular de facto measure has been widely adopted as a valid representation of the actual exchange rate regimes. Second, conducting the analysis using the traditional IMF categorization allows assessing of our findings with respect to the body of research that relied on the same measure. Last, but not least, we compared our results with those obtained from Bubula et al.'s de facto measure and found no significant major difference.

3

See Juhn and Mauro (2002) for a comprehensive survey of the literature on OCA models of exchange rate regimes. Besides variables in the OCA literature, a number of researchers include variables reflecting a country's external or internal financial condition such as changes in external debt or domestic credit with the aim of examining mainly currency crises and collapse. To include such variables in our analysis is problematic for two reasons. First, the currency crises literature mainly discusses the sustainability of fixed or managed exchange regimes and the variables are chosen to explain the speculative attack on a currency. However, the speculative attack does not always cause an exchange regime transition. In other words, the choice of exchange regime does not have a direct correlation with the speculative pressure on currencies, and with the factors causing it. Second, from a statistical point of view, these high volatility financial variables are better suited to analyze a continuous time-series dependent variable, such as interest differentials or market pressure indices, as opposed to discrete time dependent variable with a large cross-sectional dimension. 5 The latest IMF classification (1999) adopts a more detailed categorization of regimes: 1) Exchange arrangement with no separate legal tender, 2) Currency board arrangement, 3) Conventional pegged arrangement, 4) Pegged exchange rate within horizontal bands, 5) Crawling peg, 6) Crawling band, 7) Managed floating with no pre-announced path for the exchange rate, 8) Independently floating. In our analysis, we group regimes 1 to 3 under “Fixed”, 4 to 7 under “Intermediate”, and 8 as “Float”. 6 For different ways of categorizing the exchange rate regimes, see for example Levy-Yeyati and Sturzengger (2005), Calvo and Reinhart (2002), Bubula and Otker-Robe (2002) and Reinhart and Rogoff (2002). 4

I. Kato, M. Uctum / International Review of Economics and Finance 17 (2008) 436–456

439

2.2. Explanatory variables The economy size is likely to be positively related to the degree of flexibility. The smaller the economy, the more vulnerable it is to external shocks transmitted through the exchange rate, the higher the probability that it will opt for a low degree of flexibility of the regime (Heller, 1978). In our analysis, the economy size (gdp) is the natural log of PPP based gross domestic product. Openness is negatively related to exchange rate flexibility, everything else being constant. The more open an economy, the worse-off is the inflation-unemployment trade-off with a flexible exchange rate because of the ensuing depreciation of the currency, and the larger is the impact on the economy of a foreign shock (Rogoff, 1985a,b). Thus, the country will likely opt for a low degree of flexibility to circumvent the disadvantage of openness on inflation. However, a reverse causal relation may give a positive relation between the degree of openness and that of flexibility. More open economies usually are subject to frequent foreign shocks and hence need a relatively flexible exchange rate regime to absorb these shocks. In this study, openness of the country (open) is defined as the ratio of the import plus export to the GDP. Inflation differential is positively related to the degree of exchange rate flexibility. A country with a relatively high inflation rate needs to adjust its fixed exchange rate frequently to remain competitive, which is likely to lead to the abandonment of the fixed regime in favor of a flexible one. However, in high inflation countries, authorities may also use fixed rates as a nominal anchor that provides the discipline to reduce the inflation rate. 7 This would then lead to a negative relation between the degree of flexibility and inflation. We calculate the inflation differential (inf ) as the difference between the gross domestic inflation and foreign inflation rates, both in natural logarithms. Later studies also explore the effect on the regime choice of monetary and inflationary shocks, real exchange rate volatility, financial integration, measured by capital flows, and institutional/political variables. Variability of the real exchange rate is positively related to exchange rate flexibility. Higher variability is more likely to shift the country to the floating exchange regime, which is expected to offset the exchange rate volatility (Melvin, 1985; Savvides, 1990). We define this variable (rerv) as the standard deviation of the real exchange rate during the last five years, with the real exchange rate defined as the ratio of foreign price denominated in domestic currency to domestic price. Financial openness (or capital mobility) is likely to be positively related to the degree of flexibility. Countries with high capital mobility and fixed exchange rates lose their monetary policy independence, hence their ability to conduct stabilization policies. In the face of an adverse shock, countries tend to opt for flexible exchange rates to prevent a costly adjustment of the economy. However, low capital mobility requires the trade account to adjust for international imbalances, supporting the case for a flexible regime (Bosco, 1987) and thus a negative relation between capital mobility and the probability of countries choosing a flexible regime. This negative relation also goes back to the OCA discussion (Mundell, 1961). We compute capital mobility (gcf ) as the ratio of gross capital flows (assets plus liabilities) to GDP, which consists of FDI flows, portfolio investment and other flows. To proxy for institutional variables, we use a composite political risk measure (polrisk) from the International Country Risk Group, which is the only historical institutional data series consistent with our sample. The broad finding of the studies that consider the effect of political variables is that countries experiencing high political instability and weak governments will have a high probability of floating their exchange rate (e.g., Edwards, 1996). On the other hand, studies point to the existence of negative duration dependence (e.g., Blomberg, Frieden & Stein 2005). This suggests that countries experiencing balance sheet problems due to unhedged currency debt show a resistance to float, which would be stronger, the weaker the government. Thus, in countries with duration dependence, we should expect weaker governments to be associated with pegged exchange rates. Finally, the evolution of international community's perception concerning the exchange rate regime can be a factor that affects countries' choice of exchange rate regimes, independently of other fundamentals. We approximate this factor with a trend (Collins, 1996), by using time dummies to examine the change in the ideas. The advantage of time dummies is that they allow precise testing of the time period when perceptions change.

7

See Edwards (1996) for an assessment of this theory in developing countries.

440

I. Kato, M. Uctum / International Review of Economics and Finance 17 (2008) 436–456

3. Data, its description, and methodology 3.1. Data All series are annual and cover the period 1982 to 1999. Data for the dependent variable, the exchange rate regimes, are collected from the International Monetary Fund's Exchange Arrangements and Exchange Restrictions Annual Reports. The World Development Indicators is the main source for most of the independent variables. Exceptions are the German GDP and PPP, which are from the OECD's Statistical Databases and the weighted average of foreign GDP (OECD countries), from the OECD Statistical Compendium. Data for foreign liability and FDI comes from the International Monetary Fund's Balance of Payment Statistics. The political risk variable is from International Country Risk Group. It is a composite variable that comprises 12 political components.8 An increase in this variable shows an improvement in the socio- and political stability of the country. Thus a negative coefficient would indicate that deterioration in political stability of the country reduces the probability of a flexible regime. 3.1.1. Currency zones As explained above, we cover three different currency zones and we compute each zone's foreign variables (foreign inflation, foreign prices and the exchange rate) based on the anchor country's variables. More specifically, for countries from the US dollar zone, the ECU/euro zone, and the CFA franc zone the foreign variables are based on the US, German, and French variables, respectively. We define the currency zones as follows: the ECU/euro region in Europe (EU); the CFA franc zone (CFA); and the two US dollar zones, the Latin American zone, comprised of South America and the Caribbean zone (LAC) and the East Asian and the Pacific zone (EAP). We also look at a more homogenous subset of the EU area, the initial 15 members of the euro area and call it EU15. From the 200 countries that belonged to one of the three currency zones 144 were left after excluding those with missing data. Majority, 97 countries, is in the US-dollar zone, 28 in the EU zone and 19 in the CFA franc zone, with a full sample size of 2063. The categorization of currency zones is based on Levy-Yeyati and Sturzengger (2005), and the regional classifications of countries are from the World Bank development report (see Appendix for the list of countries). 3.2. Description of data 3.2.1. Data according to currency zones Fig. 1 shows the annual averages of the dependent and exogenous variables in the full sample and currency zones. The exchange rate regime is represented by the bold line in the first row and its average value varies between 1 (fixed rates) and 3 (perfectly flexible rates). The full sample reflects a general move towards more flexible exchange rates. The move tapers off in mid-1990s and declines somewhat thereafter. In parallel with the pattern in the exchange rate regimes, the volatility in the real exchange rate, rerv, and the inflation differential, inf, both increase during the 1980s and then decline after 1995. The other variables, namely trade and financial integration (open, gcf), and GDP exhibit broadly positive trends. The institutional variable, poli, shows a declining trend until late 1980s reflecting deterioration in the political risk, followed by a rise that suggests improvement in the political profile, and a deterioration again once more at the end of the 1990s. This pattern is roughly consistent with the negative relation between political stability and exchange rate regimes discussed above. The full-sample trends are not replicated in the EU zone. The area follows relatively fixed exchange regimes until 1991, switches to more flexible regimes through the 1990s as a result of new members joining the zone after the collapse of the Soviet Union, and widening of the bands following the 1992–1993 currency crisis in EU15 zone. Both the EU and EU15 revert back to more fixed regimes after 1997 in the wake of the advent of euro in 1999. The volatility of the real exchange rate and inflation differentials follow the events in the EU zone. They increase after 1991 in EU as new members grapple with nominal volatilities in their economies, but decrease gradually in EU15 as the core countries follow policies to converge their economies, in line with the requirements for the European monetary union (EMU). Openness, and financial integration increase in both the large and the small group but GDP stagnates in EU 8

The components consist of government stability, socio-economic conditions, investment profile, internal and external conflicts, military and religion in politics, law and order, ethnic tensions, democratic accountability, bureaucracy.

I. Kato, M. Uctum / International Review of Economics and Finance 17 (2008) 436–456

Fig. 1. A: Annual averages of variables by currency zones⁎. B: Annual averages of variables by currency zones.

