Journal of Development Economics 73 (2004) 55 – 74 www.elsevier.com/locate/econbase Financial development and economic growth: evidence from panel unit root and cointegration tests Dimitris K. Christopoulos a,*, Efthymios G. Tsionas b a Department of Economic and Regional Development, Panteion University, Leof. Syngrou 136, 17671, Athens, Greece b Department of Economics, Athens University of Economics and Business, Athens, Greece Received 1 November 2001; accepted 1 March 2003 Abstract In this paper we investigate the long run relationship between financial depth and economic growth, trying to utilize the data in the most efficient manner via panel unit root tests and panel cointegration analysis. In addition, we use threshold cointegration tests, and dynamic panel data estimation for a panel-based vector error correction model. The long run relationship is estimated using fully modified OLS. For 10 developing countries, the empirical results provide clear support for the hypothesis that there is a single equilibrium relation between financial depth, growth and ancillary variables, and that the only cointegrating relation implies unidirectional causality from financial depth to growth. D 2003 Elsevier B.V. All rights reserved. JEL classification: C23; O16; O40; G28 Keywords: Financial development; Growth; Panel unit roots; Panel cointegration; Threshold cointegration 1. Introduction A large and expanding literature tries to shed some light on the roles of policy or ‘‘ancillary’’ variables in the determination of economic growth. Most of this literature has mainly focused on the role of macroeconomic stability, inequality, income and wealth, institutional development, ethnic and religious diversity and financial market imperfections. For an extensive survey of this literature, see Levine (1997). Among these factors * Corresponding author. Tel.: +30-210-9224948; fax: +30-210-9229312. E-mail address: christod@panteion.gr (D.K. Christopoulos). 0304-3878/$ - see front matter D 2003 Elsevier B.V. All rights reserved. doi:10.1016/j.jdeveco.2003.03.002 56 D.K. Christopoulos, E.G. Tsionas / Journal of Development Economics 73 (2004) 55–74 the role of financial markets in the growth process has received recently considerable attention. In this framework, financial development is considered by many economists to be of paramount importance for output growth. In particular, government restrictions on the banking system (such as interest rate ceiling, high reserve requirements and directed credit programs) hinder financial development and reduce output growth, see McKinnon (1973) and Shaw (1973). Likewise, the endogenous growth literature stresses the influence of financial markets on economic growth, see among others Bencivenga et al. (1995), Greenwood and Smith (1997), and Obstfeld (1994). These authors include financial intermediaries, information collection and analysis, risk sharing etc in the proposed models. In this line of research, Benhabib and Spiegel (2000) argue that a positive relationship is expected also to exist between financial development and total factor productivity growth and investment. However, their empirical results are very sensitive to model specification. Further, Beck et al. (2000) find that financial development has a large positive impact on total factor productivity (TFP), which feeds through to overall GDP growth. See also Neusser and Kugler (1998). A problem with the previous studies is that a positive relationship between financial development and output growth can exist for different reasons. As output increases the demand for financial service increases too, which in turn has a positive effect on financial development. Other things being equal, it is financial development that follows output growth and not the opposite. This issue was considered in Robinson (1952, p. 86). Others were keener to totally dismiss the impact of financial development on economic growth. Lucas (1988, p. 6) states, for example, that ‘‘the importance of financial matters is very badly overstressed’’ while Chandavarkar (1992, p. 134) notes ‘‘none of the pioneers of development economics. . . even list finance as a factor of development’’. See also Luintel and Khan (1999). Although many empirical studies have investigated the relationship between financial depth, defined as the level of development of financial markets, and economic growth, the results are ambiguous. On the one hand, cross country and panel data studies find positive effects of financial development on output growth even after accounting for other determinants of growth as well as for potential biases induced by simultaneity, omitted variables and unobserved country-specific effect on the finance-growth nexus, see for example King and Levine (1993a,b), Khan and Senhadji (2000) and Levine et al. (2000). On the other hand, time series studies give contradictory results. Demetriades and Hussein (1996) find little systematic evidence in favor of the view that finance is a leading factor in the process of economic growth. In addition they found that for the majority of the countries they examine, causality is bi-directional, while in some cases financial development follows economic growth. Luintel and Khan (1999) used a sample of ten less developed countries to conclude that the causality between financial development and output growth is bi-directional for all countries. All these results show that a consensus on the role of financial development in the process of economic growth does not so far exist. There are a number of concerns with previous empirical work. Although the nature of I(1) variables has been recognized as critical, and proper estimation techniques (organized around unit roots and cointegration) have been used, the small samples typically used may significantly distort the power of standard tests, and lead to misguided conclusions. Thus, D.K. Christopoulos, E.G. Tsionas / Journal of Development Economics 73 (2004) 55–74 57 all efforts must be made to utilize the data in the most efficient manner in order to draw sharp inferences. This is true not only for unit root but also for cointegration inferences. Even with time series tests, however, the possibility exists that there are threshold effects in a possible cointegrating relationship between output and financial development. Indeed, it could be that below a level of financial development there is no effect on growth or a small effect, and a larger effect as financial development crosses the threshold. Important studies in the field include Levine et al. (2000) and Beck et al. (2000) who take into account the problem of simultaneity of regressors. They used a GMM dynamic panel estimator but they did not consider the integration and cointegration properties of the data. Thus, it is not clear what the estimated panel models represent: Do they represent a structural long run equilibrium relationship or a spurious one? We consider this issue of the utmost importance, and we believe the present paper throws some light on this question. More specifically, this paper contributes the following: (1) We use time series unit root tests along with panel unit root tests to examine the stationarity properties of the data. The use of panel-based tests is necessary because the power of standard time-series unit root tests may be quite low given the sample sizes and time spans typically available in economics. (2) The cointegration framework of Johansen (1988) is applied to test for multivariate cointegrating relationships. Additionally, panel cointegration tests are conducted to make sure that Johansen tests do not suffer power loss due to finite samples. (3) We take into account the possibility