Information, institutions, and banking sector development in West Africa.

Demetriades, Panicos ; Fielding, David

I. INTRODUCTION

Banks and other financial intermediaries can perform an important developmental function, especially in Africa, where alternative sources of finance are limited or nonexistent. By providing firms with essential finance, they help them to take advantage of productive investment opportunities which may not otherwise materialize. By screening loan applicants, they can help to address adverse selection in the credit market and channel funds toward productive uses. By monitoring borrowers, they can contain moral hazard behavior, for example, excessively risky investment activity that could undermine a borrower's ability to repay a loan. Through long-term bank-borrower relationships, well-established banks can address both adverse selection and moral hazard. This not only helps banks to remain solvent but also ensures that bank finance is channeled toward productive and sound investments.

There is a large body of empirical evidence which suggests that the development of banking systems goes hand in hand with economic development (see, e.g., Levine 2004). Although the evidence on causality is mixed (see, e.g., Demetriades and Hussein 1996), there is broad consensus that well-functioning banking systems promote economic growth (Demetriades and Andrianova 2005). It is, therefore, a puzzle that so many countries remain financially underdeveloped. This is particularly true of Sub-Saharan Africa, which remains one of the most financially under-developed regions in the world. A recent study by the World Bank has shown that African banking systems lack depth compared to other regions in the world, but are also excessively liquid (Honohan and Beck 2007). According to the World Bank, banks themselves complain that there is a lack of creditworthy borrowers, while at the same time households and firms find finance to be a major constraint. The evidence presented by the World Bank also suggests that the least developed banking systems are also the most liquid, suggesting that excess liquidity is a common feature of financial under-development.

This article aims to shed light on these features of financial under-development in Africa, utilizing a panel data set comprising the banks operating in the West African Economic and Monetary Union (UEMOA) during 2000-2005. The UEMOA provides a uniform financial system across eight countries; the structure of this system has changed little in the last 15 years. Therefore, we can be sure that the variations in bank behavior we observe within the UEMOA are not because of variations in the nature of public financial institutions which the banks face. This makes feasible the identification of the institutional sources of the variations in bank behavior, which are not correlated with variations in the quality of public financial institutions.

Our data set includes balance sheet information on each bank in the UEMOA, including bank characteristics such as age and ownership type, profitability, and the number of urban and rural branches. We also utilize country-level data on loan defaults, which provides information on the average quality of borrowers; we use this as a proxy for the severity of information problems faced by banks in the credit market. We combine this information with macroeconomic data including institutional quality indices constructed by the World Bank. Our data set enables us to examine the extent to which informational and institutional factors, and interactions between different factors, can explain a bank's loans to assets ratio, which is an inverse measure of bank liquidity. Our data set is also used to examine the microeconomic and macroeconomic determinants of the total volume of assets of an individual bank, which is a good micro-level indicator of overall banking sector development.

Our results suggest that to a large extent financial under development, including excess liquidity and low banking sector development, can be attributed to severe informational problems. These problems are particularly acute for younger banks; older, more established banks are less affected. These results highlight the importance of information capital in both developing banking systems and reducing excess liquidity. Our results suggest that it is not so much the lack of creditworthy borrowers that is the obstacle for financial development, but the lack of a developed infrastructure that would enable new banks to screen and monitor borrowers. This result is consistent with evidence on the importance of credit registries in reducing credit constraints (Galindo and Miller 2001). Our results also indicate that banking sector development in Africa does indeed follow economic development, but that it is also particularly sensitive to political stability and the rule of law.

This article is structured as follows. Section II reviews the institutional setting within which commercial banks in the UEMOA operate, and provides the conceptual background for our analysis. Section III describes the data and modeling strategy. Section IV presents and discusses the empirical findings. Section V summarizes and concludes.

II. COMMERCIAL BANKING IN THE UEMOA

The UEMOA is a monetary union arising from the final phase of French colonialism in West Africa (1948-1962), and encompasses most of France's former colonies in the area. The current member states are Benin, Burkina Faso, Cote d'Ivoire, Guinea-Bissau, Mali, Niger, Senegal, and Togo. It forms part of the Franc Zone, the other main component of which is a second monetary union, the Economic and Monetary Community of Central Africa (CEMAC). Both monetary unions have a central bank issuing a currency that the French Treasury guarantees to exchange for Euros at a fixed rate. The two currencies are both called the CFA Franc, (1) but they are entirely separate. The arrangements that the two monetary unions have with the French Treasury are parallel but entirely independent of each other.

The enduring institutional link with the former colonial power gives the UEMOA countries an unusually high level of financial stability, compared to other African countries with similar levels of economic development. The institutional framework is defined by a constitutional accord dating from the period in which the colonies became fully independent (1960-1962), and preserving many of the features of the financial system of post-war French colonial Africa. The main features are as follows:

(i) Guaranteed convertibility. Article 1 of the accord stipulates that France will help UEMOA member states to ensure the free convertibility of their currency. In practice, this means that the French Treasury will exchange CFA Francs for Euros on demand. Lending by the BCEAO (the UEMOA central bank) to domestic governments and to the private sector is now limited by rules designed to prevent free-tiding on the French guarantee.

(ii) A fixed exchange rate. Up until 1994, Article 2 of the accord stipulated a fixed rate of 50 CFA Francs to one French Franc. The rate has been changed only once, to 100:1, in January 1994. The entry of France into the European Monetary Union means that the rate is now defined in terms of Euros, but the current Euro rate is equivalent to 100:1 against the French Franc.

(iii) Free transferability. Article 6 of the accord describes the "freedom of financial relations between France and members of the Union." This obligation on the part of the African states is not without qualification, and the practice of member states has not always been in harmony with the principle. International capital transfers are taxed, and occasionally (especially during the run-up to the devaluation in 1993) the transferability has been suspended. Nevertheless, there is usually a reasonable degree of capital mobility between the UEMOA and France.

(iv) Harmonization of rules governing currency exchange. Article 6 of the accord notes that the "uniform regulation of the external financial relations of member states ... will be maintained in harmony with that of the French Republic." These regulations cover such things as the remittance of salaries abroad (that is, outside the Franc Zone), foreign investment, and borrowing from abroad.