441

442

I. Kato, M. Uctum / International Review of Economics and Finance 17 (2008) 436–456

while it increases in EU15. Political risk, on a slight negative trend, does not show any improvement before the second half of the 1990s. Not surprisingly, the variables in the CFA zone do not follow those of the full sample either. Their currencies have been pegged to the French franc and now to euro with a major adjustment in mid-1980s and in early 1990s. However, because this is a more homogenous group than EU, the pattern in the exchange rate regime is different. Inflation came under control in the second half of the 1980s, while the real exchange rate volatility remains fairly modest, except during the late 1980s–early 1990s. Financial integration, however, is on a declining trend, reflecting the difficulties the area is experiencing in attracting foreign investment. By contrast, openness, GDP and improvements in the political profile have been increasing throughout the sample period. The exchange rate regimes in LAC and the EAP zones are following the full sample relatively closely, though differences exist. The LAC area experienced an increase in the flexibility of the regimes in 1988–1993 and a rise in more fixed regimes since then. In EAP, the tendency to adopt more flexible rates continued throughout the 1990s until the Asian crisis when countries moved sharply towards more fixed regimes. After 1997, the exchange rate volatility drastically declined in the LAC, while it sharply increased in the EAP as expected. The inflation differentials in both zones also follow similar patterns to the volatility of the real exchange rates. They subside in the second half of the 1990s in LAT but increase in EAP. Capital flows dwindled in 1980s in LAC following the debt crisis but they have been in general stable throughout the 1990s. By contrast, they increased in the EAP region until 1996, declined sharply in 1997 triggering the crisis, and started to recover thereafter. GDP and openness, while following the full sample pattern in LAC, stagnated for a while in EAP after the Asian crisis. Political risk, while showing no clear positive trend in EAP, marks an impressive improvement in LAC. Next, we turn to the description of the data according to exchange rate regimes. 3.2.2. Data according to regime change Fig. 2 shows the patterns in the means of regressors for three different samples. The first one, X0, represents the means of regressors for countries that remained in the same regime throughout the sample period. X1 is the means of regressors for countries that changed regimes only once, and X2 refers to the means of regressors for countries that went through two regime changes. This categorization of the data helps us explore possible correlation of variables with changes between regimes. Countries that stuck to a single exchange rate regime are on average small open economies, and highly integrated in international markets. Throughout the 1980s, they had relatively low quality political institutions, but also low inflation and exchange rate volatility. During the 1990s, both inflation and volatility increased and then abruptly reversed as a result of large shocks in the second half of the decade. Political institutions enjoyed a steady improvement throughout the 90s but deteriorated at the end of the decade. By contrast, countries that

Fig. 2. Annual averages of variables by regime changes.

I. Kato, M. Uctum / International Review of Economics and Finance 17 (2008) 436–456

443

Fig. 3. Sample averages of variables by exchange rate regimes.

had three different regimes during the same period were middle sized, had relatively low openness, and low financial integration, and higher inflation and exchange rate volatility but with lower variance. In terms of improving their institutions, they ranked consistently below the other groups. We further break down the single-regime sample (X0) into total, fixed, intermediate and floating regimes (Fig. 3). The sample averages suggest that countries adhering to flexible regimes have larger and less open economies, low capital mobility, high inflation rate, large supply shocks (real exchange rate volatility) but better institutions. Countries that stuck to fixed regimes have smaller and more open economies, with lower inflation and exchange rate volatility but also higher political risk. Countries with intermediate regimes are the most financially integrated and rank about in or close to the middle in the other categories. 3.3. Methodology Previous studies used various statistical techniques to analyze the choice of exchange rate regimes. The methodology generally consists of discrete choice models such as binary probit (Collins, 1996; Eichengreen, Rose, & Wyplosz, 1996; Frankel & Rose, 1996; Savvides, 1990), binary logit (Bosco, 1987), and ordered probit or logit (Bosco, 1987; Dreyer, 1978; Eliasson & Kreuter, 2001; Melvin, 1985; Rizzo, 1998). To control for country heterogeneity we use the random effects ordered probit model, which is one of the commonly adopted methodologies to discuss categorical dependent variables, and has a natural order in panel data. In the estimated equation for the model defined as in Eq. (1), Yit⁎ is the unobserved dependent variable, vit is a normally distributed error term, vi ∼ N(0,σv) and σv = 1.9 Since Yit⁎ is unobserved, what we observe in the exchange rate regime analysis is Yit = {1,2,3} if {Yit⁎ ≤ μ1, μ1 b Yit⁎ ≤ μ2, μ2 ≤ Yit⁎}, where Yit is the choice of exchange regimes of ith country, and Yit = 1, 2, 3 for fixed, intermediate, and floating regimes, respectively, and μ is the unknown cut-point parameter. The probability of each regime being chosen is: Prob(Yit = 1) = Φ(μ1 − β′xit − ui), Prob(Yit = 2) = Φ(μ2 − β′xit − ui) − Φ(μ1 − β′xit − ui) and Prob(Yit = 3) = 1 − Φ(μ2 − β′xit − ui), where Φ(˙ ) is the normal distribution function. The model is estimated by the log-likelihood function introduced by Butler and Moffitt (1982) and the Gauss–Hermite quadrature method deals with the random effects structure in the model (Frechette, 2001). To avoid multicollinearity, we include the dummy variables only for intermediate and floating regimes. We assume that the initial regime choice at t = 0 is non-stochastic constant (Wooldridge, 2002). Two additional common problems that need to be addressed are endogeneity and cross-section dependence. Since our panel discrete-time data analysis does not allow the use of the existing methodologies, we handled these problems using different approaches. Endogeneity, caused by contemporaneous interaction between economic fundamentals and the exchange rate regimes, is a well-known drawback in this type of analysis. In cross-section studies, some of The variance of the composite error term εit = ui + vit is Var(εit) = σ2v + σ2u = 1+σ2u, and the cross-period correlation of εit is: Corr(εit,εis) = ρ = σ2u/(1 + 2 σu) if t ≠ s. If the random effects exists, εit and εis are correlated within a group, but not correlated across groups. If the effects are not significant, σ2u = 0 and ρ = σ2u/(1 + σ2u) = 0, which indicates there is no cross-period correlation with respect to εit. To test for random effects, we examine the statistical significance of ρ, using the Wald test statistics (w = ρ2 / s2ρ). If w N χ2 critical value (3.84 for a 95% critical level), we can reject the null of ρ = 0. 9

444

I. Kato, M. Uctum / International Review of Economics and Finance 17 (2008) 436–456

the independent variables are frequently instrumented to deal with this problem using time-invariant country characteristics (e.g., openness instrumented with land area or a landlocked dummy variable). However, since our data has an important time-series dimension, we cannot adopt the traditional instruments, which are devoid of time variance. More recently, in panel data the methodology by Arellano and Bond (1991) has become popular, which uses internal instruments. However, due to its discrete-time characteristic, our model does not lend itself to this approach. To tackle the question of endogeneity, we adopt the Hausman (1978) and Wu (1973) two-step test procedure to select relevant internal instruments, which is a suitable methodology for the discrete data analysis. For this, we first determine whether there is an endogeneity problem, we then identify the affected variables and finally select the instruments. To control for cross-section dependence, methodologies such as GLS applied in some studies are again not easily applicable for discrete-time models such as ours. To partly solve this problem, we draw on time dummies as an alternative methodology, which is tantamount to removing the global mean, and therefore, the average cross-section effect from each series. Although this is a crude approach (O'Connelle, 1998), time dummies could capture unobserved cross-sectional (regional or global) shocks, and are likely to reduce the contemporaneous correlation across disturbances. For the pooled analysis, we refine the methodology one step further. We explicitly control for regional time dummies in the pooled regression. For this, we generate 5 different series of currency-zone time dummies, which capture specific year shocks in each region as opposed to global shocks. When we examine each currency zone

Table 1 Pooled estimates ⁎ Full Sample 1982–99

gdp

Restricted Sample 1985–99

1982–99

1985–99

0.00 (0.1) 0.02 (2.0) 0.02 (0.9) − 0.04 (1.4) 0.87 (5.6) –

− 0.01 (0.1) 0.03 (2.5) 0.02 (0.7) − 0.03 (1.0) 0.60 (2.9) − 0.02 (3.1) 0.63 (12.5) − 704.4 234.2 0.14 187.2 134.0 34.9 70.7 34.7 51.0 882

poli ρ

–

LL LR Pseudo-R2 LRTIME LRREG % predicted Fixed Intermediate Float Total N

−1744.4 806.1 0.19 179.7 268.4

−1143.2 349.1 0.13 75.6 277.2

− 1214.5 412.6 0.15 283.5 152.1

− 838.4 207.1 0.11 129.8 224.5

0.54 (13.1) − 1059.6 496.2 0.19 335.4 123.3

74.8 62.5 18.1 59.3 2063

40.9 70.4 22.8 47.4 1210

57.1 65.9 28.4 53.8 1320

23.4 76.9 32.7 50.2 882

49.3 66.1 37.5 53.3 1320

gcf rerv

–

1985–99

0.04 (1.4) −0.03 (2.9) 0.04 (2.2) 0.00 (0.4) 0.54 (5.0) 0.01 (3.3) –

inf

0.01 (0.4) − 0.01 (0.5) 0.01a (0.4) 0.03 (1.3) 0.38 (2.7) − 0.00 (0.7) –

1982–99 0.19 (11.3) −0.02 (2.7) 0.06 (4.4) 0.01 (0.7) 0.33 (5.1) –

open

0.07 (3.4) − 0.01 (0.6) 0.03 (1.8) 0.30 (1.2) 0.42 (3.9) –

Random effects

⁎The dependent variable is the probability of choosing a relatively flexible exchange rate regime. The independent variables are: gdp = gross domestic product; open = openness; inf = inflation differential; gcf = gross capital flows; rerv = real exchange rate volatility; poli = political risk (an increase is an improvement). N is the number of observations. Variables open, inf and gcf are scaled up by 10. Full Sample contains countries with zero, one or more regime changes. Restricted Sample excludes countries with no regime change. Regressions include regional time dummies (not shown) used as a proxy to control for cross-sectional dependence. LRTIME is the likelihood ratio test statistics for time effects with 27.59 critical χ2 value. LRREG is the likelihood ratio test statistics for regional effects, with 33.92 critical χ2 value. Figures in parentheses are absolute values of z-test statistics below coefficient estimates, and the Wald test statistics for random effects below the cross-correlation estimate ρ with 3.84 critical χ2 value. a Change in significance due to sample change from 1982–99 to 1985–99, no inclusion of poli.