that the relationship between economic activity and financial development may involve a ‘‘threshold effect’’, see Berthelemy and Varoudakis (1996) and Deidda and Fatouh (2002). It could be that below a level of financial development there is no effect on growth, and a sizeable effect above the threshold. Further, the presence of a threshold affects seriously the stationarity properties of the data. According to Enders and Granger (1998) standard tests of integration and cointegration have lower power in the presence of misspecified dynamics. This calls for application of cointegration tests that account for possible threshold effects in the long run relationship if we are to obtain sensible results. (4) Cointegrating vectors are estimated using the fully modified (FM) OLS estimation technique for heterogeneous cointegrated panels (Pedroni, 2000). This methodology allows consistent and efficient estimation of cointegrating vectors. We argue below why this methodology retains the flexibility of the Levine et al. (2000) approach while at the same time: (a) allows consistency of the long-run relation with the short-run adjustment, (b) deals with the endogeneity of regressors problem, and (c) respects the time-series properties of the data in that integration and cointegration properties are explicitly taken into account. (5) We distinguish between long run and short run causality. This distinction is very important since as Darrat (1999) states, most of the benefits of higher levels of financial development could be realized in the short-run while in the long run as the economy grows and becomes mature these effects slowly disappear. Thus, testing only for long run causality would lead to the wrong conclusion, namely absence of any casual relationship between financial development and output growth. To this end, we 58 D.K. Christopoulos, E.G. Tsionas / Journal of Development Economics 73 (2004) 55–74 specify and estimate an error correction model (ECM) appropriate for heterogeneous panels. The present paper addresses the empirical relationship between financial development and economic growth for ten less developed countries over the period 1970 –2000. In Section 2, we present a brief literature review. The econometric techniques are presented in Section 3. The results are discussed in Section 4. The paper concludes with a summary of methodology and results. 2. A review on the relationship between financial development and growth There is no general agreement among economists that financial development is beneficial for growth. In a simple endogenous growth model, Pagano (1993) uses the AK model to conclude that the steady state growth rate depends positively on the percentage of savings diverted to investment, so one channel through which financial deepening affects growth is converting savings to investment. Berthelemy and Varoudakis (1996) use a theoretical model with banks acting as Cournot oligopolists to find that, in the stable equilibrium, the growth rate depends positively on the number of banks, or the degree of competitiveness of the financial system. Their results show that educational development is a pre-condition of growth, and financial underdevelopment is an obstacle when the educational system is not successful. Greenwood and Jovanovic (1990) consider a model that allows examining the relation between growth and income distribution, as well as between financial structure and economic development. The fundamental reason for a positive effect of financial structure on growth is the more efficient undertaking of investment, and more efficient capital allocation because agents can have better information about the nature of shocks (aggregate versus idiosyncratic) that hit particular projects. This is more or less consistent with the classical view on the relation between growth and financial development. Levine (1991) considers an endogenous growth model with stock markets, and shows that they accelerate growth for two reasons: First, because ownership of firms can be traded without disrupting the production process. Second, because agents are allowed to diversify portfolios. The model has the reasonable implication that in the absence of stock markets, agents would be discouraged to invest because of risk aversion. Also ‘‘they accelerate growth directly by eliminating premature capital liquidation which increases firm productivity and indirectly by reducing liquidity risk which encourages firm investment’’ (p. 1459). Singh (1997) claims that financial development may be not be beneficial for growth for several reasons. He states that ‘‘first, the inherent volatility and arbitrariness of the stock market pricing process under DC [developing countries] conditions make it a poor guide to efficient investment allocation. Secondly, the interactions between the stock and currency markets in the wake of unfavorable economic shocks may exacerbate macroeconomic instability and reduce long-term growth. Thirdly, stock market development is likely to undermine the existing group-banking systems in DC’s, which, despite their many difficulties, have not been without merit in several countries, not least in the highly successful East Asian economies’’ (pp. 779 – 780). D.K. Christopoulos, E.G. Tsionas / Journal of Development Economics 73 (2004) 55–74 59 On the empirical side, King and Levine (1993a) use IMF data and various financial indicators to conclude that there is a positive relationship between financial indicators and growth, and that financial development is robustly correlated with subsequent rates of growth, capital accumulation, and economic efficiency. They correctly emphasize that policies that alter the efficiency of financial intermediation exert a first-order influence on growth. This is a standard implication of models of endogenous growth with financial intermediation. Atje and Jovanovic (1993) examine the role of stock markets on development, and conclude that there is positive effect on the level as well as on the growth. They could not, however, establish a significant relationship between bank liabilities and growth. Levine and Zervos (1996) use various measures of stock market development, and conclude that there is a significant relationship. When they include banking depth variables in their regressions, they turn out to be non-significant. They emphasize their results are indicative of partial correlation only, and more research would be needed in the area. Arestis and Demetriades (1997) use time series analysis (whereas previous studies use cross-section data) and Johansen cointegration analysis for the US and Germany. For Germany, they find an effect of banking development growth. In the US, there is insufficient evidence to claim a growth effect of financial development, and the data point to the direction that real GDP contributes to both banking system and stock market development. In this line of research Neusser and Kugler (1998) used manufacturing data from thirteen OECD countries over the period 1970 –1991 to analyze the existence of a long-run relationship between manufacturing sector GDP and financial sector GDP as well as between manufacturing TFP and financial sector GDP. To this end, different tests were performed including Johansen maximum likelihood and residual-based panel cointegration tests. They found that for the majority of countries they examine a cointegration relationship could be established not so much between financial sectors, GDP and manufacturing GDP but mostly between financial sector GDP and manufacturing TFP. Subsequent causality tests gave mixed results. For some countries financial activity causes manufacturing GDP, for others financial sector