This seminar will introduce nonparametric and semiparametric regression for cross-sectional and longitudinal (panel) data models.

Nonparametric and semiparametric methods are useful whenever we aren’t certain what the correct functional form is for the relationship between two variables—which in many fields is most of the time. Misspecifying the functional form can lead to inconsistent estimates as well as incorrect policy prescriptions. For example, estimating a curvilinear relationship as linear would give the mistaken impression that an effect is constant, when in reality the magnitude (and possibly the direction) of the effect differs across units.  Using nonparametric and semiparametric methods can help detect and correct such problems.

We will focus both on developing intuitions about nonparametric regression and on how to program and apply these methods in practice. We will pay particular attention to how to present the results from nonparametric regressions, as this is quite different from the standard linear regression models that many researchers are familiar with.

We will begin with the simple case of nonparameteric regression with a single regressor, which will allow us to probe issues of bias, variance, and inference, with an eye toward understanding local-estimation which is foundational to more advanced nonparameteric methods. We will then move on to multivariate models. Here we will discuss the curse of dimensionality (the primary criticism of nonparametric methods), how it arises, why we should be mindful of it, and how we can avoid it. We will also discuss mixed data types (both continuous and discrete regressors). Finally, we will introduce advanced methods, including two of the most popular forms of semiparametric regression—partially linear models and varying coefficient models, as well as nonparametric methods for longitudinal (panel) data.

Along the way we will discuss practical issues, including choosing the appropriate bandwidth for your analysis and testing model assumptions. Each day will include plenty of hands-on practice, so you will leave with both a firm grasp of the theoretical underpinnings of nonparameteric and semiparametric methods and a clear understanding of how to apply them to your own work.