A Nonparametric Estimation of the Local Zipf Exponent for all US Cities

Abstract

The methodology proposed by Ioannides and Overman (2003 Regional Science and Urban Economics 33 127–137) is applied to estimate a local Zipf exponent using data for the entire 20th century of the complete distribution of cities (incorporated places) without any size restrictions in the US. First, kernel regressions are run using the Nadaraya–Watson estimator, excluding some atypical observations (5.66% of the sample). The results reject Zipf's law from a long-term perspective, but the evidence supports Gibrat's law. In the short term, decade by decade, the evidence in favour of Zipf's law is stronger. Second, to consider the whole sample the LOcally WEighted Scatter plot Smoothing (LOWESS) algorithm is applied. From a long-term perspective the evidence supporting Zipf's law increases, but the evidence supporting Gibrat's law is weaker, as small cities exhibit higher variance than the other cities. Finally, the estimated values by decade are again closer to Zipf's law.