causes both manufacturing GDP and TFP while for others a feedback exists from manufacturing to financial sector. Levine et al. (2000), using a sample of 74 developed and less developed countries over the period 1960– 1995, go beyond previous studies recognizing the potential biases induced by simultaneity, omitted variables and unobserved country-specific effect on the finance growth nexus. According to these authors, this issue is of paramount importance for settling the question of causality. To deal effectively with theses problems, they suggest the use of estimators appropriate for dynamic panels like GMM as well as cross-sectional instrumental variable estimators where legal rights of creditors, the soundness of contract enforcement and the level of corporate accounting standards are used as instruments to extract the exogenous component of financial development. Both estimation techniques correct for biases associated with previous studies of the financial development-growth relation. At the same time, they offer more precise estimates. They found that the strong positive relationship between financial development and output growth can be partly explained by the impact of the exogenous components like finance development on economic growth. Levine et al. (2000) interpreted these results as supportive of the growth-enhancing hypothesis of financial development. These results are in line with those reported by Levine (1999) in a sample of 49 countries over the period 60 D.K. Christopoulos, E.G. Tsionas / Journal of Development Economics 73 (2004) 55–74 1960– 1989 where GMM procedures are used to find that the exogenous components of financial development (national, legal, and regulatory characteristics) have a positive influence on economic growth. The results could imply that the direction of causality between financial development and economic growth runs in both directions. Beck et al. (2000) investigate not only the relationship between financial development and economic growth but also the relationship between financial development and the sources of growth in terms of private saving rates, physical capital accumulation, and total factor productivity. Once again, GMM and IV estimators were used to correct for possible simultaneity biases. They conclude that higher levels of financial development lead to higher rates of economic growth, and total factor productivity. For the remaining variables, they could not document any relationship with financial development. Further, Levine (1998) using a sample of 44 developed and less developed countries during the period 1975– 1993, examines the links between banking development and long-run economic growth. The usual GMM estimation procedure is used to account for simultaneity bias. The degree to which legal codes emphasize the rights of creditor and the efficiency of the legal system in enforcing laws and contracts are considered as instruments. The empirical evidence is supportive for a strong positive relation between the exogenous component of banking development with output growth, physical accumulation and productivity growth. Finally, Demirgücß-Kunt and Maksimovic (1998) estimate a financial planning model to find that financial development facilitates the firm’s growth. In this context an active stock, market and a well-developed legal system are crucial for the further development of the firms. The empirical studies reviewed in this section are subject to a number of limitations: (a) With time series data, although the nature of I(1) variables has been recognized as critical, and proper estimation techniques (organized around unit root and cointegration) have been used, the small samples used may significantly distort the power of standard tests, and lead to misguided conclusions. (b) The cointegration tests including panel cointegration tests have lower power in the presence of misspecified dynamics, see Enders and Granger (1998). (c) An issue of simultaneity arises: To this end, Levine et al. (2000) and Beck et al. (2000) proposed the use of GMM dynamic panel estimators. However, in this approach, the integration and cointegration properties of the data are ignored. Thus, it is not clear that the estimated panel models represent a structural long run equilibrium relationship instead of a spurious one. Within this context, the imposition of homogeneity assumptions on the coefficients of lagged dependent variable in panel data estimators could lead to serious biases (Kiviet, 1995). Neusser and Kugler (1998) and Levine et al. (2000) represent two different poles in the literature. Neusser and Krugler focuses on time series properties of the data ignoring the simultaneity issue, while Levine et al. (2000) deal with simultaneity without accounting for the time series properties of the data. An alternative is explored in this paper. This alternative consists briefly in the following: In Levine et al. (2000) estimation is conducted in two steps, first a cross-sectional regression of growth on finance and ancillary regressors, and GMM in the second stage to address simultaneity. In our estimation approach, we exploit both the cross-sectional and time-series dimension of the data by using panel cointegration techniques. In that way we can address the simultaneity issues of the regressors but we also have another important advantage relative to previous research. D.K. Christopoulos, E.G. Tsionas / Journal of Development Economics 73 (2004) 55–74 61 In Levine et al. (2000), the first-pass cross-sectional regression represents the long-run regression while the second-pass regression (estimated by GMM) captures the short-run dynamics. The two regressions, however, are not connected as they should: One would expect that the second-pass regression can be derived from the long-run model by appropriate restrictions but this does not seem possible within the Levine et al. (2000) framework. More importantly, Levine et al. (2000) do not formally test that the first-pass regression is valid so it is not certain that it represents something structural. It is, therefore, not certain whether the second-stage regression represents an adjustment to the long-run equilibrium implied by the first stage. Within the panel cointegration framework used in this paper, we are able to address these important issues, and at the same time we retain the flexibility of the Levine et al. (2000) approach in that we are able to provide long-run estimates, short-run adjustments, and address the endogeneity issues by formally treating all variables as part of a vector autoregression in the context of testing for cointegration, and estimating panel cointegrating regressions. More importantly, we can formally test whether there is indeed a structural, long run relationship between financial development and growth. We consider this of the utmost importance: Without such a structural relationship, short run dynamics will be misleading. 