In this paper Kuznets' U—curve hypothesis is tested on two unbalanced panel data sets of 47 and 62 countries, for the period 1970-93, using two-way fixed and random effects models. Several competing model specifications are estimated and the one best fitting the data is selected by appropriate model selection procedures. It is shown that there is no hard empirical evidence to support the usual econometric model formulations and the U—curve hypothesis.