(v) A common regulatory framework. Regulation of the banking system is the responsibility of the UEMOA Banking Commission, which was created in 1990 with French technical support. The commission has oversight over the day-to-day activities of all banks and other financial institutions in the UEMOA, and has the power to intervene in the operations of individual banks when its rules are infringed. In the case of serious infractions, the commission can impose disciplinary sanctions of differing degrees of severity, ranging from a formal warning to the dismissal of senior bank officials and suspension of a bank's activities. Commission staff produce regular reports on the extent of compliance with UEMOA banking regulations; the loan default data used in this article are taken from statistics compiled by the Banking Commission.

The financial stability provided by these institutions means that commercial banks in the UEMOA are free from some of the uncertainties facing financial institutions in other parts of Africa; the same is true of depositors. However, other risks remain. Firstly, many banks face a serious adverse selection problem arising from a low average level of borrowers' creditworthiness. In our sample, the average rate of default on bank loans exceeds 10%, which is very high by international standards. In theory, this should depress the equilibrium volume of loans (Stiglitz and Weiss 1981, 1983), particularly in markets where credit bureaus are in their infancy like in most of Africa. Existing evidence indicates that the magnitude of the problem varies considerably across countries and over time (e.g., Fuentes and Maquieira 2001; Koopman, Lucas, and Klaassen 2005). In our own sample, the default rate sometimes dips below 5%, while it occasionally exceeds 30%. Secondly, corruption could make loans less profitable, if it means that banks are forced to ignore the commercial worth and riskiness of projects they finance for the political elite. Direct evidence of such corruption in Kenya is discussed by Bigsten and Moene (1996), and evidence for a link between the corruption of bank officials and the productivity of investments is discussed by Beck, Demirguc-Kunt, and Maksimovic (2005). Such corruption may reduce the loans-assets ratio, and may also depress asset and liability growth.

Moreover, the quality of contract enforcement and overall political stability in the country could affect the extent of moral hazard that banks face when making loans. Institutions promoting the rule of law are likely to enhance banks' ability to enforce loan contracts and may therefore increase a bank's willingness to lend and its ability to grow (Messick 1999), even at low levels of average borrower quality. These institutions could act as a deterrent to moral hazard behavior by borrowers, helping to limit the number as well as the cost of bad loans. Governments of some UEMOA countries have enacted legislation to facilitate the recovery of bad debts of individual banks (e.g., the Banque de l'Habitat du Mali); however, such support for banks is by no means universal.

These factors must be interpreted bearing in mind that many of the banks in our sample are very young. For 25% of our observations, the age of the bank is 7 years or less. For very young banks, raising deposits is likely to be easier than identifying creditworthy borrowers. Older banks are likely to have more information capital so that their ability to screen loan applicants is likely to be better than that of younger banks. The adverse selection problem is likely to be more acute for younger banks, at any given average quality of borrowers. Very young banks may therefore opt to channel most of their resources into building up their deposit base, while their liabilities might in the first instance be transformed into foreign assets or claims on government and other domestic financial institutions rather than into business loans. Therefore, we expect that very young banks will have a lower loans-assets ratio than older, more established banks, ceteris paribus. We might also expect younger banks to exhibit more sensitivity to borrowers' propensity to default than older banks: a higher national default rate imposes more of a cost for younger banks who find it more difficult to screen customers.

Age is not the only factor that might affect banks' sensitivity to the propensity to default. Banks owned (or partly owned) by the government might have access to better ways to screen potential customers, as might foreign-owned banks. Banks that are operating intensively in provincial areas outside the financial capital of the country, where infrastructure of all kinds is likely to be weaker, may find customers more difficult to screen effectively. It is also possible that some of the idiosyncratic variation in screening efficiency is correlated with observable bank characteristics, such as profitability.

Higher levels of risk are one explanation for a relatively low ratio of loans to assets in Africa, and risk represents one channel through which corruption, rule of law, and political stability could affect banking performance. Of course, it is not the only channel. For example, Barth, Caprio, and Levine (2004) indicate that there is a positive association between a high level of government corruption and the existence of excessively strong supervisory agencies, severe restrictions on bank activities and barriers to entry that limit banking competition. However, all of these effects reinforce the mechanisms we have already described, either by reducing the profitability of loans or by creating a monopolistic incentive for banks to limit the quantity of loans in order to increase profits.

All of these factors are relevant to most African countries. However, in most African countries they are correlated with financial or monetary stability, and are therefore difficult to identify precisely. This is less of a problem in our sample as we restrict our attention to banks in the member states of the UEMOA in the period 2000-2005, where the quality of the financial system is uniform over time: there has been no major revision of UEMOA legislation in this period. It is also uniform across countries: there is a single authority--the Banking Commission--responsible for regulating all banks in the monetary union. We can therefore be confident that the effects we identify are not because of variations in financial or monetary stability but due to variations in the quality of governance.

III. DATA AND METHODOLOGY

A. Data

The loans and assets data used in our econometric model are taken from the annual BCEAO publication Bilans des Banques et Etablissements Financiers. (2) These data are used to construct two dependent variables for bank i in year t: the loans-assets ratio ([RATIO.sub.it]) and the logarithm of real assets ([ASSETS.sub.it]). Annual data are available for 113 banks in the UEMOA over the period 2000-2005: 15 in Benin, 14 in Burkina Faso, 27 in Cote d'Ivoire, 2 in Guinea-Bissau, 16 in Mali, 11 in Niger, 17 in Senegal, and 11 in Togo. This is not a balanced panel, because some banks came into existence during the sample period; with lags and differencing, 87 banks remain in the sample. [RATIO.sub.it] is constructed as the ratio of commercial loans ("creances sur la clientele") to total assets ("total de l'actif"). ln([ASSETS.sub.it]) is constructed as the log of total assets deflated by the consumer price index reported in the BCEAO Annuaire Statistique.

The econometric model also incorporates a number of explanatory variables, as follows. The countrywide default rate facing a bank in country j in year t ([DEFAULT.sub.jt]) is the ratio of the total bad debt of all commercial banks in the country to the total commercial lending of those banks. The figures for bad debt ("credits en souffrance") are taken from the UEMOA Banking Commission's Rapport Annuel. Data on the fraction of bank capital owned by the government ([GOVERNMENT.sub.it]) and foreigners ([FOREIGN.sub.it]), and on the number of years each bank has been in operation by year t ([AGE.sub.it]), are taken from the BCEAO publication Annuaire des Banques et Etablissements Financiers de l'UEMOA, as are data on the number of branches outside the financial capital ([PROVINCIAL.sub.it]). Data on bank profitability ([PROFITABILITY.sub.it]), measured as the ratio of profits to turnover, are taken from the Bilans des Banques et Etablissements Financiers. (3)

Data on the log of total real GDP in the country in which a bank is operating ([GDP.sub.jt]) are taken from the Annuaire Statistique; this is likely to be a correlate of the total asset volume of the banks of the country, because higher income will induce higher asset demand.

