POLICY RESEARCH WORKING PAPER 3 059 Do Capital Flows Respond to Risk and Return? Cesar Calder6n Norman Loayza Luis Servtn The World Bank Development Research Group Macroeconomics and Growth May 2003 I POLICY RESEARCH WORKING PAPER 3059 Abstract This paper explores empirically the role of risk and employs a dynamic panel estimation procedure allowing return in the observed evolution of net foreign asset for unrestricted short-run heterogeneity across countries, positions of industrial and developing economies. The using the pooled mean group estimator recently paper adopts a dynamic approach in which investors' developed by Pesaran, Shin, and Smith (1999). The portfolios adjust gradually to their long-run equilibrium, empirical results lend considerable support to the defined by a standard Tobin-Markowitz framework. The model when applied to countries with low capital parameters characterizing the long-run equilibrium are controls and/or high and upper-middle income. The estimated using data on foreign assets and liabilities of a results for countries with either high capital controls or large number of industrial and developing countries low per capita income are less supportive of the stock spanning the period from 1965 to 1997. The paper equilibrium model for net foreign asset positions. This paper-a product of Macroeconomics and Growth, Development Research Group-is part of a larger effort in the group to understand international capital flows. Copies of the paper are available free from the World Bank, 1818 H Street NW, Washington, DC 20433. Please contact Tourya Tourougui, room MC3-301, telephone 202-458-7431, fax 202-522- 3518, email address ttourougui@worldbank.org. Policy Research Working Papers are also posted on the Web at http:// econ.worldbank.org. The authors may be contacted at nloayza@worldbank.org or Iserven@worldbank.org. May 2003. (45 pages) The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the view of the World Bank, its Executive Directors, or the countries they represent. Produced by Partnerships, Capacity Building, and Outreach Do CAPITAL FLOWS RESPOND TO RISK AND RETURN?* Cesar Calder6n Central Bank of Chile Norman Loayza The World Bank Luis Serven The World Bank JEL classification codes: F32, F37, GI I This work was supported by the Latin American Regional Studies Program of the World Bank and the Reseach Program of the Central Bank of Chile. We are grateful to Gian Maria Milesi-Ferreti, Phil Lane, Klaus Schmidt-Hebbel, Raimundo Soto, and Alan Stockman for valuable discussions. We also thank participants at the 1999 Latin America Meetings of the Econometric Society, the 1999 LACEA Meetings and the 2000 Winter Camp on International Finance for valuable comments. George Monokroussos and Rashmi Shankar provided excellent research assistance. a Corresponding author. Address: The World Bank, 1818 H St NW, Washington DC 20433. E-mail address: Iservenalworldbank.ory. Phone 2024737451 Do CAPITAL FLOWS RESPOND TO RISK AND. RETURN? 1. INTRODUCTION One of the major puzzles in international economics is the failure of standard portfolio models to explain the observed patterns of cross-country capital allocation. The search for solutions to this puzzle has attracted a great deal of theoretical and empirical work. 1 Most of this effort has focused on explaining the 'home-bias' effect, according to which domestic investors disproportionately favor domestic over foreign asset holdings. As the literature has amply documented, individuals do not appear to do a good job at diversifying risks across countries: they hold too little of their wealth in foreign assets, much less than predicted by conventional risk-return portfolio equilibrium models.2 Rather than attempting to explain the well-documented divergence between observed portfolio shares and those predicted by theory, this paper examines the empirical validity of a weaker theoretical prediction, namely that international asset positions should systematically respond to risk and return conditions. Thus, the aim of the paper is to check whether -- and how much -- international capital flows reflect market incentives, and if the effects of the latter are similar across the world or there are significant differences among countries and/or specific country groups. In following this positive approach, we implicitly take as given the 'home bias' of international portfolios - i.e., we allow for unobserved country-specific characteristics that may affect net foreign asset (henceforth NFA) positions and, more generally, we allow for heterogeneity across countries in the response of NFA positions to risk and return fundamentals. The paper's framework is guided by a Tobin-Markowitz model of portfolio diversification in which the share of domestic investors' wealth allocated to foreign assets depends on four factors: investment returns in the home country relative to the rest of the world, investment risk in the home country relative to the rest of the world, the degree of co-movement between investment returns at home and abroad, and the ratio of foreign-owned to domestic- owned wealth. 1 Lewis (1999) provides a comprehensive overview of this literature. 2 See, for example, French and Poterba (1991) for the case of intemational equity portfolios. Tesar and Werner (1995) show that the same puzzle arises with bonds. 2 This framework characterizes long-run portfolio equilibrium. However, costs and frictions to instantaneous portfolio reallocation - arising from sources such as investors' imperfect information, congestion effects or investment adjustment costs - may drive a wedge between short-run and long-run portfolio equilibrium. 3 Further, these frictions, and hence portfolio dynamics, may differ across countries. The paper's empirical analysis focuses on the estimation of the long-run portfolio equilibrium condition, while allowing for unrestricted cross-country heterogeneity in the short-run dynamics. The paper extends previous literature along three dimensions. First, it builds on a recent strand of the literature that adopts an international portfolio equilibrium approach to the analysis of the current account (Ventura 2002).4 Our paper shares with this literature the emphasis on risk and adjustment costs as essential ingredients for explaining the observed patterns of international asset portfolios. However, that literature has focused primarily on the impact of wealth changes on capital flows (what has been termed the portfolio growth effect). In contrast, the present paper also brings into focus the determinants of intemational investors' portfolio shares, for given levels of their wealth (the portfolio rebalancing effect). Second, the paper implements empirically the portfolio diversification model using a comprehensive data set on foreign assets and liabilities that covers a large number of developing and industrial countries and spans the years from 1965 to 1997. Importantly, the data encompass not only industrial economies, which have been the focus of previous empirical literature, but also a large number of emerging markets and developing economies. Using this information, we can assess the empirical robustness of the portfolio model across different country groups and alternative measures of risk and return. Third, the paper follows a novel econometric approach to the estimation of the long-run portfolio equilibrium condition in a heterogeneous dynamic panel setting, using the Pooled-Mean Group estimator recently developed by Pesaran, Shin, and Smith (1999). This approach combines the efficiency gains from restricting long-run parameters to be the same across 3 See Bacchetta and van Wincoop (1998) for a theoretical discussion of portfolio dynamics arising from these and other sources. 4 Along similar lines, the paper's portfolio equilibrium approach to capital flows also brings it close to a strand of the literature on 'current account sustainability' that underscores the role of international investors' portfolio choices in shaping the sustainable current account (see Mann 2002 for references). By shedding light on the factors that shape international portfolio diversification and its time path, the analysis in this paper could be readily adapted to identify current-account trajectories consistent with portfolio equilibrium. 3 countries (the units in the panel) with the flexibility and consistency gains of country-specific short-run adjustment. Further, the approach allows formal testing of the long-run pooling restrictions imposed by the model - i.e., the homogeneity across countries of the parameters describing the long-run portfolio equilibrium condition. The paper's plan is as follows. Section 2 describes the analytical framework and presents the econometric strategy for estimation of the long-run relationship implied by the model. Section 3 briefly summarizes the main features of the NFA data and the measures of investment returns and risks used in the empirical analysis. Section 4 presents the empirical results from estimation of the model for various groups of countries. The model is first implemented on the full country sample, and then separately on country groups that differ in per capita income level and restrictions to international portfolio diversification. Section 5 concludes. 2. METHODOLOGY 2.1 A portfolio-diversification approach to external asset positions Our analytical framework follows recent literature underscoring the role of investment risk and adjustment costs in the allocation of agents' wealth between domestic and foreign assets, and thus in the determination of capital flows (Ventura 2002). This literature shows that those two ingredients are needed to reconcile theoretical predictions and observed facts on the dynamics of countries' asset portfolios. Specifically, we adopt a portfolio-diversification approach according to which external asset positions are driven by portfolio equilibrium in the long run and by the dynamic forces shaping asset reallocation in the short run. Long-run equilibrium obtains when domestic and foreign investors achieve the desired allocation of their asset portfolio across countries. However, imperfections and frictions in real and financial markets may prevent the instantaneous achievement of the optimal portfolio. Short-run external equilibrium is then given by the adjustment path towards investors' long-run equilibrium portfolio. In our framework, the optimal portfolio allocation follows along the lines of the standard Markowitz-Tobin model of mean-variance investors. As is well known, the model can be derived under fairly standard assumptions from intertemporal optimization by forward- looking, risk- averse agents. Such procedure can be shown to yield an optimal saving/consumption plan 4 characterized by the permanent income hypothesis, and an optimal allocation of wealth between domestic and foreign assets characterized by mean-variance portfolio optimization. The key property of mean-variance investors is that the desired share of each asset in their total wealth depends only on the distribution of asset returns and not directly on the level of wealth.6 In our context of intemational diversification, the optimal portfolio share allocated to assets in a given country can be divided into two pieces, namely, the 'speculative' component and the 'minimum variance' component (using the terminology in Adler and Dumas 1983). An increase in mean retums in the country