Abstract

Semiparametric additive regression models relax the common assumption of covariate effects to have a known functional form specified by some polynomial. These models only assume that relationships are smooth and are additively separable and thus are less restrictive than common parametric approaches. Estimation techniques for such models assuming all covariates to be exogenous, i.e. uncorrelated to the error term ruling out the presence of confounding omitted variables, for example, have become widely available. However, additive models with weaker assumptions on the error term and methods for inference such as simultaneous confidence bands and specification tests are still subject to extensive research. Thus, the objectives of the thesis are the development of flexible methods for estimation and inference and their application in various complex data situations. Thereby, particular focus is laid on the computational implementation of all proposed approaches aiming at the provision of user-friendly software packages. First, the determinants of chronic child undernutrition in Kenya are analyzed. Particular research questions include the possibility of catch-up growth, i.e. improvements of the nutritional status over age, and relevance of hypotheses on the functional forms of certain effects. In order to address these questions, simultaneous confidence bands for additive models with locally-adaptive smoothed components and heteroscedastic errors are proposed. These appropriately quantify the estimation uncertainty of function estimates and can be used for assessing the statistical significance of an effect and of certain features in a curve. Further, a powerful nonparametric specification test is introduced that allows to test for polynomial regression versus nonparametric alternatives. Next, the needs-relatedness of relief supply in earthquake-affected communities in Pakistan is studied. Here, non-random sample selection calls for the application of a sample selection model with flexible spatial and time-varying effects accounting for unobserved regional heterogeneity and for varying survivor needs over changing seasonal conditions. A flexible Bayesian approach to correct for the sample selection bias is proposed that allows to simultaneously estimate the determinants of the probability to receive relief supply and of the amount of delivered supply. Finally, the usual assumption of the unobservable error term to be orthogonal to the covariates is relaxed relying on the availability of some instrumental variable. A Bayesian nonparametric instrumental variable approach is proposed where bias correction relies on a simultaneous equations specification with flexible modeling of both the covariate effects and the joint error distribution. The approach is used to analyze the relationship between class size and scholastic achievements of students in Israel.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call