UC Berkeley

Network-based connections are pervasive in hiring and mobility patterns. While the theoretical impacts of network connections on the inequality of labor market outcomes are well known (Calvo-Armengol and Jackson, 2004), the empirical magnitude of actual network effects is less certain. A key issue is the difficulty of disentangling the causal effect of network connections from differences in the characteristics of workers in better and worse networks.

In the first chapter of my dissertation, I study this question using data on freelance workers in Hollywood, who, like many workers in the "freelance economy," are hired for short-term jobs through an informal process that relies in part on previous connections. In such a labor market, the fortunes of an individual worker are closely linked to the careers of key agents who make the hiring decisions for jobs. In Hollywood, workers who know position supervisors who manage more jobs will have more job opportunities. To measure the size of the network effects, I follow cohorts of freelance grips and lighting technicians who first work on a major movie production between 1988 and 2002. I develop two alternative models for the probability that workers are hired on subsequent productions---based on random effects and fixed effects specifications---that incorporate network effects, experience effects, and unobserved heterogeneity. Both models yield large estimates of the effect of experience-based connections on hiring outcomes. I then use the random effects model to develop simulation-based estimates of the fraction of overall inequality in job outcomes for workers in a given cohort that is attributable to inequality in the career success of the key supervisors they met during their initial year of work. I find that about half of the wide dispersion of career outcomes in Hollywood is generated by differences in the career trajectories of these initial key supervisors.

In the second chapter, I study the asymptotic properties of the fixed effects estimator I employ in the first chapter. The estimator is robust to unobserved heterogeneity across workers and movies. It is based on subgraphs of worker-movie dyads that I call pairs. Inference is non-standard, because pairs within a sample are only independent when they do not share any workers or movies in common. The underlying criterion is a two-sample U-process. I show that the U-statistic derived from the estimator's first order condition is asymptotically equivalent to a certain projection which involves summation over all the worker-movie dyads in the sample. I use this result to derive a consistent estimator of the variance of the estimator.