Stylized Fact

As shown in Figures 4 and 5, per capita GDP growth is negatively correlated with debt-equity ratio in financial and non-financial corporations. According to preliminary investigations based on an OLS fitted line, the coefficients of non-financial and financial corporations’ debt–equity ratio are –1.120 and –0.23. These values are obviously overestimated in the absence of control variables.

Figure 4.

Average leverage in non-financial corporations and per capita GDP growth.

Figure 5.

Average leverage in OECD financial corporations and per capita GDP growth.

Baseline Model

In modern economic growth theories, economic growth is considered as the outcome of various endogenous factors that vary, from the behaviour of rational optimizing agents to technology, and macroeconomic policies.Reference Turnovsky²⁷ For instance, the role of entrepreneurs, investments in research and development, fiscal policy, tax structure and stable government revenues affect economic growth. As high corporate debt levels are due to debt-biased corporate tax systems, corporate debt levels are supposed to affect economic growth endogenously.

For this purpose, without loss of generality, the Cobb-Douglas production function can be written as:

$$Y\left( t \right) = A\left( t \right)K\left( t \right)^\alpha H\left( t \right)^{1 - \alpha } $$ (1)

where Y is the output, K is the capital, A is the technological advancement, H is the labour and α and 1 − α are the capital and labor productivities respectively. Following Lee and Gordon, the production function in equation (1) can be modified for the growth in the economy’s output as:Reference Lee and Gordon²⁸

$${{\dot y} \over y} = \dot a + \alpha \left( {{\dot k} \over k} \right) + \left( {1 - \alpha } \right)\left( {{\dot h} \over h} \right) + \dot \eta $$ (2)

where $\dot y/y$ is the growth rate per unit of time in the output; mathematically, ${{\dot y} \over y} = {\rm{}}{{{\rm{d}}Y\left( t \right)/{\rm{d}}t} \over {Y\left( t \right)}}$ . Likewise, ${{\dot k} \over k} = {\rm{}}{{{\rm{d}}K\left( t \right)/{\rm{d}}t} \over {K\left( t \right)}}$ and ${{\dot h} \over h} = {\rm{}}{{{\rm{d}}H\left( t \right)/{\rm{d}}t} \over {H\left( t \right)}}$ are the growth rate per unit of time in capital and labor respectively, $\dot a = {\rm{}}{{{\rm{d}}A\left( t \right)} \over {{\rm{d}}t}}$ is the change in technology over time, and $\dot \eta $ is the residual term.

Lee and Gordon state that any positive spillovers of productivity on economic growth would appear in the term ȧ.Reference Lee and Gordon²⁸ Likewise, any spillovers of debt-bias corporate tax systems should appear through ȧ. For instance, any change in the tax system that affects the corporation’s financing choice affects economic growth. This is because any level of the debt–equity ratio or capital structure is associated not only with the tax benefits but also with the government revenue and investment level. Equation (2) requires to be controlled for capital stock, savings, education, and average hours worked by the labour force.Reference Beck, Levine and Loayza^10, Reference Lee and Gordon²⁸ In addition, government budget deficits are added in equation (2) for capturing the potential impact of fiscal variables. With the above-mentioned outline, the specification for a panel data analysis of economic growth can be written as:

$$Growth_{it} = {\rm{}}\beta _{it} Leverage + \gamma _{it} Deficits + \delta _{it} X_{it} + \mu _i + e_t + \varepsilon _{it} $$ (3)

where Growth_it refers to growth in GDP Per Capita, leverage is corporate debt levels or debt equity ratio. Deficits represents government deficits with and without interest payments, X_it is the set of all the macroeconomic factors, μ_i is the country-specific effects, e_t represents the time-specific effects, ε_it is the error term that varies with time, and finally, i and t denote the country and year, respectively.

The government budget deficits are important for two reasons, first, because the effect of budget deficits on economic growth is an inevitable factor in tax policy. Second, consistent with the objective of this study, a decrease in government tax revenue due to high corporate debt levels increases government budget deficits. The cash deficit/surplus ratio is added in order to capture the government size. The cash deficit/surplus ratios both with interest payments and without interest payments in the model not only account for the government size but also incorporate the government’s decisions, which are taken to keep the debt levels sustainable. Moreover, the government cash deficit/surplus ratio determines the fiscal sustainability, which is important for economic growth.Reference Barro²⁹ The model is controlled for factors such as inflation, education, average hours worked by the labour force, net capital stock growth, savings, and population growth.

Finally, for the robustness of results, multiple models with various sets of variables are estimated. This is because owing to the small N (number of countries) and T (number of years), the data could not be divided into two or more periods nor could the countries be divided with respect to income or other characteristics. We use multiple models with various sets of variables including population growth, savings, net capital stock growth, government deficits, labour force education, inflation, and working hours for the robustness of the results.

Preliminary Tests

The preliminary step in estimating equity (equation (3)) is to test for the unit root. The panel unit root tests play an important role not only in avoiding spurious regression due to the existence of a unit root in the series but also in devising an appropriate technique for the problem at hand. We selected panel unit root tests of the Breitung testReference Breitung³⁰ based on their highest power as well as the size that is close to normal size,Reference Hlouskova and Wagner³¹ whereas, according to Maddala and Wu,Reference Maddala and Wu³² for balanced panels, the Im-Pesaran-Shin test (IPS)Reference Im, Pesaran and Shin³³ is appropriate.