leaves unaffected the 'minimum variance' piece of the portfolio but raises the 'speculative' component and thus leads to an expansion of investors' portfolio share in that country. Analogously, a decrease in the variance of investment retums in the country, holding constant the 'speculative' component, raises the 'minimum variance' piece, thus producing an increase in investors' portfolio share in the country. The same effect occurs when the co- variation of country investment returns with those in the rest of the world decreases -holding constant the 'speculative' component, lower co-variation with the world economy implies that investments in the country provide a better hedge against systemic (world-wide) risk. Formally, let A represent world assets and W the wealth of world residents. Obviously, A = W. Let Ai represent the assets located in country i and Wi represent the wealth of country i's residents. The assets located in foreign countries and the wealth of foreigners are respectively represented by Ap A-Ai and Wf = W-Wi. Let aii be the share of wealth of country i's residents that they desire to allocate to country i's assets, and let afi represent the share of foreigners' wealth that they desire to allocate to country i's assets. Hence when actual aDd desired portfolio shares coincide, we have that Ai = a,i Wi + afi Wf As explained above, desired portfolio shares are assumed increasing in the anticipated return of country i's assets relative to those abroad, decreasing in their perceived riskiness relative to external assets, and decreasing in the co- movement of country i's returns with those in the rest of the world. We denote these three factors REiVf, RJi,f, and COivf, respectively. In (long- 5 The analytical derivations are standard, and thus for brevity they are not reproduced here. For the general case, the details can be found in Merton (1971). For an application similar to ours, see Kraay and Ventura (2000). 6 Of course, in the intertemporal optimization framework these results require (standard) simplifying assumptions such as log utility or homothetic preferences and lognormal returns (see Merton 1971). Even under such conditions, wealth and capital stocks may still affect indiiectly the return characteristics of available assets. 5 run) portfolio equilibrium, the desired holdings of country i's assets by domestic plus foreign residents should be equal to the country's total existing assets; that is, ail RE,, if ,R XOf )- Wi + aft(RE,,f ,RI,,S COuIf Wf = A, where the sign over each argument corresponds to the sign of the partial derivative. It is important to keep in mind that the a,, 0 and afii functions above may embody different preferences of domestic and foreign investors, including differential attitudes towards domestic and foreign assets - i.e., home-bias effects (Lewis 1999). The net foreign asset position of a country is the difference between the wealth owned by its residents and the assets located in the country. Therefore, in long-run equilibrium the net foreign asset position of country i will be given by: NFA = W1 - (a, 1 WI + af W) (2) For given portfolio shares aii and afi, equation (2) highlights the dependence of the net foreign asset position on wealth stocks, which is at the core of Kraay and Ventura's (2000) analysis of the current account. Normalizing by dividing both sides of (2) by country i's wealth, we get: NF, Wf -R=1-ai,-al -f (3) We can then express equation (3) as follows: NF4. - + + - =f(RE,f RI Cof 9f /WW Equation (4) defines the long-run equilibrium relationship resulting from optimal asset allocation across countries. Note that the ratio of net foreign assets to wealth depends on the relative wealth og domestic residents, even though portfolio shares are themselves independent of wealth. For empirical implementation we shall take a linear approximation to (4): (NFAI fI~I S WiJ =/30+P,'RE;,f +P2Rl;/lf +03COf +p4 Wf ) 6 where the stars denote long-run values, and the idiosyncratic intercept Bo ' captures country- specific factors that we do not model exp licitly.7 We view the above equations as characterizing long-run portfolio equilibrium, and hence expressions (4)-(5) describe the wealth share of net foreign assets in the long run. However, the dynamics of NFA along the adjustment path may show temporary departures from these long-run equilibrium rules, reflecting existing constraints to immediate portfolio adjustment.8 These may arise from various sources (Bacchetta and van Wincoop 1998; Ventura 2002): (i) investors' imperfect information (e.g., gradual learning about the state of the world, or about the permanence of reforms which affect asset returns but may initially suffer from imperfect credibility); (ii) congestion effects, such as increasing marginal costs to foreign investment due for example to its use of internationally immobile labor inputs; (iii) costs of adjusting the capital stock - such as investment irreversibility -- that make investment respond sluggishly to aggregate disturbances (Caballero 1998, Dixit and Pindyck 1996). While we do not model explicitly such dynamic effects here, in our empirical implementation we take them into account by employing a suitably expanded version of (5) allowing for lagged effects of risk, return and relative wealth. This is discussed in the next subsection 2.2 Econometric Estimation Empirical implementation of the model outlined in the previous section on a large cross- country time-series sample poses two main issues. First, the model defines a long-run relationship between the ratio of net foreign assets, wealth shares, and expected returns and risks. However, given the imperfections in international financial and factor markets, stock equilibrium does not hold at every point in time but is achieved gradually in the long run. Therefore, in the empirical analysis, the process of short-run adjustment must complement the long-run equilibrium model. Second, it seems reasonable to assume that countries can differ regarding the market imperfections and barriers to portfolio reallocation that govem short-term dynamics - and perhaps even in the parameters characterizing the long-mn equilibrium. Thus, we must allow for parameter heterogeneity across countries. We deal with each of these two issues in turn. For example, it could reflect the effects of home bias on long-run net foreign asset holdings. 8 Kraay and Ventura (2000) underscore the discrepancies between the short- and long-run patterns of change of NFA. Ventura (2002) stresses the need to take into consideration adjustment costs to account for these differences. 7 Single-country estimation The challenge we face is to estimate long- and short-run relationships without being able to observe the long- and short-run components of the variables involved. Over the last decade or so, a booming cointegration literature has focused on the estimation of long-run relationships among I(1) variables (Johanssen 1995, Phillips and Hansen 1990). From this literature, two common misconceptions have been derived. The first one is that iong-run relationships exist only in the context of cointegration among integrated variables. The second one is that standard methods of estimation and inference are incorrect. A recent literature, represented in Pesaran and Smith (1995), Pesaran (1997) and Pesaran and Shin (1999), has argued against both misconceptions. These authors show that simple modifications to standard methods can render consistent and efficient estimates of the parameters in a long-run relationship between both integrated and stationary variables and that inference on these parameters can be conducted using standard tests. Furthermore, these methods avoid the need for pre-testing and order-of- integration conformability given that they are valid whether the variables of interest are I(0) or I(1). The main requirements for the validity of this methodology are that, first, there exist a long-run relationship among the variables of interest and, second, the dynamic specification of the model be augmented such that the regressors are strictly exogenous and the resulting residual is serially uncorrelated.9 Pesaran and co-authors label this the "autoregressive distributed lag (ARDL) approach" to long-run modeling. Appendix B presents an illustration of the main assumptions and properties of the ARDL approach. In order to comply with the requirements for standard estimation and inference, we embed the long-run portfolio equilibrium condition (5) into an ARDL(p,q) model. In error-correction form, this can be written as follows: 4(N.j j | JI W | ++ xjARE,f, -j +B'jARIf +B3jACO,,f,j +B4,- ] i[{NA+i_ REj{f, +02R11jf,l+3iCOL/fPIf 4I4{ } 0 ] +71i, (6) 9 It is worth noting that the assumption of a unique long-run relationship underlies implicitly the various single- equation based estimators of long-run relationships commonly found in the cointegration literature. Without such assumption, these estimators would at best identify some linear combination of all the long-run relationships present in the data. 8 where T is the speed of adjustment, i7, is a time-varying disturbance and the term in square brackets in the second line contains the long-run equilibrium condition (5). As just discussed, it is critical that the order of the ARDL process be appropriate. Pesaran and Shin (1999) recommend a two-step procedure, whereby the lag order of the ARDL is first selected using a consistent information criterion, and then the corresponding error-correction model is estimated and tested by standard methods. As explained later, we use the Schwartz-Bayesian Criterion (SBC) to select appropriate values ofp and q in equation (6) on a country-by-country basis. Multi-country estimation Our empirical samples below are characterized by time-series (T) and cross-section (N) dimensions of roughly similar magnitude. In such conditions, there are a number of alternative methods for multi-country estimation, which allow for different degrees of parameter heterogeneity across countries. At one extreme, the fully heterogeneous-coefficient model imposes no cross-country parameter restrictions and can be estimated on a country-by-country basis -- provided the time-series dimension of the data is sufficiently large. When, in addition, the cross-country dimension is also large, the mean of long- and short-run coefficients across countries can be estimated consistently by the unweighted average of the individual country coefficients. This is the mean group (MG) estimator introduced by Pesaran, Smith, and Im (1996). At the other extreme, the fully homogeneous-coefficient model requires that all slope and intercept coefficients be equal across countries. This is the simple pooled estimator. In between the two extremes, there are a variety of estimators. The dynamic fixed effects estimator restricts all slope coefficients to be equal across countries but allows for different country intercepts. The pooled mean group (PMG) estimator, introduced by Pesaran, Shin and Smith (1999), restricts the long-run coefficients to be the same across countries but allows the short-run coefficients (including the speed of adjustment) to be country specific. The PMG estimator also generates consistent estimates of the mean of short-run coefficients across countries by taking the unweighted average of the individual country coefficients (provided that the cross-sectional dimension is large). The choice