Finance or Capabilities? Why the Finance-Growth Question Is Mismeasured in the MENAT Region

1. Introduction

Few empirical questions have been asked as often, and answered as inconsistently, as whether financial development raises long-run income in the Middle East and North Africa. Ben Naceur and Ghazouani (2007) find banking depth uncorrelated or negatively correlated with growth across the region. Khattab et al. (2015) report a positive effect for the Maghreb. Batuo et al. (2018) find that finance stabilises in some African configurations and destabilises in others. Boukhatem and Ben Moussa (2025) condition everything on institutions. The most recent studies do not converge either: Kamal et al. (2023) recover a clean positive long-run effect for sixteen regional economies, whereas Shaddady (2023) finds financial institutions eventually constraining growth, and work published in 2026 continues to split, with Idrissi (2026) reading financial integration as growth-enhancing for the region while Mo (2026) finds remittances substituting for, rather than complementing, a financial system that uses them inefficiently. The wider measurement literature now cautions that the credit ratio on which many of these results rest is the wrong instrument altogether (Boďa, 2024). The policy world has not waited for the academic dust to settle: the October 2024 Regional Economic Outlook of the IMF devotes an entire chapter to strengthening growth through financial development in the region (IMF, 2024), and the meta-analytical record shows the estimated effect weakening precisely in less developed economies and in recent decades (Valickova et al., 2015).

This paper argues that the regional disagreement is not a property of the economies but of the measurement and identification choices imposed on them, and that resolving those choices changes the question worth asking. Two disciplines protect the argument from the reviewer’s first two questions. The sample is not chosen; it is the output of a single coverage rule applied blind to the entire regional universe, which retains ten economies. And cointegration is not read from an adjustment speed but tested formally with a statistic built for panels driven by common factors, the pervasive feature these series display. The argument then proceeds in three steps on the resulting panel, 1995-2021, drawn from the World Development Indicators, the IMF Financial Development Index database and harmonised schooling series.

The first step prosecutes the proxies. Nearly every regional study measures finance with the ratio of broad money or private credit to GDP. Two constant-sample diagnostics show why such studies cannot agree. First, the ratio tracks remittances with no common sign across economies, negatively in Turkiye and Egypt, positively in Tunisia, so the popular story of a single remittance-contamination channel does not survive contact with the data; what the ratio carries in levels is only a small coefficient whose remittance partner term is itself unstable in sign and significance across samples. Second, the instability term that accompanies these ratios is significant only in the full crisis window and disappears once the recent Turkish and Egyptian years are removed, so an apparent fragility effect is a crisis-year denominator rather than the financial system. A proxy this sensitive to window, control and sample cannot support a stable conclusion, which is why any two studies drawing differently from this space reach different verdicts, and they have.

The second step shows that better measurement relocates the problem rather than solving it. The multidimensional index of Svirydzenka (2016) behaves far better than the ratios, and its institutions component in particular survives window truncation. It then collides with a deeper feature of the region: institutional finance and mean years of schooling are almost collinear within countries. They are, statistically, one slow trend. Estimator families arbitrate that trend in contradictory ways: the level-cointegration family awards it to schooling and leaves finance an economically trivial residue, the error-correction family with cross-sectionally robust errors awards it to neither, and the common-factor family absorbs it entirely. No econometric arbitration can allocate a single trend between two collinear claimants at this cross-sectional dimension, and the attempt is precisely what condemns the literature to contradiction.

The third step stops arbitrating and models the bundle. We extract the common component of institutional finance and schooling, interpret it as a capabilities factor in the spirit of Sen (1999), and estimate its long-run association with income alongside openness and standard controls. The result is the most stable object this panel produces. The capabilities elasticity is positive and significant across four estimator families that rest on different assumptions: pooled mean group, dynamic fixed effects with Driscoll and Kraay (1998) errors, group-mean fully modified least squares, and a dynamic common correlated effects estimator that reads the relationship through the common factors. It survives a pre-committed robustness programme spanning a remittance control, a governance control, window truncation and country-by-country exclusion, with a single disclosed boundary. Trade openness contributes in most specifications, though not in the reference dynamic fixed effects column, which we report rather than suppress. The estimated speeds of adjustment place the return to equilibrium on the timescale of structural reform rather than of electoral cycles.

The paper makes two contributions. It provides the first constant-sample diagnosis of why ratio-based finance measures misbehave in remittance-dependent, crisis-prone economies, and it shows that reframing the object of measurement from finance to a joint capabilities factor yields a long-run relationship robust across estimator families that disagree about almost everything else. The remainder of the paper is organised as follows. Section 2 reviews the relevant literature and identifies the gaps the paper addresses. Section 3 describes the data and the estimation strategy. Section 4 presents the results, including the robustness programme. Section 5 discusses the findings against the prior regional evidence and draws out the policy implications. Section 6 concludes.

2. Background and Related Work

This section places the paper in three literatures that it draws on and departs from: the finance-growth nexus and its measurement turn, the regional evidence for the Middle East and North Africa, and the capabilities view of development. Each supplies part of the argument, and each leaves a gap the paper fills.