In order to capture the effects of variation in country-specific institutions that may impact on contract enforcement relevant for lending, we make use of the indicators reported in the World Bank World Governance Indicators. These indicators are described and discussed by Kaufmann, Kraay, and Mastruzzi (2007). Our measure of the extent to which a country is corruption-free is the "control of corruption" index in World Governance Indicators. There are several different governance indicators that may be associated with ease of contract enforcement: "rule of law," "voice and accountability," "political stability," "government effectiveness," and "regulatory quality." These indicators are quite highly correlated with each other, so it does not make sense to include them all in a single regression equation. However, there are no strong a priori grounds for supposing that one particular indicator is an especially good measure of the extent to which banks are protected from moral hazard effects. The methodology section that follows explains how we deal with the multicollinearity of the governance indicators.

Descriptive statistics for the variables in our model are presented in Table 1, while Figures 1-3 depict some of our key variables. Note that the governance variables are normalized, so that the mean of each is equal to zero across a worldwide sample. Negative means in our sample indicate that the UEMOA countries perform below the worldwide average in terms of governance, despite their financial stability.

[FIGURE 1 OMITTED]

[FIGURE 2 OMITTED]

On average, the ratio of loans to assets is 56.7%, which is low by international standards, and the ratio of defaults to total loans is 14.3%, which is high by international standards. The standard deviations around these two means are quite high, providing useful variation in the data. Correspondingly, the range of both variables is substantial: 0.00-0.96 for the loans to assets ratio and 0.05-0.42 for the default rate. The variation in the two dependent variables is shown in the histograms in Figures 1-2. It can be seen from Figure 1 that a majority of banks lend between 40% and 70% of their assets. However, there is also a substantial fraction lending over 80%, and some lending less than 20%. Figure 2 shows a similarly wide dispersion in asset levels.

There is substantial variation in the bank ownership variables: some banks are wholly government or foreign owned, while others are owned by the domestic private sector. Somewhat surprisingly, there is also substantial variability in the governance indicators over time, as shown in Figure 3. Annual changes in individual governance variables are often a large fraction of one unit (the worldwide variance in each of the variables). Political stability is the most variable governance indicator, but annual changes in the others are not always trivially small.

[FIGURE 3 OMITTED]

B. Methodology

The discussion in Section II suggests that banks' willingness to lend depends on aggregate credit market conditions, particularly borrower creditworthiness and the quality of contract enforcement, and on individual bank characteristics that capture a bank's informational capital, such as bank age and the location of its branches. We conjecture that the loans-assets ratio (RATIO) is decreasing in the loan default rate in a country (DEFAULT), and increasing in the quality of governance (as captured by the governance indicators), bank profitability (PROFITABILITY), and bank age (AGE). Age and other bank characteristics (GOVERNMENT; FOREIGN; PROVINCIAL) may also affect the impact of borrower creditworthiness, so various interaction terms in DEFAULT are included in our RATIO regression equation. Because we are using panel data, we also allow for both fixed and time effects as well as persistence in the dependent variable. Given that the governance indicators are quite highly correlated with each other (see Table 1), we avoid fitting a model with more than one such indicator. Because we have no strong a priori view on which of these indicators best captures the contract enforcement effect, we report results with all six indicators entered one at a time in the model. Thus, our model for the loan to assets ratio is as follows:

(1) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

Here, i [member of] j indicates the ith bank in the jth country, and t indicates the year. The [alpha] and [beta] parameters capture time and bank fixed effects, and u is a regression residual. (4) GOVERNANCE is measured by one of the six governance indicators. Note that GOVERNMENT, FOREIGN, and PROVINCIAL appear only in interaction terms, not as linearly separable effects. This is because in our sample they do not exhibit any substantial variation over time, and so are collinear with the bank fixed effects. (5)

Our second model is designed to explain variations in the logarithm of real assets. Our modeling strategy is similar to the one above, but the assets model contains one additional effect: we control for the size of the economy in which a bank is operating, as measured by ln(GDP). However, interaction terms in GOVERNMENT, FOREIGN, and PROVINCIAL are never statistically significant in the assets regressions, and are excluded from the models reported below; the same is true of PROFITABILITY. Our assets regressions take the following form:

(2) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

In estimating the parameters in a dynamic panel model such as Equation (1) or (2), we need to allow for the fact that the lagged dependent variable--[RATIO.sub.it - 1] or ln[(ASSETS).sub.it - 1]--will be correlated with the error term [u.sub.it]. Moreover, [DEFAULT.sub.jt] and [PROFITABILITY.sub.it] may also be endogenous and correlated with [u.sub.it]. OLS estimates of the parameters will therefore be biased. Arellano and Bond (1991) propose a GMM estimator that takes into account the endogeneity of the lagged dependent variable. Firstly, we can take differences of Equation (1) to account for country fixed effects:

(3) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

The same can be done with Equation (2). For any bank i and any year t, [RATIO.sub.it - 2] will be a valid instrument for [DELTA][RATIO.sub.it - 1], because as long as [u.sub.it] is serially uncorrelated, [RATIO.sub.it - 2] will be orthogonal to [DELTA][u.sub.it]. In fact, all the other lags back to [RATIO.sub.i2000] will be orthogonal to [DELTA][u.sub.it], which gives us a whole set of moment conditions with which to identify the parameter [rho] : [RATIO.sub.i2000] is orthogonal to [DELTA][u.sub.i2002], [RATIO.sub.i2000] and [RATIO.sub.i2001] are orthogonal to [DELTA][u.sub.i2003], and so on. (6) Following Blundell and Bond (1998), we can take the GMM approach a step further: because [RATIO.sub.it] is stationary, we can treat Equations (1) and (3) as a system and make use of the orthogonality of [u.sub.it] to [DELTA][RATIO.sub.it - 2], and to other lags back to [DELTA][RATIO.sub.i2000], to help identify [rho].