I. Kato, M. Uctum / International Review of Economics and Finance 17 (2008) 436–456

445

Table 1a EU and EU15 areas a EU

EU15

Full Sample

Restricted Sample

Full Sample

Restricted Sample

Random effects

Random effects

1982–99 1985–99 1982–99 1985–99 1982–99 1985–99 1982–99 1985–99 1982–99 1985–99 1982–99 1985–99 gdp

poli

0.25 (4.6) − 0.11 (4.3) 0.01 (0.2) 0.08 (2.4) 0.81 (5.1) –

ρ

–

0.27 (4.3) − 0.10 (3.4) − 0.17 (1.4) 0.07 (2.1) 1.19 (3.3) − 0.14 (1.5) –

LL LR Pseudo-R2 N

− 283.0 137.8 0.20 356

− 224.2 119.7 0.21 282

open inf gcf rerv

a b

0.33 (5.6) − 0.02 (0.6) 0.07 (1.7) 0.02 (0.4) − 0.18 (0.4) – –

0.36 (5.3) − 0.01 (0.3) − 0.05 (0.6) 0.01 (0.2) 0.11 (0.2) − 0.02 (1.3) –

− 222.0 83.5 0.16 279

− 185.0 74.7 0.17 228

0.38 (3.5) −0.06 (1.1) 0.05 (1.1) −0.05 (1.1) −0.09 (0.2) – 0.28 (2.3) −211.8 54.9 0.11 279

0.46 (3.3) 0.00 (0.0) − 0.11 (0.8) − 0.05 (1.1) 0.08 (0.1) − 0.04 (2.3) 0.27 (2.6) − 176.8 55.7 0.14 228

0.50 (4.3) 0.09 (1.8) − 0.05 (0.2) − 0.03 (0.6) 10.09 (4.4) – –

0.22 (1.6) 0.01 (0.1) b −1.55 (3.4)b 0.01 (0.3) 10.32 (4.1) −0.08 (3.9) –

0.64 (5.1) 0.14 (2.8) − 0.57 (1.5) − 0.04 (0.09) 13.97 (5.1) – –

0.34 (2.3) 0.06 (1.0)b − 1.61 (3.3) 0.00 (0.0) 12.62 (4.3) − 0.08 (3.5) –

− 138.52 97.4 0.26 216

−113.9 99.2 0.30 183

− 122.7 104.9 0.30 190

− 102.3 87.2 0.33 160

0.58 (2.1) 0.09 (0.9) −0.84 (1.7) −0.06 (1.0) 13.68 (4.3) – 0.28 (1.9) −116.2 92.5 0.28 190

0.18 (0.8) 0.4 (0.5) − 0.25 (3.9)b 0.01 (0.1) 10.70 (3.0) − 0.23 (4.2) 0.58 (4.4) − 88.3 110.7 0.39 160

See legend of Table 1. Change in significance due to sample change from 1982–99 to 1985–99, not addition of poli.

separately, the regression includes one series of time dummies for each zone. Our methodology allows us to remove each zone's average cross-section effect from the pooled series. 4. Empirical results In the next subsections we examine the empirical evidence covering the period 1982 to 1999 (Tables 1, 1a–1c). The figures indicate how the probability of choosing a more flexible exchange rate regime changes in response to a change in an independent variable. Figures in parentheses are absolute values of z-statistics associated with the estimates. The likelihood ratio (LR) test statistics distributed as χ2 (10) and the pseudo-R2 are two goodness-of-fit measures we use to compare across models. The coefficient of within group error terms is denoted by ρ, which measures the significance of random effects. The first column in each panel displays the regression results in full sample, all countries and periods combined. Adding the poli variable reduces the sample size by making it start in 1985 instead of 1982 and reducing the number of countries for which the series are available. Random effect regressions require elimination from the sample of countries that did not experience any variation in the exchange rate regime. Since a change in the sample is likely to affect the results, we reran the estimations with new samples. To illustrate, in Table 1 representing results for the pooled model, the first column displays the regression results in full sample, all countries and periods combined. The second column adds the institutional variable and starts from 1985. We repeat the same exercises in the third and fourth columns with the sample restricted to economies that experienced at least one change in the regime. Columns five and six present the results with random effects, without and with the institutional variable, respectively. When a new variable or random effects are introduced, a change in the estimates can be attributed either to a change in the sample, or to a change in the model.10 Although our interest is in the first two and the last two columns, we present the intermediate columns to indicate some of the intermediate steps in the change in the coefficient estimates. 10 For example, consider the variable poli. The available data for this variable start in 1985 and cover a subsample of the countries in our original sample. To keep track of the origin of the changes in the estimates, we rerun the regression in the first column (without poli) starting from 1985, then include only countries for which the poli data is available, and finally add the series poli.

446

I. Kato, M. Uctum / International Review of Economics and Finance 17 (2008) 436–456

Table 1b CFA, and EAP areas a CFA area

EAP area

Full Sample

Restricted Sample

Full Sample

Restricted Sample

Random Effects 1982–99 1985–99 1982–99 1985–99 1982–99 gdp

poli

0.70 (6.3) − 0.07 (1.1) 0.69 (6.9) 0.15 (1.3) 0.54 (1.0) –

ρ

–

0.58 (4.0) − 0.05 (0.7) 0.61 (5.4) 0.19 (0.8) 0.91 (1.6) 0.02 (1.0) –

LL LR Pseudo-R2 N

− 103.9 101.4 0.33 328

− 83.9 55.6 0.25 164

open inf gcf rerv

a b c

0.35 (1.8) 0.05 (0.4) 0.04 (0.3) 0.42 (1.3) 4.59 (3.1) – –

0.25 (1.0) b 0.17 (1.3) 0.11 (0.6) 0.09 (0.3) 4.81 (3.0) 0.01 (0.4) –

−50.9 26.1 0.20 72

− 42.8 18.0 0.17 59

0.22 (1.0) 0.37 (2.5) 0.12 (0.7) 0.52 (1.4) 3.00 (1.8) – 0.86 (13.7) − 46.9 30.0 0.24 72

Random Effects

1985–99 1982–99 1985–99 1982–99 1985–99 1982–99 1985–99 1.24 (2.5) 0.32 (1.4)c 0.66 (1.7)c 1.00 (1.3) 4.49 (1.7) − 0.17 (2.8) 0.92 (25.4) − 25.2 36.6 0.42 59

–

− 0.19 c (2.2) − 0.01 (5.2) 1.10 (3.5) 0.04 (0.7) − 0.06 (0.1) 0.16 (7.9) –

–

−0.07 (0.6) −0.10 (4.6) 1.35 (3.7) 0.32 (2.6) −1.04 (0.8) 0.14 (6.0) –

− 190.7 74.8 0.16 209

− 85.3 137.0 0.45 143

− 138.6 46.9 0.14 155

−76.3 85.2 0.36 110

0.18 (4.0) − 0.03 (2.4) 0.11 (1.2) 0.02 (0.4) − 0.01 (0.0) –

0.16 (2.5) − 0.04 (2.6) 0.30 (1.9) 0.22 (2.2) − 0.62 (1.2) –

0.13 (1.8) − 0.04 (2.3) − 0.06 (0.3) 0.13 (1.2) 1.14 (2.1)

0.60 (9.4) − 100.4 62.9 0.24 155

0.11 (0.9)c − 0.04 (1.3)b 0.98 (2.1)c 0.14 (1.1) 2.62 (1.7) 0.07 (2.2) 0.82 (14.3) − 67.3 56.9 0.30 110

See legend of Table 1. Change in significance due to sample change from 1982–99 to 1985–99, not addition of poli. Change in significance or sign due to change of the sample size (countries).

Table 1c LAC area a Full Sample

Restricted Sample Random Effects

1982–99

1985–99

1982–99

1985–99

1982–99

1985–99

poli

0.31 (6.8) 0.01 (0.6) 0.10 (2.7) − 0.08 (2.1) 1.35 (5.8) –

0.19 (3.5) 0.05 (1.9) c 0.01 (0.2)c 0.04 (0.4) 1.46 (4.7) –

–

–

0.17 (2.4) 0.15 (3.4) 0.05 (1.0) 0.19 (1.3) 1.02 (2.4) 0.01 (0.8) –

0.24 (2.4) 0.08 (1.9) 0.01 (0.3) 0.14 (1.3) 1.73 (4.9) –

ρ

0.15 (2.3) 0.16 (4.4) b 0.13 (3.1) − 0.11 (2.5) 0.06 (0.2)b − 0.02 (1.5) –

LL LR Pseudo-R2 N

− 375.6 294.2 0.28 508

− 207.3 48.8 0.11 229

− 274.5 150.6 0.22 322

− 165.3 53.9 0.14 201

0.24 (2.1) 0.19 (2.6) 0.05 (0.8) 0.35 (2.0)b 0.91 (2.0) 0.00 (0.1) 0.22 (2.0) − 161.2 48.9 0.13 201

gdp open inf gcf rerv

a b c

See legend of Table 1. Change in significance due to change of the sample size (countries). Change in significance due to sample change from 1982–99 to 1985–99, not addition of poli.

