Central features in health-related quality of life in older adults: network analysis using nationwide survey data

Background Population ageing is a global phenomenon that necessitates consideration of health-related quality of life (HRQoL) in older adults. Previous studies have investigated related factors including mobility, social support and living situations. Aims This study aimed to provide a network perspective on factors related to HRQoL. Method Cross-sectional nationwide data were obtained from the Korean National Health and Nutrition Examination Survey conducted from 2018 to 2020 for network analyses. Data for participants aged 65 years or above were analysed, resulting in a total of 4317 eligible cases. The variables included were EQ-5D (a measure of HRQoL), household income, education, living situation, subjective perceived health, Charlson Comorbidity Index (a measure of medical comorbidities), stress, exercise per week, alcohol consumption and smoking. Three networks were produced: (a) EQ-5D dimensions network, (2) EQ-5D dimensions, lifestyle and psychosocial factors network, and (3) overall EQ-5D index, lifestyle and psychosocial factors network. Node centralities, bridge centralities and edges of the networks were examined. Results The most central EQ-5D dimension was the ability to carry out usual activities. In the second network, subjective health, stress and anxiety/depression were revealed as nodes with high bridge centralities. Subjective health, exercise, and Charlson Comorbidity Index were nodes closely linked to the overall EQ-5D index. Conclusions The results emphasise the importance of enhancing functional independence and subjective health cognition, increasing routine exercise and reducing stress as targets for interventions to improve HRQoL in older adults.


Figure S1
Bootstrapped 95% confidence intervals for all edges in Network I.

Figure S2
Case-dropping bootstrap graph for centralities, strength, closeness, and betweenness for Network I

Figure S3
Bootstrapped 95% confidence intervals for all edges in Network II

Figure S4
Bridge closeness, bridge betweenness, and bridge two-step expected influence of EQ-5D dimensions and life factors.X-axis indicates the original scores of each centrality index.Variables are ordered from largest to smallest, from top to bottom as indicated in the y-axis.

Figure S5
Case-dropping bootstrap graph for bridge betweenness, bridge closeness, and bridge expected influence for Network II

Figure S6
Network III showing EQ-5D index and life and psychosocial factors.Thicker lines indicate stronger edge-weights.Red indicates negative edge-weights and green indicates positive edge-weights.

Figure S7
Bootstrapped 95% confidence intervals for all edges in Network III

Network Analysis
As our data is cross-sectional, we estimated an undirected, weighted network using Gaussian graphical model (GGM), where edges represent partial correlation coefficients (Burger et al., 2022;Epskamp et al., 2018).GGM requires the data to have a multivariate normal property.However, in the case of ordinal data, polychoric correlations can be used instead (Epskamp, 2016).The polychoric correlation matrix undergoes a regularization process to create a parsimonious and sparse model by giving penalty for model complexity (Epskamp & Fried, 2018).For GGM with ordinal data, a regularized estimation was conducted using the graphical least absolute shrinkage and selection operator (glasso) with the Extended Bayesian Information Criterion (EBIC) (Epskamp & Fried, 2018;Friedman et al., 2008).This regularization process is included in the qgraph package, called the EBICglasso function (Epskamp et al., 2012).As a result, spurious edges represented by weak partial correlation coefficients are removed from the network, thus improving its parsimoniousness and sparsity.
For centrality analyses, strength, closeness, and betweenness centrality indices were computed using the centrality function included in the qgraph package (Epskamp et al., 2012).Strength centrality is a measure of how strong a node is related to each adjacent nodes by considering the absolute values of edge weights.Closeness refers to the inversed sum length of the shortest path of a node to all other nodes in the network.Finally, betweenness is the number of times a node of interest is passed through in the shortest route between every possible pair of nodes in the network (Bringmann et al., 2019;Opsahl et al., 2010).

Bridge Centrality Analysis
Bridge centralities, including bridge closeness and bridge betweenness were identified using the bridge function in networktools package (Jones & Jones, 2017).To account for the presence of both positive and negative edges in the network, we computed a two-step bridge expected influence.Expected influence is a similar measure to strength, but with consideration of negative relationships.A two-step expected influence takes into account not only the edge between focal node and its connected node, but also subsequently connected nodes (Robinaugh et al., 2016).Bridge centrality indices are defined similarly to extant centrality indices, but in the context of node communities.

Accuracy and Stability
For all parameters estimated, we performed bootstrapping to evaluate their accuracy and stability using the bootnet package (Epskamp et al., 2018).First, the accuracy of edge-weights of the network was evaluated by estimating the 95% confidence interval (CI) on each edge using non-parametric bootstrap method with 1000 bootstraps.Next, centrality indices were also evaluated where applicable.Stability of centrality indices can be evaluated using the correlation stability coefficient (CS-coefficient).CS-coefficient is obtained by employing the case-dropping subset bootstrap, where portions of cases (i.e.observations from the data) are dropped from the data and correlation between the subset centrality indices and original centrality indices are calculated.CScoefficient is a measure of the maximum proportion of cases that can be dropped such that the correlation is higher than the chosen value.Recommended cutoff scores for CS-coefficients is 0.5, and it is advised not to interpret centrality indices below 0.25 (Epskamp et al., 2018).

Network Visualization
Network I was visualized using the qgraph function, as Figure 2a.All 10 possible edges are described by partial correlation coefficient regularized using the EBICglasso method, with hyperparameter set to the default value, 0.5 (Epskamp et al., 2012).The strongest edge is the edge connecting EQL2 (self-care) and EQL3 (usual activities), with partial correlation coefficient value 0.55.Nodes were placed with respect to the Fruchterman-Reingold algorithm applied using spring layout, where strongly related nodes are placed closer together (Fruchterman & Reingold, 1991).The results for bootstrapped CI for each edge is supported in Supplementary File, Fig. S1.
The centrality analysis revealed that overall, largest centralities were observed for EQL3, usual activities.
Figure 2b illustrates the raw centrality scores for strength, closeness, and betweenness.CS-coefficients were 0.75, 0.75, and 0.52, respectively.As all CS-coefficients were above 0.5, all indices were treated as accurate and were thus interpreted.The results for case-dropping bootstrap for the centrality indices can be accessed in Supplementary File, Fig. S2.