among these estimators faces a general trade-off between consistency and efficiency. Estimators that impose cross-country constraints dominate the heterogeneous estimators in terms of efficiency if the restrictions are valid. If they are fa lse, however, the 9 restricted estimators are inconsistent. In particular, imposing invalid parameter homogeneity in dynamic models typically leads to downward-biased estimates of the speed of adjustment (Robertson and Symons 1992, Pesaran and Smith 1995). For our purposes, the pooled mean group estimator offers the best available compromise in the choice between consistency and efficiency. This estimator is particularly useful when, as in our case, the long run is given by country- independent equilibrium conditions while the short- run adjustment depends on country characteristics -- such as, e.g., financial development and/or relative price flexibility. Furthermore, the PMG estimator is sufficiently flexible to allow for long-run coefficient homogeneity over only a subset of variables and/or countries. Therefore, we use the PMG method to estimate a long-run relationship that is common across countries (i.,e, Pk' = p for all ij and k=1,...,4) while allowing for unrestricted country heterogeneity in the adjustment dynamics. The interested reader is referred to Pesaran, Shin and Smith (1999) where the PMG estimator is developed and compared with the MG estimator. Briefly, the PMG estimator proceeds as follows. The estimation of the long-run coefficients is done jointly across countries through a (concentrated) maximum likelihood procedure. Then the estimation of short-run coefficients (including the speed of adjustment l9), country-specific intercepts P3o', and country-specific error variances is done on a country-by-country basis, also through maximum likelihood and using the estimates of the long-run coefficients previously obtained. 0 An important assumption for the consistency of our PMG estimates is the independence of the regression residuals across countries. In practice, non- zero error covariances usually arise from omitted common factors that influence the countries' ARDL processes. We seek to eliminate these common factors and, thus, ensure the independence condition through two means. 10 The comparison of the asymptotic properties of PMG and MG estimates can be put also in terms of the general trade-off between consistency and efficiency noted in the text. If the long-run coefficients are in fact equal across countries, then the PMG estimates will be consistent and efficient, whereas the MG estimates will only be consistent. If, on the other hand, the long-run coefficients are not equal across countries, then the PMG estimates will be inconsistent, whereas the MG estimator will still provide a consistent estimate ofthe mean of long-run coefficients across countries. The long-run homogeneity restrictions can be tested using Hausman or likelihood ratio tests to compare the PMG and MG estimates of the long run coefficients. In turn, comparison of the small sample properties of these estimators relies on their sensitivity to outliers. In small samples (low T and N), the MG estimator, being an unweighted average, is very sensitive to outlying country estimates (for instance those obtained with small T). The PMG estimatorperforms better in this regard because it produces estimates that are similar to weighted averages of the respective country-specific estimates, where the weights are given according to their precision (that is, the inverse of their corresponding variance-covariance matrix). 10 First, as explained below, we construct the indices for return and risk in a way such that each observation represents the value for a country/year relative to the corresponding mean for the whole world in all time periods. Second, we allow for time-specific effects in the estimated regression; this is equivalent to a regression in which each variable enters as deviations with respect to the cross-sectional mean in a particular year. 3. DATA 3.1 NFA and Wealth The cornerstone of our data is a set of wealth, foreign asset and foreign liability stocks for a large group of industrial and developing countries spanning the period from the 1 960s to the present. Construction of this data set is documented in Kraay et al. (2000), so for the sake of brevity here we limit our remarks to a few key issues. The total wealth of country i 's residents at time t is defined as Wi, = NFA,, + Ki, + Gil 7 where NFA denotes the country's net foreign assets, K is the capital stock, and G denotes the Central Bank's gold holdings." In tum, net foreign assets are defined as NFA., = E,,, - Eft,,,+Ljf,,-Lf,;, -(8) where E4fdenotes local residents' holdings of capital abroad, Efi denotes domestic capital owned by foreigners, Lif are loans issued by domestic residents to foreigners (inclusive of foreign currency reserves held by the domestic Central Bank) and Lf, are loans from foreigners to domestic residents. All quantities are measured in 1995 US dollars at PPP. The various wealth components shown above are constructed in two steps. First, we use the limited available information on stocks of these assets to determine an initial value. The second step involves the use of flow data and estimates of changes in the value of these assets to " Thus, we abstract from other components of wealth such as natural resources and human capital. 11 extend the initial stocks forward and backward over time.12 The required information is drawn from a number of standard sources: initial stocks of domestic capital are taken from the Penn World Tables, and combined with flow data on gross domestic investment to build up capital stock series. For foreign holdings of domestic equity and domestic holdings of foreign equity, we rely primarily on data on stocks and flows of direct and portfolio equity investment reported in the IME's Balance of Payments Statistics Yearbook. Finally, stocks of borrowing and lending are obtained by combining stock data on the debt of developing countries reported in the World Bank's Global Development Finance with data on debt stocks and flows from the Balance of Payments Statistics Yearbook. To account for mismeasurement of capital flows (and hence stocks) and in order to capture unrecorded assets, we augment our measures of loan assets by adding to them the cumulative errors and omissions of the Balance of Payments. Putting together all these pieces, we arrive at estimates of the wealth stock of the countries in the sample. Using these estimated wealth stocks, we construct the foreign wealth / domestic wealth ratios of each country i. This procedure yields data on wealth and its components for a large group of industrial and developing countries. 13 For the empirical experiments in this paper, we restrict the sample to those economies possessing a number of annual observations in the period from the 1 960s to the present sufficient to allow country-specific time-series econometric estimation. We set such minimum at 20 (consecutive) years. This results in an unbalanced panel of 54 countries with time coverage ranging from 20 to 32 years. 12 The main exceptions to this procedure are gold holdings, on which complete stock data are available from the IMF's International Financial Statistics, and some specific items of loan assets and liabilities. These are foreign currency reserves of the central bank, available from IMF sources, and foreign debt of developing countries, available from the World Bank's Global Development Finance. For the remaining wealth components, complete stock data are unavailable, and hence we rely on the method of cumulating flows even for those countries with more abundant stock data in order to avoid a potential bias that could result from applying different methods to construct stocks in different countries: as longer time series of stock data are available for a few rich countries, using these as the primary source would essentially result in different methods being used to construct stocks for rich and poor countries. These differences would then contaminate our inferences regarding, for example, how net foreign assets vary with wealth. 13 We begin with a sample of 98 countries with population greater than one million and per capita GDP greater than 1000 US dollars at PPP in 1990. Of these we discard 25 countries with missing, incomplete, or inconsistent balance of payments data. Next, we also drop 5 former socialist economies, whose data we view as of uncertain reliability, and a handful of developing countries that have experienced prolonged war episodes over the sample years. Finally, we also remove a few country-year observations characterized by very small (or even negative) estimates of wealth, corresponding to countries with very large external debt. We exclude these observations of doubtful quality by limiting the sample to those where the ratio of wealth to GDP is greater than 0.5. 12 The countries in this sample are admittedly very diverse. As already explained, for some of them return and risk considerations may not be the only or most important factor behind the changes in their net foreign asset positions. Non-market forces -- related to, for instance, geopolitical interests, humanitarian aid, or developmental purposes -- may drive to some extent the transfer of capital resources across countries. In addition, some countries use capital and current account restrictions to prevent market forces from 'undoing' net foreign asset positions based on non-market factors. These considerations have the practical implication that the long- run impact of risk and return on net foreign assets may not be the same for all countries (which in turn would imply that the long-run restrictions imposed by the PMG estimator would hold only for specific country groupings). To explore this issue, we break the overall country sample according to two criteria. First, we separate high- and upper-middle-income countries from lower-income countries. Specifically, using the World Bank's World Development Report income classification we form one group consisting of 29 industrial, high- income and upper- middle income developing economies, and a matching group of 25 low and lower- middle income developing economies. Second, we separate countries that feature low capital controls from those that have high capital controls. The only source of data on this topic with broad time-series and cross-country coverage is the IMF's Exchange Rate Restrictions, which includes qualitative information on four kinds of measures that hamper international portfolio diversification. 14 We combine them into a summary measure by adding them up, and compute the average for each country over the period 1965-97. If for a country the average is greater than or equal to three (implying that, on average, restrictions exist in at least three of the four categories throughout the sample period), we classify the country as having high capital controls. This procedure yields a subsample of 20 countries with low capital controls and 34 with high capital controls. The countries included in each subsample are listed in Table Al in the Appendix. An inspection of the list of countries in each group shows that almost all countries with low capital controls belong to the group of high and upper-middle income countries (the exception is Thailand). 