Is the speed of convergence constant?

Empirical attempts to measure the speed of convergence -- the rate at which a country's per capita income approaches its steady state relative to its distance from its steady state -- have started from the assumption that it is constant. In contrast, neoclassical models of capital accumulation usually predict that the speed of convergence decreases as income approaches its steady state. Estimating a flexible functional form which allows the speed of convergence to vary suggests that the speed of convergence actually increases as income approaches its steady state. An increasing speed of convergence calls into question structural interpretations of coefficients on conditioning variables in cross-sectional growth regressions. Instead, excluding initial income from cross-sectional growth regressions allows coefficients on exogenous variables to be interpreted as measuring changes in underlying structural relationships.