3. The model and econometric techniques To investigate the relationship between growth and financial depth, we use the following model yit ¼ b0i þ b1i Fit þ b2i Sit þ b3i ṗit þ uit ð1Þ where yit is real output in country i and year t, Fit is a measure of financial depth, Sit is the output share of investment, ṗit is inflation, and uit is an error term. Since the direction of causality is not clear we also specify the model Fit ¼ b0i þ b1i yit þ b2i Sit þ b3i ṗit þ vit ð2Þ Both equations are to be considered as long run, or equilibrium relations. We may, of course, have more cointegrating relations involving inflation or investment share as the dependent variable. Provided all variables involved are integrated of order one, or I(1), valid economic inferences can be drawn only if these relations (or perhaps more, having investment share or inflation as dependent variable) are cointegrating relations, otherwise spurious inferences would result. Previous studies have examined cointegration on country by country basis by using time-series techniques, like Dickey-Fuller tests, and Johansen’s maximum likelihood cointegration methodology. However, given the short span of the data, we need to utilize information in the most efficient way, and make use of panel-based unit root and cointegration tests as well. In our empirical analysis, we will use pure time series tests and procedures as well, for comparison purposes. Regarding the data, y is the quantity of output expressed as an index number (1995 = 100), finance depth ( F) is the ratio of total bank deposits liabilities to nominal 62 D.K. Christopoulos, E.G. Tsionas / Journal of Development Economics 73 (2004) 55–74 GDP and the share of investment, S, is the share of gross fixed capital formation to nominal GDP. Inflation rate (ṗ ) is measured using the consumer price index. All data is drawn from issues of the International Financial Statistics published by the International Monetary Fund over the period 1970– 2000. We use data for ten developing countries, namely Colombia, Paraguay, Peru, Mexico, Ecuador, Honduras, Kenya, Thailand, Dominican Republic, and Jamaica. Regarding the measurement of financial deepening we have followed Luintel and Khan (1999). The selection of countries was dictated by the requirement of having continuous data records over the period 1970 – 2000, and it is the same set of countries used by De Mello (1999) in a different context. In the following, we describe the panel-based econometric procedures. 3.1. Testing for integration Before proceeding to the identification of a possible long run relationship we need to verify that all variables are integrated of order one in levels. However, since the power of individual unit root tests can be distorted when the span of the data is short (Pierse and Shell, 1995), we have used panel unit root tests due to Im et al. (1997) and Maddala and Wu (1999). These are denoted by IPS and MW, respectively. In both tests, the null hypothesis is that of a unit root. The IPS statistic is based on averaging individual Dickey-Fuller unit root tests (ti) according to pffiffiffiffi N ðt̄ E½ti j qi ¼ 0Þ tIPS ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ! N ð0; 1Þ ð3Þ var½ti j qi ¼ 0 P where t̄ ¼ N 1 Ni¼1 ti . The moments of E[tijqi = 0] and var[tijqi = 0] are obtained by Monte Carlo simulation and are tabulated in PIPS. The MW statistic is given by P ¼ 2 Ni¼1 ln pi , and combines the p-values from individual ADF tests. The P test is distributed as v2 with degrees of freedom twice the number of cross section units, i.e. 2N, under the null hypothesis. Breitung (1999) finds that IPS suffers a dramatic loss of power when individual trends are included, and the test is sensitive to the specification of deterministic trends. The MW test has the advantage over the IPS that its value does not depend on different lag lengths in the individual ADF regressions. In addition Maddala and Wu (1999) and Maddala et al. (1999) found that the MW test is superior compared to the IPS test. 3.2. Testing for cointegration The next step is to test for the existence of a long run relationship among y, F and the control variables S and ṗ. A common practice to test for cointegration is Johansen’s procedure. However, the power of the Johansen test in multivariate systems with small sample sizes can be severely distorted. To this end, we need to combine information from time series as well as cross-section data once again. In this context three panel cointegration tests are conducted. D.K. Christopoulos, E.G. Tsionas / Journal of Development Economics 73 (2004) 55–74 63 First, we use a test due to Levin and Lin (1993) in the context of panel unit roots, to estimated residuals from (supposedly) long run relations. Levin and Lin (1993) consider the model yit ¼ qi yi;t1 þ zitVc þ uit ð4Þ where zit are deterministic variables, uit is iid(0,r2) and qi = q. The test statistic is a tstatistic on q given by sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi N X T X 2 ðq̂ 1Þ ỹi;t1 i¼1 t¼1 tq ¼ ð5Þ se where ỹit ¼ yit T X ũit ¼ uit hðt; sÞyis ; s¼1 s2e ¼ ðNT Þ1 T X hðt; sÞuis hðt; sÞ ¼ ztV s¼1 N X T X T X ! zt ztV zs ; t¼1 ũ2it ; i¼1 t¼1 and q̂ is the OLS estimate of q. It can be shown that if there are only fixed effects in the model, then pffiffiffiffi pffiffiffiffi 51 N T ðq̂ 1Þ þ 3 N ! N 0; ð6Þ 5 and if there are fixed effects and a time trend, pffiffiffiffi 2895 N ðT ðq̂ 1Þ þ 7:5Þ ! N 0; 112 ð7Þ Second, we use the unit root tests developed for Eq. (4) by Harris and Tzavalis (1999). If there are only fixed effects in the model, then ! pffiffiffiffi 3 3ð17T 2 20T þ 17Þ N q̂ 1 þ ! N 0; ð8Þ T þ1 5ðT 1ÞðT þ 1Þ3 If there are fixed effects and a time trend, then pffiffiffiffi N q̂ 1 þ 15 2ðT þ 2Þ ! N 0; 15ð193T 2 728T þ 1147Þ 112ðT þ 2Þ3 ðT 2Þ ! ð9Þ It must be noted that Levin and Lin (1993) tests may have substantially size distortion if there is cross-sectional dependence (O’Connell, 1998). Also, Harris and Tzavalis (1999) find that small T yields Levin and Lin tests which are substantially undersized and have 64 D.K. Christopoulos, E.G. Tsionas / Journal of Development Economics 73 (2004) 55–74 low power. A drawback of the Levin and Lin or Harris and Tzavalis tests is that they do not allow for heterogeneity in the autoregressive coefficient, q. Finally, to overcome the problem of heterogeneity that arises in both tests we use Fisher’s test to aggregate the p-values of individual Johansen maximum likelihood cointegration test statistics, see Maddala and Kim (1998, p. 137). If pP i denotes the p-value of the Johansen statistic for the ith unit, then we have the result 2 Ni¼1 log pi fv22N . The test is easy to compute and, more importantly, it does not assume homogeneity of coefficients in different countries. 3.3. Testing for unit roots in threshold autoregressive models We use tests for unit roots from threshold autoregressive (TAR) models, following Caner and Hansen (2001) on which this section largely draws. The model is given by a TAR(k) of the form Dyt ¼ h1Vxt1 1ðzt1 VkÞ þ h2Vxt1 1ðzt1 zkÞ þ ut ð10Þ where yt is the series we consider, k is the threshold parameter, xt 1=( yt 1,rtV 1 , Dyt 1,. . .,Dyt k)V, zt u yt yt m for some m z 1 (m = 1 in this application), and rt 1 is a vector of exogenous variables, a constant in our case. The procedure can be used to test simultaneously for stationarity as well as threshold effects. First, model (10) P is estimated by OLS for a fixed ka[k,k̄], and the residual variance s2 ðkÞ ¼ T 1 Tt¼1 ûðkÞ2 is computed. The threshold parameter is estimated by k̂ ¼ agrmin: s2 ðkÞk. For the estimate k̂, the residuals ût and the residual variance s2 are computed. To test for a threshold effect, we need to test the hypothesis H0: h1 = h2 which can be tested using a Wald test W = T(s02/ s2 1) where s02 denotes the residual variance under the null (i.e. the residual variance computed using OLS in the linear model under the null hypothesis). However, the distribution of the Wald test is non-standard. Caner and Hansen (2001) provide the correct asymptotic distribution but point out that for small sample inference, model-based bootstrap approach should be used. We have used bootstrapping to compute exact pvalues of the test, using 10,000 replications. It is also possible to compute Wald tests for threshold effects in specific coefficients. To test for unit roots, we use the one-sided formulation of Caner and Hansen (2001), namely H0: q1 = q2 = 0 versus the alternative H1: q1 < 0 or q2 < 0 where qi denotes the first element of hi. The test statistic is a two sided Wald test of the form R1T = t12 + t22 where ti signifies the t-ratios for q̂i from OLS regression in the TAR model. Exact p-values for this test can be computed using the bootstrap approach. Since exact p-values are available, a panel data version of the Caner and Hansen (2001) P tests can be constructed by considering an MW formulation of the form 2 Ni¼1 log pi fv22N, where pi denotes the p-value of a given Wald test for the ith country. 