14 These are: (a) multiple exchange rate practices, (b) current account restrictions, (c) capital account restrictions, and (d) mandatory surrender of export proceeds. 13 Table 1 presents some descriptive statistics on the net foreign asset / wealth ratios for the full sample and the various country groups just defined. For the overall country sample, both the mean and median of country averages are negative, an indication of the fact that few countries possess net creditor positions. However, the figures reflect some systematic differences across country groups. As just noted, rich countries, as well as countries with less restricted capital accounts, tend to possess higher NFA/wealth ratios than poor ones. Among higher income countries, as well as countries with moderate capital account restrictions, the median NFA/wealth ratio is below the mean, reflecting the existence of a small group of large creditors. The opposite happens among lower income countries and countries with high capital controls, where the mean is below the median. Dispersion of the NFA ratios to wealth is also much higher for low-income than for high- income countries. Finally, NFA/wealth ratios of rich countries (as well as those of countries with low capital account restrictions) show only modest variation over time, while those of low-income countries display a pronounced decline in the 1980s followed by a recovery in the 1990s. The group of countries with high capital account restrictions shares this pattern. 3.2 Measures of return and risk Apart from wealth ratios, the key explanatory variables in our model of net foreign asset positions are the measures of relative risk and return for each country. In practice, these likely depend on a large variety of underlying variables reflecting relative prices, total factor productivity, transaction costs, property rights, tax regimes and so on. In order to consider as many relevant underlying variables as possible and assess the robustness of the results, we use three alternative sets of indices for the categories introduced in the theoretical discussion - namely, expected returns (REq), perceived risks (R[q), and co- movement with other countries' returns (COf). The first and most ambitious set of indices is constructed as a weighted average of several indicators of economic performance, as described below. The second set is exclusively based on the level and variance of real GDP growth per capita. The third set focuses on the profitability of- the domestic stock market, that is, on the level and variance of stock returns (calculated from constant U.S. dollar prices). The first two sets of indices reflect overall economic activity, while the third one accounts mostly for the activity of those firms traded in organized equity markets. The motivation for the composite indices is-that they summarize the information provided by several macroeconomic variables regarding the performance of investment projects in the 14 country. In contrast, the second and third sets of indices take an alternative, minimalist approach as they are based on a single-variable proxy. The advantage of the composite set of indices is its comprehensiveness while the others' advantage is their simplicity and clarity. Using all of them, we can examine whether the estimation results are robust to changes in return and risk measurement. Therefore, their respective results should be regarded as complementary. Chart 1 summarizes the three alternatives. In all three cases, co-movement was measured by the correlation of the relevant return index in a country and the rest of the world. 15 Chart 1: alternative measures of return and risk Expected return | Perceived risk' 1. Cor osite indexb Overall productivity: real per capita GDP General macro instability: standard deviation of real -growth per capita GDP growth Absence of price distortions: inverse of the Lack of intemational risk-sharing: ratio of external debt black market premium - i.e., l/(l+bmp) to debt +equity external liabilities Financial depth: quasi- liquid liabilities/GDP Nominal instability: average and standard deviation of inflation Openness: real imports plus exports / GDP Extemal instability: standard deviation of real exchange rate changes, standard deviation of terms of trade shocks, standard deviation of (imports + exports) / GDP Institutional quality: Indices of governance Low institutional quality: negatives of indices of (Kaufinann et.al.) and Gastil civil liberties govemance (Kaufmann et.al.) and Gastil civil liberties Low tax burden: negative of government Lack of financial depth: negative of quasi-liquid consumption / GDP liabilities/GDP Size and scale economies: population size 2. GDP-based Real per capita GDP growth Standard deviation of real per capita GDP growth 3. Stock market-based Real stock market return I Standard deviation of real stock market retuim Notes: aAll standard deviations are computed over the current and four preceding years. b The components listed were aggregated giving 50% weight to GDP growth in the case of return, and its standard deviation in the case of risk. In both cases, the remaining components received equal weights. Sources: World Bank World Development Indicators; IMF International Financial Statistics; Freedom House; Kauffmann et al. (1999); Standard and Poor's Emerging Markets Database: Shiller (1999. 2001). '5 Specifically, we computed rolling correlations of the return index in a country and the average for the rest of the world, considering overlapping periods spanning the current and four preceding years. 15 In the case of the composite indices of risk and return, the underlying components were selected on the basis of both their relevance in previous theoretical and empirical work and their data availability (see Milesi-Ferreti and Razin 1996, 1998; Easterly, Islam, and Stiglitz 1999; and Rodrik 1999). 16 Each individual component was standardized using its respective pooled (time- series, cross-section) mean and variance. Apart from homogenizing units across indicators, this standardization procedure allows us to control for common factors and yields measures for the performance of a country relative to the world. An issue is how to weigh the underlying indicators to construct the composite indices. Since there is no obvious weighing scheme, we decided to favor the indicators related to the level and variance of per capita GDP growth rates and assign them a large weight in the return and risk indices, respectively. The following two reasons justify this choice. The first is motivated by the new growth literature and argues that GDP growth per capita reflects the most important elements of economic policy and performance. The second reason is statistical and based on the fact that when stock- market returns are regressed on all of our underlying indicators, per capita GDP growth takes the lion share of explained variance. In practice, we assigned a 50% weight to the level and standard deviation of the per capita GDP growth rate in the return and risk indices, respectively; all remaining variables received equal weights. 17 Combining the risk and return data with the wealth and foreign asset data, we obtain an unbalanced panel covering the years 1966-97. In the case of the composite and growth-based indices, the panel includes 54 countries. In turn, for the set of indices derived from stock market returns the sample size is considerably smaller - just 33 countries -- and with a large representation of industrial economies. Tables A2-A4 in the appendix show the correlations between the composite indices, their underlying indicators, and the single-variable indices. Also, Tables A5-A10 provide descriptive statistics on the three sets of indices for selected samples of countries and time periods, in a form analogous to Table 1. It is immediately apparent from the tables that higher- income countries and 16 Note that some variables (such as financial depth and governance quality) enter in both the return and risk measures. The reason is that they may affect both the level and the degree of uncertainty of the return on the country's assets. 17 We also constructed indices giving equal weights to all variables underlying each composite index. Furthermore, we experimented with indices constructed as principal components of the underlying indicators. The empirical results (not reported to save space) were qualitatively similar to those related to our main weighing scheme. 16 countries with low capital account restrictions typically possess higher returns and lower risks than lower- income countries and countries with high capital account restrictions. 4. EMPnuCAL RESULTS The main objective of our empirical analysis is to examine whether long-run movements in the ratio of NFA/wealth for a given country are related to long-run changes in the risk, return and wealth characteristics of the country relative to the world, as a portfolio-diversification model would predict. We want to test if a country's NFA/wealth responds negatively to its (differential) mean returns and the ratio of foreign to domestic wealth, and positively to its (differential) perceived risks and co-movement with the world economy. Furthermore, we would like to explore whether these predictions hold for all countries or for particular groups of them. We use the econometric methodology outlined in section 2 based on the pooled mean group (PMG) estimator to obtain the coefficients of the long- and short-run relationships between NFA/wealth and its proposed determinants. As noted earlier, the PMG estimator forces the long- run coefficients to be homogenous across countries in the sample but allows the short-run parameters to vary from country to country. Given that we expect the portfolio-diversification model to drive the allocation of external assets mostly in the long run (that is, after an adjustment period), our focus is on the steady-state relationship. In the estimation we also allow for intercept heterogeneity by including country-specific constants. These will account for unobserved time-persistent factors that are specific to each country -- such as home-bias effects. Furthermore, in order to eliminate common factors across countries --which would induce cross-sectional correlation of the residuals--, we also allow for time (year) effects. The inclusion of country- and time-specific intercepts modifies the interpretation of the estimated coefficients. Including country-specific intercepts means that we allow the NFA/wealth ratio to vary across countries for factors not totally captured by the explanatory variables. In turn, including time-specific intercepts implies that the change in each variable should be interpreted as a change relative to the mean of all countries, as already noted. Two other important specification assumptions are that the regression residuals be serially uncorrelated and that-the explanatory variables can be treated as strictly exogenous. As noted in section 2, we seek to meet these requirements by appropriately selecting the lag order of the 17 ARDL process- for NFA/wealth in each country. We use the Schwartz Bayesian Criterion (SBC) to determine the dynamic specificationfor each country, subject to a maximum of two lags for each of the five variables in the model (NFA/wealth ratio, return, risk, co-movement, and foreign/domestic wealth ratio). The specification selected in this way varies across countries; however, for most of them the information criterion selected at least one lag for NFA/wealth and foreign/domestic wealth. In a number of cases the SBC also retained lags of the return, risk, and co-movement indicators. 