We still need to identify the [delta] and [eta] parameters in our model, bearing in mind the potential endogeneity of [PROFITABILITY.sub.it] and [DEFAULT.sub.jt]. In the results reported below, identification is achieved by using moment conditions analogous to those used to identify [rho], rather than by fitting separate equations for [PROFITABILITY.sub.it] and [DEFAULT.sub.jt] using some set of instrumental variables. We impose the restriction that lags of [PROFITABILITY.sub.it] and [DEFAULT.sub.jt] (including the interaction terms) are orthogonal to [DELTA][u.sub.it], and that lags of [DELTA][PROFITABILITY.sub.it] and [DELTA][DEFAULT.sub.jt] are orthogonal to [u.sub.it]. In the Appendix, we explore the consequences of identifying the parameters by using standard instrumental variables. This alternative strategy produces results that are qualitatively very similar to the ones that we now discuss.

IV. EMPIRICAL RESULTS

Tables 2 and 3 contain the main empirical results of this article: Table 2 reports the results of fitting Equation (1) to our data and Table 3 reports the results of fitting Equation (2).

The diagnostic statistics in Table 2 provide evidence of the appropriateness of the estimator and the validity of the instruments used. The Sargan test does not reject the over-identification restrictions. The residual autocorrelation tests reject the null of no first order serial correlation but do not reject the null of no second order serial correlation, confirming that the residual series prior to differencing are white noise processes.

In Table 2, the lagged dependent variable is positive and highly significant, suggesting considerable persistence and underlining the appropriateness of a dynamic panel model. The coefficients on the lagged dependent variable (in the range 0.7-0.75) indicate that the half-life of a temporary shock to the loans to assets ratio is about 2 years. When the following paragraphs mention coefficient magnitudes, these indicate the immediate impact of each explanatory variable. With a lagged dependent variable coefficient of 0.75, the eventual effect of a permanent change in an explanatory variable would be four times higher.

There is strong evidence that default rates represent a major obstacle to bank lending in the UEMOA. Estimates of the [[eta].sub.0] coefficient in Equation (1) are negative and significant in all six versions of the model (one for each governance indicator). This coefficient indicates the effect of the default rate on the loans to assets ratio of a privately owned bank in its first year of operation with branches only in the financial capital of its country. The coefficient is very large: estimates of [[eta].sub.0] range from -1.51 to -1.22. In our sample, the standard deviation of DEFAULT is 8.5 percentage points, so a two standard deviation increase in this variable entails a 20-25 percentage point reduction in the loans to assets ratio. Over the sample period, average default rates range from 8.3 percentage points in Burkina to 34.6 percentage points in Togo; using our estimates of [[eta].sub.0], this spread entails a 30-40 percentage point difference in the average loans to asset ratio. In fact, the difference between Burkina's average ratio and Togo's is only 13 percentage points, because the [[eta].sub.0] coefficient does not represent the effect of DEFAULT on the average bank. The various interaction terms in DEFAULT indicate that the effects of loan defaults (i) diminish with bank age (ii) increase with the number of provincial branches, and (iii) are much smaller for government-owned or foreign-owned banks.

Because so many of the interaction terms in Table 2 are statistically significant, individual default coefficients are not in themselves very meaningful. However, computation of different linear combinations of the interaction terms (and of corresponding standard errors using the Delta Method) permits the following observations. Firstly, consider a private bank with no provincial branches. If the bank is completely new, then the DEFAULT effect is given by the coefficient [[eta].sub.0]. As the bank ages the DEFAULT effect diminishes, and for banks older than 35-40 years the effect is insignificantly different from zero. Now consider a bank with 35 provincial branches (the maximum observed in the sample). For a completely new bank, the DEFAULT effect is approximately twice as large as for a bank with no provincial branches, and although age diminishes the effect, it is still statistically significant at age 50. At this age, the DEFAULT effect is roughly the same as for a completely new bank with no provincial branches, that is, [[eta].sub.0].

Now consider a bank that is partly government or foreign owned. Because the GOVERNMENT and FOREIGN interaction terms have positive coefficients, the DEFAULT effect for a young bank with some government or foreign ownership is smaller, although still significantly negative. However, the effect is positive for very old banks with a very high share of government or foreign ownership. For a bank owned completely by the government, or for a bank with at least a 70% foreign ownership share, the DEFAULT effect is significantly greater than zero at age 50. In other words, there may be some banks which increase their loans to assets ratio in the presence of high default rates, although most do not. A riskier environment deters most banks from lending a large share of their assets to domestic customers, but there are a few banks--perhaps the ones with most informational capital--which partially fill the resulting vacuum in the market.

As anticipated, all of the governance indicators have a positive and highly significant effect. The coefficients range from 0.04 (political stability) to 0.09 (government effectiveness). That is, a unit increase in the indicator is associated with an increase in the loans to assets ratio of between 4 and 9 percentage points. One unit corresponds to one standard deviation in the worldwide sample, although the standard deviations in our sample are a little smaller. At the mean sample value of the default rate (14 percentage points), a one standard deviation increase in the governance indicator has roughly the same effect on the loans to assets ratio as transferring the bank from private control to government or foreign control.

The impact of age on the loans to assets ratio depends on the value of DEFAULT. At the minimum sample default rate (5%), the derivative of the ratio with respect to age is insignificantly different from zero; at the maximum sample value of DEFAULT (42%), the derivative is slightly below 0.01, and significantly greater than zero. Age matters only when default rates are high. At the highest default rates, one extra year of bank life is associated with a loans to assets ratio that is just under 1 percentage point higher.

When we use government effectiveness or control of corruption to measure governance, the profitability coefficient is significantly greater than zero at the 10% level, and using regulatory quality instead increases the significance level to 5%. Otherwise, the effect of profitability is insignificantly different from zero, so we have no robust evidence that more profitable banks tend to lend more to domestic customers, ceteris paribus. (7)

Table 3 presents results for the models of the total real volume of assets. In this case, the Sargan test rejects the overidentification restrictions at the 5% level in one model (the one using regulatory quality); it also rejects the restrictions at the 10% level in three other models. Relaxing some of the orthogonality conditions used to identify the model reduces these significance levels, but also reduces the precision of our estimates somewhat. The most reliable results are the ones using political stability or rule of law to measure governance, for which the Sargan test statistics are insignificant at the 10% level. The residual autocorrelation tests are uniformly satisfactory, indicating first-order but not second-order serial correlation in all six cases. Again, the lagged dependent variable is positive and highly significant, and takes a value of just less than 0.9 (implying a half-life of around 5 years). Therefore, if there were any permanent change in an explanatory variable, the eventual impact on asset volumes would be an order of magnitude greater than the immediate effect indicated below.

