Análise do desempenho de diferentes modelos de predição de eventos binários em epidemiologia – estudo com dados de retenção de placenta.

Abstract

Statistical modeling for classification is basically aimed at predicting certain behavior of an event based on its characteristics that, in some way, are correlated with the response variable of interest. Predictive models can be developed using traditional methods of regression analysis, for example logistic regression or even with more sophisticated methods like classification tree models (CART) and Random Forest. This study aimed to analyze the performance of three different methods of data modeling: a classic GLM method, Logistic Regression and two methods based on machine learning techniques (Machine Learning), the Classification Tree (CART) and the Random Forest Model. The applicability of the models was evaluated using classification performance indices such as the AUC statistic (Area Under the Curve). The database is related to the retained placenta in 3 dairy farms in Unaí, Minas Gerais. Retained placenta was analyzed by self-completed epidemiological forms by the owners, generating a total of 699 observations. Due to a very specific profile of the database, the analysis was conducted to compare the results of two databases, called as dt01 and dt02. We aimed to verify how an imbalance data, and a problem of missing data related to the lactation order and number of lactations, specific to a property, could affect the performance of the models. Thus, evaluating each model individually, the performance of the logistic regression and the CART model, less complex models with a more direct understanding, obtained higher AUCs than the Random Forest. Based on the previous discussion, the complexity of the event studied, and the studie variables, we can consider that the performance of the simpler models was minimally satisfactory and superior to the more complex models.