Essentials of Preprocessing Data in Improving Logistic Regression Performance Based on Rough Sets Theory: A Case Study of Stunting in West Sumatra, Indonesia

Various studies have considered a logistic regression model for investigating stunting and its determinants. However, some models above fall short of the researcher's expectations, such as the fact that there is no significant variable, making them very challenging to interpret. In this paper,we are interested in handling the problem above by applying data reduction strategies using rough set theory in the preprocessing phase and implementing them for stunting data sets in Solok Regency, West Sumatra Province, Indonesia. There are three types of data reduction: removing inconsistent samples, irrelevant attributes, or both, so three kinds of modified models will be used in this paper, namely Logistic Regression Reduction Rough Set (LR3S) type I, II, and III. The classic and modified models were compared using performance model criteria, which include precision, recall, F1-score, accuracy, ROC curves, and AUC values. It was determined that the stunting data set was unsuitable for the classical logistic model, and the best model was built by removing inconsistent observations (LR3S type I). At a 5% significance level, the best model indicates the factors significantly influencing stunting incidence: exclusive breastfeeding, birth weight, smoking family, immunization, and gender. Stunting classes are not considerably differentiated by factors such as worms, clean water, comorbidities, hygienic latrines, and health assurance in Solok Regency. A priority program and policies for stunting prevention in this regency can be developed by considering the extent of these factors' major influence.