What drives financial development? A Meta-regression analysis

Abstract

1. Introduction

There is a large literature showing that finance promotes economic development (at least up to a point). A well-developed financial system channels savings into value-creating investments, monitors borrowers to increase efficiency, facilitates to pool, share and diversify risk, and enables trade. King and Levine (1993a,b) and Rajan and Zingales (1998) were among the first to argue that financial development (FD) is related to economic development. Most studies in this line of research report evidence that FD stimulates economic growth (Levine, 2005), although some reach more ambivalent conclusions (e.g. Gries et al., 2009).¹

There is also a large literature on the relationship between FD and income inequality. FD can provide poor households with greater access to resources to meet their financial needs and may thereby lower income. Theory, however, suggests that FD could also benefit the rich (Greenwood and Jovanovic, 1990). Indeed, several recent studies have found that FD increases income inequality (de Haan and Sturm, 2017).

In view of the importance of FD for economic growth and income inequality, it is not surprising that the drivers of FD have been extensively researched. Determinants of FD that have been analyzed include trade and financial openness, legal rules and the quality of their enforcement, institutional quality, remittances, and inflation. Despite reporting close to two thousand estimates of the determinants of FD (discussed below), research has not resolved the issue of which factors are the most important. The reported findings are very diverse and often conflicting and fragile. What is missing from the literature is a systematic evaluation of these determinants.

This article offers the first quantitative review of the literature on the drivers of FD. For this purpose, we use meta-regression analysis (MRA), which is a set of statistical techniques that have been developed to identify and quantify associations drawn from an existing body of literature (Stanley and Doucouliagos, 2012; Schmidt and Hunter 2015). MRA has become increasingly popular, also in economics (Huang and Sim, 2018; Havránek and Sokolova, 2020; Heimberger, 2021). It is based on a focused examination of the role of methodological, specification, and data factors on, in our case, the reported effects of the drivers of FD. MRA can help to resolve divergent theoretical debates and conflicting empirical findings and draw robust inferences from the evidence base. Meta-analysis is particularly useful where there is disagreement over the qualitative (e.g. is there a negative or positive association) and/or quantitative inferences (the magnitude of the association). But, meta-analysis can also shed new light even on qualitatively well-established results in the literature. For example, many (but not all) economists would agree that institutions matter for FD and that inflation can be harmful. The evidence base includes qualitative divergent results ranging from institutions and inflation being good to being harmful to FD, thereby allowing to test whether the established view is really supported by available research. An additional benefit of meta-analysis is that it can be used to quantify the relative importance of institutions (or inflation) vis-à-vis other determinants, especially when the evidence base reports widely varying magnitudes of these key associations.

The aims of our meta-regression analysis are to: (i) provide a statistical synthesis of the existing research on the drivers of FD; (ii) assess the competing claims made about the impact of drivers of FD; (iii) explore the sensitivity of reported empirical results; and (iv) investigate and correct the evidence base for publication and misspecification biases. It is well known that methodological, specification, and data differences affect empirical estimates. These choices can create heterogeneity in reported estimates making it very difficult for conventional narrative reviews to make robust and valid inferences. Without some method of research synthesis, such as MRA, it becomes very difficult to draw robust inference of the effects of data, measurement, estimation, and empirical methodology on the reported results. Meta-regression analysis helps to explain why results systematically differ within and between studies; meta-regression analysis identifies and quantifies model misspecification and omitted variable biases.

MRA enables us to assess whether the evidence base is credible, strong, and decisive. An important dimension of research credibility is publication selection, where authors report results that are statistically significant, or consistent with some prior. Publication selection bias is impossible to identify from individual empirical studies (Stanley and Doucouliagos, 2012). However, when the various studies and estimates are assessed through the lens of meta-analysis, publication and model misspecification bias can be identified and thereby the literature can be corrected for such biases, producing more credible and robust estimates of policy relevant parameters.

