Aplicações de métodos de seleção de variáveis em modelos de regressão

• Autor(a) principal: Duarte, Alice Silva
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UFLA
• Texto Completo: http://repositorio.ufla.br/jspui/handle/1/55736
• topic: Alta dimensionalidade; Importância de variáveis; Lasso; Floresta aleatória; Regressão logística; High dimensionality; Variables importance; Random forest; Logistic regression; Stepwise; Regressão e Correlação

Regression models are appliedtostudy a cause/effectrelationshipbetween a response variable and oneor more explanatory variables. Withtechnologicaladvances, the volume and dimension of theanalyzed data canbeincreasing. Whilethelargenumber of variables canincreasethepredictivecapacity of the model many of them variables cancontributelittle and generate a high computational cost. Then it maybenecessarytoselect variables and search for thosethathavethegreatestimpact in the model. In thisworkweevaluatethe use of variable selection methods in two case studies. The firstonewascarried out toevaluatethefrequency and food security of preschoolers in thecity of Lavras, MG. The responses analyzed in thisfirststage are data fromcategories of theBrazilianScale of Food Insecurity (EBIA) and the Food Frequency Questionnaire (FFQ), analyzedthroughlogistic models. Data werecollectedfrom 581 preschoolers and refertoabout 50 variables of differenttypes. The methods Stepwise, Lasso, the Purposeful Selection of Covariates (PSV) and Random Forest wereconsidered for the selection of variables. Subsequently, thelogistic models wereobtainedwiththe variables selectedbythesemethods. The models wereevaluated in terms of AIC. Amongtheevaluatedmethods, theonethatproducedthebestperforming model was Stepwise. The secondapplicationinvolved a high-dimensional data scenario, obtainedwiththe use of NIRS (Near infraredspectroscopy) in a problem of predicting food consumption, fromfeces of dairycows. The methods Stepwise, lasso and Random Forest wereconsidered for the selection of variables. Lasso performedwell in thecross-validationstudy. However, thisstudyislimitedtothe use of themethodsindependently. Other authorsobtainedgoodresultsapplying more thanonemethodsimultaneously. The contributions of this case study are thecomparisonamong lasso and Random Forest, usedseparately for the selection of variables in NIRS and thecomparisonbetweendifferenttypes of validations for the models obtainedusing lasso.