To address the second and third goal, a thorough exploratory and descriptive preliminary analysis was conducted by assessing bivariate plots and univariate models by using the ggplot2 R package
38 to examine for possible significant associations between explanatory and outcome variables, which guided how multivariate modeling was used. Pearson correlation was examined to ensure there were no issues with collinearity during multivariate modeling. Owing to the variability in interobserver reliability noted in dry eye assessment, investigators were coded as separate categorical variables and factored in all statistical models to account for interobserver variability.
39 Aggregate values (i.e., summation of values from all grading zones) for bulbar and limbal redness, conjunctival staining, and corneal staining type, extent, and depth were used for analysis. In addition, aggregate values from the upper and lower lids for blepharitis, Line of Marx, meibum quality and quantity, and meibomian gland count and atrophy score were used. The use of these aggregate values in statistical modeling was decided a priori to minimize the possibility of type 1 error and because our previous studies have found none-to-minimal benefits in sectoral analysis.
21,40–42 In addition to conducting the analysis using the mean of LWE length and width to generate one value when assessing LWE, we also opted to examine LWE length and width separately because of studies that have found that greater LWE width is associated with symptoms.
4,5,43,44 Linear mixed effects models using the nlme R package
45 were used to account for potential within-subject correlations related to measurements done on the right and left eye. The use of the right and left eye in the statistical analysis was designed to account for the intereye variability in ocular surface parameters, and in symptoms of ocular discomfort and dryness for each subject.
18 Upon examining residual plots, FTBUT and NITBUT were natural-log transformed to better approximate normality to meet key assumptions for univariate and multivariate modeling. For multivariate analysis, a stepwise regression procedure with consideration of
F test
P values and examination of residual and other diagnostic plots was used to determine accurate models. Interaction terms based on the final set of significant explanatory variables chosen in each multivariate model were considered, using the jtools R package,
46 but none were found to be significant.