8 The PMG estimator does not require the variables to be stationary or have the same order of integration. Nevertheless, given the novelty of this estimator, there may be some lingering doubts as to whether its properties prevail in the presence of integrated series. These doubts, however, do not apply in our case given that all the series involved in our econometric model appear to be stationary. First, on conceptual grounds, we work with ratios, rates of growth, and normalized indices that are naturally bounded (see Cochrane 1991). Second, on statistical grounds, we conduct panel unit-root tests whose results reject the null hypothesis of nonstationarity for each of the series included in our empirical model (see Table 2). This strongly suggests that we are working with stationary series. - Tables 3-5 present the estimates of the long-run coefficients for different groups of countries. In Table 3 we use the composite indices of risk and return, in Table 4 we use the indices based on the growth rate of GDP per capita, and in Table 5 we use the indices derived from stock- market returns. In all cases, the results are broadly supportive of the empirical specification when the model is estimated -on the high-income and/or low-capitalcontrol samples. When using the composite indices of risk and return (Table 3),- all the explanatory variables carry the expected sign and their coefficients are statistically significant for the sample of high- income countries; the results are similar for the sample of low-capita}control countries, except that the comovement index is no longer significant. In turn, when using the indices based on per capita GDP growth and stock market returns (Tables 4 and 5), we find that the return and risk measures as well as the relative wealth ratio carry significant coefficients of the expected 18 We also experimented with imposing common dynamic specifications across countries; this obviously alters the short-run estimates but has a relatively minor effect on the long-run parameters. 18 sign for the samples of high- income and low-capitalcontrol countries; on the other hand, the comovement index is not statistically significant. For these samples of countries, the main results are not only statistically significant but also economically relevant. Focusing onthe sample of high and upper-middle income countries, we can draw some estimates and comparisons for the long-run effect on net foreign asset positions of changes in the portfolio indices and relative wealth. A one-standard-deviation increase in the composite return index leads to a reduction in the ratio of NFA to domestic wealth of about 0.28 standard deviations, and an analogous increase in the composite risk index produces a decline of twice that magnitude. A corresponding increase in the composite comovement index produces a rise of about 0.11 standard deviations of NFA/domestic wealth; thus, the effect of the comovement index is not only statistically weak but also economically small in relative terms. An increase of one-standard deviation in the ratio of foreign to domestic wealth leads to a decrease in NFA/domestic wealth of about 0.28 standard deviations, quite similar to the corresponding effect of the return index. When we use the indices based on GDP growth, the results are similar with two exceptions. First, the effect of the risk index drops but still remains above that of the return index; and second, the effect of the comovement index falls to one-tenth of the effect of the other variables. Finally, when we use the indices based on stock returns, the ratio of foreign to domestic wealth becomes the most important variable. The effects of return and risk are smaller than in the previous cases, but the ranking of their relative strength (first risk, then return) is preserved. Focusing on the economic impact on NFA/domestic wealth, we can draw the following conclusions. First, the return and risk indices based on several macroeconomic variables and GDP growth rate have a larger effect on the net foreign asset position than those based solely on stock-market returns. Second, changes in the risk index appear to cause stronger effects than those of changes in the return index. Third, the effect of the cormiovement index is of a much smaller magnitude than those of the other variables. And fourth, although the effect of relative wealth varies somewhat with the type of indices used, its magnitude is always of the same order as the effect of the return and risk indices. The results change considerably when we consider other samples of countries. In the full sample, as well as for the groups of low and lower- middle income and high capital control countries, the risk and return proxies are in most cases insignificant and in some cases carry the 19 wrong sign. The same occurs with the co-movement indicator. Only the coefficient on the ratio of foreign to domestic wealth remains consistently negative and significant for all groups of countries and for the three types of retum/risk measurements. 19 For countries with high capital controls, the weaker performance of the model might be viewed as evidence that capital controls achieve some degree of success - they dampen the effects of risk and return factors on portfolio decisions. For the lower income countries, the likely reason is the limited role that optimal diversification decisions play in the observed evolution of net foreign assets, which may be dominated instead by other considerations such as the willingness of donor governments to extend, and forgive, concessional lending. In summary, our portfolio-diversification approach seems to apply for some, but not all, groups of countries. For countries where market forces are likely to dominate other considerations, our results indicate that when a country becomes more productive (greater mean returns) and more stable (lower perceived risk), its net foreign asset position relative to wealth declines. The effect of providing a better hedge for worldwide risks (lower co- movement) appears to go in the same direction, but our results in this respect are less significant and robust. Finally, note that the effects of return, risk, and co-movement on the NFA ratio hold when we control for relative wealth. Wealth per se has a significant influence over the NFA ratio in the sense that when domestic residents' wealth grows faster than that of foreigners, the fraction of net foreign assets in wealth increases. Tables 6-8 display additional results for the samples of high and upper-middle income countries and low capital control countries, for which the empirical model is more successful. In Table 6 we use the composite indices of risk and return, while in Tables 7 and 8 we use the indices based on per capita GDP growth and stock market returns, respectively. In these tables we present the estimation of the full error-correction model using both the Pooled Mean Group estimator and its Mean Group counterpart that allows for unrestricted long-run parameter heterogeneity across countries. Comparison between both sets of estimates allows the construction of formal tests of the long-run pooling restrictions imposed by the Pooled Mean Group estimator. As explained in section 2, we can test the maintained assumption in the PMG estimator that the long-run coefficients are the same across countries through Hausman-type tests. 19 This robust effect of wealth is in agreement with the stylized fact underscored by Kraay et al.(2000) that foreign 20 Specifically, we can compute individual test statistics for each one of the long-run coefficients. These are reported, along with the associated p- values, in columns 3 and 6 of Tables 6-8. We find that the cross-country homogeneity of long-run coefficients is never rejected in the cases of the return, co-movement, and relative wealth variables. This is also the case for the risk index in four out of the six instances considered. Cross-country homogeneity of the risk- related coefficient is rejected in the sample of low capital controls with composite indices and in the sample of high- income countries with stock-return indices. It is also apparent from Tables 6-8 that the long-run coefficients estimated with the alternative Mean Group method suffer from very poor precision. Of 24 coefficients (6 exercises with 4 explanatory variables each), only six are statistically significant, and only the coefficient on relative wealth shows a consistent (negative) sign across all exercises. This lack of precision and robustness across different samples and return/risk measures reflects the sensitivity of the MG estimator to outliers in the country.-specific estimates. The bottom half of Tables 6-8 reports the average estimates of the speed of adjustment (denoted as q in equation (6) above) and the short-run parameters. As required for dynamic stability, the coefficient on the error-correction term (i.e., the speed of adjustment) is negative and significant in all six exercises. It is also somewhat smaller in magnitude in the PMG than in the MG specification, in accordance with the theoretical prediction that pooling in the presence of heterogeneity tends to increase inertia (Robertson and Symons 1992). Focusing on the PMG estimates, the average short-run parameters obtained for the two samples and three sets of return/risk indices reveal significant lagged effects of the dependent variable and contemporaneous effects of the foreign/domestic wealth ratio. In addition, there are also significant contemporaneous effects of the return variable when the composite indices are used and lagged effects of the foreign/domestic wealth ratio for the sample of low capital controls. On the whole, the explanatory power of the PMG estimates is rather satisfactory, and the average of the country-specific adjusted RI is over 30% for the high and upper-middle income countries and over 40% for the low capitalcontrol countries (R2s are larger for the MG estimates). This is encouraging particularly in view of the large sample size (828 and 577 assets show a strong positive association with wealth levels. 21 observations for high income and low capital control samples, respectively) and the simplicity of the model. 