2.3. Statistical Analysis

We performed all statistical analyses with Mplus Statistical Software, version 8 (Muthén & Muthén, Los Angeles, CA, USA). This software uses by default a Full Information Maximum Likelihood (FIML) (34) estimation approach to handle missing values. Study subjects who did not complete 40% or more of the 46 non-dental items of OHIP-49, PROMIS v.1.1 Global Health, and v.1.0 Emotional Distress - Depression were excluded from the study. Ninety-eight percent of patients in the analysis sample were missing less than 5% of 46 OHIP items, and 96% were missing less than 5% of the ten global health items. For the Depression score, the average item score from non-missing responses was imputed for these patients missing less than 40% of items; 93% of patients were not missing any Depression items, and 97% were missing less than 5% of 28 Depression items.

We analyzed the following three models with Structural Equation Modeling (SEM) methodology.

Model 1: We measured HRQoL with the two HRQoL factors, i.e., Physical Health and Mental Health. In this model, only the Physical and Mental Health of HRQoL were used because PROMIS documentation does not indicate a global HRQoL factor that combines the two HRQoL factors.

We measured OHRQoL with four first-order factors, i.e., the dimensions of OHRQoL. We introduced a second-order OHRQoL factor comprising the four first-order factors. We determined two correlation coefficients for the HRQoLOHRQoL association: one for HRQoL Physical Health - OHRQoL and one for HRQoL Mental Health - OHRQoL.

Model 2: Even if the PROMIS authors do not suggest forming a global factor for PROMIS v.1.1 Global Health PROM, we added a second-order global HRQoL factor in Model 2 because we wanted to derive a single correlation coefficient characterizing the association between OHRQoL and HRQoL constructs. Model 2 is otherwise identical to Model 1.

Model 3: To be comparable with previous analysis (10), three independent variables, i.e. age, gender, and level of depression, were included in Path Analysis (35). Age and gender were self-reported by the patients. This model allows us to compute the association between OHRQoL and HRQoL controlled for potential confounders. We first approached the path analysis model by adding the three independent variables to Model 2. However, we encountered problems fitting the second model, and have, therefore, added the three independents to Model 1 composed of one global OHRQoL factor and two HRQoL Physical and Mental Health factors, which we regressed upon patients’ age, gender, and depression score.

SEM-based second-order confirmatory factor analysis was used to test the model fit scale of OHIP and both PROMIS PROMs and to assess correlations between the global OHRQoL factor and its four dimensions with HRQoL Physical and Mental Health dimensions. In addition, correlation coefficients between OHRQoL dimensions were also estimated from the SEM second-order confirmatory factor analysis model.

Two model fit indices that account for model complexity are the Comparative Fit Index (CFI) and Tucker-Lewis Index (TLI). For both indices, an index value higher than 0.95 indicates a good fit of the model to the data. The Root Mean Square Error of Approximation (RMSEA) of less than 0.08 indicates a good fit (36). More recently, RMSEA (37) of less than 0.06 or a stringent upper limit of 0.07 is the current consensus for a good fit (38). The Weighted Root-Mean-Square Residual (WRMR) uses a variance-weighted approach (39). The WRMR statistic of less than 1.0 indicates a good fit. We judged the magnitude of correlation coefficients according to Cohen (40), where the effect size of 0.1 is considered small, 0.3 medium, and 0.5 large.