2.1. Finance, Growth and the Measurement Turn

The theoretical case for finance rests on functions rather than balance-sheet sizes: mobilising savings, allocating capital, monitoring managers and diversifying risk (Schumpeter, 1911; McKinnon, 1973; Shaw, 1973; Levine, 2005). The first generation of cross-country evidence attached those functions to observable depth, private credit and liquid liabilities relative to GDP, and read them as growth-promoting (King and Levine, 1993; Rajan and Zingales, 1996; Levine et al., 2000; Beck et al., 2000). A second generation complicated the verdict. The effect is non-monotonic, turning negative beyond a threshold of credit deepening (Arcand et al., 2015; Samargandi et al., 2015); it separates sharply into a positive long-run and a fragile short-run component (Loayza and Rancière, 2006); and it shrinks, sometimes to nothing, in meta-analytical samples dominated by developing economies (Valickova et al., 2015).

The common thread through these qualifications is measurement. If a single depth ratio stands in for a multidimensional financial system, threshold effects and horizon reversals are as likely to reflect the proxy as the phenomenon. The response, associated with Sahay et al. (2015) and Svirydzenka (2016), replaced depth ratios with indices spanning the depth, access and efficiency of both financial institutions and financial markets. A parallel critique cautions against the heuristic quality-adjustments layered onto the credit-to-GDP ratio and builds distance-to-frontier measures of depth, access, efficiency and stability in their place (Boďa, 2024), while cross-country work using the multidimensional indices confirms that financial institutions, more than markets, carry the growth signal (Nguyen and Su, 2021), and recent threshold evidence shows the growth effect concentrated in less financially developed economies and amplified by institutional quality (Özmen and Taşdemir, 2026). Panel cointegration evidence built on the same pooled mean group and dynamic-panel toolkit continues to qualify the finance-growth link outside the region: financial development is itself an outcome of capital-account openness rather than a purely domestic process in emerging Europe (Bayar and Gavriletea, 2018), the depth-growth relationship remains fragile in low-income Southern Africa even under favourable estimator assumptions (Mlambo, 2024), and the long-run growth payoff of financial inclusion in Sub-Saharan Africa is conditional on institutional quality rather than automatic (Golpet et al., 2026), the same conditionality this paper documents for MENAT under a different measure. The gap this leaves is diagnostic: the literature has improved the measure without establishing what the older proxy captures when it misleads, or how much of the recorded disagreement is measurement rather than substance.

2.2. The Regional Evidence and Its Instability

Regional applications inherited the ratios and the contradictions together. Depth measures underpin the negative or insignificant banking results of Ben Naceur and Ghazouani (2007), the positive Maghreb estimates of Khattab et al. (2015), the finance-stability interactions of Batuo et al. (2018) and the institution-conditioned effects of Boukhatem and Ben Moussa (2025). The contradiction persists in the most recent work: using the same multidimensional indices this paper adopts, Kamal et al. (2023) report a positive long-run finance effect for sixteen MENA economies, Shaddady (2023) finds an inverted-U in which financial institutions eventually constrain growth, Soltani et al. (2021) attribute the frequently insignificant finance coefficient to the underdevelopment of regional financial systems, and Dahmani et al. (2023) obtain an outright negative long-run coefficient under a cross-sectionally augmented specification. The 2026 vintage does not settle matters: Idrissi (2026) finds financial integration growth-enhancing for the region over 1995-2021 with openness and investment positive and instability negative, while Mo (2026) reports that remittances and financial development act as substitutes rather than complements in MENA, local financial sectors using diaspora inflows inefficiently. That the same region yields opposite signs across studies is usually read as heterogeneity or as sensitivity to method. Two regional features are rarely confronted directly and are candidates for that disagreement. Remittance inflows, among the highest in the world relative to GDP in Morocco, Egypt and Tunisia, enter deposit money and private credit without constituting intermediation of domestic saving, a channel documented for financial aggregates by Guillaumont Jeanneney and Kpodar (2011) among others, and one that recent nonlinear evidence outside the region shows to be asymmetric rather than proportional, with reductions in remittances corroding financial development more than equivalent increases build it (Biyase and Naidoo, 2023). And recurrent macroeconomic crises move the GDP denominator violently, so that ratio variation in crisis windows tracks the business cycle rather than financial development. Neither has been isolated at a constant sample in the regional literature, which is the gap the empirical section fills.

Two gaps in this regional literature motivate what follows. The first is diagnostic. Reviews attribute the conflicting regional results to heterogeneity, to sample composition or to estimator choice, but rarely ask what the finance proxy itself registers when it misbehaves; the remittance and denominator channels are named in passing, if at all, and never quantified at a fixed sample. The second is conceptual. Studies that add human capital as a control treat it as a competitor to finance in a single regression, on the implicit assumption that the two can be told apart in the data. In economies where financial and educational development have advanced along one reform trajectory, that assumption is untested and, as the results below show, unwarranted. The regional evidence has therefore been asking finance and human capital to arbitrate a contest they may not be separable enough to hold.