5. CONCLUSIONS The determinants of intemational portfolio diversification have attracted considerable attention in the literature. Empirical studies have examined mostly equity holdings across a small number of industrial economies, and in most cases conclude that the extent of intemational diversification falls short of what would be predicted by standard portfolio equilibrium models - the home bias puzzle. This paper explores empirically the role of risk and. return factors in the observed evolution of net foreign asset positions of a large number of industrial and developing economies. Its objective is to examine whether intemational capital flows respond to market incentives and, if so, whether this conclusion can be generalized to all countries in the world or only particular subsets of them. Thus, the paper does not attempt to reconcile the observed extent of diversification with theoretical predictions, but instead tries to assess empirically the role of changing fundamentals in the actual evolution of intemational portfolios, taking implicitly as given their 'home bias'. The paper adopts a dynamic approach according to which intemational and domestic investors achieve in the long run their desired portfolio allocation of assets across countries. Frictions and adjustment costs, however, can make short-run portfolios differ from their long-run counterparts. Based on a standard Markowitz-Tobin portfolio diversification framework, the paper develops a reduced-form model of net foreign asset positions. The model yields a long-run equilibrium condition in which the ratio of NFA to the total wealth of domestic residents depends on four factors: investment returns in the home country relative to the rest of the world, investment risk in the home country relative to the rest of the world, the degree of co- movement between investment returns at home and abroad, and the ratio of foreign-owned to domestic- owned wealth. The paper fo'cuses on the empirical estimation of this long-run equilibrium condition, using data on foreign asset and liability stocks for a large number of industrial and developing countries spanning the period from the 1960s to the present. With these data and capital stock estimates, the wealth of each country's residents can be computed. In addition, the paper 22 develops measures of country returns and risks - in three versions: composite indices construc ted using a comprehensive set of macroeconomic, policy, and institutional variables; indices based on the rate of economic growth; and indices based on stock market returns. The econometric approach is derived from the Pooled Mean Group estimator recently developed by Pesaran, Shin, and Smith (1999). This approach is well-suited to the paper's objective, as it provides a dynamic setting imposing a long-run relationship common to all countries but allows for heterogeneous short-run adjustment across countrie s. On the whole, the estimation results lend support to the model when applied to high and upper- middle income countries and/or countries with moderate capital account restrictions. The estimated long-run parameters on relative wealth and the two alternative measures of risk and return are correctly signed and always significant. Thus, as predicted by the theoretical model, net foreign assets (as a ratio to total wealth) are negatively related to the measures of domestic investment returns and the ratio offoreign to domestic wealth, and positively to the measures of investment risk. Our measure of co- movement also shows an association with the NFA/wealth ratio, but not as robust as with the other explanatory variables. Finally, the long-run parameter homogeneity across countries imposed by the PMG estimator is supported in most cases by Hausman specification tests. The results for countries characterized by high capital controls and, especially, lower income levels, are less supportive of the portfolio equilibrium model. For the former countries, this might be viewed as evidence that capital controls achieve some degree of success - they dampen the effects of risk and return factors on portfolio decisions. For the lower income countries, the likely reason is the limited role that optimal diversification decisions play in the observed evolution of their net foreign assets. To a large extent, these consist of official concessional debt, whose pattem across countries and over time may be dominated instead by nonr-market considerations related to geopolitical interests, humanitarian aid, and development purposes. 23 REFERENCES Adler, Michael and Bernard Dumas (1983). "International Portfolio Choice and Corporate Finance: A Synthesis." Journal of Finance, 38:925-84. Bacchetta, Philippe and Eric Van Wincoop (1998). "Capital flows to emerging markets: liberalization, overshooting, and volatility." National Bureau of Economic Research Working Paper Series No. 6530:1-43, April. Berg, Andrew and Catherine Patillo (1999). "Are currency crises predictable? A test." IMF Staff Papers 46: 107-38. Caballero, Ricardo J. (1998). "Aggregate Investment." National Bureau of Economic Research. Working Paper Series No. 6264 :1-5 8, November. Cochrane, John H. (1991). "What Should Macroeconomists Know About Unit Roots? Comments on 'Pitfalls and Opportunities: What Macroeconomists Should Know About Unit Roots"' by John Campbell. NBER Macroeconomics Annual. Dixit, Avinash and Robert S. Pindyck (1994). Investment under Uncertainty. Princeton University Press. Easterly, William; Roumeen Islam and Joseph E. Stiglitz (1999). "Shaken and Stirred: Volatility and Paradigms for Rich and Poor countries." Michael Bruno Memorial Lecture. Given at the XII World Congress of the IEA. Buenos Aires, August 1999. Easterly, William and Aart Kraay (1999). "Small Countries, Small Problems?" Washington, D.C.: The World Bank Manuscript French, Kenneth R. and James M. Poterba (1991). "Investor Diversification and International Equity Markets." American Economic Review, 81:222-26. Ghosh, Atish and Jonathan D. Ostry (1995). "The Current Account in Developing Countries: A perspective from the Consumption-Smoothing Approach." The World Bank Economic Review, 9: 305-334. Hsiao, Cheng (1997). "Statistical properties of the two-stage least squares estimator under cointegration." Review of Economic Studies64(3): 385-98. Johanssen, Soren (1995). Likelihood based Inference in Cointegrated Vector Autoregressive Models. Oxford: Oxford University Press. Kraay, Aart et al. (2000). "Country Portfolios." National Bureau of Economic Research Working Paper Series No. 7795: 1 -61, July. Kraay, A. and J. Ventura (2000). "Current accounts in debtor and creditor countries." Quarterly Journal of Economics 95, 1137-1166. Lane, Philip and Gian Maria Milesi-Ferreti (1999). "The External Wealth of Nations: Measures of Foreign Assets and Liabilities for Industrial and Developing Countries." Washington, D.C.: International Monetary Fund. Manuscript. August. Lewis, Karen K (1999). "Trying to Explain Home Bias in Equities and Consumption". Journal of Economic Literature XXCVII: 571-608. Mann, C. (2002). "Perspectives on the U.S. current account deficit and sustainability." Journal of Economic Perspectives 16, 131-152. Merton, R. (1971). "Optimum consumption and portfolio rules in a continuous time model." Journal of Economic Theory 3, 373-413. Milesi-Ferreti, Gian Maria and Assaf Razin (1996). "Current account Sustainability: Selected East Asian and Latin American experiences." National Bureau of Economic Research Working Paper Series No. 5791: 1-61, October. Pesaran, M. Hashem (1997). "The role of economic theory in modeling the long run." Economic Journal 107:178-91, January. Pesaran, M. Hashem and Yongcheol Shin (1997). "Long Run Structural Modeling." University of Cambridge, DAE Working Paper No. 9419. *24 Pesaran, M. Hashem and Yongcheol Shin (1999). "An Autoregressive Distributed Lag Modelling Approach to Cointegration Analysis." Econometrics and Economic Theory in the 20th Century: The Ragnar Frisch Centennial Symposium, chapter 11, (ed.) S. Strom, , Cambridge University Press, Carnbridge. Pesaran, M. Hashem, Yongcheol Shin and Ron P. Smith (1999). "Pooled Mean Group Estimation of Dynamic Heterogeneous Panels." Journal of the American Statistical Association, 94: 621-634. Pesaran, M. Hashem, Yongcheol Shin and Richard J Smith (2000). "Bounds Testing Approaches to the Analysis of Level Relationships."University of Cambridge, DAE Working Paper No.9907. Pesaran, M. Hashem and R.P. Smith (1995). "Estimating Long-Run Relationships from Dynamic Heterogeneous Panels."Journal of Econometrics 68: 79-113. Pesaran, M. Hashem, R.P.Smith and K.S.Im (1996). "Dynamic Linear Models for Heterogenous Panels." In The Econometrics of Panel Data, (eds.) L. Matyis and P. Sevestre, Kluwer Academic Publishers. Phillips, Peter C.B., and B. Hansen (1990). "Statistical Inference in Instrumental Variables Regression with I(1) Processes". Review of Economic Studies 57: 99-125. Robertson, D. and J. Symons (1992). "Some strange properties of Panel Data Estimators". Journal of Applied Econometrics, 7: 175-189. Rodrik, Dani (1999). "Why is there so much insecurity in Latin America?". Washington, D.C.: The World Bank Manuscript. Shiller, R.J. (1989). Market Volatility. Cambridge, MA: MIT Press. Shiller, R.J. (2001). "Market Volatility Update". www.econ.vale.edu/-shiiler/data/chapt26.html Standard and Poor's (2000). Emerging Markets Database. New York: McGraw-Hill. Tesar, Linda L. and Ingrid M. Werner (1995). "Home Bias and High Turnover." Journal of International Money and Finance, 14:467-92. Ventura, J. (2002): 'Towards a theory of current accounts", NBER Working Paper 9163. 25 TABLES Table 1. Descriptive Statistics on the Ratio of Net Foreign Assets to Wealth Period 1966-97 1966-79 1980489 1990-97 1. All Countries Mean -15.1% -10.5% -19.5% -17.4% Median -10.4% -8.7% -12.8% -9.6% Standard Deviation 27.6% 18.1% 33.6% 30.8% No. Observations 1597 684 540 373 2. High and Upper Middle Income Countries Mean -5.0% -4.4% -5.3% -5.6% Median -5.8% -4.2% -7.9% -6.3% Standard Deviation 16.4% 18.0% 18.6% 8.7% No. Observations 886 378 290 218 3. Low and Lower Middle Income Countries Mean -27.8% -18.0% -35.9% -34.1% Median -17.9% -15.1% -22.1% -19.6% Standard Deviation 32.9% 15.2% 39.2% 41.4% No. Observations 711 306 250 155 4. Countries with Low Capital Restrictions Mean -1.8% -1.5% -0.9% -3.7% Median -3.3% -1.8% -4.7% -3.9% Standard Deviation 17.8% 19.7% 19.9% 9.1% No. Observations 617 267 200 150 5. Countries with High Capital Restrictions Mean -23.5% -16.2% -30.4% -26.7% Median -15.4% -12.3% -18.5% -15.5% Standard Deviation 29.3% 14.4% 35.2% 36.4% No. Observations 980 417 340 223 26 Table 2. Panel Unit Root Tests Im, Pesaran and Shin (1995): The tin Statistic Sample Levels Levels Variables Period without trend with trend Ratio of Net Foreign Assets to Wealth 1966-97 -5.8122** -2.2997** Ratio of Foreign to Domestic Wealth 1966-97 -2.7424** -2.5149** 1. Composite Indices Index of Returns (RE) 7 1966-97 -2.4663** | -6.0693*i Index of Risks (RI) 1 1966-97 -1.8974** -3.0823** Comovement of Returns (CO) 1966-97 -2.2861** -2.2721** 11. Growth-based Indices _2_5565_ _ Growth per capita (DY) 1961-97 -2.7424** -2.5565** Std. Dev. Growth per capita (SDY) 1964-97 -2.0089** -2.5149** Comovement of Growth rates (COY) -1961-97 -2.0754** -2.5395** Il. Stock Retum Indices Stock Returns (SR) 1960-97 -3.1680** -3.9104** Std. Dev. Stock Returns (SDR) 1960-97 -2.2373** -2.7288** Comovement of Stock Returns (COS) 1960-97 -2.6190** -2.8013** Notes: Before performing the ADF regressions for individual countries, we remove the common time dummies from all variables. The ADF regression in levels includes the time trend, whereas the ADF regression in differences does not. In the latter case, the altemative hypothesis is that series is stationary around a constant since any time trend in levels will be removed by differencing. This table reports the tbar (t,N ) statistic, defined as the sample average of the t-statistics obtained from the ADF regressions of individual countries. For 85 countries dwing the 1960-97 period, the approximate sample crtical values of the tF. statistic are: (i) Without deterministic trend: -1.73, -1.67, and -1.64 at the 1, 5, and 10 percent significance level; (ii) With deterministic trend: -2.37, -2.31, and -2.28 at the 1, 5, and 10 percent significance level. In addition, note that for the stock market indicators we have data only for 40 countries. In this case the approximate critical values of the tF. statistic are: (Q wlo deterministic trend: -1.81, -1.73, and - 1.68 at the 1, 5, and 10 percent significance level, (ii) With deterministic trend: -2.44, -2.36, and -2.32 at the 1, 5, and 10 percent significance level. For more details, see Table 4 in Im, Pesaran and Shin (1995). * (*) indicates that the test is significant at the 10 (5) percent level. 