The article is structured as follows. Section 2 offers a summary of the main drivers of FD as identified in the literature and explains the reasons why these variables may affect FD. This section also describes the proxies used to test these main drivers of FD. Section 3 describes the data used in our analysis. Section 4 outlines the methodology and Section 5 presents our main results. Section 6 concludes.

2. Drivers of FD

2.1 Law and finance: creditor rights, investor protection, enforcement, information, and legal origin

In their seminal contributions, La Porta et al. (1997, 1998) argue that differences in financial systems around the world can be traced in part to the differences in the protection of shareholders and creditors, as reflected by legal rules and the quality of their enforcement. For instance, if lenders feel that regulations do not protect them and that their chance of taking control over the assets pledged as collateral is uncertain, they are likely not to extend credit since the implicit bankruptcy risk will severely reduce their expected earnings.

Furthermore, according to La Porta et al. (1997, 1998), protection of legal rights and contract enforcement vary systematically by legal origin, which is either English, French, German, or Scandinavian. La Porta et al. argue that the English common law tradition protects the rights of shareholders and creditors best, while the French civil code is associated with less efficient contract enforcement and weaker protection of shareholders and creditors. Countries with German or Scandinavian legal origins are said to have intermediate levels of protection, but the highest level of contract enforcement.

The law and finance theory can be tested by examining whether legal origin is related to a measure of FD, or by examining whether proxies for creditor and shareholder protection and contract enforcement are related to financial deepening, or both. Supplementary Appendix  Table A1 summarizes the evidence in support of the law and finance view of FD.

La Porta et al. (1997) consider legal origin, contract enforcement (the ‘law and order’ index from ICRG; see below), and some proxies for investor protection in their cross-section model for stock market capitalization. They conclude that ‘our shareholder rights variables account for some of the difference between relative market capitalizations of different legal families, but that the family effects are also significant.’ (p. 1142). Levine et al. (2000) use differences in the legal rights of creditors, the efficiency of contract enforcement, and accounting system standards in a panel model to explain cross-country differences in the level of financial intermediary development, and find that countries with high scores on these variables tend to have better developed financial intermediaries.

However, some authors argue that other factors may trump legal origin. Four alternatives have been put forward: endowments; political factors; political instability; and culture (Beck et al., 2003; Keefer, 2007; Roe and Siegel, 2011). For example, Roe and Siegel (2011) and Girma and Shortland (2008) point out that institutions of investor protection—such as legal rules, courts, and regulators—cannot function well in an unstable political environment.

2.2 Institutional quality: political and financial risk, democracy, governance, and economic freedom

The law and finance theory points to the importance of the quality of legal institutions, notably when it comes to contract enforcement/rule of law. Other studies employ different indicators of institutional quality, which are argued to affect FD. The reasoning why institutional quality matters is very similar as under the law and finance theory.

Several studies employ International Country Risk Guide (ICRG) data (see Supplementary Appendix  Table A1). ICRG classifies country risk into three different categories: political risk (including corruption, law and order, and bureaucratic quality), financial risk (including repudiation of contracts by the government, and risk of expropriation, as also used in the law and finance literature), and economic risk (including inflation).

Other authors focus on the extent to which democratic institutions are linked to FD. Pagano and Volpin (2001) argue that countries with closed and static political regimes tend to resist the availability of external financing, since the ensuing competition would threaten the entrenched powers of the political elite. Several studies report support for this view (see Supplementary Appendix  Table A1), mostly using the Polity index.

A few studies consider constraints on the power of the executive, which also comes from the Polity database. This variable captures the official (de jure) discretionary leeway that the executive branch has in changing and implementing new policies. Acemoglu and Johnson (2005) report that constrains on the executive are significant in their estimates of FD.

Some studies focus on governance, using the average of the World Bank’s Worldwide Governance indicators (Kaufmann et al., 2009). These studies find evidence that better governance enhances FD. Finally, a few studies use (components of) some economic freedom measure to proxy institutional quality (see Supplementary Appendix  Table A1).

There is strong support for the view that institutional quality fosters financial deepening (see Supplementary Appendix  Table A1), although some studies report mixed or non-linear effects.

