Purposeful variable selection framework

A structured, step-wise model-building framework was applied to identify independent biomarkers of taste dysfunction:– Stage 1: univariable regression screening: Individual candidate biomarkers were initially evaluated for association with taste dysfunction using univariate Complex Samples Logistic Regression (CSLOGISTIC) models. To prevent the premature exclusion of clinically relevant background markers, a relaxed significance threshold of p<0.25 was used to qualify candidates for multivariable consideration.– Stage 2: multicollinearity diagnostics for related continuous variables: To evaluate multicollinearity among conceptually related biochemical markers (e.g., liver enzymes, kidney function indices, or serum fatty acids), all variables within each marker group were entered simultaneously into a Complex Samples Linear Regression (CSLINEAR) model. Collinearity was assessed using the Variance Inflation Factor (VIF), with values >5.0 indicating substantial multicollinearity and resulting in exclusion from further modeling. Variables that also demonstrated weak contribution in the collinearity model (p>0.25) were removed. For serum fatty acids, several candidates met univariable significance criteria and required additional data transformations, as described below.– Transformation of fatty acid components: Fatty acids that demonstrated statistical significance in the univariate logistic regression screening were further evaluated for multicollinearity. Several of these candidate fatty acids showed severe collinearity (VIF>10), so they were first aggregated into biologically coherent families – saturated fatty acids (SFAs), monounsaturated fatty acids (MUFAs), and omega-3 fatty acids – by summing their individual concentrations. When reassessed, substantial collinearity persisted between the SFA and MUFA groups. To resolve this, each fatty-acid family was converted into a proportional measure relative to the total fatty-acid pool.

For example:

After applying these proportional transformations, collinearity diagnostics indicated that the three fatty-acid groups no longer exhibited meaningful multicollinearity (VIF<1.1).