0.22 (2.7) −263.0 158.1 0.23 322

I. Kato, M. Uctum / International Review of Economics and Finance 17 (2008) 436–456

447

Finally, we add a superscript a to the z-statistics we report when the change in the coefficient estimate is only due to a change of sample and a superscript b when the change is due to a change in the country sample. Throughout the tables, we use the following notation: Full Sample consists of countries with zero, one or more regime changes, while Restricted Sample excludes countries with no regime variation. Pooled designates all countries and contrasts with specific regions. In a probit model the magnitude of the coefficient estimates are hard to interpret. The sign of the coefficient gives the information about the choice. Here, a positive (negative) sign means that as the independent variable increases, the probability of choosing a relatively flexible regime over fixed regime rises (falls). To go beyond these coefficients, we also report the marginal effects for the pooled analysis (Appendix B, Table 3A) and discuss quantitatively the meaning of the coefficients in the text. The LR test statistics (LRTIME) in pooled estimates suggest that time effects are significant across groups and models. We report estimation results controlling for time effects in all currency areas. The coefficient of crosscorrelation ρ is strongly significant across samples, suggesting significant heterogeneity effect even within currency zones, which requires control of country-specific effects. An exception is the regression results involving EU15 where random effects are rejected. Comparing across currency zones, the pseudo-R2 is the highest in the full sample results of the CFA and the LAC regions, suggesting that the model explains the choice of currency regimes best in these areas. Endogeneity can be a problem for any of the explanatory variables. To see if this is the case, we conduct the Hausman–Wu test (Appendix B, Table 1A). We run the regression of the dependent variable Y on Xi and Xfiti, where Xi is the independent variable i, and Xfiti is the predicted value of Xi obtained by regressing it on the instrumental variable(s). As instruments, we chose Xj, with j ≠ i, and Xi, t−1. Under the null hypothesis of endogeneity, the coefficient of Xfiti should be significantly different from zero. We repeated the same procedure for each potential endogenous variable in each currency zone. Evidence suggests that endogeneity is not a problem in the pooled analysis for any of the variables. Although this outcome broadly holds for the regional analysis, some exceptions arise. Test results cannot reject the hypothesis that there is no reverse causality for openness in EU, political risk in EAP and EU15, capital flows in LAC, inflation and volatility in CFA, and gdp in EAP (Table 1A, column 1). In the restricted sample, endogeneity is rejected in CFA and LAC but not for openness in pooled and EU, risk in EU15, and for gdp, inf and gcf in EAP (Table 1A, column 2). We then repeated the test using instrumental variables, starting with the first lag for each variable. Overall, we observe an improvement in the results, in particular in the regional data with restricted sample. In the full sample (Table 1A, column 3), the pooled test results again reject endogeneity. In the regional data, the number of independent variables with possible reverse causality is now reduced to gcf in EU, LAC and EAP, gdp in EU15. Results remain unchanged in CFA. In the restricted sample test results reject endogeneity for any variable in all regions except EAP (Table 1A, column 4). Since larger lags do not improve the results, while introducing additional issues such as lost information, we decided to instrumentalize the variables with their first lag. Although this does not entirely resolve the difficulty of reverse causality, it has the merit of alleviating it. In addition, the instruments are in general highly correlated with the independent variables and insignificantly correlated with the error term, indicating that they are good instruments.11 4.1. Pooled estimates Table 1 presents the estimation results for the pooled model, all countries and periods combined. Results with full sample estimates show that most independent variables are significant at the 95% confidence level and come in with the expected sign (first two columns). Consistent with the theory, the probability of choosing a flexible over a fixed exchange rate regime increases the larger the economy, the higher the inflation rate, the real exchange rate volatility and the lower the openness (first column). The effect of financial integration is insignificant. Introducing the political risk variable does not affect the significance or sign of these estimates except gdp. In the full sample result, the sign is positive and significant, showing that an improvement in the political institutions increases the probability of flexible rates in full sample. This finding is not surprising since the full sample contains several regime dependent countries. 11

One problem with using lagged variables is that if there are any significant feedback effects from the contemporaneous variables, this information might be lost. However, since the instruments are correlated with the explanatory variables, we expect this problem to be relatively minor. Using a Granger causality test to compare the feedback effects of contemporaneous explanatory variables with those of lagged ones, we find that overall the loss of information is statistically very small.

448

I. Kato, M. Uctum / International Review of Economics and Finance 17 (2008) 436–456

This gives support to the view that in countries with duration dependence, weak governments are associated with relatively fixed rates, and affect the overall sign of the coefficient estimate. Our result is also consistent with Meon and Rizzo (2002) who find a negative relation between a time-variant political instability variable and flexible rates. The estimates presented in this table control for regional time dummies (not reported). To examine explicitly the evolution of perceptions about the exchange rate regimes we reran the estimations with simple time dummies (Appendix, Table 2A). Simple time dummies have significant positive signs from 1990 onwards, reflecting a general move away from relatively more fixed regimes towards relatively flexible regimes. We also wanted to see how the de jure estimates of IMF data compare with the de facto estimates of Bubula and Otker-Robe (2002). We find that the conclusion concerning the evolution of the regimes remains unchanged (last two columns). Results are mostly robust to sample size and model. Restricting the sample size and controlling for random effects improves the LR statistics, but changes the sign of openness against the theory's predictions. Although a positive relation between openness and the degree of flexibility looks puzzling, it is consistent with the experience of some currency zones, as we will see below. It reflects, at least partially, the fact that throughout the 1990s many emerging economies adopted floating regimes while opening up their economies. Political risk variable is still significant but this time with the expected sign. This is not surprising, since to obtain the restricted sample we eliminated all countries that did not change their exchange rate regimes (i.e., countries exhibiting duration dependence) and this helped to get the negative sign. Another point of interest is that the inclusion of the political variable affects neither the significance nor the sign of the other estimates. The loss of significance of inflation differential is entirely due to change in the sample from full to restricted sample, which eliminates countries that did not change their exchange rates. Controlling for unobservable country effect did not affect the estimate of inflation. In contrast the loss of significance of gdp is due to random effects. We checked for potential bias caused by multicollinearity. For this, we reran the regressions by excluding consecutively highly correlated variables. Results are largely unaffected by this exercise, except in the random effect regression, where omitting rerv increases the significance and the magnitude of the inflation effect. To see the meaning of these coefficients, we calculated the marginal effects on the choice of exchange rate regimes of a change in the independent variables (Appendix 3A). In the full sample, a one-unit rise in GDP increases the probability of choosing a flexible regime by 0.002 percentage points, intermediate regime by 0.003% and thus reduces the probability of choosing the fixed regime by 0.004% (left panel). With the random effects model, the signs are less consistent and the magnitudes are smaller (right panel). This is not unexpected since introduction of random effects modifies the cross-sectional variation between countries, and therefore, reduces the significance (or the coefficient size) of variables with high cross-sectional and low time series variation. A one-unit rise in the real exchange rate volatility in the full sample increases the probability of the flexible and intermediate regimes by 0.004 and 0.05 percentage points, respectively, and decreases that of the fixed regime by 0.06 percentage points. This time, controlling for the random effect increases the estimates. The probability of choosing a flexible regime when the volatility rises by one-unit increases by 0.22 and that of the fixed regime declines by 0.071 percentage points. Although in our pooled model we controlled for country and regional heterogeneity by introducing random effects and regional dummies respectively, we conducted additional tests to check the existence of regional effect on coefficients of explanatory variables. One of these is the LRREG test statistics that we computed by taking the ratio of the likelihood ratio of the pooled sample regression with that of the sum of regional likelihood ratios still suggests some heterogeneity. All these results taken together highlight the importance of analyzing the regions separately. Each currency zone consists of countries that share common characteristics. In each currency zone, we include time dummies. We expect that grouping the countries according to the currency zones and controlling for time dummies will help to reduce the contemporaneous correlation within group's disturbances and thus cross-sectional dependence. We turn to the disaggregated analysis below. 4.2. European (EU) area Table 1a shows the regression results for the 28 European countries, EU (left panel), and the 15 countries, EU15 (right panel) constituting the more integrated core of the European monetary union (EMU). The two groups display a fair amount of differences, largely due to the disparity in the development of their respective markets and economies.