3.4. Estimating the long run relationship Having established that the dependent variable is structurally related to the explanatory variables, and thus a long run equilibrium relationship exists among these variables, we D.K. Christopoulos, E.G. Tsionas / Journal of Development Economics 73 (2004) 55–74 65 proceed to estimate the Eq. (1) by the method of fully modified OLS appropriate for heterogeneous cointegrated panels (Pedroni, 2000). This methodology addresses the problem of non-stationary regressors, as well as the problem of simultaneity bias raised by Levine et al. (2000): OLS estimation yields biased results because, in general, the regressors are endogenously determined in the I(1) case. We consider the following cointegrated system for panel data yit ¼ ai þ xitVb þ uit ð11Þ xit ¼ xi;t1 þ eit ð12Þ where nit=[uit,eitV] is stationary with covariance matrix Xi. Following Phillips and Hansen (1990) a semi-parametric correction can be made to the OLS estimator that eliminates the second order bias caused by the fact that the regressors are endogenous. Pedroni (2000) follows the same principle in the panel data context, and allows for the heterogeneity in the short run dynamics and the fixed effects. Pedroni’s estimator is !1 ! N T N T X X X X 2 2 1 1 b̂FM b ¼ ðxit x̄t Þ ðxit x̄t Þu*it T ĉi X̂22i X̂11i X̂22i i¼1 t¼1 i¼1 t¼1 ð13Þ ûit * ¼ uit X̂1 22i X̂21i ; 0 ĉi ¼ Ĉ21i þ X̂021i X̂1 22i X̂21i ðĈ22i þ X̂22i Þ ð14Þ where the covariance matrix can be decomposed as Xi = Xi0 + Ci + Ci where Xi0 is the contemporaneous covariance matrix, and Ci is a weighted sum of autocovariances. Also, X̂i0 denotes an appropriate estimator of Xi0. 4. Empirical results Time series ADF tests are reported in Table 1 for all 10 countries. All time series involved contain unit roots according to the ADF test, save for output in Ecuador and Jamaica, finance depth in Thailand, the share of investment in Colombia, Paraguay, Peru, and the Dominican Republic, as well as inflation in Thailand. ADF tests in first differences show that for most of these series, their first differences are stationary so the conclusion of the ADF test is not safe. Possible exceptions are output series for Columbia, Thailand and Dominican Republic finance depth for Thailand and Honduras, and investment share for Paraguay, Kenya and Thailand. However, panel unit roots tests (both IPS and MW), reported in Table 2, support the hypothesis of a unit root in all variables across countries, as well as the hypothesis of zero order integration in first differences. Country by country Johansen maximum likelihood cointegration results are reported in Table 3. The hypothesis of no cointegration is rejected for all countries, and the hypothesis of one cointegrating vector is accepted. Panel cointegrating tests are reported in Table 4. The results are fairly conclusive: Fisher’s test supports the presence of one cointegrating 66 D.K. Christopoulos, E.G. Tsionas / Journal of Development Economics 73 (2004) 55–74 Table 1 ADF unit root tests Country Output ( y) Levels Colombia 0.90 Paraguay 1.84 Peru 2.54 Mexico 2.19 Ecuador 4.94*** Honduras 2.10 Kenya 0.63 Thailand 2.57 Dominican Republic 1.29 Jamaica 3.24* Finance depth ( F) Investment share (S) Inflation (ṗ) Diff Levels Diff Levels Diff Levels Diff 1.86 3.26* 3.72** 3.15* 3.51** 3.88** 5.85*** 2.56 2.89 3.32* 0.36 1.34 2.31 2.21 1.38 1.75 1.20 3.99*** 1.33 2.12 3.31* 3.65** 3.72** 5.32*** 5.21*** 2.19 3.78** 2.97 5.83*** 3.78** 3.35* 3.69** 3.22* 2.47 2.08 1.14 2.12 2.01 3.39* 2.53 3.30* 2.84 5.20*** 3.99** 3.83** 3.99* 2.64 2.63 4.71*** 2.96 2.98 1.93 2.63 1.49 2.28 2.59 2.13 4.04** 1.71 3.05 3.29* 4.95*** 4.07** 3.68** 3.28* 4.17** 3.41** 3.98** 3.46** 4.32*** Levels and Diff denote the augmented Dickey-Fuller t-tests for a unit root in levels and first differences respectively. Number of lags was selected using the AIC criterion. Boldface values denote sampling evidence in favour of unit roots. (***), (**) and (***) signify rejection of the unit root hypothesis at the 1%, 5% and 10% levels, respectively. vector. The Harris and Tzavalis tests support the hypothesis of a cointegrating relation. Finally, the Levin and Lin test with fixed effects supports the hypothesis of a cointegrating relation but this is not the case when both fixed and time effects are included. When financial depth is used as dependent variable, all tests are in agreement that cointegration does not exist. Therefore, all in all both time series and panel-based tests agree that there is a single cointegrating vector, and long run causality is unidirectional from financial depth to growth. Next, we consider Hansen’s TAR(1) model. Wald tests for the hypothesis of no threshold effects are reported in the second column of Table 5. The hypothesis can be rejected only for Peru and Thailand but judging from p-values (0.043 and 0.054) it cannot be rejected at 4% or lower. So we do not have strong evidence in favor of threshold effects. Wald tests for threshold effects in individual coefficients (namely c, b, and c) are reported in columns 3 – 5. The evidence is again rather weak: At the 4% level we cannot reject that these coefficients are the same in the two regimes defined by the threshold, as expected from the results of the threshold effect Wald statistic in the second column. We have only three marginal exceptions which, however, are significant at the 4% level. Finally, t-tests Table 2 Panel unit root tests Variables Output ( y) Finance depth ( F) Investment share (S) Inflation (ṗ) Levels First differences IPS MW IPS MW 0.18 2.71 0.04 0.47 27.12 14.77 30.37 26.37 4.52*** 6.63*** 5.81*** 5.19*** 58.33*** 83.64*** 62.98*** 74.29*** IPS and MW are the Im, Pesaran and Shin t-test and Maddala and Wu v2 test for a unit root in the model. The critical values for MW test are 37.57 and 31.41 at 1% and 5% statistical levels, respectively. Boldface values denote sampling evidence in favour of unit roots. ***Signifies rejection of the unit root hypothesis at the 1% level. D.K. Christopoulos, E.G. Tsionas / Journal of Development Economics 73 (2004) 55–74 67 Table 3 Johansen Cointegration tests Country Max eigenvalue statistic H0: rank = r Colombia Paraguay Peru Mexico Ecuador Honduras Kenya Thailand Dominican Republic Jamaica r = 0 (62.99) r V 1 (42.44) r V 2 (25.32) r V 3 (12.25) 65.58a 115.12a 106.88a 117.87a 125.14a 141.06a 98.83a 66.27a 141.83a 97.14a 25.83 22.44 33.83 26.64 25.39 38.27 29.02 36.99 40.86 36.39 8.02 11.86 13.74 13.52 9.16 14.21 14.47 18.61 17.32 14.25 0.25 5.47 4.83 5.84 3.45 5.46 6.11 5.65 6.50 6.51 r denotes the number of cointegrating vectors. The optimal lag lengths for the VARs were selected by minimising the Schwarz criterion. Numbers in parentheses next to r = 0, r V 1, r V 2 and r V 3 represent the 5% critical values of the test statistic. An (a) indicates rejection of the null hypothesis of no-cointegration at 5% level of significance. for stationarity are reported in the two last columns. Stationarity of the error correction term (taking account of possible thresholds) is the important task of this analysis. Given that we could not reject absence of threshold effects, it would be reasonable to expect that our previous