This dissertation consists of three separate papers in macroeconomics. The first chapter deals with role of institutions in explaining cross-country productivity differences. I pick one measure of weak contract enforceability for which we have reliable micro-measurements, the recovery rate, and show that differences in the recovery rate take us far in explaining cross-country productivity differences. I calibrate the model to U.S. data, and feed observed recovery rates across countries. I find that the ensuing misallocation of talent generates aggregate TFP differences of 1.5, about a fourth of observed TFP differences. I also find that complementarities between contract enforceability and aggregate endowments are quantitatively important.

In the second chapter, joint work with Marek Kapicka, we show that human capital accumulation decisions have important implications for the design of tax systems. We show that the no distortion at the top result from the Mirrleesean literature may not apply if discouraging labor supply increases incentives to invest in human capital. We show that, under certain conditions, the inverse of the intratemporal wedge follows a random walk. This result is, to our knowledge, novel. We calibrate a two period economy and find several notable results. First, to elicit learning effort, it is efficient to make the consumption process risky for high ability agents while insuring low ability agents. Second, high ability agents face the largest expected increase in the intratemporal wedge. Third, high ability agents face higher intertemporal wedge. Those normative prescriptions differ significantly from the existing literature that abstracts from human capital. We also find large welfare gains for the U.S. from switching to an optimal tax system.

In the third chapter, joint work with Kyle Klein, we propose a method to implement a workhorse optimization algorithm, the Nelder-Mead Simplex Method, in parallel using a distributed memory implementation. The simplex' vertices are equally distributed among processors at each iteration. Each processor then updates its worst local vertices, communicates the results, and a new simplex is formed with the vertices from all processors. In simulations, our implementation exhibits large speedups and is scalable to large problem sizes.