We should however be cautious about this interpretation concerning the evolution of inequality within and between groups. We have interpreted each member of the mixture as representing a social group. This is already a strong interpretation. And we have no guarantee that the mixture members identify exactly the same group of people in each period. In order to fully describe social mobility, we would need panel data and dynamic models.
4.5. Comparing classical and Bayesian standard deviations
Several methods were proposed in the classical literature to compute the standard deviation of a Gini index. The question is complex because the Gini index is based on ordered data. Building on the fact that a Gini index can be seen as a covariance between observations and their rank, Giles (2004) proposes computing the Gini index using a linear regression, which is correct, and to compute its associated standard error using the standard error of the regression, which appears to be misleading because the usual assumptions underlying the OLS are not satisfied. Davidson (2009) proposed an asymptotic method based on the natural estimator of the cumulative distribution. We are now in a position, for the particular samples of the FES, to compare these methods with our Bayesian approach. According to our computations reported in Table 9, it appears that all methods give comparable results for the value of the index, but that they differ in their standard deviations. Presumably because of the strong prior information we introduced, our Bayesian standard deviations for the Gini index are in general slightly lower than their asymptotic counterpart. The regression method, on the other side, gives much higher classical standard errors. It is well known that Giles’ regression corrects for heteroscedasticity, but not for autocorrelation, which leads to biased standard errors.