Table 1 provides the Breitung and IPS tests results. The results for the Breitung test in column 2 show that all series are stationary except for the financial corporation leverage, non-financial corporation leverage, population growth, average hours worked by the labour force, and secondary education. The IPS tests in column 3 reveal that all series are stationary in levels, except for non-financial corporation leverage, population growth, average hours worked in the labour force, secondary education and savings. This indicates that the unit root tests might have different results.

Table 1.

Panel unit root test results.

() parenthesis contain p-values, IPS represents Im-Pesaran-Shin panel unit root test.

As can be observed in column 4, the data are a mixture of stationary I(0) and first-differenced stationary I(1) variables. For instance, GDP per capita growth, bank debt-equity ratio, net capital stock growth, deficit/surplus ratio including interest payment, deficit/surplus ratio excluding interest payments and inflation are stationary in level. The rest of the variables, including financial corporation leverage, non-financial corporations leverage, population growth, average hours worked, secondary education and saving as a percentage of GDP are non-stationary in levels. The panel ARDL estimation technique is appropriate for data that contain stationary and first-difference stationary series.Reference Im, Pesaran and Shin^33, Reference Im, Pesaran and Shin³⁴ The panel ARDL cannot be applied to second-difference series I(2). As shown in Table 1, all the series are either stationary at level or stationary at first difference. The two tests of Breitung and IPS were applied to all the series that were non-stationary at level, which confirmed that the series are stationary at first difference. To save space, the results of the panel unit root tests applied at first difference are omitted.

Model Specification

Without loss of generality, following Blackburne and Frank, an ARDL (p, q₁, …, q_k) dynamic panel specification can be written in the following formReference Blackburne and Frank³⁵:

$$y_{it} = \mathop \sum \limits_{j = 1}^p \lambda _{ij} y_{i,t - j} + \mathop \sum \limits_{j = 0}^q \delta _{ij}^{\prime} X_{i,t - j} + \mu_i + \epsilon_{it} $$ (4)

where i is the number of countries, t is the number years, X_{i, t} is the k × 1 vector of explanatory variables, $\delta _{ij}^{\prime} $ is the k × 1 vector of coefficients, λ_ij are the scalars, and μ_i is the country specific effect, p and q are the lags of dependent and independent variables respectively. To incorporate short-run disequilibrium, the specification in equation (4) can be reparametrized into an error correction equation as:

$$\Delta y_{it} = \phi _i \left( {y_{i,t - 1} - \theta _i^{\prime} X_{it} } \right) + \mathop \sum \limits_{j = 1}^{p - 1} \lambda _{ij}^{\rm{*}} \Delta y_{i,t - j} + \mathop \sum \limits_{j = 0}^q \delta _{ij}^{{\prime*}} \Delta X_{i,t - j} + \mu_i + \epsilon_{it} $$ (5)

where, $ \phi _i = \bigg(1 - \mathop \sum \limits_{j = 1}^p \lambda _{ij} \bigg),\,\theta _i = {{\mathop \sum \nolimits_{j = 0}^q \delta _{ij} } \over {\big(1 - \mathop \sum \nolimits_k \lambda _{ik} \big)}},{\rm{}}\lambda _{ij}^{\rm{*}} = \, - \,\mathop \sum \limits_{m = j + 1}^p \lambda _{im} ,\,j = 1,2, \ldots ,p - 1$

$$\delta _{ij}^{{\prime*}} = \mathop \sum \limits_{m = j + 1}^p \delta _{im} ,\,j = 1,2, \ldots , q - 1$$.

In the above error correction model, the parameter of interest is Φ. If Φ = 0, it implies that there is no long-run relationship. If the value for Φ is negative, statistically significant and less than 1 in absolute terms, it implies that there exists a relationship between economic growth and corporate debt levels. However, $\left| \phi \right| \ge \left| { - 1} \right|$ implies that there is overcorrection, which is meaningless. The models given in equation (4) or equation (5) can be estimated either as Pooled Mean Group (PMG), Mean-Grouped (MG), and Dynamic Fixed Effects (DFE) models. The mean-group (MG) estimates the model for each country separately and then simply calculates an arithmetic average of the coefficients.Reference Pesaran and Smith ³⁶ As the model is estimated for each country separately, the intercepts, slope coefficients, and error variances are allowed to differ across countries. The dynamic fixed effect approach on the other hand pools the data and only allows the intercept to vary across countries by introducing dummies for each country. Finally, the pooled mean group rests somewhere in between the MG and DFE because it combines the pooling of DFE and averaging of MG.Reference Im, Pesaran and Shin ^{33 ,} Reference Im, Pesaran and Shin ³⁴ PMG allows the intercept, short-run coefficients, and error variances to differ across the groups by mimicking the MG and restricts the long-run coefficients to remain constant across countries, as in the case of DFE.

The PMG estimator is efficient and consistent because it restricts long-run elasticities to be the same across countries; however, it is prone to a problem of heterogeneous slope parameters across countries. In this case, PMG becomes inconsistent whereas MG remains consistent whether the slope parameter is heterogeneous or otherwise. The consistency of PMG regarding homogeneity can be checked by the Hausman test.Reference Hausman ³⁷ The null hypothesis of the Hausman test states that both MG and PMG are consistent but only one of them is efficient. Given that PMG is consistent and efficient provided that the slope parameters for countries are homogeneous, the acceptance of a null hypothesis implies that PMG is consistent and efficient. DFE, on the other hand, is prone to simultaneous equation bias or endogeneity problems. Again, using the Hausman test, the extent of endogeneity problems can be checked. If the endogeneity problem is nominal, DFE is preferred over MG for its efficiency.