The first chapter studies how households respond to school quality. I combine large-scale administrative and survey data from Chile to estimate parental and child time investment responses to classroom inputs and teachers from fourth to tenth grade. Since classroom inputs are not directly observable, I estimate a dynamic skill formation technology that provides classroom and teacher effects as a by-product, in a similar fashion as value-added models. I address selection by leveraging repeated observations of students and rich data on factors involved in household decisions. Parents of fourth graders compensate for low quality teachers and classroom inputs, while parents of high school students reinforce the quality of these inputs. Students, on the other hand, increase time self-investment if their classroom environment improves at every grade, but the responses are larger for older children. The heterogeneous responses by grade found in Chapter 1 motivate the analysis of optimal resource allocation policies across education levels. Chapter 2 builds on Chapter 1 to understand how the differential impact by grade of school resources and home investment can be used to design the optimal allocation of the school resources across grades. To that end, I build and estimate a child development model using an indirect inference approach. I use the estimated model to simulate counterfactuals of the dynamics of the cognitive skills of students and I characterize the optimal allocation of school resources across grades. The results suggest that, on average, it is optimal to allocate relatively more resources in lower grades than in upper grades with respect to the allocation observed in the data. Moreover, the behavioral response of households plays a key role in the characterization of the optimal allocation. In the last chapter, I develop an empirical test for employer asymmetric learning about the productivity gains of On-the-Job (OTJ) training programs. I developed a model of OTJ training and employer learning. I solve the model under two types of learning: (i) asymmetric, current employer learns faster than potential employers and (ii) symmetric, the whole market learns simultaneously. The solution suggests different wage profile predictions under each form of learning. I build an empirical test based on these predictions and implement it on the Chilean Social Protection Survey by estimating a wage equation with interactions of training variables and tenure on the job. The results provide evidence of employer asymmetric learning.