27 Table 3. Long-Run Relationship between Net Foreign Assets and Measures of Risk and Return (I): Composite Indices - Dependent variable: ratio of net foreign assets to wealth (NFAIW) - Estimation method: Pooled Mean Group estimator (Pesaran, Shin and Smith 1999), controlling for country and time effects. - Samples: All countries and groups formed on the basis of income levels and capital controls. - Period: 1966-97, Annual Data Income Level Capital Controls All High and Upper- Lower and Lower- Low High Variables Countries Middle Middle Income Controls Controls Income A. Long-Run Parameters Return (RE) 0.03212 -0.10164 0.00829 -0.11792** 0.04486 (0.03) (0.02) (0.02) (0.02) (0.03) Risk (RI) 0.01494 ** 0.19106 0.01548 0.23639 ** -0.00683 (0.01) (0.02) (0.02) (0.02) (0.01) Comovement (CO) -0.01222 0.03590 ** -0.02387 0.01219 -0.00139 (0.01) (0.01) (0.02) (0.01) (0.01) Foreign / Domestic -0.00015 -0.00030 -0.00014** -0.00030 ** -0.00010 ** Wealth (WfIWi) (0.00) (0.00) (0.00) (0.00) (0.00) No. Countries 54 29 25 20 34 No. Observations 1,495 828 667 577 918 Average RBarSq 0.3272 0.3200 0.4792 0.4280 0.3918 Observations: * Significant at the 10 percent level, ** Significant at the 5 percent level Numbers In parenthesis below coefficient estimates are standard errors. 28 Table 4. Long-Run Relationship between Net Foreign Assets and Measures of Risk and Return (Il): Indices Based on GDP Growth Dependent variable: ratio of net foreign assets to wealth (NFAIW) - Estimation method: Pooled Mean Group estimator (Pesaran, Shin and Smith 1999), controlling for country and time effects. - Samples: All countries and groups formed on the basis of income levels and capital controls - Period: 1966-97, Annual data Income Level Capital Controls All High and Lower and Low High Upper- Lower- Variables Countries Middle Income Middle Income Controls Controls A. Long-Run Parameters Growth in GDP -0.07490 -1.46684** -0.42531 -1.12810** 0.41484 per capita (DY) (0.16) (0.32) (0.39) (0.34) (0.21). Std. Dev. in GDP per 0.02935 2.39211 1.18297 * 2.64142 ** 0.87326 capita Growth (SDY) (0.14) (0.35) (0.35) (0.37) (0.17) Comovement (COY) -0.01724 -0.00832 -0.02904 * 0.01218 -0.01866 (0.01) (0.01) (0.02) (0.01) (0.01) Foreign / Domestic -0.00015 ** -0.00031 ** -0.00012 -0.00030 -0.00011 Wealth (Wf/Wi) (0.00) (0.00) (0.00) (0.00) (0.00) No. Countries 54 29 25 20 34 No. Observations 1495 828 667 577 918 Average RBarSq 0.2298 0.3209 0.4768 0.4110 0.3103 Observations: * Significant at the 10 percent level, ** Significant at the 5 percent level Numbers in parenthesis below coefficient estimates are standard errors. 29 Table 5. Long-Run Relationship between Net Foreign Assets and Measures of Risk and Return (III): Indices based on Stock Retums - Dependent variable: ratio of net foreign assets to wealth (NFAIW) - Estimation method: Pooled Mean Group estimator (Pesaran, Shin and Smith 1999), controlling for country and time effects. - Samples: All countries and groups formed on the basis of income levels and capital controls - Period: 1966-97 Income Level Capital Controls All High and Upper- Lower and Low High Lower- Variables Countries Middle Income Middle Income Controls Controls A. Long-Run Parameters Stock -0.03355** -0.04801 ** -0.06073* -0.03520** 0.02154 Returns (SR) (0.009) (0.012) (0.034) (0.007) (0.014) Std. Dev. of Stock 0.12929 ** 0.11677 ** 0.00588 0.06946 ** 0.05067 ** Returns (SDR) (0.017) (0.023) (0.060) (0.018) (0.020) Comovement of 0.00014 -0.00581 0.01318 0.00486 0.00461 Stock Retums (COS) (0.005) (0.007) (0.027) (0.006) (0.009) Foreign / Domestic -0.00017 ** -0.00062 -0.00013** -0.00004 ** -0.00017 ** Wealth (Wf/Wi) (0.000) (0.000) (0.000) (0.000) (0.000) No. Countries 33 26 7 19 14 No. Observations 875 699 176 534 341 Average RBarSq 0.5927 0.3900 0.8857 0.4589 0.7779 Observations: * Significant at the 10 percent level, ** Significant at the 5 percent level Numbers in parenthesis below coefficient estimates are standard errors. 30 Table 6. Long- and Short-Run Relationship between Not Foreign Assets and Measures of Risk and Retum (1): Composite Indices - Dependent variable: rato of net foreign assets to wealth (NFAIW) - Estimation method: Pooled Mean Group and Mean Group estimators, controlling for country and tme effects - Samples: Groups of countries with high and upper-middle income and lowcapital controls - Period: 1966-97 High and Upper-Middle In;om-e Low Capita' Controls -"Poole9d" Mean Hausman "Pooled" Mean Hausman Variables Mean Group Group Test Mean Group Group Test A. Long-Run Parameters Retum (RE) -0.10164 " 0.41900 1.37 -0.11792 -0.07300 0.02 (0.02) (0.44) [0.241 (0.02) (0.33) [0.891 Risk (RI) 0.19106 " 0.36500 0.85 0.23639 " -0.08200 3.87 (0.02) (0.19) [0.361 (0.02) (0.16) (0.051 Comovement (CO) 0.03590 '" -0.02000 0.23 0.01219 0.05500 0.6 (0.01) (0.12) (0.631 (0.01) (0.06) [0.441 Foreign / Domestic -0.00030 " -0.00001 0.00 -0.00030 -0.00100 1.71 Wealth (Wf/Wi) (0.00) (0.00) (0.971 (0.00) (0.00) [0f191 Error Correcion -0.074i" -0.18. -0.092 -0.154 Coeffident (0.03) (0.04) (0.05) (0.05) B. Short-Run Parameters d[NFA(-1)1 0.161 *- 0.172 * 0.200 * 0.185 (0.042) (0.043) (0.057) (0.053) dRE 0.012 *- 0.011 0.014 0.013 (0.005) (0.006) (0.006) (0.008) dRE(-1) 0.003 0.003 0.001 4.323E-05 (0.004) (0.004) (0.003) (0.004) dRI -0.002 0.001 0.0001 0.007 (0.009) (0.010) (0.011) (0.013) dRI(-1) -0.0069 * -0.005 -0.007 -0.006 (0.004) (0.005) (0.006) (0.007) dCO -0.002 0.001 0.002 0.002 (0.003) (0.006) (0.001) (0.003) dCO(-1) 0.0004 0.001 -0.004 -0.003 (0.001) (0.003) (0.003) (0.002) dWf/Wi 0.0001 0.0002 *- 0.0002 0.0003 * (0.00005) (0.0001) (0.0001) (0.0001) dWf/Wi(-1) 0.0001 0.0002 0.0027 0.0022 (0.001) (0.001) (0.001) (0.001) Constant 0.017 0.021 0.022 0.015 (0.024) (0.024) (0.031) (0.032) No. Countries 29 29 20 20 No. Observations 828 828 577 577 Average RBarSq 0.3200 0.6214 0.4280 0.6680 Observations: I Significant at the 10 percent level, - Signilicant at the 5 percent level Numbers in parenthesis below coefficient es0mates are standard errors. Numbers in parenthesis below Hausman Tests are pvalues 31 Table 7. Long- and Short-Run Relationship between Net Foreign Assets and Measures of Risk and Retum (11): Indices based on GDP Growth - Dependent variable: ratio of net foreign assets to wealth (NFAIW) - Estimation method: Pooled Mean Group and Mean Group estimators, controlling for country and time effects - Samples: Groups of countnes with high and upper-middle income and low.capital controls - Period: 1966-97 High and Upper-Middle Inoome Low Capital Controls "Pooled" Mean Hausman "Pooled" Mean Hausman Vauiables Mean Group Group Test Mean Group Group Test A. Long-Run Parameters Growth in GDP -1.46684 - *7.89800 * 2.04 -1.12810 * -7.19100 0.58 per capita (DY) (0.32) (4.52) [0.15 (0.34) (7.96) [0.451 Std. Dev. in GDP per 2.39211 * 2.70800 0.01 2.64142 " -3.07500 1.33 capita Growth (SDY) (0.35) (3.48) [0.931 (0.37) (4.97) [0.251 Comovement (COY) -0.00832 0.37200 1.41 0.01218 -0.11600 1.37 (0.01) (0.32) [0.231 (0.01) (0.11) [0.241 Foreign / Domestic -0.00031" -0.00002 0.12 -0.00030 " -0.00001 * 0.11 Wealth (WfIWi) (0.00) (0.00) [0.731 (0.00). (0.00) [0.74] Error Correction 40.094 - -0.239 * -0.110 o -0.165 Coeffident (0.04) (0.04) (0.05) (0.06) B. Short-Run Parameters d[NFA(-1)1 0.121 * 0.121 - 0.144*" 0.126 (0.035) (0.043) (0.054) (0.049) DDY 0.043 0.043 0.099 0.091 (0.052) (0.080) (0.070) (0.063) dDY(-1) -0.016 -0.028 0.00025 0.00043 (0.016) (0.028) (0.0005) (0.0007) dSDY -0.112 0.023 -0.069 -0.018 (0.089) (0.137) (0.125) (0.158) dSDY(-1) -0.052 -0.011 -0.066 -0.047 (0.055) (0.062) (0.090) (0.105) dCOY -0.004 * 0.002 0.0008 0.0006 (0.002) (0.002) (0.001) (0.002) dCOY(-1) 0.0008 -0.0004 0.0006 -0.0010 (0.0012) (0.001) (0.001) (0.001) dWf/Wi -0.0001 * -0.0006 - -0.0002 - -0.0026 - (0.0000) (0.0001) (0.0001) (0.0010) dWf/Wi(-1) 0.00004 -0.000003 0.00006 -0.00005 (0.00011) (0.000010) (0.00004) (0.00000) Constant 0.021 0.024 0.032 0.026 (0.027) (0.028) (0.038) (0.039) No. Countries 29 29 20 20 No. Observations 828 828 577 577 Average RBarSq 0.3209 0.5807 0.4110 0.6380 Observations: I Significant at the 10 percent level, *- Significant at the 5 percent level Numbers in parenthesis below coefficient estimates are standard errors. Numbers in parenthesis below Hausman Tests are pvalues 32 Table 8. Long- and Short-Run Relationship between Net Foreign Assets and Measures of Risk and Return (111): Indices based on Stock Retums - Dependent variable: ratio of net foreign assets to wealth (NFAtW) - Estimation method: Pooled Mean Group and Mean Group estimators, controlling for country and time effects - Samples: Groups of countries with high and upper-middle income and low-capital controls - Period: 1966-97 High and Upper-Middle Income Low Capital Controls "Pooled" Mean Hausran 'Pooled" Mean Hausman Variables Mean Group Group Test Mean Group Group Test A. Long-Run Parameters Stock Retums -0.04801- -0,35900" 0.77 -0.03520 * -0.41000 1.04 (SR) (0.012) (4,52) [0.441 (0.007) (0.17) [0.281 Std. Dev. in Stock 0.11677 * 0.86200 * 5.76 0.06946 * 1.219 1.5 Retums (SDR) (0.023) (3.48) [0.02] (0.018) (0.94) [0.221 Comovement (COS) -0.00581 -0.19700 0.96 0.00486 0.158 1.66 (0.007) (0.195) [0.331 (0,006) (0.12) [0.201 Foreign / Domestic -0.00062 -0.00400 0.33 -0.00004 * 40.002 0.38 Wealth (Wf/Wi) (0.000) (0.005) [0.57] (0.000) (0.003) [0.541 Error Correcton -0.099 -0.161 -0.083 - -0.133 Coeffident (0.03) (0.04) (0.03) (0.04) B. Shorl-Run Parameters d[NFA(-1)1 0.112 0.109 0.136 * 0.119 (0.030) (0.028) (0.060) (0.051) dSR 0.004 0.0003 0.001 -0.001 (0.004) (0.003) (0.001) (0.003) dSR(-1) 0.003 0.0028 0.00005 0.0001 (0.005) (0.006) (0.002) (0.005) dSDR -0.003 0.00144 -0.00355 -0.00817 (0.006) (0.012) (0.009) (0.018) dSDR(-1) 40.0065 -0.0045 -0.0056 -0.0032 (0.005) (0.003) (0.005) (0.002) dCOS -0.003 -0.00359 * -0.00232 -0.00286 (0.001) (0.002) (0.002) (0.003) dCOS(-1) 0.0006 40.001 -0.001 -0.0004 (0.003) (0.004) (0.005) (0.003) dWf/Wi -0.00006 -0.0004 -0.000003 -0.00001 (0.0000) (0.0001) (0.000) (0.000) dWfUWi(-1) 0.00001 -0.000003 0.00006 * -0.00005 - (0.0002) (0.00002) (0.00004) (0.00000) Constant -0.003 0.003 -0.00085 -0.00012 (0.003) (0.005) (0.002) (0.008) No. Countries 26 26 19 19 No. Observatons 699 699 534- 534 Average RBarSq 0.3900 0.5527 0.4589 0.5313 Observations: * Significant at the 10 percent level, *- Significant at the 5 percent level Numbers in parenthesis below coefficient estimates are standard errors. Numbers in parenthesis below Hausman Tests are pvalues 33 APPENDIX A: Sample and descriptive statistics Table Al: Sample of Countries Per Capita Income ' Capital Controls Code Country Name Region High Low Low - igh Stock Market Returns ARG Argenina AMER X X X AUS Australia IND X X X AUT Austria IND X X X BEN Benin SSA X X BGD Bangladesh SA X X BRA Brazil AMER X X X CAF Central African Republic SSA X X CAN Canada IND X X x CHL Chile AMER X X X CIV CSte d Ivoire SSA x x COL Colombia AMER X X X CRI Costa Rica AMER X X DEU Germany IND X X X DNK Denmark IND X X X DOM Dominican Republic AMER X X ECU Ecuador AMER X X ESP Spain IND X X X FIN Finland IND X X X FRA France IND X X X GBR United Kingdom IND X X X GHA Ghana SSA x x GRC Greece IND X X X IND India SA X X X ISR Israel MENA X X X ITA Italy IND X X X JAM Jamalca AMER X X X JOR Jordan MENA X X X JPN Japan IND X X X KEN Kenya SSA X X KOR Korea EAP X X X LKA Sri Lanka SA X X X MEX Mexico AMER X X X MLI Mall SSA X X MWi Malawi SSA X X NER Niger SSA X X NGA Nigeria SSA X X X NLD Netherlands IND X X X PAK Pakistan SA X X X PAN Panama AMER X X PER Peru AMER X X PHL Philippines EAP X X X PRT Portugal IND X X X SAU Saudi Arabia MENA X X X SEN Senegal SSA X x SGP Singapore EAP X X SWE Sweden IND X X X THA Thailand EAP X X X TTO Trinidad and Tobago AMER X X TUN Tunisia MENA X X TUR Turkey MENA X X X URY Uruguay AMER X X USA United States IND X X X VEN Venezuela AMER X X X ZAF South Africa SSA X X Total 54 29 25 20 34 33 Notes: I/ The classification of countnos by income level Is based on the criterlon used by the World Bank's World Development Report. 