Chapter 1 estimates the correlation of household advertising elasticities across various brands of chocolate and laundry detergent. We leverage a unique data set consisting of matched ad views and household purchases across a two year period. This data allows us to address endogenity issues such as the correlation between ad exposure and household ad elasticities and the distinction between unobserved heterogeneity and state dependence. We extend the dynamic panel methods of Arellano and Bond (1991) to allow time varying random coefficients that can be correlated with regressors (advertising exposure) and correlated across equations in a seemingly unrelated regressions system. We also address two specific puzzles. The first involves the high advertising spending in these industries, which we estimate to be well above the Dorfman-Steiner level of optimal spending. The second looks at the strength of the store brand and why consumers purchase them despite their lack of television advertising. Chapter 2 examines how increases in competitive intensity (and associated increases in coordination costs) affect negative service-quality outcomes in the railroad (Study 1) and airline (Study 2) industry. In Study 1, we leverage an exogenous cost shock to the market through the Rail Safety Improvement Act of 2008, which enforced strict employee hours of service. The results of Study 1 show a 1% decrease in competitive intensity led to a 1.1% decrease in accidents. Study 2 shows a 1% decrease in competitive intensity in the airline industry at the route level led to a .2% decrease in average minutes of delays. These findings are relevant for firms, regulators, and consumers across all industries that suffer from service quality lapses. Chapter 3 expands school choice matching models by incorporating endogenous selection. I show that a general model of selection is identified without any additional assumptions from the standard school choice framework. I perform a monte carlo to show that this model pins down the correct parameter values, while not incorporating selection may severely bias the estimates. I also present a counterfactual where an alternative is removed, and the model with selection is able to correctly predict the results.