I. Kato, M. Uctum / International Review of Economics and Finance 17 (2008) 436–456

449

Country size is robust to sample specification and mostly to model specification and affects the choice of the exchange rate regime in the expected direction. In the large EU group most results are quite sensitive to sample changes and not to changes in the model. Whereas in the full sample low openness, high capital mobility and real exchange rate volatility increase the probability of choosing a relatively flexible regime, these factors are not significant in the restricted sample with or without random effects. Some of the insignificant results are due to multicollinearity. Although inflation appears insignificantly across models and samples, omitting rerv increases the significance in full sample (column 1). The institutional variable is significant and of expected sign when country specific effects are controlled for, suggesting that country variation dampens the time series effect. This interpretation is corroborated with the results from the EU15. In this homogenous group country effect does not change the significant and negative sign of the institutional variable. Turning back to EU15, several differences emerge with the larger area. Unlike in EU, the real exchange rate volatility, rerv, is positively and significantly associated with flexible exchange rates in both samples and models, while gcf is not significant. Openness is significant in the full sample (columns 1, right panel), but its significance decreases with a change in sample (column 2) or random effects are added (columns 3, 4). In contrast, inflation becomes significant when sample size starts from 1985. At first, the negative relation between inflation differential and the degree of flexibility looks odd. Careful inspection of the data reveals that this is not a causal relation but a result of convergence criteria observed by the countries that joined the EMU at the end of the 1990s (Fig. 1, column 3). While the fluctuation bands were widened in response to the European crisis, the EU15 economies continued their efforts to comply with the Maastricht criteria, among which was the convergence of inflation rates. This explains largely the negative sign of the inflation estimate in EU15. Although its introduction does not affect the other coefficient estimates, the institutional variable is a significant determinant of the regime choice in EU15, and enters the regressions with the expected sign independent of country effects. 4.3. CFA zone This is the area with the highest pseudo-R2 in the full sample results (Table 2b, left panel). Time effects were not significant and affected the estimated coefficients little. Given the nature of the exchange rate arrangement in this zone, the restricted sample is substantially smaller than the full sample once all countries with no variation in the dependent variable have been removed, making the small sample bias hard to avoid. In full sample, all variables except rerv are significant and their sign is consistent with the theory. We checked if the low significance of the volatility of the real exchange rate is due to multicollinearity. To be sure, in most cases omitting inf changed the sign to positive and raised the significance and the magnitude of the estimate, in line with the theory. Table 2 Sample analysis⁎ Pooled RS gdp

0.07 (3.4) open − 0.01 (0.6) inf 0.03 (1.8) gcf 0.03 (1.2) rerv 0.42 (3.9) LR 412.6 Pseudo-R2 0.15 N 1320

EU

LAC

EAP

RSP

RSF

RSI

RS

RSP

RSF

RSI

RS

RSP

RSF

RSI

RS

RSP

RSF

RSI

0.17 (9.3) − 0.02 (2.3) 0.09 (5.9) − 0.01 (0.8) 0.18 (2.3) 667.1 0.18 1844

0.11 (5.8) −0.01 (1.5) 0.02 (1.2) 0.04 (1.6) 0.53 (6.0) 443.8 0.14 1428

0.07 (3.6) 0.00 (0.1) 0.02 (1.5) 0.00 (0.4) 0.36 (3.6) 406.2 0.13 1431

0.33 (5.6) − 0.02 (0.6) 0.07 (1.7) 0.02 (0.4) − 0.18 (0.4) 83.5 0.16 279

0.34 (5.8) − 0.05 (1.8) 0.05 (1.6) 0.03 (0.9) − 0.31 (1.0) 118.2 0.21 304

0.25 (4.4) − 0.07 (2.2) − 0.00 (0.0) 0.07 (1.9) 1.07 (4.9) 105.2 0.17 299

0.33 (5.6) −0.02 (0.8) 0.07 (1.9) 0.01 (0.2) −0.26 (0.6) 88.0 0.16 311

0.19 (3.5) 0.05 (1.9) 0.01 (0.2) 0.04 (0.4) 1.46 (4.7) 150.6 0.22 322

0.29 (6.4) 0.02 (0.8) 0.09 (2.5) − 0.02 (2.5) 1.46 (6.0) 284.8 0.28 494

0.20 (3.6) 0.04 (1.7) 0.01 (0.2) 0.05 (0.5) 1.40 (4.6) 154.1 0.22 329

0.19 (3.5) 0.16 (0.8) 0.09 (2.5) − 0.04 (0.8) 1.20 (4.3) 147.0 0.21 329

0.16 (2.5) − 0.04 (2.6) 0.30 (1.9) 0.22 (2.2) − 0.62 (1.2) 46.9 0.14 155

0.17 (3.3) −0.02 (1.4) 0.26 (1.8) 0.01 (0.2) −0.36 (0.7) 55.8 0.16 166

0.16 (3.0) − 0.05 (3.5) 0.19 (1.6) 0.24 (2.4) − 0.19 (0.4) 61.9 0.15 190

0.15 (2.6) − 0.04 (2.7) 0.21 (1.8) 0.23 (2.3) − 0.49 (1.0) 48.2 0.14 163

⁎RS correspond to samples consisting of restricted sample (countries with at least two regime changes); RSP = RS plus pure pegs; RSF = RS plus pure floats; RSI = RS plus intermediates. N = number of observations. All regressions are inclusive of time dummies, except in pooled sample where they are regional time dummies.

450

I. Kato, M. Uctum / International Review of Economics and Finance 17 (2008) 436–456

The reduced sample reduces most variables' significance with the exception of rerv. However, collinearity between rerv and inf conceals the real effect of inflation. Omitting rerv, the inflation differential enters the regression with a significant and positive coefficient. We checked if the weakness of the other estimates is due to multicollinearity, and verified that this was not the case and concluded that it is most likely caused by small sample bias. 4.4. LAC area Most variables in this area enter the regression equation significantly and in line with the theory (Table 1c). Among these, the positive sign of openness reflects a policy decision rather than a causal relation. Throughout the 1980s, LAC economies reduced their trade barriers while embracing more flexible exchange rate regimes. By contrast, during the 1990s the switch to more flexible regimes stopped and more countries started adopting intermediate regimes, while opening up to trade reached a ceiling and stopped progressing (Fig. 1B). The experience of these economies during both episodes presumably created a positive, noncausal, correlation between openness and the degree of flexibility of exchange rates. After the 1980 debt crisis, LAC countries took drastic measures and established responsible policies. The policy change restored international investors' confidence and contributed to a gradual rise in financial integration of these economies. The increase in gcf occurred simultaneously with a gradual return to more intermediate regimes. Thus, in the full sample, we observe a negative relation with the degree of exchange rate flexibility and capital flows. However, the relation is not strong and weakens with sample changes, presumably reflecting the structural changes in several countries. Inflation is positively related to the degree of flexibility of the exchange rate regime as expected, but the relation also weakens in the restricted sample. Institutional factors do not appear to make a difference in regime choice in any sample. Time effect is significant and positive in particular after 1990, reflecting the trend in the area of a move towards more flexible regimes, and a higher volatility of inflation and capital flows. The positive sign of the time dummies is consistent with the evolution of the “climate of ideas” in Collins (1996). 4.5. EAP area In parallel with the other areas, the economy size is significant and positive in this region, suggesting that the larger the economy, the higher the probability of choosing a more flexible exchange rate regime. Openness is also robust to model specifications and sample change and comes with the expected sign. But both variables lose their significance in the subsample of countries containing poli. Inflation is a significant determinant of the exchange rate regime in the same subsample, but not in the other subsamples. Capital flows appear to be not significant overall. A close inspection of Fig. 1B gives some insight into this result. The positive relation between gcf and the degree of exchange rate flexibility is present over the period 1982–1986. The Asian crisis, however, introduces major changes in most variables, in particular in exchange rate regimes and gcf. To examine these changes we excluded the last two years from our sample and reran the regressions. Interestingly, while the changes in the other variables were mostly insignificant, capital flows became positive and highly significant, while rerv turned insignificant. This result reflects what we observe in Fig. 1B. Until the crisis, supply shocks, represented by real exchange rate volatility, were negligible and hence, did not enter the choice of exchange rate regime, while capital flows were an important component of this decision. 4.6. Discussion of the results Stepping back from the regional analysis, several points of interest arise for comparative studies on different country samples. The pooled results indicate that among the six explanatory variables analyzed, openness, exchange rate volatility and the institutional variables are the most likely to affect the exchange rate choice of a country. At the regional level, results vary according to the zone. GDP is an important determinant in almost all regions, whereas regional time dummies accounted for it in the pooled sample and made it insignificant. Openness is an important determinant in EAP and LAC and exchange rate volatility in EU15 and LAC. Whereas inflation is a significant factor in the full sample of the pooled regressions, this result is reproduced in the full samples of CFA and LAC, but becomes significant after 1985 in EAP and EU15. The lack of significance of gcf is replicated in EU15, CFA and EAP but not in EU and LAC, at least in the full sample. The significance of poli in the pooled model is mimicked only in EU15 and EAP.

