Comparative Performance Analysis for Generalized Additive and Generalized Linear Modeling in Epidemiology

Abstract

Abstract-Most environmental-epidemiological researches emphasize modeling as the causal link of different events (e.g., hospital admission, death, disease emergency). There has been a particular concern in the use of the Generalized Linear Models (GLMs) in the field of epidemiology. However, recent studies in this field highlighted the use of non-parametric techniques, especially the Generalized Additive Models (GAMs). The aim of this work is to compare performance of both methods in the field of epidemiology. Comparison is done in terms of sharpening the relation between the predictors and the response variable as well as in predicting outbreaks. The most suitable method is then adopted to elucidate the impact of bioclimatic factors on the emergence of the zoonotic cutaneous leishmaniasis (ZCL) disease in Central Tunisia. Monthly epidemiologic and bioclimatic data from July 2009 to June 2016 are used in this study. Akaike information criterion, R-squared and F-statistic are used to compare model performance, while the root mean square error is used for checking predictive accuracy for both models. Our results show the potential of GAM model to provide a better assessment of the nonlinear relations and to give a high predictive accuracy compared to GLMs. The results also stress the inaccurate estimation of risk factors when linear trends are used to model nonlinear structured data.

Spatio-temporal modeling of visceral leishmaniasis in Midwest Brazil: An ecological study of 18-years data (2001–2018)

“…Particularly, GAM models showed better fit and good prediction accuracy when compared to generalized linear models, which supports the use of this technique in the field of epidemiology where a causal link needs to be assessed [25]. The practical use of this method has been demonstrated through a real data analysis [18,25].…”

“…These studies have provided important information for health surveillance, such as monitoring and mapping of public health impact risk factors, as well as allowing a better description, understanding, and prediction of risk areas for different diseases [ 67 , 80 – 82 ]. Particularly, GAM models showed better fit and good prediction accuracy when compared to generalized linear models, which supports the use of this technique in the field of epidemiology where a causal link needs to be assessed [ 25 ]. The practical use of this method has been demonstrated through a real data analysis [ 18 , 25 ].…”

“…Among several methods, generalized additive models (GAMs) [ 23 ] were recently used to study the relationship between cutaneous leishmaniasis occurrence and possible risk factors [ 18 ] and to predict the potential distribution of Leishmania vectors [ 24 ]. GAMs are semi-parametric regression methods that relate the response variable to smoothed functions of potential explanatory variables via a link function [ 23 , 25 ]. Thus, this study aimed to describe the demographic features of human VL and model the spatio-temporal distribution of reported cases of VL using GAM, covering an 18-year period from the first report of an autochthonous case in 2001 to 2018 in an urban area endemic for VL.…”

The Pelagic Habitat of Swordfish (Xiphias gladius) in the Changing Environment of the North Indian Ocean

“…AIC is an estimator of the relative quality of statistical models. The model with the lowest AIC can be considered as the best model [33,34]. Log (CPUE) were used in both models because the log transformation made the data normalized.…”