Entrepreneurship and income inequality: a spatial panel data analysis

Abstract

The main objective of this paper is to empirically analyze the relationship between entrepreneurship and income inequality. We use a spatial panel data analysis for both 33 high-income countries and 39 middle- and low-income countries over a period of 11 years. Estimation results and rigorous diagnostic analysis suggest that: (i) there is a strong support for the existence of an inverted U-shaped relationship between entrepreneurship and income inequality espoused by the Kuznets Curve hypothesis; (ii) the relationship between entrepreneurship and income inequality is negatively moderated by country’s level of economic development; (iii) regardless of income inequality levels, entrepreneurship has a non-linear relationship with income per capita; (iv) gross domestic expenditure on research and development exhibits significant negative impacts on entrepreneurship; (v) significant mixed effects on the likelihood of entrepreneurial activity are observed with governance, globalization, population growth rate, and competitiveness variables; (vi) there are significant mixed feedback effects on entrepreneurship; and (vii) there are statistically significant, positive as well as negative spatial spillovers to country-level entrepreneurial activity.

Keywords:

• Entrepreneurship
• income inequality
• spatial panel data model
• feedback effect
• direct effect
• spillover effect

JEL classification:

• C21
• C23
• D31
• L26
• R12

Notes

1. With the exception of the matrix based on the common border countries, the spatial weight matrix used in this paper is based on the calculation of distances using the spherical distance between geographic centroids of the countries. Formally, the spherical distance (in kilometers) between the centroids of two countries is defined as follows: . X_i denotes the latitude of the centroid of country i, while Y_i is the longitude of the centroid of country i.

2. In this paper, we use the matrix of k-nearest neighbors (Pace and Barry 1997; Pinkse and Slade 1998; Baller et al. 2001). Its general form is defined as follows:

where is an element of the weight matrix, is an element of the row-standardized matrix, is the threshold value defined for each country i. More precisely, is the smallest distance of order k between country i and country j, such as the country i has exactly k neighboring countries.

3. According to LeSage and Pace (2009), the LM test statistics only require estimation of a non-spatial panel model associated with the null hypothesis that the SAC coefficient is equal to zero.

4. Elhorst (2014a, 2014b) argue that one may use conditional LM tests which test for the existence of one type of spatial dependence conditional on the other. For a mathematical derivation of these tests for a spatial panel data model with spatial fixed effects, see Debarsy and Ertur (2010). The difference between the robust and conditional LM tests is that the first are based on the residuals of non-spatial panel data models and the second on the ML residuals of the panel SAR or panel SEM model.

5. The PCA methodology consists of seven steps: constructing a data matrix, standardizing variables, computing the correlation matrix, finding eigenvalues (to rank a new set of variables or PCs) and eigenvectors, selecting PCs (based on stopping rules), interpreting the results and calculating scores.

6. The fundamental idea of PCA is to reduce the dimensionality of a data-set consisting of a large number of interrelated variables, while retaining as much as possible of the variation present in the data-set. This is achieved by transforming it into a new set of variables or principal components (PCs), which are uncorrelated and ordered so that the first few retain most of the variation present in all of the original variables. In summary, it can be said that PCA is a variable reduction technique that can be used when variables are highly correlated. It reduces the number of observed variables to a smaller number of PCs that account for most of the variation of the observed variables and is a large sample procedure (Manly 1994; Sharma 1996; Joliffe 2002).

7. For the panel of HIC, KMO = 0.895 and Bartlett’s test p-value = 0.0000 (Approx. Chi-Square = 2891.245; df = 15). For the panel of MLIC, KMO = 0.858 and Bartlett’s test p-value = 0.0000 (Approx. Chi-Square = 1944.535; df = 15). These results indicate that PCA fits the data well.

8. The list of HIC includes Austria, Belgium, Canada, Chile, Croatia, Cyprus, Czech, Denmark, Estonia, Finland, France, Germany, Greece, Hungary, Iceland, Ireland, Italy, Japan, Latvia, Lithuania, Luxembourg, Netherlands, Norway, Poland, Portugal, Slovakia, Slovenia, Spain, Sweden, Switzerland, United Kingdom, United States of America, and Uruguay. The list of MLIC includes Argentina, Armenia, Bolivia, Bosnia and Herzegovina, Botswana, Brazil, Bulgaria, Burkina Faso, Cambodia, China, Colombia, Congo Democratic Republic, Costa Rica, Dominica, Dominican Republic, El Salvador, Ethiopia, Georgia, Ghana, Guatemala, India, Indonesia, Jamaica, Jordan, Kazakhstan, Kyrgyz Republic, Macedonia, Malawi, Malaysia, Mauritius, Moldova, Montenegro, Panama, Peru, Romania, Russian Federation, Tunisia, Turkey, and Ukrain.

9. For more details, see Marquardt (1970, 1987), Mason and Perreault (1991), and O’Brien (2007).

10. To conserve space, the results are not reported here but can be available from the authors upon request.

11. The Global Moran’s I statistic is given as where n is the number of observations (countries). w_ij is the spatial weight of the link between i and j, and S₀ is a standardization factor which equals the sum of the spatial weights, i.e. . z_i is, in our case, the annual entry density (in natural logarithms) for country i and is the mean value of logged entry density., .

12. The results of the panel SDM model without the bias correction are almost identical.

13. Alvarez, Barbero, and Zofío (2017) use a generalized spatial two-stage least squares (GS2SLS) estimator for spatial panels following Kapoor, Kelejian, and Prucha (2007) and Baltagi and Liu (2011).

14. The turning point is calculated by , where β₁ is the coefficient on the log of GINI and β₂ is the coefficient on the log of GINI, whole quantity squared.

15. Unlike the other variables though, the coefficients estimates on income inequality and per capita income could not be interpreted as elasticities since we specify a quadratic polynomial of both inequality and income.

16. The turning point is computed by , where β₃ is the coefficient on the log of GDP and β₄ is the coefficient on the log of GDP, whole quantity squared.

17. Necessity entrepreneurship arises when people lack sufficient labor market opportunities.