I. Kato, M. Uctum / International Review of Economics and Finance 17 (2008) 436–456

451

Does controlling for random effects always improve the explanatory power of the equations? The log-likelihood (LL) test statistics results in the restricted sample indicate that accounting for country heterogeneity improves the explanatory power of the models in general. However, the improvement in EU and EU15 is marginal, consistent with low significance of cross-correlation coefficients in these areas. How is the significance of the explanatory variables affected by unobservable country heterogeneity? A close examination of estimates across models shows that, in contrast to pooled results, controlling for random effects in general has little effect on the estimates and their significance regionally. We also compared the rate of correctly predicted regimes across models. In the pooled model, the model predicts best the regime choice in the full sample (Table 1). The prediction is the highest for the intermediate regime in countries belonging to the restricted sample and for which the series poli is available (last column). Among the currency zones (not shown), the model predicts best the regime choice in the full sample of the CFA zone, and the restricted sample of EU15. In terms of regimes, the model predicts best the fixed regime in full sample in CFA, and the intermediate regimes in EU15. As for the random effects, we find that controlling for country heterogeneity does not significantly improve the predictions in general. Next, we conduct various sensitivity checks. First, we consider the effect of changing our sample, by including or excluding certain regimes. Second, we compare our results based on IMF's de jure classification with those on de facto classification by Bubula and Otker-Robe (2002). In the last section we analyze regime persistence. 4.7. Robustness checks Since the introduction of the political variable reduces the sample size considerably in several cases, we will omit it for various checks we perform below. This allows us to maximize the degree of freedom for the tests and gives more reliable results. 4.7.1. Sample analysis Due to change in the samples, the full sample and the restricted sample results Tables 1, 1a–1c are hard to compare. We observe differences in going from the full sample to restricted sample in each region, and from pooled data to regional data. In this section we examine the data to understand whether these differences are caused by the frequency of regime changes or by regional factors. We take the restricted sample as the benchmark in each region and add to it groups of countries with a single regime, one at-a-time. To illustrate, suppose we would like to examine the effect of pure floaters on the estimates obtained from a regression with the restricted sample. We compare the restricted sample results with those obtained from a sample consisting of the restricted sample plus pure floaters. Table 2 summarizes the results. In each panel, in the first column we replicate the results from RS without random effects from Tables 1, 1a–1c. The next columns, in the order of appearance, are regression results with the restricted sample plus pegs (RSP in column 2), restricted sample plus floats (RSF in column 3), restricted sample plus intermediates (RSI in column 4). The full sample pool results are fairly robust to sample specification (panel 1). Adding the fixers, floater or the intermediate regimes does not affect the significance or the sign of gdp, gcf and rerv. Openness becomes significant when peggers are added, and inf becomes insignificant when nonpeggers are added. We repeat the same exercise with the regional samples and conclude that the differences between pooled results and regional results are due to regional idiosyncrasies, not the frequency of regime changes. In the EU zone, including the floaters increases the significance of openness and the volatility of the real exchange rate (panel 2). Inclusion of fixers or intermediates affects the results little. In the LAC area, adding the fixed regimes increases the significance of the inflation differential and capital flows (panel 3). Adding only the intermediate regime makes inf significant but does not affect the coefficient of gcf. This is consistent with the common view that currency boards and hard fixes provide low inflation and lead to greater integration with the international capital markets because of lower risk of exchange rate uncertainty. Adding the floaters does not affect the full sample results qualitatively. By contrast to the LAC area, adding the fixed regimes to the restricted sample in the EAP region reduces the significance of openness and capital flows substantially (panel 4). One explanation for this puzzling result is presumably the Asian crisis. To check this, we reran the regression by excluding the period 1997–99. Doing so increased the significance of these variables back to the level of restricted sample in column 1 and the magnitude of the coefficient of gcf above its original level. This result reflects the detrimental effect of the crisis on the real and

452

I. Kato, M. Uctum / International Review of Economics and Finance 17 (2008) 436–456

Table 3 Comparison of de facto classification with de jure classification of regimes a sample: 1990–1999 Full Sample

Restricted Sample

De jure

De facto

De jure

De facto

Random Effects De jure

gdp

poli

0.11 (5.3) − 0.04 (4.3) 0.05 (3.1) 0.01 (0.01) 0.22 (3.3) –

LL LR Pseudo-R2 N

− 1163.1 398.2 0.15 1248

open inf gcf rerv

−0.04 (1.4)b −0.03 (3.5) 0.04 (1.9) −0.00 (0.2) 0.36 (3.3) 0.01 (2.8) −884.9 230.5 0.12 880

0.03 (1.3) −0.06 (6.3) 0.01 (1.1) −0.01 (0.5) 0.04 (1.0) – −987.1 528.0 0.21 1248

−0.13 (4.2)b −0.05 (4.5) 0.00 (0.2) −0.02 (1.4) 0.30 (2.6)b 0.02 (3.7) −647.6 332.9 0.20 880

− 0.02 (1.0) − 0.01 (0.6) 0.01 (0.4) − 0.02 (0.7) 0.24 (2.1) – − 769.4 150.0 0.09 787

− 0.05 (1.6) − 0.01 (0.5) 0.00 (0.0) − 0.00 (0.3) 0.19 (1.3)b − 0.00 (0.5) − 609.0 124.3 0.09 638

− 0.14 (4.8) − 0.03 (2.4) − 0.03 (1.5) − 0.05 (1.5) 0.05 (1.2) – − 488.2 151.5 0.13 787

− 0.18 (5.1) − 0.05 (3.3) − 0.07 (2.4) b − 0.01 (0.3) 0.23 (1.5) 0.01 (1.0) − 396.7 134.4 0.15 638

− 0.09 (1.8) 0.02 (1.3) 0.00 (0.2) − 0.10 (2.1) 0.49 (2.7) – − 657.2 158.8 0.11 787

De facto − 0.10 (1.7) 0.06 (3.4) 0.01 (0.4) − 0.06 (1.3)b 0.26 (1.2)b − 0.01 (1.2) − 515.4 131.2 0.11 638

− 0.26 (2.9) − 0.06 (1.9) − 0.02 (0.7) − 0.07 (1.2) 0.12 (1.5) – − 373.1 154.7 0.17 787

− 0.30 (3.2) − 0.06 (2.0) −0.05 (1.2) − 0.03 (0.5) 0.08 (0.7) 0.01 (0.6) − 313.5 117.8 016 638

Regression results inclusive of regional-time effects to capture the cross-sectional variation. a See legend of Table 1. b Change in significance due to change of country sample for poli.

financial integration in this region and in particular on economies with pegged regimes. Adding the floaters or the intermediate regimes does not change the original results significantly. Finally, we do not report the results for the CFA zone since they either replicate the full sample or the restricted sample results. This is because all countries in this zone have either fixed or intermediate regimes and it is not possible to create distinct samples, as is the case with the other regions. 4.7.2. Comparison of de jure and de facto regime categorizations Our objective is to see if there is a significant difference in the estimated coefficients of the independent variables using two different categorizations of exchange rate regimes. Because Bubula et al. sample does not start before 1990, we confine the analysis to 1990–99 (Table 3). We first perform the ordered probit regression over the full sample, and compare the de jure and de facto exchange rate regimes (left panel). We then restrict the samples by eliminating countries with no change in their exchange rates, and repeat the ordered probit analysis with random effects (right panel). In all regressions we control for regional trend dummies. It is interesting to note that both sample results have more in common than differences. Most estimates in both categorizations have the same sign and roughly the same magnitude. Openness in the full sample and the institutional variable in all samples have the same sign and significance in both categorizations. Inflation in the restricted sample and capital flows in all samples are insignificant in both measures. Exchange rate volatility significantly and positively affects the regime choice in the de jure measurement and it is positive but often not significant in the de facto measurement. GDP enters both de jure and de facto measurements in full sample (no poli) with a positive sign and all the other samples negatively, though it has lower significance in the subsample containing poli with the de jure measurement. A similar exercise without time trends gave even more comparable outcomes. Overall, results suggest that de facto and de jure categorizations lead to fairly equivalent results. At this point we can also tie our discussion to another strand of the literature that examines “the fear of floating” behavior. This theory argues that de jure floaters actually intervene heavily to stabilize their currency, exhibiting “fear of floating” (Calvo & Reinhart, 2002). According to this theory we should continue to expect a positive relation between de jure floaters and high volatility since countries stay in this category when volatility is high even though they actually intervene. When we switch to the de facto measurement, however, we should see a weaker positive relation between volatility and the probability of floating since many countries with the fear of floating now would be

I. Kato, M. Uctum / International Review of Economics and Finance 17 (2008) 436–456

453

categorized as fixers. When we compare both measurements, evidence shows that the volatility estimates for de facto categorization in Table 3 are consistently smaller (or insignificant or both) than the estimates for de jure categorization. This result lends support to the “fear of floating” hypothesis. 4.7.3. Regime persistence In most panel analyses it is customary to examine the state dependence of the data. In our case, this would amount to examining how the exchange rate regime in the previous period affects the probability distribution of current exchange rate regime choice. The estimation equation is then modified to Yit⁎ = β′xit + θ′wit + γ′dit−1 + εit where dit−1 is a vector for dummy variables, which represents the country's choice of exchange regimes in the previous period, and we test for the significance of γ′. Often in micro analyses regime persistence is a significant component of the choice and in our case, it would not be surprising to find the same result. The problem with exchange rate regimes, however, is that the dependent variable changes infrequently due to high costs of changing the regime. In several cases it does not change at all, as is shown by our analysis distinguishing full sample from restricted sample. Thus, we expected a lagged dependent variable to dominate the effect of the other explanatory variables. We reran the regressions including the previous regime state. As predicted, in all specifications this effect is large and significant, suggesting that countries are more likely to remain in their initial regime rather than move away from it, regardless the other independent variables. 5. Conclusion In this paper we examine countries' choice of exchange rate regimes with an ordered choice variable analysis. Within a framework of optimal currency area, we consider three different currency blocks, the US dollar–consisting of the Latin America and the Caribbean (LAC), and the East Asian and Pacific (EAP) regions–, the EU (ECU/Euro) area, and the CFA franc region. By breaking down the data in currency blocks, we control for one type of regional heterogeneity in the sample. We run various robustness checks, including different sample length, different country samples, comparison of de jure and de facto measures of exchange rate regimes. Our methodology, data span and the comparison of different currency zones provide a framework that allows a detailed analysis of the exchange rate regimes. Our findings suggest that, even when we control for factors such as unobservable country heterogeneity, time dummies, and different samples, substantial regional differences remain, a result that calls for disaggregation of the data such as currency zones. The pooled results indicate that among the explanatory variables analyzed, openness, exchange rate volatility and the institutional variables are the most likely to affect the exchange rate choice of a country. Inflation and GDP are significant only in the full sample. However, these results mask the differences at the regional level. The GDP is consistently an important determinant of the exchange regime determination in all currency zones. Openness is an important determinant in EAP, exchange rate volatility in EU15 and LAC, and inflation in the full samples of CFA and LAC, after 1985 in EAP and EU15. We also show that samples incorporating countries with a single regime often lead to different conclusions than samples consisting of only countries with variation in the regimes. In the EU zone, adding the floaters to the sample increases the significance of openness and the volatility of real exchange rate. In the LAC area, adding pure pegs increases the significance of inflation differential and capital flows, consistent with the view that currency boards and hard fixes provide low inflation and lead to greater international financial integration due to lower risk of exchange rate uncertainty. By contrast, adding the pegs to the sample in the EAP region reduces the significance of openness and capital flows if the period after 1997 is included, reflecting the disturbance caused by the Asian crisis. Appendix A. List of countries A.1. USD Zone LAT: Latin America and the Caribbean Antigua and Barbuda⁎, Argentina⁎, Bahamas The⁎, Barbados⁎, Belize⁎, Bolivia, Brazil, Chile, Colombia, Costa Rica, Dominica⁎, Dominican Republic, Ecuador, El Salvador, Grenada⁎, Guatemala, Guyana, Haiti, Honduras,

