ABSTRACT
Employing the U.S. data for the 1990’s, we disentangle state dependence from unobserved heterogeneity in earnings mobility. The fixed effects dynamic multinomial logit utilized to model earnings quintiles dynamics fits the data very well, much more so than the autocorrelated dynamic ordered probit. Unordered models should thus be preferred to study earnings mobility. State dependence is found to be significant. It has a protective (anti-fall) effect at the top of the distribution and a detrimental (anti-rise) effect at the bottom. Consequently, the difficulty of climbing the ladder is not exclusively a question of individual characteristics. Public policies could thus be implemented to improve information about individuals’ ability and job offers, and to design more efficient wage-setting institutions.
Acknowledgements
I thank an anonymous referee and the editor for very useful remarks and suggestions. I also thank Jean-Marc Robin, Bruno Crépon, Thierry Kamionka, Francis Kramarz, Edwin Leuven, Thierry Magnac, Abla Safir, seminar participants at CREST, PSE, UPEC, BETA and conference participants at COST in Sankt Gallen, JMA in Fribourg, IZA in Bonn, RTN in Paris, AFSE in Paris, EALE in Oslo and TEPP in Aussois for useful suggestions and comments. The usual disclaimer applies.
Disclosure statement
No potential conflict of interest was reported by the author.
Notes
1 Earnings mobility is an increasing function of its step: when mobility is measured with a larger step, workers have more time to move.
2 This threshold has been chosen such that 25% of the workers work less than this value.
3 −2*log likelihood + 2*number of parameters.
4 −2*log likelihood + log (number of observations)*number of parameters.
5 One exception is Browning, Ejrnaes, and Alvares (Citation2010) who find that MA parameters are negative for only 30% of the population.
6 These features justify the use of a two-step procedure rather than one step, which would require for the estimation of the ’s, to specify the law of the unobserved heterogeneity and to approximate the correlation between the latter and the initial conditions.
7 To do so, the extended method of Honore and Kyriazidou (Citation2000) could have been applied, but the sample size of the PSID is too small for that, which explains why the model has no time varying covariates.
8 These regressions are performed year by year.