Modeling Dengue Severity in Pregnancy

i)

Model selection exploring the moderation effect of DENV serotype on dengue severity across pregnancy status

A multicollinearity test was performed and tolerance for hypertension (0.93) and diabetes (0.93) showed no multicollinearity between them. Similarly, there was no multicollinearity observed among the other covariates. Year and region were considered as potential random effects and were evaluated in a two-level hierarchical model. The Wald tests for random effects for the region (p = 0.049) and year (p = 0.032) were both significant. A comparison between random and fixed effects model using the Bayesian information criteria (BIC) revealed that the random-effects model performed better. Hence, both region and year were retained in the model as random effects. Based on the hierarchical model (1), a multiple logistic regression was performed with severe or non-severe DF as the outcome variable and pregnancy as the exposure variable. $$y=X\ast \beta +Z\ast u+\mathit{\epsilon }$$
where X a matrix (N *p) with p predictor variables and Z is a matrix (N*q) for q random effects [19 ].
Two-way interactions were explored, and multiple interactions were found to be statistically significant and retained. Other covariates in the model included dengue serotype, diabetes, hypertension, and age. The final model was chosen by comparing AIC and BIC between models. The AUC for the predicted probabilities of the final model was 0.7156 (Additional file 1: Fig. S1).