We used the Gibbs algorithm described above to generate 10,000 draws from the posterior density in order to select the optimal number of mixture components. The chain was run with 10 000 draws plus 1000 draws for warming. We computed various indicators: the BIC (evaluated at the maximum a posteriori estimator), the marginal likelihood using Chib’s method (computed at the prior mean), and two deviance information criteria, View the MathML source and View the MathML source. Most of the time, BIC and Chib were in agreement to select a model with 3 components. The deviance information criterion provided the same answer only for 1988. Otherwise it continues to decrease as an inverse function of k. With those data sets, we cannot get a general and clear answer for selecting mixture components using a DIC, corroborating the extensive results of Celeux et al. (2006). Chib’s method does not seem to be too influenced by label switching in this example, because adding logk! to the log of the marginal likelihood does not change the ordering. Finally, we undertook a sensitivity analysis. The last block of Table 3 is devoted to analysing a variant consisting in adopting a unique prior mean for μ equal to the sample mean of the logs, and this only for the year 1988. We get an identical conclusion for Chib’s predictive and BIC. However, we can no longer get a conclusion with the DIC. 1988 was the only period for which the DIC gave a similar answer. Just by changing the prior, we lose this unique result, which shows the fragility of the DIC in the case of mixtures. We can however conclude that the best approximating model is a finite mixture of three log-normal densities.