Various interaction terms were tested in the multivariable model for statistical significance.
Backwards elimination was used to identify a parsimonious multivariable model (SAS statistical analysis package, Cary, NC).
Only DRB1*03, DRB1*0701, and DRB1*0804 could be retained as independent factors (adjusted P = 0.005 0.029) in the reduced multivariable model, after statistical adjustments for the clear, negative impact of severe and moderate immunodeficiency (P < 0.001).
In the analyses featuring poisoning as the outcome of interest, we also found marked attenuation after adding education to the multivariable model, although statistical significance at conventional levels was retained (hazard ratio (HR) = 1.31, 95% CI: 1.18, 1.46).
Although discharge hemoglobin and destination were examined in a descriptive manner only, hospital mortality, length of stay (LOS) and costs were looked at in multivariable models (see Statistical analyses, below).
Two multivariate multivariable statistical models were created: The UMod included males and females without variables related to oestrous cycle, while the FMod included only females with all analysed variables.
The final complete multivariable model allowed the statistical independence of the predictive variables to be established.
Due to the REGARDS sampling methods, the variables race, region, and sex were included in the multivariable model regardless of statistical significance.
Variables included in the final multilevel multivariable model showed either statistical significance (α ≤ 0.05), or evidence of acting as a confounding variable (20%% or greater change to other statistically significant model coefficients when removed).
To determine which covariates were included in the multivariable models, descriptive statistics and statistical tests were evaluated to identify characteristics that differed between cohorts and also may be associated with adherence or persistence (i.e., were potential confounders).
