A panel data test for poverty traps

Abstract

This article develops a threshold panel data nonlinearity test for poverty traps. The new testing strategy extends the work on nonlinearity tests for panel data by considering threshold nonlinearities in the fixed-effects components. Monte Carlo simulations are conducted to evaluate the finite-sample performance of these tests. The tests are applied to the relationship between Gross Domestic Product (GDP) per capita and capital stock per capita. Our application to a panel of countries for the period 1973 to 2007 uncovers the presence of two regimes determined by the level of capital stock per capita. The conclusions from our test also support the existence of a poverty trap determined by a capital stock per capita level at the 11% quantile of its pooled worldwide distribution.

Keywords:

• nonlinearity tests
• panel data
• poverty traps
• threshold models

JEL Classification::

• C33
• C12
• C13
• O1

Acknowledgements

We are grateful to the excellent research assistance provided by Alexander Dentler. We are also grateful to an anonymous referee and participants of the Econometric Society World Congress held in Shanghai, 2010. This project was supported by a 2009/2010 Grant from City University London.

Notes

¹ While our work is framed within the dual-equilibrium paradigm, the existence of poverty traps is also contested on the empirical growth literature. For instance, Jones and Olken (2008) claim that ‘almost all countries in the world have experienced rapid growth lasting a decade or longer’, and that ‘economic growth can be easily reversed, often leaving countries no better off than they were prior to the expansion’ (p. 584).

² Threshold nonlinearity tests for panel data have been recently applied by Nautz and Scharff (2012) for the case of inflation and relative price variability.

³ We use a very stylized model. Different production functions can be used here for controlling education levels (in a Mincerian sense) or other observable variables. In the following subsection, we argue that these variables could be captured by the country-specific fixed effect.

⁴ We thank an anonymous referee for pointing this out.

⁵ The same methodology can be applied to study the existence of convergence clubs in growth rates. The linear model is

with Δy _it  = y _it  − y _i0, and y _it and y _i0 denoting terminal and initial per capita income of country i, respectively. The alternative model is

The rate of convergence for each ‘club’ is characterized by the parameters α₁₁ and α₁₂.

⁶ The sample contains Afghanistan, Albania, Algeria, Angola, Antigua and Barbuda, Argentina, Australia, Austria, Bangladesh, Barbados, Belgium, Belize, Benin, Bhutan, Bolivia, Botswana, Brazil, Bulgaria, Burkina Faso, Burundi, Cambodia, Cameroon, Canada, Cape Verde, Central African Republic, Chad, Chile, Colombia, Comoros, Dem. Rep. Congo, Costa Rica, Cuba, Cyprus, Denmark, Djibouti, Dominica, Dominican Republic, Ecuador, El Salvador, Equatorial Guinea, Ethiopia, Fiji, Finland, France, Gabon, The Gambia, Germany, Ghana, Greece, Grenada, Guatemala, Guinea, Guinea-Bissau, Guyana, Haiti, Honduras, Hungary, Iceland, India, Indonesia, Ireland, Israel, Italy, Jamaica, Japan, Jordan, Kenya, Kiribati, Kuwait, Lebanon, Lesotho, Liberia, Libya, Luxembourg, Madagascar, Malawi, Malaysia, Maldives, Mali, Malta, Mauritania, Mauritius, Mexico, Fed. Sts. Micronesia, Mongolia, Morocco, Mozambique, Namibia, Nepal, The Netherlands, New Zealand, Nicaragua, Niger, Norway, Oman, Pakistan, Panama, Papua New Guinea, Paraguay, Peru, The Philippines, Poland, Portugal, Puerto Rico, Qatar, Romania, Rwanda, Saudi Arabia, Senegal, Seychelles, Sierra Leone, Singapore, Solomon Islands, Somalia, South Africa, Spain, Sri Lanka, St. Lucia, Sudan, Suriname, Swaziland, Sweden, Switzerland, Tanzania, Thailand, Togo, Tonga, Tunisia, Turkey, Uganda, United Arab Emirates, United Kingdom, United States, Uruguay, Vanuatu, Vietnam, Zambia and Zimbabwe.

⁷ For GDP per capita, it has a Levin et al. (2002) pooled Dickey–Fuller test coefficient of −0.04685 without lags and −0.05258 augmented one lag. For capital stock per capita, it has a Levin et al. (2002) pooled Dickey–Fuller test coefficient of −0.02272 without lags and −0.05623 augmented one lag. In all cases the null hypothesis is rejected with a p-value < 0.01.