This thesis is composed of two chapters. Chapter 1 presents an empirical investigation of changes in worker-attachment in the United States labor market over the last thirty years. Chapter 2 combines both empirical and theoretical work to investigate one theory behind worker-attachment, namely match-specific human capital. Chapter 1 extends the analysis in Farber (2008) from the Current Population Survey to include both the Survey on Income and Program Participation and the Panel Study of Income Dynamics. There is general agreement among the three surveys for males, who see flattening age-tenure profiles and declining probabilities of long-term employment relationships among private sector workers, while male public sector worker-firm attachment appears fairly stable. The comparison across surveys for women shows substantial disagreement, leading to a muddier picture than in previous papers. The chapter concludes analyzing employer investment in worker training, hypothesizing that this should vary with changes in expectations of worker attachment. While in the cross section investment is higher for groups that tend to have long-term employment relationships, the longitudinal trends in training do not reflect the changes in worker attachment, suggesting that other factors are driving the observed trends. Chapter 2 estimates the returns to on-the-job training and, in particular, asks how portable these acquired skills are from one job to the next. I build a model of a labor market with undirected, on-the-job search and counteroffers where firms may sign long-term contracts with workers that jointly determine wages and costly training (human capital acquisition). For any amount of training, some fraction of the human capital acquired is job-specific and is lost upon job termination. Estimating returns to training in this environment is complicated because wages do not reflect productivity; long-term contracts limit the degree to which training passes through to wages. Further, job transitions are endogenous; workers choose to leave only when a new job compensates them for lost skills. In this environment, even job transitions that induce human capital loss commonly feature wage gains. Estimates indicate a substantial specific component of training, between 32%-47%. Results also imply that specific skills play an important role in reducing worker separation rates (around 30% at 6 years of tenure). In spite of the existence of the job-specific component, wage regressions estimated on simulations would suggest training is quite general, as in the data.