Manual labelling of training examples is common practice in supervised\nlearning. When the labelling task is of non-trivial difficulty, the supplied\nlabels may not be equal to the ground-truth labels, and label noise is\nintroduced into the training dataset. If the manual annotation is carried out\nby multiple experts, the same training example can be given different class\nassignments by different experts, which is indicative of label noise. In the\nframework of model-based classification, a simple, but key observation is that\nwhen the manual labels are sampled using the posterior probabilities of class\nmembership, the noisy labels are as valuable as the ground-truth labels in\nterms of statistical information. A relaxation of this process is a random\neffects model for imperfect labelling by a group that uses approximate\nposterior probabilities of class membership. The relative efficiency of\nlogistic regression using the noisy labels compared to logistic regression\nusing the ground-truth labels can then be derived. The main finding is that\nlogistic regression can be robust to label noise when label noise and\nclassification difficulty are positively correlated. In particular, when\nclassification difficulty is the only source of label errors, multiple sets of\nnoisy labels can supply more information for the estimation of a classification\nrule compared to the single set of ground-truth labels.\n