2.3. A Capabilities Reading

The identification problem this paper documents, near-collinearity between financial-institutional development and human capital, has a natural interpretation in the capabilities tradition of Sen (1999). Banking networks and schooling systems are joint outputs of the same state capacity and social investment, facets of one expansion of what people and firms are able to do rather than independent policy dials. The growth-accounting habit of assigning each a separate coefficient presumes a separability the data of a small regional panel need not honour. Reading finance and education as one capabilities bundle does more than handle collinear regressors. It is the substantive claim that the two develop together because they are produced together, and it reframes the policy question from which lever to pull toward how to build the joint capability. A separate strand already finds the two entangled: financial development contributes to human capital accumulation with bidirectional causality in emerging Asian panels (Hong Vo et al., 2021), the long-run finance-human capital link is asymmetric (Ha and Ngoc, 2022), human capital and education mediate the finance-growth relationship internationally (Ngo et al., 2022), and recent panel evidence asks directly whether human capital is the hidden engine of financial-sector development in developing economies (Gherghina et al., 2026). What the regional literature has lacked, and this paper supplies, is the step from that co-development to its measurement consequence: near-collinear components cannot be given separate coefficients, so the bundle is the estimable object.

3. Materials and Methods

3.1. Sample and Sources

The population is defined before any estimation and independently of results. It is the set of emerging and developing economies of the Middle East and North Africa in the World Bank regional classification, augmented with Turkiye to form the established MENAT grouping, twenty economies in all. A single mechanical rule then decides membership: an economy enters if and only if it offers at least eighteen consecutive years of jointly complete data on the full set of estimation variables, a floor dictated by the six-regressor long-run specification. The rule is applied blind to every candidate, including economies never previously studied in this literature, and the sample is whatever it returns. Table 1 reports the outcome for all twenty candidates. Ten economies qualify: Algeria, Bahrain, Egypt, Iran, Libya, Morocco, Saudi Arabia, Tunisia, Turkiye and Yemen. Ten fail the coverage floor, most by discontinued or absent national-accounts and monetary series (Djibouti, Iraq, Jordan, Kuwait, Lebanon, Oman, the Palestinian territories, Qatar, Syria, the United Arab Emirates). Macroeconomic series come from the World Development Indicators; the financial development indices come from the IMF Financial Development Index database (Svirydzenka, 2016); schooling comes from the UNDP-based harmonised series distributed by Our World in Data; governance from the Worldwide Governance Indicators. Estimation retains, per economy, the longest contiguous span with complete data. Two boundaries are dictated by the sources rather than chosen. The multidimensional financial development indices are published with a lag and currently extend to 2021 in their latest vintage, which fixes the upper bound of every index-based specification; the World Development Indicators reach further, but the financial-development measure that the paper turns on is the binding series, and recent regional studies using these indices share the same endpoint. The lower bound of 1995 reflects the first year in which the governance and schooling series are jointly available across the sample.

Table 1. Mechanical sampling funnel over the MENAT universe.

Economy	Joint span	From	To	Decision
Algeria	27	1995	2021	Retained
Bahrain	21	1995	2015	Retained
Djibouti	9	2013	2021	Excluded (< 18 yr)
Egypt	27	1995	2021	Retained
Iran	22	1995	2016	Retained
Iraq	0	n.a.	n.a.	Excluded (< 18 yr)
Jordan	13	1995	2007	Excluded (< 18 yr)
Kuwait	12	2010	2021	Excluded (< 18 yr)
Lebanon	9	2009	2017	Excluded (< 18 yr)
Libya	27	1995	2021	Retained
Morocco	27	1995	2021	Retained
Oman	17	2001	2017	Excluded (< 18 yr)
Palestinian terr.	0	n.a.	n.a.	Excluded (< 18 yr)
Qatar	0	n.a.	n.a.	Excluded (< 18 yr)
Saudi Arabia	23	1995	2017	Retained
Syria	17	1995	2011	Excluded (< 18 yr)
Tunisia	27	1995	2021	Retained
United Arab Em.	2	2008	2009	Excluded (< 18 yr)
Yemen	19	1995	2013	Retained
Turkiye	27	1995	2021	Retained

Notes: Population: emerging and developing MENA economies (World Bank) plus Turkiye. An economy is retained if its longest contiguous span of jointly complete data on all estimation variables reaches eighteen years. The rule is blind to estimation output.

Table 1. Mechanical sampling funnel over the MENAT universe.

Economy	Joint span	From	To	Decision
Algeria	27	1995	2021	Retained
Bahrain	21	1995	2015	Retained
Djibouti	9	2013	2021	Excluded (< 18 yr)
Egypt	27	1995	2021	Retained
Iran	22	1995	2016	Retained
Iraq	0	n.a.	n.a.	Excluded (< 18 yr)
Jordan	13	1995	2007	Excluded (< 18 yr)
Kuwait	12	2010	2021	Excluded (< 18 yr)
Lebanon	9	2009	2017	Excluded (< 18 yr)
Libya	27	1995	2021	Retained
Morocco	27	1995	2021	Retained
Oman	17	2001	2017	Excluded (< 18 yr)
Palestinian terr.	0	n.a.	n.a.	Excluded (< 18 yr)
Qatar	0	n.a.	n.a.	Excluded (< 18 yr)
Saudi Arabia	23	1995	2017	Retained
Syria	17	1995	2011	Excluded (< 18 yr)
Tunisia	27	1995	2021	Retained
United Arab Em.	2	2008	2009	Excluded (< 18 yr)
Yemen	19	1995	2013	Retained
Turkiye	27	1995	2021	Retained

Notes: Population: emerging and developing MENA economies (World Bank) plus Turkiye. An economy is retained if its longest contiguous span of jointly complete data on all estimation variables reaches eighteen years. The rule is blind to estimation output.

