In this paper we show that the double Pareto lognormal (DPLN) parameterization provides an excellent fit to the overall US city size distribution, regardless of whether "cities" are administratively defined Census places or economically defined area clusters. We then consider an economic model that combines scale-independent urban growth (Gibrat's law) with endogenous city creation. City sizes converge to a DPLN distribution in this model, which is much better in line with the data than previous urban growth frameworks that predict a lognormal or a Pareto city size distribution (Zipf's law).