Table 3 provides additional evidence that high loan defaults are a major obstacle to financial development in the region. It shows that a higher default rate is associated with a significantly lower level of total assets. This means that the impact of default on the total volume of loans is even larger than what is suggested by Table 2. The coefficient on the default rate in Table 3 is negative and highly significant, ranging from -1.25 to -1.98. Once again, the negative effect of loan defaults is mitigated by bank age. The model that uses political stability to measure governance implies that although a 1 percentage point increase in the default rate reduces the loan volume of a very young bank by nearly 2%, it increases the loan volume of a 50-year old bank by over 1%. Both effects are statistically significant. The tipping point is at about 35 years of age. Again, some of the reduction in the asset volumes of younger banks is offset by the expansion of older banks, which raise not only their loans to assets ratio, but also their total asset base.

All the governance indicators have a positive and significant effect on total asset volumes. The coefficient itself ranges from 0.10 (regulatory quality) to 0.22 (rule of law).

There is only one model in which GDP per capita is statistically significant, namely the model with political stability, in which case the coefficient is positive, as anticipated; the elasticity is just under 0.08.

V. CONCLUDING REMARKS

Our results suggest that a major factor explaining why most banks in Africa choose to remain excessively liquid is a high default rate among borrowers. The same factor appears to be a serious obstacle to the growth of bank balance sheets. Our results also suggest that older government-owned and foreign banks suffer less from this problem; this is consistent with an information capital story in which banks without sufficient information capital are unwilling to lend and unable to grow their assets. Young, privately owned banks suffer the most. It is therefore unrealistic, in current circumstances, to expect much financial development to come from the emergence and growth of new banks. Such banks will have little or no information capital, unless there is more effort to establish credit bureaus and other mechanisms that improve information on prospective borrowers (IMF 2001; Sacerdoti 2005). Given that high default rates and limited public information about borrowers serve to bolster the market share of the oldest government-owned and foreign banks, the political will to engage in such reform may well be lacking.

Our findings also suggest that good governance, however measured, has a uniformly positive on both banks' willingness to lend and their ability to grow their balance sheets. While all aspects of governance are important, government effectiveness appears to have the largest economic impact on the loans to assets ratio, closely followed by control of corruption. Rule of Law, on the other hand, appears to have by far the largest impact on the volume of bank business.

Our results relate to a region of Africa across which there is a high degree of homogeneity in financial and monetary systems. This makes it relatively straightforward to identify the impact of variations in governance and default rates on bank behavior. Future research might examine the extent to which these results are more widely applicable in the rest of Africa, and also in other developing regions where default rates are high. However, such research will need to deal with the challenge of identifying the effects of governance and credit risk when there is also substantial variation in financial and monetary systems.

ABBREVIATIONS

CEMAC: Economic and Monetary Community of Central Africa

UEMOA: West African Economic and Monetary Union

APPENDIX

In the model presented in the main text, DEFAULT and PROFITABILITY are potentially endogenous regressors; their effect on RATIO, as presented in Table 2, is identified using moment conditions of the kind outlined by Blundell and Bond (1998). In this appendix, we present some alternative results in which the set of identifying restrictions is supplemented by modeling DEFAULT and PROFITABILITY as a function of two indicators of macroeconomic conditions in each country. These indicators affect the default rate and profitability, but are unlikely to have any direct impact on the loans to assets ratio.

The two indicators are a terms of trade index, taken from the World Bank World Development Indicators, and the consumer price inflation rate. The terms of trade index ([TOT.sub.jt]) is measured as the logarithm of the ratio of export prices to import prices. It represents a source of exogenous shocks to domestic income: improvements in the terms of trade may stimulate domestic demand, increasing bank profitability and making defaults less likely. The inflation rate ([INF.sub.jt]) is measured using the consumer price index discussed in the main text. Higher inflation reflects a more unstable macroeconomic environment in which profitability may be lower and default more likely. Using these instruments, we first fit a regression equation for [DEFAULT.sub.jt]:

(A1) [DEFAULT.sub.jt] = [[beta].sub.j] + [rho] x [DEFAULT.sub.jt - 1] + [theta] x [TOT.sub.jt] + [xi] x [INF.sub.jt] + [u.sub.jt]

Here again, the lagged dependent variable is correlated with the error, so the parameters in Equation (A1) are estimated using the Blundell-Bond method. The parameter estimates are: [rho] = -0.499 (significant at 10%), [theta] = -0.174 (significant at 1%), and [xi] = 0.886 (significant at 1%). Default rates do rise significantly when the terms of trade deteriorate or when inflation rises. We then fit a regression equation for [PROFITABILITY.sub.it]:

(A2) [PROFITABILITY.sub.it] = [[beta].sub.j] + [rho] x [PROFITABILITY.sub.it - 1] + [theta] x [TOT.sub.jt] + [xi] x [INF.sub.jt] + [u.sub.jt]

In this case, the parameter estimates are [rho] = 0.072, [theta] = -0.014, and [xi] = -0.478. The estimate of [theta] is insignificantly different from zero, but [xi] is significant at the 5% level. Profitability as we measure it does not appear to be affected by the terms of trade, but higher profits are associated with lower inflation.

Interestingly, none of the governance indicators is statistically significant when added to Equations (A1) and (A2): the quality of governance has no impact on bank profitability or the propensity to default on loans. Similarly, when we include as a regressor the level of bank assets, we do not produce a statistically significant coefficient: default rates and profitability do not depend on bank size. (This result is the same whether we use moment restrictions on lags of the asset level or the exogenous variables in Table 3 to instrument the current level of bank assets.) Nevertheless, the default rate does respond significantly to macroeconomic conditions, as captured by TOT and INF, and profitability does respond significantly to INF, so these variables will be strong instruments in a regression of RATIO on DEFAULT and PROFITABILITY.

Table A1 therefore presents the results of estimates of the parameters in Equation (1)--the equation for RATIO--using TOT and INF as instruments for DEFAULT and PROFITABILITY. In these estimates, the effects of DEFAULT and PROFITABILITY are identified by both the moment conditions outlined in the main text and the exclusion of TOT and INF from the RATIO equation. Using the Sargan Tests in the table, we cannot reject the null that the exclusion restrictions are valid. Comparison of the parameter estimates in Table A1 with those in Table 2 of the main text shows very little difference between the two sets of results, either in terms of the size of the estimated coefficients or in terms of their level of significance.

