The papers in this thesis cover a variety of ideas. They are united by the common theme of carefully observing existing relationships in data. In the first chapter, I find that looking at averages can be insuffcient. On average, there is not a strong relationship between an individual’s ability and her tolerance of risk. However, the joint distribution of the two characteristics shows a great deal of heterogeneity that the average masks. The highest ability individuals are most likely to report middle levels of risk tolerance, whereas those of lower ability are also likely to report extreme values. For those of high ability, therefore, insuffcient risk tolerance may prevent them from starting their own business. In the second chapter, I illustrate and quantify the effciency gain that results from accounting for nonlinearities in the first stage of 2SLS when estimating the second stage parameters of interest. Additionally, I show that in some cases estimating nonlinearly can prevent incorrect inference resulting from misestimation of the underlying first stage data generating process when using a linear method. In the third chapter, I observe and attempt to explain a behavioral puzzle. A standard screening question for rational decision-making screens out a large minority of the sample. Since so many people are affected, it is unlikely to be due to noise or measurement error. I take the answers seriously and develop an explanation.