Symptom Patterns and Contributing Factors by Age

Statistical analyses were performed by using SPSS Statistics for Windows, version 26.0 (IBM Corp., Armonk, NY, USA), R version 4.1.0 and Mplus version 8.0. Descriptive analysis was used for the distribution of sociodemographic, clinical, symptoms, and function characteristics. Categorical variables were presented as frequencies and percentages, and continuous variables as means and SDs. A symptom network analysis was used to identify the most central symptom in the entire sample and in each age group. In the symptom networks, a node indicates an independent symptom, an edge indicates the conditional relationships between two symptoms, and the edge thickness shows the strength of the relationship between them [16 (link)]. Thus, two centrality indices (strength and closeness) were output to quantify the relationship. The strength value represents the probability of one symptom and other symptoms occurring together, and the closeness value represents the path from one symptom to all other symptoms [16 (link)].
The questionnaires were scored according to the PROMIS Scoring Manual, and were dichotomized as 0 or 1 according to the cutoff scores for clinical differences (https://www.healthmeasures.net/). After data processing, LCA was performed to identify clusters of individuals displaying similar patterns of symptoms by age groups (15–39, 40–59, and over 60 years). Models with an increasing number of latent classes were assessed until the best fitting model was determined. To select the optimal LCA model, the following indices were included: the Akaike information criterion (AIC), Bayesian information criterion (BIC), and adjusted BIC (aBIC) were used to assess information criteria; and the Lo-Mendell-Rubin (LMR) test and bootstrapped likelihood ratio test (BLRT) were used to improve the model fit, with significant values indicating a better fit for the k-class model than the k-1-class model. Entropy values that exceed 0.80 indicate a satisfactory classification accuracy [17 (link)]. Among the LCA models with different numbers of latent classes, a lower AIC, BIC, aBIC, larger entropy, and significant LMR-LRT and BLRT p values were indicative of good model fit [18 (link)]. Clinical interpretability was also considered to decide the best option. After the optimal model was determined, between-group difference was examined using Chi-square tests, Fisher’s exact tests or analysis of variance (ANOVA) where appropriate. Only statistically significant variables were entered into the stepwise logistic regression model. The regression was conducted separately by age groups to determine the contributing factors of symptoms for each group. P < 0.05 was considered statistically significant.