TABLE A1
Dynamic Panel Estimation of the Loans-Assets Ratio (RATIO)
of 87 West African Banksa with Supplementary Instruments

 GOVERNANCE INDICATOR

 VOICE & POLITICAL
 ACCOUNTABILITY STABILITY

[RATIO.sub.-1] 0.7114 *** 0.7054 ***
 0.0536 0.0448
AGE -0.0009 ** -0.0006
 0.0004 0.0004
GOVERNANCE 0.0532 *** 0.0276 ***
 0.0115 0.0079
PROFITABILITY 0.0229 0.0531
 0.0407 0.0392
DEFAULT -1.3207 *** -1.1740 ***
 0.1409 0.1318
DEFAULT x AGE 0.0241 *** 0.0220 ***
 0.0039 0.0036
DEFAULT x GOVERNMENT 0.6367 *** 0.4280 ***
 0.1741 0.1654
DEFAULT x FOREIGN 0.5877 *** 0.4270 ***
 0.1596 0.1430
DEFAULT x PROVINCIAL -0.0306 *** -0.0300 ***
 0.0066 0.0067
Number of observations 304 304
Sargan test p-value 0.4108 0.4567
Residual AR(1) test p-value 0.0045 0.0029
Residual AR(2) test p-value 0.8668 0.8102

 GOVERNANCE INDICATOR

 GOVERNMENT REGULATORY
 EFFECTIVENESS QUALITY

[RATIO.sub.-1] 0.7238 *** 0.7175 ***
 0.0510 0.0487
AGE -0.0003 -0.0008 ***
 0.0003 0.0003
GOVERNANCE 0.0839 *** 0.0783 **
 0.0177 0.0399
PROFITABILITY 0.0702 * 0.0923 **
 0.0364 0.0371
DEFAULT -1.0164 *** -1.0717 ***
 0.1274 0.1344
DEFAULT x AGE 0.0180 *** 0.0214 ***
 0.0036 0.0038
DEFAULT x GOVERNMENT 0.5692 *** 0.4551 ***
 0.1548 0.1650
DEFAULT x FOREIGN 0.5431 *** 0.3601 **
 0.1432 0.1424
DEFAULT x PROVINCIAL -0.0291 *** -0.0326 ***
 0.0060 0.0059
Number of observations 304 304
Sargan test p-value 0.4423 0.3321
Residual AR(1) test p-value 0.0024 0.0045
Residual AR(2) test p-value 0.7166 0.5163

 GOVERNANCE INDICATOR

 CONTROL OF
 RULE OF LAW CORRUPTION

[RATIO.sub.-1] 0.7013 *** 0.7198 ***
 0.0483 0.0514
AGE -0.0004 -0.0006
 0.0004 0.0004
GOVERNANCE 0.0430 *** 0.0619 ***
 0.0138 0.0164
PROFITABILITY 0.0689 * 0.0782 **
 0.0382 0.0383
DEFAULT -1.0242 *** -1.0877 ***
 0.1264 0.1242
DEFAULT x AGE 0.0170 *** 0.0213 ***
 0.0038 0.0037
DEFAULT x GOVERNMENT 0.4879 *** 0.4088 ***
 0.1583 0.1336
DEFAULT x FOREIGN 0.4542 *** 0.3273 **
 0.1482 0.1483
DEFAULT x PROVINCIAL -0.0291 *** -0.0372 ***
 0.0065 0.0054
Number of observations 304 304
Sargan test p-value 0.3576 0.4422
Residual AR(1) test p-value 0.0033 0.0027
Residual AR(2) test p-value 0.7113 0.6796

(a) Standard errors are in italics. Estimates are
obtained using the xtdpd command in Stata 10.0.
The regression also includes time fixed effects.

* Significant at 10%; **significant at 5%; ***significant at 1%.

REFERENCES

Arellano, M., and S. Bond. "Some Tests of Specification for Panel Data: Monte Carlo Evidence and an Application to Employment Equations." Review of Economic Studies, 58, 1991, 277-97.

Barth, J. R., G. Caprio, and R. Levine. "The Regulation and Supervision of Banks around the World: A New Database," in Integrating Emerging Market Countries into the Global Financial System, Brookings-Wharton Papers on Financial Services, edited by R. E. Litan, and R. Herring. Washington, DC: Brookings Institution Press, 2001.

--. "Bank Regulation and Supervision: What Works Best?" Journal of Financial Intermediation, 13, 2004, 205-48.

Beck, T., A. Demirguc-Kunt, and V. Maksimovic. "Financial and Legal Constraints to Growth: Does Firm Size Matter?" Journal of Finance, 40, 2005, 137-77.

Bigsten, A., and K. O. Moene. "Growth and Rent Dissipation: The Case of Kenya." Journal of African Economies, 5, 1996, 177-98.

Blundell, R., and S. Bond. "Initial Conditions and Moment Restrictions in Dynamic Panel-Data Models." Journal of Econometrics, 87, 1998, 115-43.

Demetriades, P., and S. Andrianova. "Sources and Effectiveness of Financial Development: What We Know and What We Need to Know." Research Paper RP2005/76, World Institute for Development Economic Research, Helsinki, 2005.

Demetriades, P., and K. Hussein. "Does Financial Development Cause Economic Growth? Time Series Evidence from 16 Countries." Journal of Development Economics, 51, 1996, 387-411.

Fuentes, R., and C. P. Maquieira. "Why Borrowers Repay: Understanding High Performance in Chile's Financial Market," in Defusing Default: Incentives and Institutions, edited by M. Pagano. Washington, DC: Inter-American Development Bank/Johns Hopkins University Press, 2001.

Galindo, A., and M. Miller. "Can Credit Registries Reduce Credit Constraints? Empirical Evidence on the Role of Credit Registries in Firm Investment Decisions." Presentation to Towards Competitiveness: The Institutional Path, Annual Meeting of the Board of Governors, Inter-American Development Bank and Inter-American Investment Corporation, Santiago, Chile, March 16, 2001. Accessed January 31, 2010. http://socsci2.ucsd.edu/~aronatas/project/academic/ Credit%20Registries.pdf.

Honohan, P., and T. Beck. Making Finance Work for Africa, Washington, DC: The World Bank, 2007.

