Measuring segregation when units are small: a parametric – yet useful – approach
1CREST, France, 2DARES, France
The usual concentration indices – dissimilarity, Theil or Gini – do not behave well when the units on which their computations are based contain few individuals. This study starts by assessing the extent of the bias, that increases when the sample size decreases and when the proportion of the population is closer to zero or one. Assuming a beta model on the probabilities for an individual of a given unit to belong to the group of interest, a method is proposed to compute unbiased estimates of the concentration indices, along with intervals of confidence. Simulations show that this new method is robust to misspecification, and performs well with, for instance, truncated normal or even discrete models. An application to residential segregation in France by parents' nationalities is then proposed.
View full paper