We test implications of economic geography models for location, size and growth of cities with US Census data for 1900 û 1990. Our tests involve non-parametric estimations of stochastic kernels for the distributions of city sizes and growth rates, conditional on various measures of market potential and on features sizes of neighbors. We show that while these relationships change during the twentieth century, by 1990 they stabilize such that the size distribution of cities conditional on a range of spatial variables are all roughly independent of these conditioning variables. In contrast, similar results suggest that there is a spatial element to the city wage distribution. Our parametric estimations for growth rates against market potential, entry of neighbors, and own lagged population imply a negative effect of market potential on growth rates, unless own lagged population is also included, in which case market potential has a positive effect and own lagged population a negative one. Cities grow faster when they are small relative to their market potential. In total, our results support some theoretical predictions, but also provide a number of interesting puzzles.