Table 2. Variables, definitions and sources.

Variable	Definition	Source
ln y	Log GDP per capita, constant 2015 USD	WDI, NY.GDP.PCAP.KD
FI	Financial institutions index (0-1), standardised	IMF FDI database
Schooling	Mean years of schooling, interpolated, in logs	UNDP via Our World in Data
Capabilities	Common component of FI and schooling (mean of pooled z-scores)	Authors
Instability	Five-year rolling SD of M2/GDP and bank credit/GDP log-growth, pooled z	WDI, authors
Openness	Trade in goods and services, % GDP, in logs	WDI, NE.TRD.GNFS.ZS
Inflation	ln(1 + CPI inflation/100)	WDI, FP.CPI.TOTL.ZG
Government	Final consumption expenditure, % GDP, in logs	WDI, NE.CON.GOVT.ZS
Investment	Gross fixed capital formation, % GDP, in logs	WDI, NE.GDI.FTOT.ZS
Remittances	ln(1 + personal remittances received, % GDP)	WDI, BX.TRF.PWKR.DT.GD.ZS
Governance	WGI government effectiveness, interpolated	WGI
Ratio index	Mean of pooled z-scores of M2/GDP and bank credit/GDP	WDI, authors

Table 2. Variables, definitions and sources.

Variable	Definition	Source
ln y	Log GDP per capita, constant 2015 USD	WDI, NY.GDP.PCAP.KD
FI	Financial institutions index (0-1), standardised	IMF FDI database
Schooling	Mean years of schooling, interpolated, in logs	UNDP via Our World in Data
Capabilities	Common component of FI and schooling (mean of pooled z-scores)	Authors
Instability	Five-year rolling SD of M2/GDP and bank credit/GDP log-growth, pooled z	WDI, authors
Openness	Trade in goods and services, % GDP, in logs	WDI, NE.TRD.GNFS.ZS
Inflation	ln(1 + CPI inflation/100)	WDI, FP.CPI.TOTL.ZG
Government	Final consumption expenditure, % GDP, in logs	WDI, NE.CON.GOVT.ZS
Investment	Gross fixed capital formation, % GDP, in logs	WDI, NE.GDI.FTOT.ZS
Remittances	ln(1 + personal remittances received, % GDP)	WDI, BX.TRF.PWKR.DT.GD.ZS
Governance	WGI government effectiveness, interpolated	WGI
Ratio index	Mean of pooled z-scores of M2/GDP and bank credit/GDP	WDI, authors

3.2. Descriptive Statistics

Table 3 summarises the estimation variables over the estimation window. The panel is unbalanced: macroeconomic series run the full length for every economy, but the multidimensional financial-institutions index and the governance series are shorter, and remittances are recorded for only a subset of country-years, which is itself informative about how peripheral formal transfer channels are in several of these economies. Two features of the cross-section shape everything that follows. First, the economies are heterogeneous in level: the oil-rich members average roughly two and a half times the income per capita of the diversified members, and inflation ranges from mild deflation to episodes above eighty per cent, so any pooled estimator must be robust to wide dispersion in the nuisance variables. Second, and more consequentially for identification, the two capabilities components move within a narrow joint band: the financial-institutions index stays between 0.08 and 0.60 and tracks schooling closely within each country, so it is the within-country variation in the capabilities factor, not its cross-sectional spread, that the long-run estimators exploit.

The moments repay a closer reading. Openness averages seventy per cent of GDP but reaches almost twice that in the trade-entrepot economies, and its dispersion is the reason it later behaves as a first-rank rather than a control variable. The financial-institutions index has a modest mean of 0.30 on its zero-to-one scale, confirming that these are middle-income financial systems rather than frontier or advanced ones, while its standard deviation of 0.09 is small relative to the schooling range, which foreshadows the collinearity that motivates the capabilities reframing. Remittances, observed for one hundred and sixty-two country-years, average three per cent of GDP but exceed six per cent in Egypt and Morocco, the same economies where the ratio-based finance proxies later prove most fragile. The capabilities and instability factors are reported as standardised scores by construction, so their means sit near zero; what matters is their spread, and the capabilities factor varies over more than five standard deviations across the panel, which is the variation the long-run estimators draw on.

3.3. Estimation Strategy

The workhorse is an error-correction ARDL(1,1) panel. For economy i in year t, with y the log of income per capita and x the vector of the capabilities factor, instability, openness, inflation, government consumption and investment, the estimating equation is

$$\Delta y_{it}=c_i+{\phi }_i(y_{i,t-1}-{\beta }^{\text{'}}{x}_{i,t-1})+{\gamma }^{\text{'}}\Delta x_{it}+{\epsilon }_{it}$$

(1)

where φ is the speed of adjustment, β the vector of long-run elasticities that is the object of interest, and γ the short-run responses. Lag order is chosen by the corrected Akaike criterion on identical samples, which selects (1,1) unanimously; (2,1) is reported as a sensitivity. Four estimators recover β under different assumptions and are read for agreement.