analysis would apply, and error correction terms would be stationary. Indeed, only for Mexico the t1 statistic has a p-value equal to 0.003 (the test value is 5.04) but the t2 statistic clearly indicates that we must accept stationarity in the other regime. Given that the two regimes do not have a large number of observations, this finding could be an artifact. In the case of Colombia, t1 is 4.02 with p-value equal to 0.03 so we can reject stationarity at 5% but not at 3% so this is, again, very weak evidence. MW panel unit root tests provide additional evidence in favor of cointegration: The mw test rejects cointegration in the case of t1 statistic because of the small p-value for Mexico (0.003). Given that the Wald test indicates absence of threshold effects, and the vast majority of t1 and t2 Table 4 Panel cointegration tests Levin-Lin F.E. Fisher v2 cointegration test Harris-Tzavalis F.E.T. F.E. F.E.T. Dependent variable: output (y) 8.36*** 0.89 77.13*** 5.57*** Dependent variable: financial depth (F) 1.20 0.50 0.85 1.65 r=0 76.09 rV1 30.73 rV2 28.91 rV3 23.26 F.E. denotes the Levin and Lin and Harris and Tzavalis t-tests with fixed effects but no time effects in the fitted equation while F.E.T. includes both fixed and time effects in the fitted regression. ***Signifies rejection of the null hypothesis of no-cointegration at 2% significance level. Boldface values denote sampling evidence in favour of unit root. The critical values for Fisher’s v2 test are 37.57 and 31.41 at 1% and 5% statistical level, respectively. Fisher’s test is computed based on p-values from Johansen’s maximum likelihood cointegration methodology. Therefore, it applies regardless of the dependent variable. 68 D.K. Christopoulos, E.G. Tsionas / Journal of Development Economics 73 (2004) 55–74 Table 5 Testing for threshold unit roots in the error correction term (Ut) Country Wald test for Wald test for threshold effect in threshold effect Intercept Ut 1 DUt 1 t-test for stationarity Colombia Paraguay Peru Mexico Ecuador Honduras Kenya Thailand Domin. Republic Jamaica MW panel test 7.12 16.0 19.7 9.22 8.29 3.97 8.18 16.0 8.64 4.02 2.82 0.52 5.04 1.83 3.30 2.38 3.25 2.83 (0.504) (0.074) (0.043) (0.302) (0.408) (0.901) (0.395) (0.054) (0.343) 7.63 (0.457) 28.65 t1 0.078 (0.950) 5.50 (0.140) 0.018 (0.940) 10.1 (0.075) 0.898 (0.598) 10.1 (0.042) 1.78 (0.743) 7.61 (0.084) 1.21 (0.544) 2.26 (0.655) 5.45 (0.128) 1.81 (0.429) 4.54 (0.376) 0.435 (0.723) 1.77 (0.456) 1.08 (0.799) 2.22 (0.387) 0.023 (0.931) 6.77 (0.158) 2.95 (0.314) 1.04 (0.558) 6.17 (0.215) 0.470 (0.708) 0.009 (0.956) 0.015 (0.982) 6.84 (0.069) 0.335 (0.740) 3.13 (0.553) 17.4 t2 (0.030) (0.131) (0.844) (0.003) (0.305) (0.052) (0.168) (0.052) (0.093) 0.965 (0.582) 3.00 (0.301) 1.58 (0.39) 26.01 15.34 47.88 0.236 3.80 2.85 0.323 1.26 0.456 2.45 1.90 1.55 (0.755) (0.030) (0.111) (0.758) (0.503) (0.692) (0.152) (0.340) (0.974) 0.997 (0.539) 21.85 Bootstrap p-values are reported in parentheses. For bootstrapping 10,000 replications have been used. Boldface figures indicate rejection of the relevant null hypothesis at the 5% level or higher. MW is Maddala and Wu v2 test for a unit root in the model. The critical value for MW test is 37.57 at the 1% statistical level. A boldface figure indicates non-rejection of a unit root. statistics favour cointegration. We conclude in favour of panel cointegration. Therefore, we can conclude that stationarity of error correction terms seems a reasonable hypothesis, and we can claim that our estimated relations are indeed cointegrating, long run equilibrium relations. The next step is estimation of such relationships. Fully modified OLS estimates of the cointegrating relationship are reported in Table 6 on a per country basis as well as for the panel as a whole. For the panel, the coefficient of financial depth is 14.13 with t-statistic of 2.57 so it is statistically significant, and the effect is positive. The share of investment has a positive effect (0.20), and inflation has a negative impact on growth ( 0.04). However, inflation does not seem to be statistically significant for growth but investment share is statistically significant at the 10% level, and marginally so at the 5% level. On a per country basis, financial depth has a positive impact Table 6 Fully modified OLS estimates (dependent variable is output, y) Country Finance depth ( F) Investment share (S) Inflation (ṗ) Colombia Paraguay Peru Mexico Ecuador Honduras Kenya Thailand Dominican Republic Jamaica Panel 3.21*** [3.00] 51.50*** [4.33] 40.32*** [3.14] 3.08 [1.62] 18.55 [1.50] 30.40*** [3.76] 36.55*** [3.72] 83.11* [1.68] 25.40*** [3.28] 39.17*** [3.83] 14.13*** [2.57] 0.01 0.56* 0.74 0.82*** 0.67*** 0.28 3.13*** 3.05*** 0.02 0.56 0.20* 0.01 0.02 0.01*** 0.03 0.004 0.02 0.07 0.008 0.03 0.07 0.04 [0.07] [1.75] [0.87] [2.70] [2.56] [1.15] [4.21] [2.96] [0.49] [0.56] [1.91] [1.19] [0.36] [2.87] [0.53] [0.01] [1.09] [1.07] [0.58] [0.50] [0.25] [1.28] Figures in brackets are t-statistics. (***) and (*) indicate statistical significance at the 1% and 10% level, respectively. D.K. Christopoulos, E.G. Tsionas / Journal of Development Economics 73 (2004) 55–74 69 on output but the relation does not seem to be statistically significant in Mexico and Ecuador. The t-statistics are 1.62 and 1.50, respectively, so statistical significance is marginal. Investment’s share and inflation are only exceptionally statistically significant. Under the assumption of parameter homogeneity, that we may have to accept in view of the short span, the panel results should be more reliable. These results leave little ground for assuming that financial depth has no impact on output. On the contrary, we get a clear positive relation for the panel as a whole, as well as on a per country basis. Our results are in line with King and Levine (1993a,b), Levine et al. (2000), Beck et al. (2000), Levine (1999) and Khan and Senhadji (2000) who find positive effects of financial depth on growth. The results contradict the time series evidence in Demetriades and Hussein (1996), as well as Luintel and Khan (1999) who find bi-directional causality. Taking account of the fact that panel unit root test and cointegration tests utilize the data in a more efficient way, the panel results (as well as the majority of time series based tests) provide clear evidence that there is a fairly strong long run relationship between financial depth and output, that a long run causal relationship running from output to financial depth in unlikely (since the equation with financial depth (F) as the dependent variable does not show cointegration) and, therefore, the causal relationship runs from financial depth to output. Another important issue is whether causality between output and financial deepening is short run as well. To investigate this issue, we have specified error correction models (ECM) of the form Dyt ¼ c þ m X bi DFti þ i¼1 m X Dxt1 V g i þ kðyt1 xt1 V d d0 Ft1 Þ þ vt ð15Þ i¼1 where yt 1 xtV 1 d d0Ft 1 represents the equilibrium error, that is the deviation from the long run relationship. The first important issue we consider is whether k p 0. If this is not the case, the cointegration finding would not be reliable. The second important issue is whether H0: bi = 0 (all i = 1,. . .,m) can be rejected. If it can be rejected, there is no evidence of short run causality. The v2 tests of short run causality as well as diagnostic statistics (normality, autocorrelation and functional form misspecification) for the VEC model are depicted in Tables 7 and 8. According to these results, the VEC model seems to be data congruent and free from specification error for all countries we examine. The hypothesis of short run causality can be