2.3 Trade and financial openness

Several papers examine whether trade and/or financial openness affect FD. Two reasons have been suggested why trade openness may matter. First, Svaleryd and Vlachos (2002) emphasize the role of risk diversification. As openness may be associated with greater risks, such as increased exposure to external demand shocks or foreign competition, it will create new demands for external finance. Firms will need credit to overcome short-term cash flow problems and adverse shocks.

Second, an open economy may weaken the incentives and the political power of interest groups to resist financial deepening. Rajan and Zingales (2003) argue that in particular industrial and financial incumbents frequently stand to lose from FD as it creates opportunities for new firms to become established, which breeds competition and erodes incumbents’ rents. Rajan and Zingales (2003) suggest that the simultaneous opening of trade and capital accounts holds the key to successful FD.

Others disagree. For instance, Chin and Ito (2006) stress the impact of capital account liberalization on FD through several channels. First, financial liberalization may mitigate financial repression in protected financial markets, allowing the (real) interest rate to rise to its competitive market equilibrium. Second, removing capital controls allows investors to engage in more portfolio diversification. Third, the liberalization process usually increases the efficiency level of the financial system by weeding out inefficient financial institutions and creating greater pressure for reform of the financial infrastructure.

The results of studies focusing on the impact of trade openness on FD are rather mixed (see Supplementary Appendix  Table A1).²

Several studies have examined the impact of capital account openness on FD. Chin and Ito (2006) conclude that financial openness contributes to equity market development, but only when a threshold level of general development of legal systems and institutions has been attained. However, other studies do not (always) confirm the (conditional) positive effect of capital account openness on financial depth.

Apart from capital account openness, other proxies for financial openness have been used. Some studies use capital flows for this purpose, while others consider foreign direct investment (FDI), but results are mixed. Several studies employ the data of Lane and Milesi-Ferretti (2007) on countries’ foreign assets and liabilities to proxy financial openness, while others use the database of Abiad et al. (2010) of financial liberalization to proxy financial openness.³ Finally, the presence of foreign banks is sometimes used as proxy for financial openness (cf. Claessens and van Horen, 2014).

Some studies examine whether both trade and financial openness are jointly linked to FD as suggested by Rajan and Zingales (2003). The results are mixed (see Supplementary Appendix  Table A1).

2.4 Remittances

Remittances, that is, funds received from migrants working abroad, have become the second largest source of external finance for developing countries after FDI. Furthermore, remittances are more stable than other international financial flows. Several studies examine whether remittances are related to FD. However, the theoretically expected effect is ambiguous.

On the one hand, remittances may have a positive impact on FD. Several reasons have been put forward to explain this. First, as remittances are typically lumpy, recipients might have a need for financial products that allow for the safe storage of these funds (i.e. bank deposits) even if most of these funds are not received through banks. Furthermore, to the extent that banks provide remittance transfer services, they may reach out to recipients with limited or no financial intermediation. Banks may also be more willing to extend credit to remittance-receiving households, given that remittances are often a significant and stable income source. Moreover, increased loanable funds created by banked remittance transfers can increase overall credit to other, non-remittance-receiving households (Aggarwal et al., 2011).

On the other hand, remittances may have a negative impact on FD. First, remittances might not increase bank deposits if they are immediately consumed or saved in other ways than via financial intermediaries. Furthermore, migrants tend to rely heavily on informal transfer channels rather than bank transfers for various reasons, like lower transaction costs and greater efficiency of informal transfer methods, and distrust of banks and financial authorities in the recipient countries. Because remittances can help relax individuals’ financing constraints, they might lead to a lower demand for credit and have a dampening effect on credit market development. Remittance-receiving households might choose to rely on migrants abroad rather than on domestic banks for credit, in which case remittances become a substitute for credit. By becoming a substitute for inefficient or inexistent credit markets, remittances help alleviating credit constraints, thereby contributing to an improved allocation of capital and to higher economic growth.