Two of the regressors are composite factors, and their construction is part of the identification strategy rather than a preliminary. The capabilities factor is the standardised average of the standardised multidimensional financial-institutions index and standardised log schooling, so it loads the common developmental trend that the two share while discarding the scale on which they differ; entering the index and schooling separately, which the regional literature does by default, is the specification the paper shows to be unidentified. The instability factor is the standardised average of the standardised five-year rolling standard deviations of broad-money and bank-credit growth, a denominator-free measure of financial turbulence that deliberately avoids the ratio levels whose behaviour the forensic diagnostics later question. Both factors are built on the estimation sample so that no look-ahead enters the standardisation.

The choice of estimators is dictated by two properties of the panel that are tested before any long-run coefficient is read. Cross-sectional dependence is assessed with the Pesaran (2021) CD statistic on each series; it rejects independence overwhelmingly for every trending variable, which rules out first-generation panel methods and mandates estimators and cointegration tests that are robust to common factors. Integration order is then established with the cross-sectionally augmented CIPS test of Pesaran (2007), which is valid under the dependence just documented and places the panel on the I(1) boundary in levels with stationary first differences, a few members stationary in levels and no evidence of I(2) behaviour. A single common factor accounts for roughly four-fifths of the differenced variance of income, so the equilibrium relation runs substantially through a shared regional trend rather than through purely idiosyncratic co-movement. These two facts, pervasive dependence and an I(1) panel governed by a common factor, are what the estimation strategy is built to accommodate.

The dynamic common correlated effects estimator handles the common factors the pre-tests reveal by augmenting (1) with cross-sectional averages of the dependent variable and the regressors and their lags,

$$\Delta y_{it}=c_i+{\phi }_i{y}_{i,t-1}+{\delta }^{\text{'}}{x}_{i,t-1}+{\displaystyle \sum _{\mathcal{l}=0}^{p_T}}{\lambda }^{\text{'}}{z̄}_{t-\mathcal{l}}+{\gamma }^{\text{'}}\Delta x_{it}+{\epsilon }_{it}$$

(2)

with ̄z the cross-sectional averages of income and the regressors, whose inclusion filters the latent factors so that β is recovered net of the shared regional movement. Cointegration is not inferred from φ but tested directly. Because the Pesaran (2021) CD statistics reveal strong common factors in every trending series, we apply the factor-based test of Banerjee and Carrion-i-Silvestre (2017): the differenced system is defactored by principal components, and a cross-sectionally augmented unit-root test on the idiosyncratic cointegrating residuals ê assesses the null of no cointegration,

$$\Delta {ê}_{it}={\rho }_i{ê}_{i,t-1}+{\displaystyle \sum _{j=1}^{k_i}}{\varphi }_{ij}\Delta {ê}_{i,t-j}+{\nu }_{it}$$

(3)

where rejection of a unit root in ê signals cointegration net of the common factor. The pooled mean group estimator of Pesaran et al. (1999), dynamic fixed effects with the covariance of Driscoll and Kraay (1998), and group-mean fully modified least squares (Phillips and Hansen, 1990; Pedroni, 2001) complete the set. The mean group estimator of Pesaran and Smith (1995) and system GMM are set aside as uninformative or infeasible at ten cross-sections, and dynamic OLS by degrees of freedom. Country-level bounds tests of Pesaran et al. (2001) with exact small-sample critical values corroborate the idiosyncratic dimension.

Inference and the robustness programme are fixed in advance of estimation. Long-run standard errors are the heterogeneity-robust group-mean errors for the mean-group and FMOLS families, the Driscoll and Kraay (1998) covariance for the dynamic fixed effects estimator, and the cross-sectionally augmented errors of the dynamic common correlated effects estimator; none relies on the assumption of cross-sectional independence the data reject. The robustness battery is pre-committed and spans a remittance control estimated on the five remittance-complete economies, a governance control that guards against the capabilities factor proxying generic institutional quality, a window truncated at 2019 that removes the pandemic and the most recent crises, and a full leave-one-out over the ten economies. A separate set of diagnostics interrogates the bank-ratio proxies directly, the within-country correlation of the ratio with remittances and its long-run coefficient in levels, so that the case for reframing the finance measure rests on estimates rather than assertion. The complete grid, including sub-index decompositions of the multidimensional measure, quadratic terms and alternative instability measures, is estimated but reported selectively in the interest of space.

4. Results

4.1. Pre-Testing and Cointegration

Two properties of the panel must be settled before any long-run coefficient can be trusted: whether the cross-sections move together, and whether the series are integrated of an order that admits a level relationship. Table 4 reports the diagnostics that answer both.

The pre-tests fix the two facts the strategy rests on. Cross-sectional dependence is overwhelming: the CD statistic rejects independence for every trending variable, above eighteen for income, capabilities and its components, which is what mandates cross-sectionally robust inference and a factor-based reading of cointegration. Integration is clean: the cross-sectionally augmented CIPS tests, which are robust to the dependence just documented, place most series on the I(1) boundary in levels and reject the unit root for all of them in first differences, an I(1) panel with a few stationary members and no I(2) trace. A single common factor accounts for about four-fifths of the differenced variance of income, so the equilibrium relation runs substantially through the common factor rather than through purely idiosyncratic co-movement.

Table 5. Panel cointegration evidence.

