This paper examines the consequences of specification error when transition matrices are used to analyse patterns of intergenerational mobility. We show that classical measurement error in both the child's and parent's earnings can lead to biased results, with summary mobility measures biased by as much as 20% in some cases. Furthermore our results suggest that the extent of the bias is most severe in the tails of the distribution. Omitted conditioning variables appear to have a modest effect on transition matrices in our model. © Springer Science+Business Media, Inc. 2006.