IN OUR STRUCTURAL ESTIMATION, we assume throughout that ε t is distributed as a Type 1 extreme value. The computational advantages of parameterizing G(ε) this way are most evident from Lemma A.1 below, where we provide formulas for A t (h) and B t (h), the value of human capital on and off the equilibrium path, and also an expression for marginal disturbances, q jk [p t (h)]. LEMMA A.1: If ε jkt is independently and identically distributed as a Type I extreme value with location and scale parameterswhere p 0t (h) is the probability that the optimal choice is retirement,The IIA property of Type 1 extreme values implies that the marginal idiosyncratic shock for a manager who is indifferent between the best job match (j k) and retiring is the log-odds ratio of the probability that a manager with characteristics (t h) who accepts employment in (j k) versus retiring. The logodds ratio does not depend on the other components of the conditional-choice probability vector. The greater the probability of retirement observed in equilibrium, the less important is the human-capital component, and the higher is the unobserved shock for the marginal person.PROOF OF LEMMA A.1: The formula for q jk [p t (h)] given by (S-1) is well known (e.g., Hotz and Miller (1993)). Denoting the probability density function of ε * jkt ≡ d jk ε jkt by dG(ε * jkt ), we first derive an expression for E[exp(−ε * jkt /b t )] and then use it in our derivation of the formula for A t (h t ):

We develop a pure moral hazard model, and a closely related hybrid one, where there are both hidden actions and hidden information, to derive the restrictions from optimal contract theory that characterize set identi…cation. In pure moral hazard models, the expected utility of managers is equalized across states, whereas in a hybrid model the optimal contract equates the expected utility of truth telling with the expected utility of lying. These restrictions are testable. Our identi…cation analysis establishes sharp and tight bounds on the identi…ed set. Our tests and estimators are based on these bounds. We apply semiparametric methods to test the models, estimate the structural parameters, and quantify the e¤ects of hidden actions versus hidden information. The pure moral hazard model is rejected on a large panel data set measuring the compensation of chief executive o¢ cers and the …nancial and accounting returns of the publicly traded …rms they manage. We do not, however, reject the restrictions of the hybrid model, and our structural estimates for that model show the degree of private information varies considerably across sectors and over …rm size.We would like to thank Maher Said, Melissa Tartari and Steve Tadelis for their comments. We have also bene…tted from presentations at the