2/ The sub-sample of counirtes according to the presence of capital controls was based on the sum of capital contmols dummies (1 for the presence of the restriction, and 0 otherwise) collected from the IMF's Exchange Anrangements and Exchange Restrictions. These dummies capture the presence of: (a) multiple exchange rate practices, (b) current account restrictons, (c) capital account restrictions, and (d) surrender of export proceeds. If the sum of these fourcategories was higher than or equal to three (i.e. presence of restrictions in at least three categories) on average over the 1965-97 period, we considerit a mmntrv with hinh r-snitPlr tntmr.q (tharwi.n it is Inheeriaa rnr:ntrv with Inw ranitat rnntmlx 34 Table A2 Index of Retums Correlation Analysis Correlation between the Indicator and: Indicator Composite Index Growth per capita Stock Returns Growth in GDP per capita 0.52432** 1 0.2701 (0.0184) (3.0305) Population (in billions) 0.06106** 0.0325* 0.05628 (0.0184) (0.0181) (0.0303) Degree of Openness 0.49400-' 0.0355* 0.1048 (0.0184) (0.0183) (0.0302) Financial Depth 0.67685** 0.0677*- 0.18516 (0.0184) (0.0182) (0.0303) Black Market Premium 0.46835** 0.1086** 0.15624** (inverse) (0.0184) (0.0181) (0.0303) Govemance Index (scaled to 0-1) 0.68742** 0.1269** 0.09721 (0.0184) (0.0181) (0.0303) Gastil Civil Liberties Index (scaledtoO-1) 0.62396** 0.0658** 0.10715** (0.0184) (0.0181) (0.0303) Public Consumption as % of GDP (negative oo -0.10911 ** 0.0434** 0.05511 (0.0184) (0.0182) (0.0303) Composite Index 1 0.52432** 0.47123 (0.0184) (0.0303) Stock Retums 0.47123** 0.2701 ** 1 (0.0303) (0.0305) (0.0303) Alternative Composite Index (with 0.93624** 0.19169** 0.15902 equal weights to all indicators) (0.0184) (0.0184) (0.03186) 35 Table A3 Index of Risk Correlation Analysis Correlation between the indicator and: Composite Std Dev. Growth Std Dev. Stock Indicator Index per capita Returns CPI Inflation Rate 0.21073** 0.10431 ** 0.37621 ** (0.0202) (0.0191) (0.0308) Standard Deviation (S.D.) 0.65427** 0.61391 ** 0.32303** of the inflation rate (0.0202) (0.0190) (0.0308) S.D. of the Growth in 0.97324** 1 0.26841 ** GDP per capita (0.0202) (0.0311) S.D. of the Real Exchange Rate 0.56850** 0.47326 ** 0.40324 ** Changes (0.0202) (0.0193) (0.0311) S.D. of the Terms of Trade 0.33383** 0.20854** 0.13484** Changes (0.0202) (0.0193) (0.0311) S.D. of the Degree of Openness -0.00249 -0.05350* 0.01026 (0.0202) (0.0191) (0.0308) Govemance Index (negative of) 0.26520** 0.07529** 0.30647** (0.0202) (0.0179) (0.0308) Gastil Civil Liberties Index 0.20292** 0.01404 0.29032** (negative of) - (0.0202) (0.0179) (0.0308) Financial Depth (negative of) 0.27788** 0.08019 0.18034* (0.0202) (0.0181) (0.0308) Debt to Equity Ratio 0.20697 0.12934 0.15273 ** (0.0202) (0.0199) (0.0317) Composite Index 1 0.97324 0.45339** (0.0202) (0.0317) Std Dev Stock Returns 0.45339** 0.26841 ** 1 (0.0317) (0.0311) Ajternative Composite Index with 0.77837* 0.61327 0.45339^ equai weights to all indicators (0.0202) (0.0202) (0.0317) 36 Table A4 Index of Comovements Correlation Analysis Comovement Indicator Correlation between the Indicator and: derived from: Composite Index GDP Growth per capita Stock Retums Composite Index 1 0.7974** 0.0730** (0.019) (0.032) GDP Growth per capita 0.7974** 1 0.0979** (0.019) (0.032) Stock Retums 0.07303** 0.0979** 1 (0.032) (0.032) 37 Table A5 Composite Index of Retums Descriptive Statistics Period 1966-97 1966-79 1980-89 1990-97 1. All Countries Mean 0.0597 0.1086 -0.0560 0.1363 Median 0.1209 0.1550 0.0231 0.1658 Standard Deviation 0.5250 0.5127 0.5653 0.4562 No. Observations 1603 684 540 379 2. High and Upper Middle Income Countries Mean 0.2049 0.2405 0.1022 0.2801 Median 0.2571 0.2927 0.2110 0.2809 Standard Deviation 0.4457 0.3891 0.5314 0.3862 No. Observations 886 378 290 218 3. Low and Lower Middle Income Countries Mean -0.1198 -0.0543 -0.2395 -0.0584 Median -0.0800 0.0183 -0.1935 -0.0273 Standard Deviation 0.5592 0.5944 0.5486 0.4724 No. Observations 717 306 250 161 4. Countries with Low Capital Restrictions Mean 0.2710 0.2972 0.1887 0.3339 Median 0.3032 0.3059 0.2647 0.3290 Standard Deviation 0.4150 0.3421 0.5111 0.3742 No. Observations 617 267 200 150 5. Countries with High Capital Restrictions Mean -0.0725 -0.0121 -0.2000 0.0069 Median -0.0184 0.0432 -0.1500 0.0358 Standard Deviation 0.5432 0.5651 0.5464 0.4594 No. Observations 986 417 340 229 38 Table A6 Growth in Real GDP Per Capita Descriptive Statistics Period 1966-97 1966-79 1980-89 1990-97 1. All Countries Mean 1.99% 3.03% 1.00% 1.42% Median 2.03% 3.00% 1.32% 1.41% Standard Deviation 2.92% 2.84% 2.86% 2.53% No. Observations 1728 756 540 432 2. High and Upper Middle Income Countries Mean 2.61% 3.79% 1.49% 1.94% Median 2.42% 3.46% 1.78% 1.74% Standard Deviation 2.78% 2.54% 2.79% 2.35% No. Observations 928 406 290 232 3. Low and Lower Middle Income Countries Mean 1.28% 2.16% 0.43% 0.81% Median 1.43% 2.21% 0.49% 0.91% Standard Deviation 2.92% 2.93% 2.83% 2.60% No. Observations 800 350 250 200 4. Countries with Low Capital Restrictions Mean 2.52% 3.62% 1.49% 1.87% Median 2.41% 3.37% 1.92% 1.62% Standard Deviation 2.63% 2.30% 2.77% 2.22% No. Observations 640 280 200 160 5. Countries with High Capital Restrictions Mean 1.69% 2.69% 0.71% 1.15% Median 1.74% 2.69% 0.87% 1.34% Standard Deviation 3.04% 3.06% 2.87% 2.66% No. Observations 1088 476 340 272 39 Table A7 Index of Stock Retums Descriptive Statistics Period 196697 1966-79 1980489 199047 1. All Countries Mean 0.0218 -0.0304 0.0719 . 0.0283 Median 0.0152 -0.0310 0.0686 0.0351 Standard Deviation 0.3236 0.2863 0.3528 0.3233 No. Observatfons 1031 370 344 317 2. High and Upper Mlddle Income Countries Mean 0.0312 -0.0226 0.0793 0.0448 Median 0.0211 -0.0268 0.0944 0.0589 Standard Deviation 0.3146 0.2932 0.3567 0.2782 No. Observations 798 298 268 232 3. Low and Lower Middle Income Countries Mean -0.0104 -0.0629 0.0461 -0.0168 Median -0.0376 -0.0531 -0.0027 -0.0453 Standard Deviation 0.3514 0.2552 0.3399 0.4214 No. Observations 233 72 76 85 4. Countries with Low Capital Restrictions Mean 0.0201 -0.0307 0.0781 0.0250 Median 0.0225 -0.0206 0.0975 0.0511 Standard Deviation 0.2484 0.1943 0.2708 0.2745 No. Observations 596 236 192 168 5. Countries with High Capital Restrictions Mean 0.0242 -0.0300 0.0642 0.0320 Median -0.0166 -0.0552 0.0163 0.0026 Standard Deviation 0.4049 0.4009 0.4358 0.3717 No. Observatons 435 134 152 149 40 Table A8 Composite Index of Risks Descriptive Statistics Period 1966-97 1966-79 1980489 1990-97 1. All Countries Mean -0.1048 -0.0755 -0.0569 -0.2258 Median -0.1976 -0.1684 -0.0898 -0.3455 Standard' Deviation 0.5595 0.6087 0.5309 0.4856 No. Observations 1603 684 540 379 2. High and Upper Middle Income Countries Mean -0.3063 -0.3307 -0.2522 -0.3362 Median -0.4491 -0.4045 -0.4444 -0.5291 Standard Deviation. 0.4824 0.4393 0.5321 0.4808 No. Observations 886 378 290 218 3. Low and Lower Middle Income Countries Mean 0.1443 0.2397 0.1697 -0.0763 Median 0.0342 0.0780 0.1048 -0.2305 Standard Deviation 0.5480 0.6410 0.4308 0.4523 No. Observations 717 306 250 161 4. Countries with Low Capital Restrictions Mean -0.4354 -0.4604 -0.3887 -0.4530 Median -0.5468 -0.4971 -0.5698 -0.6158 Standard Deviation 0.4081 0.3769 0.4453 0.4070 No. Observatons 617 267 200 150 5. Countries with High Capital Restrictions Mean 0.1021 0.1709 0.1384 -0.0770. Median 0.0033 0.0358 0.0640 -0.2295 Standard Deviation 0.5419 0.6013 0.4781 0.4760 No. Observations 986 417 340 229 41 Table A9 Standard Deviaton of the Growth in Real GDP per capita Descriptive Statistics Period 1966497 1966-79 1980489 1990407 1. All Countries Mean 3.75% 4.07% 3.82% 3.09% Median 3.01% 3.34% 3.16% 2.36% Standard Deviation 2.72% 3.09% 2.52% 2.11% No. Observations 1726 756 540 430 2. High and Upper Middle Income Countries Mean 3.12% 3.04% 3.30% 3.03% Median 2.50% 2.73% 2.41% 2.27% Standard Deviation 2.13% 1.68% 2.66% 2.07% No. Observations 926 406 290 230 3. Low and Lower Middle Income Countries Mean 4.47% 5.25% 4.42% 3.17% Median 3.76% 3.96% 4.15% 2.43% Standard Deviation 3.13% 3.84% 2.21% 2.15% No. Observations 800 350 250 200 4. Countries with Low Capital Restrictions Mean 2.76% 2.62% 2.92% 2.80% Median 2.26% 2.35% 2.33% 2.14% Standard Deviation 1.82% 1.40% 2.16% 2.00% No. Observations 639 280 200 159 5. Countries with High Capital Restrictions Mean 4.33% 4.92% 4.35% 3.27% Median 3.68% 3.88% 4.00% 2.52% Standard Deviation 2.99% 3.47% 2.57% 2.15% No. Observations 1087 476 340 271 42 Table A1O Standard Deviation of Real Stock Retums Descriptive Statistics Period 196647 1966-79 198049 1990-97 1. All Countries Mean 0.2523 0.1964 0.2740 0.2925 Median 0.1893 0.1541 0.2239 0.2168 Standard Deviation 0.2265 0.2343 0.2175 0.2147 No. Observations 1013 359 339 315 2. High.and Upper Middle Income Countries Mean 0.2448 0.1970 0.2750 0.2701 Median 0.1816 0.1541 0.2219 0.1857 Standard Deviation 0.2404 0.2537 0.2284 0.2280 No. Observations 787 290 265 232 3. Low and Lower Middle Income Countries Mean 0.2783 0.1939 0.2706 0.3552 Median 0.2354 0.1573 0.2335 0.3574 Standard Deviation 0.1671 0.1247 0.1742 0.1571 No. Observations 226 69 74 83 4. Countries with Low Capital Restrictions Mean 0.1922 0.1559 0.2080 0.2245 Median 0.1603 0.1398 0.1893 0.1724 Standard Deviation 0.1269 0.0896 0.1262 0.1567 No. Observatons 590 232 191 167 5. Countries with High Capital Restrictions Mean 0.3361 0.2704 0.3592 0.3693 Median 0.2371 0.1804 0.2690 0.3099 Standard Deviation 0.2974 0.3644 0.2742 0.2440 No. Observations 423 127 148 148 43 APPENDIX B: An illustration of the ARDL approach to long-run modelling As an example, consider the following simple bivariate model: y, =a+by,, +cX_, +v, (BJ) XI =.y + pX,1l +e1 EB2) where y is the decision variable and X is the forcing variable. Furthermore, assume that the residuals (or shocks) have the following distributional properties: (')iid (O,£T), (BV (6 JJ) The first point to note is that X does not depend on past values of y. If a more general process for X were allowed, the long-run relationship between the two variables would not be unique. That is, both variables would be endogenous and additional identification assumptions would be needed to discern between various long-run relationships.20 Since multiple long-run relationships are beyond the scope of this paper, we restrict the dynamic process for Xto be purely autoregressive. The second point to note is that the existence of a long-run relationship requires the process fory to be stable, which in this simple example entails that IbI