454

I. Kato, M. Uctum / International Review of Economics and Finance 17 (2008) 436–456

Jamaica, Mexico, Nicaragua⁎, Panama⁎, Paraguay, Peru⁎, St. Kitts and Nevis⁎, St. Lucia⁎, St. Vincent and the Grenadines⁎, Suriname, Trinidad and Tobago, Uruguay, Venezuela RB; EAP: East Asia and the Pacific Australia, Cambodia⁎, China, Hong Kong, China⁎, Indonesia, Japan⁎, Korea, Rep., Lao PDR, Malaysia, Mongolia⁎, New Zealand⁎, Papua New Guinea, Singapore, Solomon Islands⁎, Thailand; Other regions Algeria, Angola⁎, Armenia⁎, Azerbaijan, Bahrain⁎, Bangladesh⁎, Belarus⁎, Burundi, Canada⁎, Congo Dem. Rep.⁎, Egypt, Arab Rep., Ethiopia, Gambia The, Georgia⁎, Germany, Ghana, Guinea, Hungary, India, Iran Islamic Rep., Israel, Jordan⁎, Kenya, Kyrgyz Republic⁎, Lebanon, Lithuania⁎, Malawi, Maldives, Mauritania, Mauritius, Mozambique, Nepal, Nigeria, Pakistan, Romania, Russian Federation, Rwanda, Saudi Arabia, Sierra Leone, South Africa⁎, Sri Lanka⁎, Syrian Arab Republic⁎, Tanzania, Turkey⁎, Turkmenistan, Uganda, Ukraine⁎, Yemen Rep., Zambia, Zimbabwe; A.2. CFA Franc Zone Benin⁎, Burkina Faso⁎, Cameroon⁎, Cape Verde⁎, Central African Republic⁎, Chad⁎, Comoros⁎, Congo, Rep.⁎, Cote d’Ivoire⁎, Equatorial Guinea⁎, Gabon⁎, Guinea-Bissau, Madagascar, Mali⁎, Morocco, Niger⁎, Senegal⁎, Togo⁎, Tunisia; A.3. Europe: ECU, DM and the euro zone Albania⁎, Austria, Belgium, Bulgaria, Croatia, Cyprus, Czech Republic, Denmark⁎, Estonia⁎, Finland, France, Greece⁎, Iceland, Ireland, Italy, Macedonia FYR, Malta⁎, Moldova⁎, Netherlands, Norway, Poland, Portugal, Slovak Republic, Slovenia⁎, Spain, Sweden, Switzerland⁎, United Kingdom. Note: ⁎ indicates the country did not experience any regime change. Appendix B Table 1A Hausman–Wu tests Pooled FS gdp

1.08 0.31 [0.3 0.6] open 1.69 2.65 [0.2 0.1] inf 0.07 3.02 [0.8 0.1] gcf 0.06 0.05 [0.8 0.8] rerv 0.22 0.54 [0.6 0.5] poli 0.44 0.01 [0.5 0.5] N 1210

EU

EU15

CFA

LAC

EAP

RS

FS

RS

FS

RS

FS

RS

FS

RS

FS

RS

0.40 0.25 [0.5 0.6] 4.48 5.81 [0.0 0.0] 0.12 1.56 [0.7 0.2] 1.12 3.58 [0.3 0.1] 1.69 0.00 [0.2 1.0] 1.10 3.96 [0.3 0.1] 882

2.53 2.03 [0.1 0.2] 4.72 1.46 [0.0 0.2] 2.47 1.82 [0.1 0.2] 2.29 6.01 [0.1 0.0] 0.73 2.81 [0.4 0.1] 0.36 2.06 [0.5 0.2] 282

0.39 0.03 [0.5 0.9] 7.98 2.61 [0.0 0.1] 0.20 0.05 [0.7 0.8] 0.65 2.71 [0.4 0.1] 1.06 1.84 [0.3 0.2] 0.03 0.03 [0.9 0.9] 228

0.22 4. 44 [0.6 0.0] 1.04 0.06 [0.3 0.8] 0.32 1.36 [0.6 0.2] 0.11 1.72 [0.7 0.2] 0.00 0.14 [0.9 0.7] 6.81 0.62 [0.0 0.4] 183

0.01 3.68 [0.9 0.1] 1.37 0.39 [0.2 0.5] 0.06 2.14 [0.8 0.1] 0.00 0.67 [0.9 0.4] 0.00 0.01 [0.9 0.9] 4.74 0.10 [0.0 0.8] 160

0.56 0.58 [0.5 0.4] 0.03 0.01 [0.9 0.9] 11.22 12.8 [0.0 0.0] 0.00 0.01 [0.9 0.9] 4.25 4.91 [0.0 0.0] 0.01 0.04 [0.9 0.8] 164

1.12 0.33 [0.3 0.6] 0.10 0.08 [0.8 0.8] 1.04 0.32 [0.3 0.6] 0.04 0.00 [0.8 0.9] 1.77 1.03 [0.2 0.3] 0.12 0.44 [0.7 0.5] 59

0.11 0.83 [0.7 0.4] 0.4 0.23 [0.5 0.6] 1.09 1.27 [0.3 0.2] 5.41 4.21 [0.0 0.0] 0.16 0.04 [0.7 0.8] 0.10 0.91 [0.7 0.3] 229

0.06 0.87 [0.8 0.3] 0.40 0.30 [0.5 0.6] 0.07 0.13 [0.8 0.7] 0.07 0.00 [0.8 0.9] 1.44 1.12 [0.2 0.3] 0.11 1.08 [0.7 0.3] 201

8.56 2.39 [0.0 0.1] 2.87 0.00 [0.1 0.9] 3.10 0.68 [0.1 0.4] 0.50 9.64 [0.5 0.0] 0.52 1.73 [0.5 0.2] 5.68 3.83 [0.0 0.1] 143

6.15 2.44 [0.0 0.1] 1.84 0.08 [0.2 0.8] 4.15 0.00 [0.0 0.9] 5.86 10.05 [0.0 0.0] 1.30 4.36 [0.3 0.0] 3.07 2.33 [0.1 0.1] 110

The independent variables are: gdp = gross domestic product; open = openness; inf = inflation differential; gcf = gross capital flows; rerv = real exchange rate volatility; poli = political risk (an increase is an improvement), N is the number of observations. Variables open, inf and gcf are scaled up by 10. Full Sample contains countries with zero, one or more regime changes. Restricted Sample excludes countries with no regime change. The first entry in each cell is the χ2(1) statistics for the corresponding contemporaneous independent variable, the second entry is the same statistics for the lagged value of the independent variable with 3.84 critical value at the 95% confidence level. Figures in square brackets are the corresponding p-values.

I. Kato, M. Uctum / International Review of Economics and Finance 17 (2008) 436–456

455

Table 2A Trend estimates⁎ Pooled

t83 t84 t85 t86 t87 t88 t89 t90 t91 t92 t93 t94 t95 t96 t97 t98 t99

EU

E15

FCA

LAC

EAP

Pooled

FS

RSRE

FS

RSRE

FS

RSRE

FS

RSRE

FS

RSRE

FS

RSRE

De jure

De facto

0.06 (0.3) 0.01 (0.0) 0.07 (0.4) 0.13 (0.7) 0.13 (0.7) 0.08 (0.4) 0.21 (1.2) 0.32 (1.8) 0.42 (2.4) 0.69 (3.9) 0.68 (3.9) 0.82 (4.7) 0.73 (4.3) 0.70 (4.1) 0.70 (4.1) 0.61 (3.6) 0.60 (3.5)

0.1 (0.4) 0.04 (0.2) 0.16 (0.7) 0.27 (1.2) 0.29 (1.2) 0.18 (0.8) 0.35 (1.5) 0.61 (2.7) 0.79 (3.5) 1.37 (6.1) 1.39 (6.2) 1.72 (7.6) 1.6 (7.1) 1.59 (7.1) 1.56 (7.0) 1.28 (5.8) 1.22 (5.6)

− 0.04 (0.1) 0.01 (0.0) 0.02 (0.0) 0.00 (0.3) − 0.12 (0.3) − 0.21 (0.5) − 0.28 (0.6) − 0.27 (0.6) − 0.23 (0.5) 0.84 (2.0) 0.61 (1.5) 0.74 (1.8) 0.59 (1.5) 0.47 (1.2) 0.42 (1.1) − 0.05 (0.1) 0.03 (0.1)

0.0 (0.0) − 0.03 (0.1) 0.01 (0.0) 0.00 (0.0) − 0.08 (0.2) − 0.08 (0.2) − 0.40 (0.8) − 0.29 (0.6) − 0.31 (0.6) 1.00 (2.1) 0.84 (1.7) 1.04 (2.2) 0.65 (1.4) 0.59 (1.3) 0.52 (1.1) − 0.26 (0.6) − 0.07 (0.1)

− 0.11 (0.2) − 0.02 (0.0) 0.15 (0.3) 0.28 (0.5) 0.27 (0.5) 0.49 (0.9) 0.2 (0.4) 0.09 (0.2) 0.17 (0.3) 1.38 (2.4) 1.47 (2.6) 1.58 (2.7) 1.48 (2.6) 0.99 (1.8) 1.12 (2.0) − 0.78 (1.4) − 0.94 (1.6)