International Monetary Fund. Senegal: Financial System Stability Assessment. Country Report 01/189, Washington, DC, 2001.

Kaufmann, D., A. Kraay, and M. Mastruzzi. "Governance Matters VI: Aggregate and Individual Governance Indicators 1996-2006." World Bank Policy Research Working Paper 4280, 2007.

Koopman, S. J., A. Lucas, and P. Klaassen. "Empirical Credit Cycles and Capital Buffer Formation." Journal of Banking and Finance, 29, 2005, 3159-79.

Levine, R. "Finance and Growth: Theory and Evidence." NBER Working Paper 10766, Cambridge, MA, 2004.

Messick, R. "Judicial Reform and Economic Development: A Survey of the Issues." World Bank Research Observer, 14, 1999, 117-36.

Sacerdoti, E. Access to Bank Credit in Sub-Saharan Africa: Key Issues and Reform Strategies. Washington, DC: International Monetary Fund, Working Paper WP/05/166, 2005.

Stiglitz, J., and A. Weiss. "Credit Rationing in Markets with Imperfect Information." American Economic Review, 71, 1981, 393-410.

--. "Incentive Effects of Terminations: Applications to Credit and Labor Markets." American Economic Review, 73, 1983, 912-27.

(1.) CFA originally stood for Colonies Francaises en Afrique. It now stands for Communaute Financiere Africaine (for the UEMOA currency) and Cooperation Financiere en Afrique (for the CEMAC currency).

(2.) All publications mentioned in this section are available online at www.bceao.int. Other international studies of banking sector performance use data sources different from ours, for example the Doing Business indicators of the World Bank or the data set of Barth, Caprio, and Levine (2001). However, for our countries the Doing Business data are available in 2004 at the earliest, and in some cases not until 2008; the Barth, Caprio, and Levine (2001) data set does not contain any francophone developing countries.

(3.) It makes very little difference to the results if profits are measured as a fraction of total bank assets.

(4.) There will be some heterogeneity in the performance of banks that is difficult to measure or observe. Some banks lend almost exclusively to firms in a specific sector; for example, a number of agricultural banks in the Sahelian countries are highly exposed to the cotton-producing sector.

(5.) When an interaction term in DEFAULT and PROFITABILITY is included, the effect is statistically insignificant. Similarly, it was not possible to find any robust significant interaction effects in GOVERNANCE. We also tested the sensitivity of RATIO to bank size by including ln(ASSETS) in Equation (1) as an endogenous regressor; the effect of ln(ASSETS) is statistically insignificant.

(6.) After taking lags and differences, our sample comprises 4 years. If for every t we use all lags back to [RATIO.sub.i2000] as instruments for [DELTA][RATIO.sub.it - 1] then we have dozens of moment conditions, and there is a risk of generating spurious results as our degrees of freedom diminish. In the main results reported in the article, we use only moment conditions involving [RATIO.sub.it - 2]. However, increasing the number of conditions does not make an enormous difference to our results. For the model of ln(ASSETS), in which there are fewer regressors and more degrees of freedom, we use lags up to t - 3.

(7.) All of our other results are qualitatively similar if the profitability coefficient is set to zero.

PANICOS DEMETRIADES and DAVID FIELDING *

* We would like to thank Chris Haig for her outstanding assistance with this article.

Demetriades: Department of Economics, University of Leicester, University Road, Leicester LE1 7RH, U.K. Phone +44 116 252 2835, E-mail pd28@le.ac.uk

Fielding: Department of Economics, University of Otago, P.O. Box 56, Dunedin 9054, New Zealand. Phone +64 3479 8653, Fax +64 3479 8174, E-mail david.fielding@ otago.ac.nz

doi: 10.1111/j.1465-7295.2011.00376.x

TABLE 1
Descriptive Statistics
 Obs. Mean SD
(i) Univariate statistics
RATIO 482 0.567 0.199
In(ASSETS) 487 5.515 1.577
AGE 588 17.306 16.075
PROFITABILITY 392 0.076 0.103
DEFAULT 588 0.143 0.085
GOVERNMENT 582 0.159 0.250
FOREIGN 582 0.544 0.360
PROVINCIAL 522 3.736 6.355
VOICE & ACCOUNTABILITY 588 -0.447 0.654
POLITICAL STABILITY 588 -0.448 0.821
GOVERNMENT 588 -0.678 0.401
 EFFECTIVENESS
REGUALTORY QUALITY 588 -0.444 0.245
RULE OF LAW 588 -0.669 0.436
CONTROL OF CORRUPTION 588 -0.565 0.369

 VOICE & ACC- POLITICAL GOVERNMENT
 OUNTABILITY STABILITY EFFECTIVENESS

(ii) Weighted Correlation Coefficients for the Governance Variables
(with Weights for the Number of Banks in Each Country)

POLITICAL STABILITY 0.80
GOVERNMENT 0.81 0.56
 EFFECTIVENESS
REGUALTORY QUALITY 0.65 0.57 0.76
RULE OF LAW 0.88 0.83 0.84
CONTROL OF CORRUPTION 0.47 0.49 0.63

 Min. Max.
(i) Univariate statistics
RATIO 0.00 0.96
In(ASSETS) 1.43 8.37
AGE 0.00 106.00
PROFITABILITY 0.00 0.59
DEFAULT 0.05 0.42
GOVERNMENT 0.00 1.00
FOREIGN 0.00 1.00
PROVINCIAL 0.00 35.00
VOICE & ACCOUNTABILITY -1.54 0.41
POLITICAL STABILITY -2.45 0.71
GOVERNMENT -1.44 0.04
 EFFECTIVENESS
REGUALTORY QUALITY -1.00 -0.06
RULE OF LAW -1.57 -0.04
CONTROL OF CORRUPTION -1.24 0.12

 REGUALTORY RULE OF
 QUALITY LAW

(ii) Weighted Correlation Coefficients for the Governance Variables
(with Weights for the Number of Banks in Each Country)

POLITICAL STABILITY
GOVERNMENT
 EFFECTIVENESS
REGUALTORY QUALITY
RULE OF LAW 0.73
CONTROL OF CORRUPTION 0.79 0.68

TABLE 2
Dynamic Panel Estimation of the Loans-Assets Ratio (RATIO)
of 87 West African Banks (a)

