Web Appendix for “The Value of Reputation in an Online Freelance Marketplace”

1 Details of Second-Step Estimation All the estimates from state transitions (θp, θg, θt), ηis, and non-parametric joint distributions of bid attributes are assumed to be known at this point, i.e., consistent estimates of these parameters and distributions from Step 1 are used. Throughout we estimate CCPs using weighted nested logits that are flexible functions of {xit, η̂i, si}. When doing so, we treat ηis as if they were observed by substituting η̂i in their lieu. However, because si is unobserved, at the each step of the EM algorithm, we weigh the nested logits with current estimates of ρs. We use flexible logits to estimate CCPs because commonly used non-parametric estimators are infeasible in our context because we have many continuous state variables and an extremely large state space. With a sufficiently large sample and an exhaustive set of covariates, CCP estimates from weighted flexible logits are similar to those obtained using non-parametric estimators (Arcidiacono & Miller 2011, Bajari et. al., 2010). In our flexible logit, we incorporate all the state variables, their higher order terms, and interactions. Step 1: Let the number of unobserved types be 2, and the initial guess of population probabilities be π1 = {π1 1, π1 2} = {0.5, 0.5}. Let {α1, β1, δ1, σ1} be the first guess of the structural parameters. We start with estimates from the model without persistent unobserved heterogeneity and a random αb11. Step 2: Update the ρs as follows: