Comparison of predictive models using logistic regression, classification tree, and structural equation model for severe dengue.

Abstract

Background: The aim of this study was to compare the predictive performance of three statistical models-logistic regression, classification tree, and structural equation model (SEM)-in predicting severe dengue illness.

Methods/findings: We adopted modified classification of dengue illness severity based on WHO 1997 guideline. Predictive models were constructed using demographic factors and laboratory indicators on the day of fever occurrence. We developed statistical predictive models using data from two hospital cohorts in Thailand, consisting of 257 Thai children. Different predictive models for each category of severe dengue illness were developed employing logistic regression, classification tree, and SEM. The probability of discrimination of each model for severe output of disease was analyzed with external validation data sets from 55 and 700 patients not used in model development. From external validation using predictors on the day of presentation to the hospital, the area under the receiver operating characteristic curve was between 0.65 and 0.84 for the regression model. It was between 0.73 and 0.85 for SEM models. Classification tree models showed good results of sensitivity, ranging from 0.95 to 0.99. However, they showed poor specificity ranging from 0.10 to 0.44.

Conclusions: Our study showed that SEM is comparable to logistic regression or classification tree, which was widely used for more severe form of dengue prediction.