Statistical analyses
Binary logistic regression models were computed to assess the associations between demographics, olfactory-related factors and clinically relevant changes in overall olfactory function (TDI) and the sub-dimensions threshold (T), discrimination (D), and identification (I). Clinically relevant changes were defined based on the following cut-off scores: (i) for overall olfactory function: TDI improvement greater or equal 5.5 points at follow up visit, (ii) for threshold function: T improvement greater or equal 2.5 points at follow up visit, and (iii) for discrimination and identification function: improvement greater or equal 3 points at follow up visit.16Olfactory-related variables included: age (years), gender (male and female), olfactory function at first visit (baseline olfactory function, TDI), duration of olfactory training (weeks), duration of smell loss (month), reason for OD (postinfectious, posttraumatic, and idiopathic), and presence of parosmia or phantosmia at first visit. All demographics and olfactory-related variables were entered in the models, and statistical estimates were generated to calculate adjusted odds ratios (aOR) with 95% confidence interval. Hierarchical cluster analysis and the associated dendrogram were computed based on the Ward clustering method and the Squared Euclidian distance to identify possible groupings between changes after OT in T, D and I in terms of similarity. Data were analyzed using SPSS (SPSS version 23.0 for Windows; IBM Corp., Armonk, NY, USA). This study employed a level of significance of 0.05. According to the previously reported and widely used sample size calculation-criterion of ten events per variable in logistic regression analysis, we needed at least 80 patients with parosmia. Because we included 81 patients with parosmia, this study is sufficiently powered to conduct the described analysis for parosmia as predictive value.17