rejected for all countries with the exception of the Dominican Republic. However, the p-value of the test in this case is 0.03 suggesting that the hypothesis can be rejected at the 5% level but not at 1%. Therefore, not even in this case we have a definite result. Moreover, estimates of the speed of adjustment, k, have pvalues consistent with statistical significance, which leaves little doubt that the estimated long run relationships are indeed structural. The same conclusions are supported for the panel as a whole based on the Fisher test that aggregates the individual p-values. Additionally, we estimate a VEC model allowing for panel data. The formulation is as follows. Dyit ¼ ci þ m X l¼1 bl Dyi;tl þ m X Dxi;tl V g l þ kðyi;t1 xi;t1 V d d0 Fi;t1 Þ þ vit l¼1 ð16Þ 70 D.K. Christopoulos, E.G. Tsionas / Journal of Development Economics 73 (2004) 55–74 Table 7 Short run causality tests between Output ( y) and Finance Depth ( F): Error Correction Models (ECM) Country Lags of financial deepening, v2 p-values of speed of adjustment, k Colombia Paraguay Peru Mexico Ecuador Honduras Kenya Thailand Dominican Republic Jamaica Panel Fisher test 2.29 [0.32] 2.04 [0.36] 4.76 [0.09] 0.54 [0.76] 0.47 [0.78] 4.09 [0.12] 1.13 [0.57] 2.82 [0.24] 7.08 [0.03] 1.20 [0.74] 22.60 [0.07] [0.02] [0.02] [0.06] [0.06] [0.01] [0.05] [0.03] [0.02] [0.03] 67.07 Figures in brackets represent asymptotic p-values associated with the tests. Fisher’s test is computed based on p-values from individual tests. The critical value for Fisher test is 37.57 at the 1% statistical level. Boldface values indicate statistical significance at the level 7% or higher. where ci represents fixed country effects. This model can be estimated using instrumental variables. Since this is a dynamic panel data model, it is well known that standard estimation techniques like LSDV yield biased and inconsistent estimators in the panel data case. For this reason, we must use an instrumental variables estimator to deal with the correlation between the error term and lagged dependent variables Dyi,t 1. We have found that m = 2 is necessary to satisfy the classical assumptions on the error term, so we use Dyi,t 3 and Dyi,t 4 as instruments for the lagged dependent variables. Estimates as well as diagnostic statistics for the VEC model are presented in Table 9. Again the VEC model seems to be data congruent and free from specification error for all Table 8 Diagnostic tests for the Vector Error Correction (VEC) model Countries Jarque-Bera Test (JB) Lagrange Multiplier Test (LM2) Ramsey Specification Test Colombia Paraguay Peru Mexico Ecuador Honduras Kenya Thailand Dominican Republic Jamaica Panel Fisher test 0.47 [0.79] 1.06 [0.59] 0.63 [0.73] 2.29 [0.12] 5.82 [0.06] 0.46 [0.78] 0.63 [0.73] 0.84 [0.35] 1.85 [0.39] 0.33 [0.84] 17.48 1.82 [0.39] 1.95 [0.37] 5.56 [0.06] 1.35 [0.51] 2.17 [0.34] 1.41 [0.49] 4.91 [0.08] 2.58 [0.27] 4.26 [0.09] 1.12 [0.52] 28.10 2.93 [0.08] 1.36 [0.50] 0.01 [0.99] 2.19 [0.33] 4.59 [0.10] 0.36 [0.83] 3.60 [0.13] 1.76 [0.47] 2.93 [0.23] 2.63 [0.26] 27.19 Figures in brackets represent asymptotic p-values associated with the tests. Jarque-Bera (JB) denotes the JarqueBera normality Test of errors. Lagrange Multiplier Test (LM) tests the null hypothesis that there is no second order autocorrelation. The Ramsey Test tests the null hypothesis that there is no functional form misspecification. The critical value for Fisher test is 37.57 at the 1% statistical level. Boldface values indicate statistical significance at the level 7% or higher. D.K. Christopoulos, E.G. Tsionas / Journal of Development Economics 73 (2004) 55–74 71 Table 9 Panel vector error correction model Variable Estimate Dyt 1 Dyt 2 DFt 1 DFt 2 DSt 1 DSt 2 Dṗt 1 Dṗt 2 Error Correction Term (ECTt 1) Jarque -Bera Test (JB) Likelihood Ratio Test LR(2) Ramsey Test 0.17 0.09 2.05 3.60 1.29 2.26 0.002 0.0003 0.32 0.30 1.45 4.06 [0.003] [0.08] [0.99] [0.84] [0.63] [0.56] [0.18] [0.86] [0.01] [0.87] [0.46] [0.13] Figures in brackets represent asymptotic p-values associated with the tests. Jarque-Bera (JB) denotes the JarqueBera normality Test of errors. The Likelihood ratio Test (LR) tests for the null hypothesis that there is no second order autocorrelation. The Ramsey Test tests the null hypothesis that there is no functional form misspecification. Boldface figures denote statistical significance at the 1% level. Fixed effect estimates are not reported but are available from the authors upon request. countries we examine. The v2-test for the hypothesis that lags of financial development do not contribute to output is not rejected, therefore there is no evidence of short run causality. The most important implication of our findings is a policy recommendation: If policy makers want to promote growth, then attention should be focused on long run policies, for example the creation of modern financial institutions, in the banking sector and stock markets. From that point of view, our findings conform to earlier findings of empirical studies that report routinely statistically significant coefficients of financial proxy variables on output growth, for example Gelb (1989), Ghani (1992), King and Levine (1993a,b), Levine and Zervos (1996) and Beck et al. (2000). These findings, as well as the findings in the present study, stand against empirical evidence in Ireland (1994) and Demetriades and Hussein (1996) that are consistent with the view that financial deepening is an outcome of the growth process. Not only the evidence on cointegration and the statistical significance of financial development is quite strong but the same pattern is confirmed by panel-based tests as well. This is particularly important because although time series tests allow the possibility to examine causality contrary to cross-country regressions, their power could be low given typical small sample sizes. Panel cointegration tests combine the ability of time series studies to yield causality inferences with the increase in sample size afforded by using cross-sectional data. One notable implication of our findings is that results are not dramatically countryspecific (as in Demetriades and Hussein, 1996 for example). This offers a justification for using panel-based unit root and cointegration tests. Another important implication of the absence of short run causality, and the strong nature of long run causality, is the one emphasized by Darrat (1999), namely that since the effect of financial development on growth is realized in the short run, policy makers may be deceived to believe that there is no effect at all. The long run nature of the effect, however, is a necessary implication of the fact that financial markets affect the cost of external finance to the firm and, therefore, their 72 D.K. Christopoulos, E.G. Tsionas / Journal of Development Economics 73 (2004) 55–74 effect materializes through facilitating the investment process itself. Unless conditions for low-cost investment are created, long run growth is impossible. 