Several studies report that remittances have a positive impact on FD. However, others find no support for this view (see Supplementary Appendix  Table A1).⁴

2.5 Inflation

It is widely believed that inflation distorts the incentives for FD. In particular, moderate to high inflation may discourage financial intermediation, and encourage saving in real assets. An increase in the rate of inflation reduces the real return on assets. This reduction worsens existing credit market frictions, which results in less credit. The financial sector makes fewer loans; there is increased inefficiency in resource allocation, and a diminishing of intermediary activities (Boyd et al., 2001). Likewise, lower inflation reduces the overall level of uncertainty about the ability of borrowers to meet their obligations, and the resulting less risky environment facilitates financial deepening.

As shown in Supplementary Appendix  Table A1, some studies report evidence suggesting that inflation hampers financial depth. At the same time, the negative effects of inflation may be nonlinear. Some models suggest that only when inflation exceeds a certain critical level, will it have a negative effect on FD. Boyd et al. (2001) report that FD has a strong negative relationship with inflation, but only for countries with low-to-moderate rates of inflation. Khan et al. (2006) find that the threshold level of inflation after which it has a negative effect on FD is generally about 3%–6% per annum, depending on the specific measure of financial depth that is utilized.

3. Data

Our search strategy and reporting closely follows the guidelines provided by Havranek et al. (2020). The search for relevant studies involves several databases: Econlit, Web of Science, Repec/Ideas, and Google Scholar. We used combinations of the following broad keywords: ‘FD’, ‘stock market capitalization’, ‘private credit’, ‘bank credit’, ‘inflation’, ‘institutions’, ‘creditor rights’, ‘civil law’, ‘rule of law’, ‘law and finance’, ‘governance’, ‘democracy’, ‘economic freedom’, ‘institutional quality’, ‘trade liberalization’, ‘trade openness’, ‘financial liberalization’, ‘capital account liberalization’, ‘financial openness’, ‘bank reform’, ‘remittances’, and ‘trust’. Once a study had been found, we used its references to search for more studies.

To be included in our meta-analysis, a study had to meet three criteria. First, we include only studies using data at the macro-level (thereby ignoring micro-based studies) and studies covering several countries (thereby ignoring single-country studies). The reason is that many of the drivers of FD as discussed in the previous section are based on cross-country differences, which, by definition cannot be included in micro- and single-country studies. Second, studies had to provide sufficient information from which we could quantify a comparable size effect. That is, studies had to report sample size, regression coefficients, and their standard errors or p-values or t-statistics. Studies that did not provide this information could not be included. Third, studies had to report the effect of one of the hypothesized determinants of FD, i.e. institutional quality, trade openness, financial openness, inflation, remittances, and trust. The search for studies was terminated November 2019. All coding was conducted and checked by three independent coders.

Even though it is widely used, the private-credit-to-GDP ratio does not capture several dimensions of FD, like access to credit and efficiency of the financial system.⁶ The International Monetary Fund (IMF) provides data on these and other dimensions of FD and aggregates them into an index that summarizes how developed financial institutions and financial markets are in terms of their depth, access, and efficiency (see Svirydzenka, 2016 for further details).⁷ Unfortunately, research on the drivers of FD has not or hardly used this and alternative indicators of (lack of) FD, such as dollarization. As a consequence, it is not possible to do a meaningful meta-regression analysis.

We collect estimates of ${\alpha }_m$ and transform these into a comparable measure. Unfortunately, descriptive statistics are frequently either poorly reported or not reported in studies, so that we have little confidence in extracting elasticities or percentage change from reported statistics and estimates.⁸ Instead, we opt to use the partial correlation, which measures the strength of the association between a given variable and FD, holding constant all other factors.⁹

The partial correlations for private credit are illustrated in Figures 1–5, in the form of funnel plots.¹⁰ The funnel plots illustrate the distribution in reported estimates. There is significant heterogeneity in the evidence base, with both positive and negative effects reported for all hypothesized determinants. The studies offer independent replications of the various determinants of FD. Figures 1–5 suggest that many studies fail to replicate the findings of others. By applying meta-analysis, we are able to synthesize the evidence base and analyze these distributions and thereby assess what conclusions can be established from the literature.

