Since correlations between some explanatory variables are very strong (eg, >0.9 between ultraviolet B exposure and latitude), when conducting the multivariable analysis, highly correlated variables (correlations >0.5) were not entered into the model at the same time.
A correlation analysis was conducted in order to avoid multi-collinearity; only one variable in highly correlated variable sets was selected for multiple logistic regression analysis.
To avoid the inclusion of highly correlated variables into the cluster analysis (CA) we performed a principal component analysis (data not shown).
With this new framework, the variance based global sensitivity analysis for models with correlated variables can be conducted conveniently and the variance contributions of a correlated variable to the variance of model output can be identified and interpreted distinctly.
To test for colinearity, we performed analyses with all correlated variables in the model, and separately for each correlated variable.
In addition, the inclusion of correlated variables in a multivariate analysis may result in lower ORs.
Due to the low number of children with lice in the screening (low statistical power), only the most significant of the correlated variables in the univariate analysis was entered into the full model.
In Phase 3, principal components analysis grouped correlated variables.
In this case, prior to SVM analysis, highly correlated variables were removed.
Briefly, factor analysis aggregates correlated variables, and each obtained factor (dietary pattern) is a linear combination of all food groups that are weighted by their factor loadings and explains as much variation in the food groups as possible (12).
The analysis model was based on the univariate analysis, and no significant interactions were found between correlated variables in a correlation matrix.
