Semiparametric Regression Models for Location, Scale and Shape

• Applicant: Professor Dr. Thomas Kneib
• Subject Area: Statistics and Econometrics
• Term: from 2018 to 2022
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 397587368

Generalised additive models for location, scale and shape (GAMLSS) allow one to relate not only the conditional expectation of a response variable in a regression context to potential explanatory variables but all parameters of the conditional distribution of the response. As a consequence, flexible regression relations beyond the mean can be detected such that more realistic and more informative regression specifications are obtained. In this project, different semiparametric extensions of GAMLSS will be developed and applied in complex case studies. The first part of the project will consider the analysis of time series data with regime switching based on Markov switching and smooth transition models combined with the flexibility of GAMLSS along the specific application of operational losses in the banking industry. The second part will deal with efficiency analysis based on stochastic frontier analysis which will be embedded in the context of GAMLSS to allow for a multivariate formulation with multiple output equations. A third model class considers censored GAMLSS specifications where only incomplete information is available. The most typical example are duration times, e.g. for labour market histories of individuals, but we will also consider censored data on income from observational studies. To increase the general applicability of GAMLSS, we will in addition study difference flexible extensions of the predictor structure (regularization and variable selection approaches, complex interaction types, functional explanatory variables), develop variable importance measures and investigate novel ways of interpreting and visualizing GAMLSS estimates. In summary, this project will provide important contributions to increase the applicability of GAMLSS in challenging economic applications. It will therefore increase the general awareness of such methods such that GAMLSS will turn into a standard tool for applied economists interested in regression effects beyond the mean.

• DFG Programme: Research Grants