Fig. 1

Institutions and private credit

Source: Authors’ calculations.

Notes: 503 estimates. Hollow circles are statistically insignificant correlations. Full circles are statistically significant. The solid line indicates the weighted average correlation (r = 0.07). The dotted line indicates a zero correlation.

Open in new tabDownload slide

Fig. 2

Financial openness and private credit

Source: Authors’ calculations.

Notes: See notes to Fig. 1 (305 estimates). The solid line indicates the weighted average correlation (r = 0.04).

Open in new tabDownload slide

Fig. 3

Trade openness and private credit

Source: Authors’ calculations.

Notes: See notes to Figure 1 (239 estimates). The solid line indicates the weighted average correlation (r = 0.05).

Open in new tabDownload slide

Fig. 4

Remittances and private credit

Source: Authors’ calculations.

Notes: See notes to Figure 1 (69 estimates). The solid line indicates the weighted average correlation (r = 0.03).

Open in new tabDownload slide

Fig. 5

Inflation and private credit

Source: Authors’ calculations.

Notes: See notes to Figure 1 (288 estimates). The solid line indicates the weighted average correlation (r = −0.07).

Open in new tabDownload slide

The funnel plots become narrower for estimates reported with greater precision. Nevertheless, for all determinants, there is significant spread in the more precise estimates. For example, in the case of institutions, the estimates with precision >40 (6% of all estimates for institutions) range from −0.09 to +0.12. We use MRA to model the variation or heterogeneity in estimates (across all levels of precision). However, one factor that our MRA does not address is attenuation bias. Schmidt and Hunter (2015) point out that measurement error attenuates correlations. This could be a particularly important issue for the measurement of the key variables for developing countries. To the extent that this is an issue in this literature, our meta-averages are likely to provide lower bound estimates. We have no reason to suspect that attenuation bias is differentially larger for the more precise estimates nor differentially for the various determinants.

4. Methods

Drawing credible inference from the evidence base is a major challenge when the evidence base consists of studies with low statistical power, when there is a high degree of heterogeneity in reported findings, and when some findings have been selectively reported e.g. on the basis of their statistical significance or because they match researchers’ priors (Ioannidis et al., 2017). Hence, Equation (2) will rarely adequately represent the evidence base.

If enough researchers select evidence on the basis of its statistical significance, then this will cause a correlation between ${\mathrm{SE}}_{\mathrm{ij}}$ and ${\mathrm{r}}_{\mathrm{ij}}$⁠.¹¹ The coefficient on ${\mathrm{SE}}_{\mathrm{ij}}$ provides a test for the presence and magnitude of publication selection bias. Equation (4) then provides conditional estimates of the effects of the determinants of FD after correcting for publication selection bias (Stanley and Doucouliagos, 2012).¹²

Equation (5) is known as the Precision Effect Test-Precision Effect Estimate with Standard Error (PET-PEESE) model and is the primary meta-regression model used in this article. PET-PEESE is one of several methods for identifying and correcting publication bias in an evidence base. Other notable methods include ‘Trim-and-Fill’ (Duval and Tweedie, 2000), selection models such as the three-parameter selection model (Hedges and Vevea, 1996) and the recent model proposed by Andrews and Kasy (2019), the p-curve and p-uniform methods (van Assen et al., 2015), and the weighted average of the adequately powered studies (Ioannidis et al., 2017). We here use PET-PEESE as this method has the lowest bias of all meta-analysis methods, the lowest rate of false positives, and the lowest mean squared error (Stanley and Doucouliagos, 2014; Kvarven et al., 2019).¹³