 GOVERNANCE INDICATOR

 VOICE & POLITICAL
 ACCOUNTABILITY STABILITY

[RATIO.sub.-1] 0.7259 *** 0.7126 ***
 0.0576 0.0479
AGE -0.0014 *** -0.0011 **
 0.0004 0.0005
GOVERNANCE 0.0538 *** 0.0356 ***
 0.0117 0.0080
PROFITABILITY 0.0214 0.0463
 0.0436 0.0416
DEFAULT -1.5077 *** -1.4427 ***
 0.1587 0.1542
DEFAULT x AGE 0.0308 *** 0.0292 ***
 0.0049 0.0047
DEFAULT x GOVERNMENT 0.5920 *** 0.4611 ***
 0.1682 0.1601
DEFAULT x FOREIGN 0.5321 *** 0.4085 ***
 0.1667 0.1529
DEFAULT x PROVINCIAL -0.0340 *** -0.0345 ***
 0.0070 0.0070
Number of observations 304 304
Sargan test p-value 0.4002 0.5304
Residual AR(1) test p-value 0.0013 0.0008
Residual AR(2) test p-value 0.8755 0.8589

 GOVERNANCE INDICATOR

 GOVERNMENT REGULATORY
 EFFECTIVENESS QUALITY

[RATIO.sub.-1] 0.7457 *** 0.7428 ***
 0.0525 0.0493
AGE -0.0009 ** -0.0013 ***
 0.0004 0.0004
GOVERNANCE 0.0878 *** 0.0754 *
 0.0178 0.0397
PROFITABILITY 0.0679 * 0.0942 **
 0.0396 0.0393
DEFAULT -1.2174 *** -1.2249 ***
 0.1522 0.1467
DEFAULT x AGE 0.0252 *** 0.0270 ***
 0.0050 0.0046
DEFAULT x GOVERNMENT 0.5308 *** 0.4264 ***
 0.1455 0.1566
DEFAULT x FOREIGN 0.4524 *** 0.2810 *
 0.1521 0.1489
DEFAULT x PROVINCIAL -0.0328 *** -0.0362 ***
 0.0064 0.0060
Number of observations 304 304
Sargan test p-value 0.4568 0.3219
Residual AR(1) test p-value 0.0007 0.0010
Residual AR(2) test p-value 0.7296 0.5068

 GOVERNANCE INDICATOR

 RULE OF CONTROL OF
 LAW CORRUPTION

[RATIO.sub.-1] 0.7435 *** 0.7426 ***
 0.0508 0.0520
AGE -0.0011 ** -0.0013 ***
 0.0005 0.0005
GOVERNANCE 0.0493 *** 0.0600 ***
 0.0133 0.0166
PROFITABILITY 0.0600 0.0789 *
 0.0411 0.0409
DEFAULT -1.3437 *** -1.3040 ***
 0.1523 0.1505
DEFAULT x AGE 0.0272 *** 0.0292 ***
 0.0051 0.0051
DEFAULT x GOVERNMENT 0.4709 *** 0.3952 ***
 0.1457 0.1280
DEFAULT x FOREIGN 0.3647 ** 0.2224
 0.1563 0.1610
DEFAULT x PROVINCIAL -0.0357 *** -0.0427 ***
 0.0068 0.0059
Number of observations 304 304
Sargan test p-value 0.4263 0.5086
Residual AR(1) test p-value 0.0010 0.0006
Residual AR(2) test p-value 0.7471 0.7054

(a) Standard errors are in italics. Estimates are obtained using the
xtdpd command in Stata 10.0. The regression also includes time
fixed effects.

* Significant at 10%; ** significant at 5%; *** significant at 1%.

TABLE 3
Dynamic Panel Estimation of the Log of Real
Assets (ln(ASSETS)) of 87 West African Banks (a)

 GOVERNANCE INDICATOR

 VOICE & POLITICAL
 ACCOUNTABILITY STABILITY

ln(ASSETS.sub.-1) 0.8693 *** 0.8669 ***
 0.0313 0.0300
AGE -0.0105 *** -0.0103 ***
 0.0012 0.0011
GOVERNANCE 0.1038 *** 0.1374 ***
 0.0337 0.0199
ln(GDP) 0.0295 0.0774 **
 0.0305 0.0308
DEFAULT -1.6472 *** -1.9750 ***
 0.4055 0.3964
DEFAULT x AGE 0.0596 *** 0.0723 ***
 0.0129 0.0127
Number of observations 385 385
Sargan test p-value 0.0642 0.2947
Residual AR(1) test p-value 0.0350 0.0332
Residual AR(2) test p-value 0.8407 0.8438

 GOVERNANCE INDICATOR

 GOVERNMENT REGULATORY
 EFFECTIVENESS QUALITY

ln(ASSETS.sub.-1) 0.8773 *** 0.8835 ***
 0.0323 0.0306
AGE -0.0088 *** -0.0090 ***
 0.0010 0.0010
GOVERNANCE 0.0934 ** 0.1001 **
 0.0425 0.0472
ln(GDP) -0.0335 -0.0116
 0.0268 0.0257
DEFAULT -1.4257 *** -1.2531 ***
 0.3644 0.3814
DEFAULT x AGE 0.0431 *** 0.0418 ***
 0.0103 0.0105
Number of observations 385 385
Sargan test p-value 0.0728 0.0493
Residual AR(1) test p-value 0.0346 0.0334
Residual AR(2) test p-value 0.7721 0.7654

 GOVERNANCE INDICATOR

 RULE OF CONTROL OF
 LAW CORRUPTION

ln(ASSETS.sub.-1) 0.8996 *** 0.8895 ***
 0.0302 0.0330
AGE -0.0081 *** -0.0086 ***
 0.0010 0.0010
GOVERNANCE 0.2208 *** 0.1061 **
 0.0387 0.0486
ln(GDP) 0.0262 -0.0122
 0.0284 0.0263
DEFAULT -1.4708 *** -1.2550 ***
 0.3839 0.3806
DEFAULT x AGE 0.0466 *** 0.0381 ***
 0.0108 0.0105
Number of observations 385 385
Sargan test p-value 0.1923 0.0798
Residual AR(1) test p-value 0.0440 0.0318
Residual AR(2) test p-value 0.8169 0.7977

(a) Standard errors are in italics. Estimates are obtained using
the xtdpd command in Stata 10.0. The regression also includes
time fixed effects.

* Significant at 5%; *** significant at 1%.