The relationship between the postoperative (at 12 months) GAT-IOP and 41 preoperative parameters consisted of six systemic parameters (age, height, BMI, SBP, DBP, and history of smoking), 24 blood examination scores (WBC count, RBC count, Hb, Ht, Plt count, TP, AST, ALT, LDH, T-bil, γGTP, UA, BUN, Cre, eGFR, Na, K, Cl, Ca, IP, TG, TC, BS, and CRP), 11 ocular parameters (preoperative GAT-IOP, CCT, type of glaucoma [POAG, PACG, EG, and SOAG], and use of PGF2a analog, use of beta-blocker, use of CAIs, use of brimonidine tartrate, and use of ripasudil, and three ORA parameters (IOPcc, CH, and CRF) was investigated by using the univariate linear mixed model because we included both eyes from the same patients. The measurement of two eyes from the same patients can be highly correlated. The linear mixed model adjusts for the hierarchical structure of the data, modeling so that each patient is registered as having a random effect; that is, measurements are grouped within participants. Ignoring this measurement grouping would result in underestimating the standard errors of regression coefficients. The blood examination scores at baseline are shown in
Supplementary File S1. Subsequently, the multivariate linear mixed model was used to analyze the relationship between the postoperative GAT-IOP and the 44 parameters. An optimal model for the postoperative GAT-IOP was identified through a two-stage model selection. First, 15 parameters were predetermined from 41 parameters by using the least absolute shrinkage and selection operator regression (LASSO).
46,47 Subsequently, the model selection using the second-order bias-corrected Akaike information criterion (AICc) index (all models) was conducted. We calculated the AICc values of all possible combinations of the 15 parameters (2
15 patterns), and the model with the minimum AICc was identified as the optimal model. Moreover, the AICc is a corrected form of the common statistical measure of the Akaike information criterion that provides an accurate estimate even in small sample sizes.
48 Any reduction in AICc indicates an improvement of the model,
49,50 and the relative likelihood that a model with AICc
x is minimizing information loss compared with the model with the smallest AICc (AICc
min) is calculated as exp((AICc
min AICc
x)/2).
51