Cointegration test	Statistic	Inference	p
Banerjee-Carrion-i-Silvestre (2017), factor-based	-2.251	Z = -0.42	0.337
Dynamic common correlated effects, speed of adjustment	-0.377	t = -2.10	0.035
Mean group speed of adjustment	-0.651	t = -5.78	0.000

Notes: The Banerjee and Carrion-i-Silvestre (2017) test defactors the differenced system by principal components and applies a cross-sectionally augmented unit-root test to the idiosyncratic cointegrating residuals; its non-rejection is expected when the equilibrium operates through the common factor, and is the counterpart of the dynamic common correlated effects estimate, which recovers the relation net of that factor. Speeds of adjustment are the pooled and heterogeneous error-correction coefficients from the factor-augmented and mean-group error-correction models.

Table 5. Panel cointegration evidence.

Cointegration test	Statistic	Inference	p
Banerjee-Carrion-i-Silvestre (2017), factor-based	-2.251	Z = -0.42	0.337
Dynamic common correlated effects, speed of adjustment	-0.377	t = -2.10	0.035
Mean group speed of adjustment	-0.651	t = -5.78	0.000

Notes: The Banerjee and Carrion-i-Silvestre (2017) test defactors the differenced system by principal components and applies a cross-sectionally augmented unit-root test to the idiosyncratic cointegrating residuals; its non-rejection is expected when the equilibrium operates through the common factor, and is the counterpart of the dynamic common correlated effects estimate, which recovers the relation net of that factor. Speeds of adjustment are the pooled and heterogeneous error-correction coefficients from the factor-augmented and mean-group error-correction models.

Cointegration is read here in the manner these data require. The factor-based test of Banerjee and Carrion-i-Silvestre (2017), applied to the defactored system, does not reject the null once the common factor is removed, which is the expected signature of an equilibrium that runs through that factor rather than through idiosyncratic residuals. The dynamic common correlated effects estimator, which absorbs the same factor through cross-sectional averages, recovers a significant speed of adjustment, and the heterogeneous mean-group speed corroborates it. This is cointegration via common factors in the sense of Chudik and Pesaran (2015), not residual cointegration, and it is exactly what a region moving on shared reform and price cycles should display. The country-level bounds tests reported later supply the idiosyncratic complement.

4.2. What the Ratio Proxy Measures

Before adopting the multidimensional index, it is worth establishing why the bank-ratio proxies that dominate the regional literature produce such unstable results. Two frozen diagnostics on the five oil-diversified economies, where such ratios are economically meaningful, make the case. The first asks whether the ratio tracks remittances within countries, the mechanism most often invoked to explain regional anomalies; the second asks whether the ratio carries any long-run signal at all once remittances and openness are controlled.

Table 6 dispatches the most popular explanation for regional instability. If remittances contaminated the ratio in a common way, the within-country correlation would share a sign; instead it runs from -0.78 in Turkiye to 0.67 in Tunisia. There is no single contamination channel; each economy blends deposits, credit and diaspora flows in its own way, which is already enough to make a pooled ratio coefficient unstable from one sample to the next.

Table 7 completes the diagnosis. The ratio does carry a long-run coefficient, 0.054 in the high-remittance trio and 0.103 across the five, so it is not empty; but it is small, and the remittance coefficient beside it swings from an insignificant 0.153 in the trio to a significant 0.035 across the five, confirming that the ratio and remittances are entangled differently in each configuration. A proxy whose partner term changes sign and significance with the sample cannot anchor a stable estimate, which is exactly the instability the regional literature reports. None of this afflicts the multidimensional index used below, whose capabilities reading is stable across four estimator families, and that contrast is the case for changing the measure.

4.3. What the Finance Measure Captures

Figure 1 shows the two components of the measurement problem moving together in most economies. Their within-country correlation is 0.40 across the ten economies, lower than in the smaller core but still high enough that estimators forced to separate institutional finance from schooling disagree sharply. A specification entering the two separately illustrates the point: the level-cointegration family loads the common movement onto schooling (0.521) and leaves finance a small positive residue (0.030), whereas the cross-sectionally robust dynamic fixed effects family reverses the ranking (0.187 for finance against 0.137 for schooling, the latter insignificant). The two are facets of one developmental trend, and forcing a split manufactures exactly the estimator-dependent contradictions the regional literature reports. The response is to model the bundle.

4.4. The Capabilities Resolution

Modelling the bundle rather than its collinear parts yields the central result. Table 7 estimates the long-run relation with the capabilities factor in place of separate finance and schooling terms, across the four estimator families whose assumptions were set out in the estimation strategy.

Table 8 is the destination, and its columns agree. The capabilities factor carries a long-run elasticity of 0.167 under pooled mean group, 0.242 under dynamic fixed effects with cross-sectionally robust errors, 0.181 under group-mean fully modified least squares and 0.163 under the factor-augmented dynamic common correlated effects estimator. Four families that disagree about almost every nuisance parameter place the capabilities elasticity between roughly 0.16 and 0.24: a one standard deviation advance of the joint institutional and human capital trend is associated with about sixteen to twenty-four per cent higher long-run income per capita. Openness is positive throughout and significant in three of the four families, its single insignificant reading appearing in the reference dynamic fixed effects column, which we report as it stands. Investment and inflation enter with the expected signs where significant, while instability carries no robust long-run role. The pooled and heterogeneous speeds of adjustment, near -0.15 and -0.38, place the return to equilibrium at a horizon of several years.

