Essays on structured additive regression models with applications in development economics
Doctoral thesis
Date of Examination:2022-12-20
Date of issue:2023-03-17
Advisor:Prof. Phd Inmaculada Martínez-Zarzoso
Referee:Konstantin PhD Wacker
Referee:Prof. Dr. Thomas Kneib
Files in this item
Name:ediss_TorresMunguía.pdf
Size:4.83Mb
Format:PDF
Abstract
English
Structured additive regression models are a particular class of models that provide a flexible framework to deal with a wide class of effects, including linear, nonlinear, random, spatial, and interaction effects, which enables the specification of more complex but more realistic models. The goal of this dissertation is to use these models to address practical issues in three relevant topics in the field of development economics. First, a Gaussian model is used to study gendered inequalities in time allocation to unpaid housework among partnered women and men. In the second study, we are confronted with the problem of identifying the risk factors associated with emotional intimate partner violence, for which a probit model is used. In the third study, quantile models are applied to examine heterogeneous gendered effects of a set of risk factors associated with the income-to-poverty ratio of the poor and extremely poor families. Given the complex structure of the models used in the three abovementioned cases, an estimation cannot be computed by traditional inference techniques. To overcome this issue, it is implemented a three-step strategy consisting on the use of the boosting algorithm, complementary pairs stability selection with per-family error rate control, and the calculation of pointwise bootstrap confidence intervals. From a statistical standpoint, the methodology helps to overcome common issues in regression in development economics, such as dealing with different types of response variables, the inclusion of potential nonlinear (or even a priori unknown) effects of continuous covariates on the response, select the relevant variables at their most suitable functional form, dealing with hierarchical data, to account for spatially correlated observations, to introduce complex interaction effects, and to avoid multicollinearity. From an empirical perspective, the method applied allows to illustrate how the utilization of the structured additive models contributes to enhancing knowledge on these phenomena by providing new relevant insights on the matter. Findings in the three studies not only yield evidence about significant covariates that were either hitherto unknown, understudied, or that have not yet been tested empirically, but they are also relevant for the design of public policies, such as the identification of the relevance of the individual, household, communities, and regional factors in these studies, the existence of age-varying effects, the determination of the circumstances in which women and men face particular disadvantages, and the identification of some specific risk subgroups of the population that are generally overlooked.
Keywords: Structured additive regression models; Boosting algorithm; Development economics