Equation (5) is estimated by weighted least squares, using inverse variance weights. Inverse variance weights can be either fixed effects or random effects.¹⁴ Fixed effects weights assume that there is a single underlying effect that all studies are estimating. Random effects weights allow for a distribution of effect sizes. We present the results for both sets of weights. However, random effects weights have been shown to produce more biased estimates when there is publication selection (Stanley and Doucouliagos, 2015). Further, Kvarven et al. (2019) show that random effects exaggerate meta-averages by nearly three-fold and have high rates of false positives; they find evidence of an empirical effect when there is none. Simulations show that unrestricted weighted least squares (UWLS) produce meta-averages with smaller bias and smaller Mean Squared Error, especially in cases where there is heterogeneity in the evidence base and some of the reported estimates have been preferentially chosen on the basis of their statistical significance (Stanley and Doucouliagos, 2015).¹⁵

5. Results

5.1 Unconditional meta-averages

The meta-analyses results are presented in Tables 1 and 2, for institutional quality and all other variables of interest, respectively. For both tables, Panel A reports the results for private credit and Panel B for stock market capitalization (but note that the evidence base for private credit is larger than that for stock market capitalization). As discussed in the methods section above, while we report random effects meta-averages, we rely on the UWLS results for inference as these are less biased (Stanley and Doucouliagos, 2014; Stanley et al., 2017; Kvarven et al., 2019).¹⁶

Table 1 reports the meta-analyses for estimates relating to institutional quality. Column (1) combines all estimates of the effects of institutional quality using various measures for the quality of institutions. These results suggest a small positive effect of institutional quality on private credit (r = 0.07) and a slightly higher effect on stock market capitalization (r = 0.08).

Columns (2)–(8) present meta-analyses of the main specific measures for institutional quality. Columns (2) and (3) focus on measures relating to creditor protection and French civil law, respectively. Column (4) presents the results for estimates that use an ‘aggregate’ measure of institutional quality, i.e. one that combines several individual components such as control of corruption, government effectiveness, and rule of law (see Section 2). Columns (5) and (6) focus on two of these components: the rule of law and governance quality, respectively. Column (7) looks at democracy and column (8) at economic freedom. Except for French law, all the other dimensions of institutions have a positive effect on FD, proxied by private credit. French law has a small negative effect, but this is only weakly statistically significant.

With regard to stock market capitalization, creditor protection, overall institutional quality, and democracy have positive and statistically significant correlations. The other dimensions also have positive correlations (except for French civil law), but they are estimated with low precision, most likely because of the thin evidence base.

Comparing the magnitude of the estimated partial correlations, the rule of law appears to be nearly five times more important, and governance twice as large, for private credit than for stock market capitalization. Creditor protection is also more important for private credit than it is for stock market capitalization. In contrast, democracy appears to be more important for stock market capitalization.

Table 2 suggests that both trade and financial openness have a positive effect on private credit and stock market development, whilst inflation has an adverse effect. Trade and financial openness appear to be more important for stock market development than for private credit. The effect of inflation on both measures for FD is about the same. The extant studies on remittances show a positive effect, but this is not statistically significant and this finding needs to be replicated by additional studies. We conclude that the literature has not yet established that remittances matter for FD.

5.2 Meta-regression analysis

Tables 1 and 2 also report I², a measure of between study heterogeneity. Simulations show that when I² exceeds 80% conventional meta-average methods reported in Tables 1 and 2 become less reliable at representing the evidence base (Stanley, 2017). In such cases, meta-regression analysis is recommended to accommodate the heterogeneity in estimates. That is, with large heterogeneity, there is a distribution of effects and an unconditional meta-average will not adequately represent the research record. However, as evidenced by Tables 1 and 2 the evidence base is thin (i.e. only a limited number of regressions) for several of the potential determinants. This poses a challenge, as regression-based tests will have low power to detect empirical effects for these thin-evidence-based dimensions. Hence, our meta-regressions focus on the areas where the evidence base is deepest. Thus, we do not attempt meta-regression for the effects of remittances and trust. The meta-regression results are presented in Tables 3–6, for institutions, trade openness, financial openness, and inflation, respectively.¹⁷