4.5. Where the Relation Is Nationally Visible

The panel estimate pools economies that differ widely in how far their capabilities have moved. Country-level bounds tests show where a long-run relation is individually detectable and where the span is simply too short, which explains what the pooled coefficient is averaging. Table 9 reports them.

Country evidence supplies the idiosyncratic complement to the factor-based cointegration established in the pre-tests and explains the panel geometry. Egypt, Libya, Morocco and Turkiye reject the absence of a level relationship individually, and Egypt, Morocco and Turkiye carry positive and significant capabilities multipliers (0.649, 0.234 and 0.490). These are among the economies with the largest within-country movements in the capabilities components; economies whose institutional and schooling series barely move over the window cannot reject individually at their span lengths, which is the standard rationale for pooling and the reason the panel estimates of Table 7 are the disciplined reading of the country-level pattern.

4.6. Robustness and Boundaries

The robustness programme is fixed before estimation. The capabilities coefficient stays positive and significant in the non-pooled families under a remittances control on the five remittance-complete economies (0.168 under dynamic fixed effects, 0.307 under FMOLS), under a governance control that answers the objection that capabilities merely proxy generic institutional quality (0.264 and 0.186), in the window closing at 2019 that excludes the pandemic and the recent crises (0.208 and 0.207), and under exclusion of any single economy. The one disclosed boundary is the exclusion of Libya, where the dynamic fixed effects reading softens to 0.097 while the pooled mean group and FMOLS families hold at 0.177 and 0.214; the effect is carried by the level and pooled families where one estimator weakens, which is the pattern expected when identification leans on the economies that moved most. Openness is more fragile than capabilities, significant under pooled mean group and FMOLS throughout but not in the reference dynamic fixed effects column, and we present it as such rather than selecting the specifications that flatter it.

Two further boundaries are disclosed rather than argued away. Cross-sectional dependence, documented in the pre-tests, is the reason cointegration is assessed through the factor-based test and the relationship estimated with the factor-augmented dynamic common correlated effects model; removing year means outright would extinguish the trend variables, because the shared regional movement is part of what is being measured, and separating common shocks from common reform waves would require a cross-sectional dimension the region does not offer. The pooled mean group column is read alongside the others rather than alone, its likelihood admitting more than one interior optimum at this dimension. The full estimation grid, including the sub-index decompositions of the multidimensional measure, quadratic terms and alternative instability measures, is available from the corresponding author.

5. Discussion

The findings speak directly to the disagreements catalogued in the background, and largely dissolve them. The negative or insignificant banking coefficients of Ben Naceur and Ghazouani (2007) and the positive Maghreb estimates of Khattab et al. (2015) look irreconcilable as substantive claims, but Table 6 and Table 7 show they are separated by exactly the fragility the ratio proxy carries: the same economies deliver a small and unstable long-run ratio coefficient whose remittance partner term flips sign and significance from one sample to the next. Studies that omit remittances and studies that include them are estimating different things while appearing to estimate the same thing. Recent regional work reaches the same junction from the other side: Mo (2026) finds remittances and financial development acting as substitutes in MENA, the local financial sector using diaspora inflows inefficiently, which is consistent with the sign-unstable, economy-specific relation between the ratio and remittances that our forensics document. The horizon reversals of Loayza and Rancière (2006) find a parallel here: the ratio-based instability term switches sign with the sample window, which recasts several short-run instability results in the regional literature as denominator artefacts rather than Minskyan dynamics.

Against the measurement literature of Sahay et al. (2015) and Svirydzenka (2016), the results both endorse and qualify the multidimensional turn. The institutions index behaves far better than the ratios, consistent with those authors, but the paper shows that even the index does not identify a separate finance effect, because financial-institutional development and schooling move as one trend with a within-country correlation of 0.40 across the ten economies. This is where the paper departs from the entire prior regional literature, which enters finance and human capital as distinct regressors: the capabilities elasticity of roughly 0.16 to 0.24 in Table 7 is what that literature has been approximating, imprecisely, whenever its finance and education coefficients traded significance across specifications. That the two co-develop is not a regional peculiarity, appearing as bidirectional causality in emerging Asian panels (Hong Vo et al., 2021) and as an asymmetric long-run link elsewhere (Ha and Ngoc, 2022); what the prior work stops short of is the inference that near-collinear co-development leaves a separate finance coefficient unidentified rather than merely imprecise. The threshold evidence of Özmen and Taşdemir (2026), in which the growth effect of financial development is strongest in less financially developed economies and rises with institutional quality, is consistent with reading finance as one facet of a capability that institutions and human capital jointly constitute. Read this way, the institution-conditioned effects of Boukhatem and Ben Moussa (2025) are a special case, institutions being one facet of the same capability bundle. The apparent conflict between Kamal et al. (2023), who find finance growth-promoting, and Dahmani et al. (2023), who find it growth-retarding, is on this account a measurement disagreement rather than a substantive one: the former reads the multidimensional index while the latter reads a cross-sectionally augmented credit specification, and the two proxies capture different things in the way the ratio forensics predict. The inverted-U of Shaddady (2023), in which financial institutions eventually constrain growth, is consistent with a capabilities reading in which the institutional facet alone, detached from the human capital it normally travels with, loses its association with income.