5. Concluding remarks In this paper we have combined cross-sectional and time series data to examine the relationship between financial development and growth in ten developing countries. Previous studies have used either cross-sectional or time series data but both approaches have drawbacks. Using cross-sectional data leaves open the question of spurious correlation arising from non-stationarity, and does not permit an examination of the direction of causality. Using time series data, may yield unreliable results due to short time spans of typical data sets. We have made use of panel unit root tests, and panel cointegration analysis to conclude that there is fairly strong evidence in favor of the hypothesis that long run causality runs from financial development to growth, that the relationship is significant, and that there is no evidence of bi-directional causality. We have used fully modified OLS to estimate the cointegrating relation, a method that deals with the problem of endogeneity of regressors. Time series evidence is also supportive to the idea that there exists a unique cointegrating vector between growth, financial development and ancillary variables (investment share and inflation). The empirical evidence also points to the direction that there is no short run causality between financial deepening and output, so the effect is necessarily long run in nature. The important policy implication is that policies aiming at improving financial markets will have a delayed effect on growth, but this effect is significant. Acknowledgements We wish to thank an anonymous referee for useful comments in a previous version of this paper. Thanks also go to Peter Pedroni for providing his code to implement fully modified estimation. References Arestis, P., Demetriades, P., 1997. Financial development and economic growth: assessing the evidence. Economic Journal 107, 783 – 799. Atje, R., Jovanovic, B., 1993. Stock market and development. European Economic Review 37, 623 – 640. Beck, T., Levine, R., Loyaza, N., 2000. Finance and the sources of growth. Journal of Financial Economics 58, 261 – 300. Bencivenga, V., Smith, B., Starr, R., 1995. Transaction costs, technological choice and endogenous growth. Journal of Economic Theory 67, 153 – 177. Benhabib, J., Spiegel, M.M., 2000. The role of financial development in growth and investment. Journal of Economic Growth 5, 341 – 360. Berthelemy, J., Varoudakis, A., 1996. Economic growth, convergence clubs, and the role of financial development. Oxford Economic Papers 48, 300 – 328. Breitung, L., 1999. The Local Power of Some Unit Root Tests for Panel Data, Discussion Paper. Humboldt University, Berlin. D.K. Christopoulos, E.G. Tsionas / Journal of Development Economics 73 (2004) 55–74 73 Caner, M., Hansen, B.E., 2001. Threshold Autoregression with a Unit Root. Econometrica 69, 1565 – 1596. Chandavarkar, A., 1992. Of finance and development: neglected and unsettled questions. World Development 22, 133 – 142. Darrat, A., 1999. Are financial deepening and economic growth causally related? Another look at the evidence. International Economic Journal 13, 19 – 35. Deidda, L., Fatouh, B., 2002. Non-linearity between finance and growth. Economics Letters 74, 339 – 345. De Mello, J., 1999. Foreign direct investment-led growth: evidence from time series and panel data. Oxford Economic Papers 51, 577 – 594. Demetriades, P.O., Hussein, A.K., 1996. Does financial development cause economic growth? Time series evidence from 16 countries. Journal of Development Economics 51, 387 – 411. Demirgücß -Kunt, A., Maksimovic, V., 1998. Law, finance, and firm growth. The Journal of Finance LIII, 2107 – 2137. Enders, W., Granger, C.W.J., 1998. Unit root tests and asymmetric adjustment with an example using the term structure of interest rates. Journal of Business of Economic and Statistics 16, 304 – 312. Gelb, A.H., 1989. Financial policies, growth and efficiency. World Bank Working Paper, vol. 202. June. Ghani, E., 1992. How financial markets affect long run growth: a cross-country study. World Bank Working Paper, vol. PPR 843. January. Greenwood, J., Jovanovic, B., 1990. Financial development and economic development. Economic Development and Cultural Change 15, 257 – 268. Greenwood, J., Smith, B., 1997. Financial markets in development and the development of financial market. Journal of Economic Dynamic and Control 21, 181 – 1456. Harris, R.D.F., Tzavalis, E., 1999. Inference for unit roots in dynamic panels where the time dimension is fixed. Journal of Econometrics 91, 201 – 226. Im, S.K., Pesaran, H.M., Shin, Y., 1997. Testing for Unit Roots in Heterogeneous Panel. Department of Applied Econometrics, University of Cambridge. Ireland, P.N., 1994. Money and growth: an alternative approach. American Economic, 47 – 65. Johansen, S., 1988. Statistical analysis of cointegration vectors. Journal of Economics Dynamic and Control 12, 231 – 254. Khan, S.M., Senhadji, A.S., 2000. Financial Development and Economic Growth: An Overview. IMF Working Paper. International Monetary Fund, Washington. King, R.G., Levine, R., 1993a. Finance and growth: Schumpeter might be right. Quarterly Journal of Economics 108, 717 – 737. King, R.G., Levine, R., 1993b. Finance, entrepreneurship and growth. Journal of Monetary Economics 32, 30 – 71. Kiviet, J.F., 1995. On bias, inconsistency and efficiency of various estimators in dynamic panel models. Journal of Econometrics 68, 53 – 87. Levin, A., Lin, C.F., 1993. Unit root tests in panel data: asymptotic and finite sample properties. Mimeo (September). Levine, R., 1991. Stock markets, growth, and tax policy. Journal of Finance XLVI, 1445 – 1465. Levine, R., 1997. Financial development and economic growth: views and agenda. Journal of Economic Literature 35, 688 – 726. Levine, R., 1998. The legal enviroment, banks, and long-run economic growth. Journal of Money, Credit, and Banking 30, 596 – 613. Levine, R., 1999. Law, finance and economic growth. Journal of Financial Intermediation 8, 8 – 35. Levine, R., Zervos, S., 1996. Stock market development and long-run growth. World Bank Economic Review 10, 323 – 339. Levine, R., Loyaza, N., Beck, T., 2000. Financial intermediation and growth: causality and causes. Journal of Monetary Economics 46, 31 – 71. Lucas Jr., R.E., 1988. On the mechanics of economic development. Journal of Monetary Economics 22, 3 – 42. Luintel, B.K., Khan, M., 1999. A quantitative re-assessment of the finance-growth nexus: evidence from a multivariate VAR. Journal of Development Economics 60, 381 – 405. Maddala, G.S., Kim, I.-M., 1998. Unit Roots, Cointegration, and Structural Change. Cambridge Univ. Press, Cambridge. 74 D.K. Christopoulos, E.G. Tsionas / Journal of Development Economics 73 (2004) 55–74 Maddala, G.S., Wu, S., 1999. A comparative study of unit root tests with panel data and a new simple test. Oxford Bulletin of Economics and Statistics 61, 631 – 652. Maddala, G.S., Wu, S., Liu, P., 1999. Do panel data rescue Purchasing Power Parity (PPP) theory? In: Krishnakumar, J., Ronchetti, E. (Eds.), Panel Data Econometrics: Future Directions. Elsevier. McKinnon, R.I., 1973. Money and Capital in Economic Development. Brooking Institution, Washington, DC. Neusser, K., Kugler, M., 1998. Manufacturing growth and financial development: evidence from OECD countries. The Review of Economics and Statistics, 638 – 646. Obstfeld, M., 1994. Risk-taking, global diversification, and growth. American Economic Review 84, 10 – 29. O’Connell, P.G.J., 1998. The overvaluation of purchasing power parity. Journal of International Economics 44, 1 – 19. Pagano, M., 1993. Financial markets and growth: an overview. European Economic Review 37, 613 – 622. Pedroni, P., 2000. Fully modified OLS for heterogeneous cointegrated panelsNon-Stationary Panels, Panel Cointegration and Dynamic Panels, vol. 15. Elsevier, pp. 93 – 130. Phillips, P.C.B., Hansen, B.E., 1990. Statistical inference in individual variables regression with I(1) process. Review of Economic Studies 57, 99 – 125. Pierse, R.G., Shell, A.J., 1995. Temporal aggregation and the power of tests for unit root. Journal of Econometrics 65, 335 – 345. Robinson, J., 1952. The Rate of Interests and Other Essays. Macmillan, London. Shaw, E.S., 1973. Financial Deepening in Economic Growth. Oxford Univ. Press, NY. Singh, A., 1997. Financial liberalization, stock markets and economic development. Economic Journal 107, 771 – 782.