− 0.13 (0.2) 0.01 (0.0) 0.29 (0.5) 0.29 (0.5) 0.33 (0.5) 0.55 (0.8) 0.29 (0.4) − 0.02 (0.0) − 0.24 (0.4) 1.09 (1.5) 1.2 (1.7) 1.32 (1.9) 1.17 (1.7) 0.64 (0.9) 0.87 (1.3) − 1.49 (2.1) − 1.55 (2.1)

−0.05 (0.1) 0.58 (0.8) 0.06 (0.1) 0.86 (1.2) 0.42 (0.6) 0.21 (0.3) 0.14 (0.2) −0.14 (0.2) 0.37 (0.1) 0.12 (0.2) 0.1 (0.1) 0.44 (0.6) −0.89 (1.2) −0.65 (0.8) −0.34 (0.4) −0.03 (0.0) 0.18 (0.2)

0.1 (0.1) 0.94 (1.0) 0.32 (0.3) 1.77 (1.8) 2.02 (1.8) 1.03 (0.8) 0.48 (0.4) −0.21 (0.2) 0.51 (0.5) 1.13 (1.1) 0.93 (0.9) 1.86 (1.8) 1.77 (1.7) 0.95 (0.9) 1.28 (1.2) 0.82 (0.8) 1.4 (1.3)

0.15 (0.4) −0.04 (0.1) 0.18 (0.5) −0.2 (0.5) −0.24 (0.6) −0.33 (0.8) 0.23 (0.6) 0.52 (1.4) 0.96 (2.5) 0.9 (2.3) 0.9 (2.4) 0.95 (2.6) 0.79 (2.1) 0.74 (2.0) 0.82 (2.2) 0.81 (2.2) 0.75 (2.0)

0.34 (0.8) 0.28 (0.6) 0.54 (1.2) 0.1 (0.2) 0.09 (0.2) − 0.15 (0.3) 0.61 (1.4) 1.01 (2.3) 1.64 (3.7) 2.1 (4.5) 1.95 (4.3) 1.96 (4.3) 1.47 (3.3) 1.44 (3.2) 1.34 (3.1) 1.3 (3.0) 1.2 (2.8)

0.15 (0.3) − 0.1 (0.2) − 0.14 (0.3) 0.06 (0.1) 0.15 (0.3) 0.11 (0.2) 0.42 (0.8) 0.41 (0.8) 0.43 (0.9) 0.58 (1.1) 0.58 (1.1) 0.86 (1.7) 1.08 (2.1) 0.95 (1.9) 1.38 (2.7) 1.3 (2.5) 1.03 (1.9)

0.38 (0.6) − 0.19 (0.3) − 0.27 (0.4) 0.14 (0.2) 0.36 (0.5) 0.26 (0.4) 0.53 (0.8) 0.46 (0.7) 0.58 (0.9) 0.96 (1.4) 0.93 (1.4) 1.64 (2.4) 2.01 (2.9) 1.6 (2.3) 2.67 (3.7) 2.47 (3.5) 2.11 (2.7)

–

–

–

–

–

–

–

–

–

–

–

–

–

–

–

–

0.14 (0.8) 0.59 (3.1) 0.62 (3.3) 0.87 (4.6) 0.77 (4.2) 0.77 (4.2) 0.70 (3.8) 0.47 (2.6) 0.41 (2.2)

0.16 (0.7) 0.38 (1.8) 0.45 (2.1) 0.58 (2.7) 0.59 (2.8) 0.59 (2.8) 0.69 (3.3) 0.79 (3.8) 0.53 (2.6)

⁎FS, RSRE refer to full sample and restricted sample random effects, respectively. Full Sample contains countries with zero, one or more regime changes. Restricted Sample excludes countries with no regime change. Estimates are based on the model with simple time dummies. De jure and de facto analysis is based on restricted sample without random effects.

Table 3A Marginal effects pooled estimates⁎ Full Sample

gdp open inf gcf rerv poli

Restricted Sample

Restricted sample random effects

Fixed

Intermediate

Flexible

Fixed

Intermediate

Flexible

Fixed

Intermediate

Flexible

− 0.004 0.003 − 0.005 0.000 − 0.056 − 0.001

0.003 − 0.003 0.004 0.000 0.053 0.001

0.002 0.000 0.000 0.000 0.004 0.000

− 0.004 0.002 − 0.003 − 0.013 − 0.150 0.001

0.003 −0.001 0.002 0.008 0.096 −0.000

0.001 − 0.001 0.001 0.005 0.054 − 0.001

0.000 − 0.004 − 0.002 0.004 − 0.071 0.003

0.002 −0.009 0.005 0.008 −0.151 0.006

− 0.002 − 0.013 0.007 − 0.011 0.222 − 0.01

⁎Full Sample contains countries with zero, one or more regime changes. Restricted Sample excludes countries with no regime change. All samples start from 1985 due to inclusion of poli in the regression. The independent variables are: gdp = gross domestic product; open = openness; inf = inflation differential; gcf = gross capital flows; rerv = real exchange rate volatility; poli = political risk (an increase is an improvement). N is the number of observations. Variables open, inf and gcf are scaled up by 10.

456

I. Kato, M. Uctum / International Review of Economics and Finance 17 (2008) 436–456

References Arellano, M., & Bond, S. (1991). Some tests of specification for panel data: Monte Carlo evidence and an application to employment equations. Review of Economic Studies, 58(2), 277−297. Blomberg, B., Frieden, J., & Stein, E. (2005). Sustaining fixed rates: The political economy of currency pegs in Latin America. Journal of Applied Economics, 8(2), 203−225. Bosco, L. (1987). Determinants of the exchange rate regimes in LDCs: Some empirical evidence. Economic Notes, 1, 110−143. Bubula, A., & Otker-Robe, I. (2002). Evolution of exchange rate regimes since 1990: Evidence from de facto policies. IMF WP/02/155. Butler, J., & Moffitt, R. (1982). A computationally efficient quadrature procedure for the one factor multinomial probit model. Econometrica, 50, 761−764. Calvo, G., & Reinhart, C. (2002). Fear of floating. Quarterly Journal of Economics, 117, 379−408. Collins, S. (1996). On becoming more flexible: Exchange rate regimes in Latin America and the Caribbean. Journal of Development Economics, 51, 117−138. Dreyer, J. S. (1978). Determinants of exchange-rate regimes for currencies of developing countries: Some preliminary results. World Development, 6, 437−445. Edwards, S. (1996) The determinants of the choice between fixed and flexible exchange-rate regimes. Inc, NBER Working Paper No. 5756. Eichengreen, B., Rose, A., & Wyplosz, C. (1996). Contagious currency crises: First tests. Scandinavian Journal of Economics, 98, 463−484. Eliasson, A. C., & Kreuter, C. (2001, December). On currency crisis: A continuous crisis definition. Conference paper, “X International “Tor Vergata” Conference on Banking and Finance [http://www.economia.uniroma2.it/ceis/conferenze_convegni/banking2001/papers/mercoledi/ Eliasson-Kreuter.pdf] Frankel, J. A., & Rose, A. (1996). Currency crashes in emerging markets: An empirical treatment. Journal of International Economics, 41, 351−366. Frechette, G. (2001). Random-effects ordered probit. STATA Technical Bulletin: STATA Corporation. Hausman, J. (1978). Specification tests in econometrics. Econometrica, 46, 1251−1271. Heller, H. R. (1978). Determinants of exchange rate practices. Journal of Money, Credit and Banking, 10, 308−321. Holden, P., Holden, M., & Suss, E. C. (1979). The determinants of exchange rate flexibility: An empirical investigation. Review of Economics and Statistics, 61, 327−333. Juhn, G., Mauro, P. (2002). Long-run determinants of exchange rate regimes: A simple sensitivity analysis. IMF Working Papers, 02/104. Levy-Yeyati, E., & Sturzengger, F. (2005). Classifying exchange rate regimes: Deeds vs. words. European Economic Review, 49(6), 1603−1635. Masson, P. (2000). Exchange Rate Regime Transitions. IMF Working Paper IMF No. 00/134. Melvin, M. (1985). The choice of an exchange rate system and macroeconomic stability. Journal of Money, Credit and Banking, 17, 467−478. Meon, P. -G., & Rizzo, J. -M. (2002). The viability of fixed exchange rate commitments: Does politics matter? A theoretical and empirical investigation. Open Economies Review, 13(2), 111−132. Mundell, R. A. (1961). A theory of optimum currency areas. American Economic Review, 51, 657−665. O'Connelle, J. (1998). The overvaluation of purchasing power parity. Journal of International Economics, 44, 1−19. Poirson, H. (2001). How do countries choose their exchange rate regime? IMF Working Paper, WP/01/46. Reinhart, C. M., Rogoff, K.S. (2002). The modern history of exchange rate arrangements: A reinterpretation. NBER Working Paper No. 8963. Rizzo, J. M. (1998). The economic determinants of the choice of an exchange rate regime: A probit analysis. Economics Letters, 59, 283−287. Rogoff, K. (1985). Can international monetary policy cooperation be counterproductive? Journal of International Economics, 8, 199−217. Rogoff, K. (1985). Can international monetary policy cooperation be counterproductive? Journal of International Economics, 18(3–4), 199−217. Savvides, A. (1990). Real exchange rate variability and the choice of exchange rate regime by developing countries. Journal of International Money and Finance, 9, 440−454. Wooldridge, J. M. (2002). Econometric analysis of cross section and panel data. Cambridge: MIT Press. Wu, D. (1973). Alternative tests of independence between stochastic regressors and disturbances. Econometrica, 44, 733−750.