Resumen de Estimation and Inference for 2k-p Experiments with Beta Response
Luis Fernando Grajales Hernández
Fractional factorial experiments are widely used in industry and engineering. The most common interest in these experiments is to identify a subset of the factors with the greatest effect on the response. With respect to data analysis for these experiments, the most used methods include linear regression, transformations, and the Generalized Linear Model (GLM). This thesis focuses on experiments whose response is measured continuously in the (0,1) interval (if y ∈(a,b), then (y-a)/(b-a) ∈ (0,1)). Analyses for factorial experiments in (0,1) are rarely found in the literature. In this work, advantages and drawbacks of the three mentioned methods for analyzing data from experiments in (0,1) are described. Here, as the beta distribution assumes values in (0,1), the beta regression model (BRM) is proposed for analyzing these kinds of experiments. More specifically, the necessity of considering variable dispersion (VD) and using linear restrictions on parameters are justified in data from 2k and 2k and 2k-p experiments. Thus, the first result in this thesis is to propose, develop, and apply a restricted VDBRM. The restricted VDBRM is developed from frequentist perspective: a penalized likelihood (by means of Lagrange multipliers), restricted maximum likelihood estimators with their respective Fisher Information Matrix, hypothesis tests, and a diagnostic measure. Upon applying the restricted VDBRM, good results were obtained for simulated data, and it is shown that the hypothesis related to 2k and 2k-p experiments are a special case of the restricted model. The second result of this thesis is to explore an integrated Bayesian/likelihood proposal for analyzing data from factorial experiments using the (Bayesian and frequentist) simple BRM's. This was done upon employing at prior distributions in the Bayesian BRM. Thus, comparisons between confidence intervals (frequentist case) and credibility intervals (Bayesian case) on the mean response are done with good and promisory results in real experiments. This work also explores a technique for choosing the best model among several candidates which combine the Half-normal plots (given by the BRM) and the inferential results. Starting from the active factors chosen from each plot, subsequently the respective regression models are fitted and, finally, by means of information criteria, the best model is chosen. This technique was explored with the following models: normal, transformation, generalized linear, and simple beta regression for real 2k and 2k- p experiments: into the greater part of the examples considered for the Bayesian and frequentist BRM's, results were very similar (using at prior distributions). Moreover, four link functions for the mean response in the BRM are compared: results highlight the importance to study each problem at hand.
