From the omitted variables bias formula, the coefficient from b. is cov(b, E(y | s) − y) cov(b,η(s)) +. If training is unrelated to the signal, the coefficient estimate var(b) var(b) captures the part of the earnings adjustment due to learning. [...] The employee is either promoted to a more senior position ahead of the cohort that entered the labour market with him or else he is moved laterally to another position within the firm, or is dismissed from the firm altogether. [...] If all training differences were due to initial firm placement, the model predicts that the coefficient from test 2A on s. is zero, while the coefficient from test 2B on b. is the coefficient from the auxiliary regression of [y − E(y | s)]. [...] In summary, the null hypothesis that firms statistically discriminate, under the model’s assumptions, implies that changes in earnings and changes in firm quality should be uncorrelated 7. This prediction might not hold for the lowest or highest quality firms, since workers from the firms with the lowest quality in first period can only move up and vice versa. [...] The regressions require a panel on workers with earnings and firm-level information each period, signal variables that we are interested in testing for signalling, and variables correlated with productivity that are initially unobservable to the firm, but available to the researcher.