Two results run against expectations and are worth stating as such. Trade openness, which much of the growth literature treats as secondary to finance, is here the more consistently significant lever, surviving in three estimator families where the finance-only proxies survive in none. That openness dominates is echoed in the most recent regional evidence: Idrissi (2026) likewise finds trade openness and investment significant for the region over the same window while instability enters negatively, a pattern our estimates reproduce. And the meta-analytic finding of Valickova et al. (2015) that the finance effect fades in developing economies is, on this evidence, less a fading of the effect than a fading of the ratio that measures it: reframed as capabilities, the long-run relationship is present and precisely estimated. That the relation is sharpest in the economies whose capabilities moved most, Egypt and Turkiye, is not a caveat but a confirmation of the mechanism: a capabilities effect should register precisely where the capability shift is large, and the panel identifies it exactly there, then transmits it to the rest of the region through the common factor the estimators share.

Policy Implications

The implications below follow from specific estimates rather than from the finance-growth literature in general. First, on measurement: because the ratio coefficient in Table 6 moves from statistical insignificance to 0.049 with a single remittance control and vanishes when the crisis years are excluded, credit and money ratios cannot be used to score financial-development progress in economies where remittances exceed several per cent of GDP or where the denominator is volatile. Regional surveillance should read the multidimensional financial-institutions index, whose institutions component is the part that carries the long-run signal in the decomposition, and should discount ratio-based rankings for exactly the economies, Morocco, Egypt, Tunisia, where remittances are largest.

Second, on where to invest: the capabilities elasticity of 0.16 to 0.24 per standard deviation, together with the 0.40 within-country correlation between financial-institutional development and schooling, means the two cannot be costed as independent projects with additive returns. A ministry that funds bank branches while schooling stagnates is moving one coordinate of a bundle whose payoff is estimated jointly; the return to financial-institutional access is not identified separately from the return to education at this scale. Planning should therefore target the joint capability, and where a single lever must be chosen, the estimates give no basis for preferring financial deepening over schooling, which reverses the usual regional emphasis on financial-sector reform as a standalone growth strategy.

Third, on sequencing and horizon: trade openness carries a long-run coefficient between 0.26 and 0.37 in the estimators that identify it, larger and more consistently signed than any finance-only proxy in the paper, so openness belongs in the first rank of levers rather than as a control. And the speeds of adjustment near -0.15 to -0.38 imply that a capabilities reform closes only a tenth to a third of the gap to the new equilibrium each year, so evaluations set on a three-to-five year political cycle will understate returns that accrue over a decade. The two economies where the relation is individually detectable, Egypt and Turkiye, are those whose capabilities moved most, which tells a reforming government that effects become visible only once the underlying capability shift is large.

6. Conclusions

For three decades the question of whether financial development raises long-run income in the region has produced a literature that contradicts itself, study by study, with signs that flip and significance that migrates across specifications. This paper has argued that the contradiction is manufactured rather than substantive, and that it dissolves once two disciplines are imposed. The sample stops being a choice: a single coverage rule applied blind to the entire regional universe fixes the ten MENAT economies, so the estimates cannot be attributed to a convenient selection of countries. And the finance proxy stops being taken at face value: at a constant sample the bank ratio carries only a weak long-run signal, entangled with remittances in a way that has no common sign across economies, so it cannot anchor the stable estimate the literature has read into it. The disagreement, on this evidence, lives in the measure and the sample rather than in the economies.

What survives the correction is a single, well-identified object. Institutional financial development and schooling move as one slow trend within countries, with a within-country correlation of 0.40, and no estimator can allocate that trend between two near-collinear claimants at this cross-sectional dimension. Modelling the bundle rather than its parts yields a long-run income elasticity near 0.16 to 0.24 per standard deviation that holds across four estimator families resting on different assumptions, from pooled mean group and dynamic fixed effects to group-mean fully modified least squares and a dynamic common correlated effects estimator that reads the relation through the common movements the region so plainly shares. Cointegration is established in that same spirit, through a factor-based test rather than an adjustment speed, and the capabilities elasticity is the most stable quantity the panel produces. Trade openness emerges as the more consistently identified lever than any finance-only proxy, a result the region’s recent evidence independently corroborates.

The implications reach beyond the region. For measurement, the results caution against scoring financial-development progress with credit and money ratios wherever remittances are large or denominators volatile, and they favour the multidimensional financial-institutions index whose institutions component carries the durable signal. For the finance-growth debate more broadly, the paper reframes a long-standing puzzle: where finance and human capital co-develop, a separate finance coefficient is unidentified rather than simply imprecise, and the honest estimand is the joint capability. That reframing turns a catalogue of conflicting regional results into a coherent picture in which the apparent conflicts are measurement artefacts, and it suggests that the growth payoff to financial deepening cannot be costed independently of the schooling it travels with.

The remaining questions are matters of scope rather than of doubt about the result. A wider cross-section, drawing on economies outside the region that share the same capabilities structure, would help separate common global shocks from common regional reform waves, which the present dimension cannot fully disentangle. Later vintages of the multidimensional indices, once published, will bring the recent crisis years inside the estimation window rather than at its edge, allowing the capabilities relation to be tested through a period of unusual volatility. Neither extension threatens the central finding; both would sharpen its edges. The region has been asking whether finance or education drives growth, and the answer this paper returns is that the question itself, posed as a choice, is the mistake.