Introduction
Diabetes is among the most common chronic conditions worldwide and can lead to disability, complications, and a reduced life expectancy (Heald et al., Reference Heald, Stedman, Davies, Livingston, Alshames, Lunt, Rayman and Gadsby2020, Vos et al., Reference Vos, Lim, Abbafati, Abbas, Abbasi, Abbasifard, Abbasi-Kangevari, Abbastabar, Abd-Allah and Abdelalim2020). The worldwide prevalence of diabetes in 2021 was 10.5% among 20–79-year-olds, with projections that the prevalence would rise to 12.2% by 2024 (Sun et al., Reference Sun, Saeedi, Karuranga, Pinkepank, Ogurtsova, Duncan, Stein, Basit, Chan, Mbanya, Pavkov, Ramachandaran, Wild, James, Herman, Zhang, Bommer, Kuo, Boyko and Magliano2022). Depression is the most common mental illness worldwide, a growing public health challenge (Liu et al., Reference Liu, He, Yang, Feng, Zhao and Lyu2020), and a condition for which people with diabetes are at an increased risk (Nouwen et al., Reference Nouwen, Winkley, Twisk, Lloyd, Peyrot, Ismail and Pouwer2010). There is a two to three times higher incidence of depression in people with diabetes compared to people without diabetes (Roy & Lloyd Reference Roy and Lloyd2012). For those living with diabetes, depression is associated with difficulties in keeping up with diabetes management and treatment (Gonzalez et al., Reference Gonzalez, Peyrot, Mccarl, Collins, Serpa, Mimiaga and Safren2008), macro- and microvascular complications (de Groot et al., Reference De Groot, Anderson, Freedland, Clouse and Lustman2001, Lustman et al., Reference Lustman, Anderson, Freedland, De Groot, Carney and Clouse2000), higher health care costs (Egede et al., Reference Egede, Bishu, Walker and Dismuke2016), accelerated cognitive decline (Schmitz et al., Reference Schmitz, Deschênes, Burns, Danna, Franco, Ikram, Kivimäki, Singh-Manoux and Tiemeier2018), and early mortality (Katon et al., Reference Katon, Rutter, Simon, Lin, Ludman, Ciechanowski, Kinder, Young and Von Korff2005, Zhang et al., Reference Zhang, Norris, Gregg, Cheng, Beckles and Kahn2005, Black et al., Reference Black, Markides and Ray2003).
While symptoms of depression include feelings of sadness and worthlessness, insomnia/hypersomnia, loss of interest or pleasure, and suicidal ideation (APA 2013), there is considerable heterogeneity in the depressive symptoms experienced by individuals (Fried Reference Fried2017, Wakefield & Schmitz Reference Wakefield and Schmitz2013) as well as evidence that individual depressive symptoms differ in their risk factors, associations with biomarkers, impact on functioning, and responsiveness to antidepressants (Fried & Nesse Reference Fried and Nesse2015). The heterogeneity in depressive symptomology across individuals may limit our understanding of the association between depression and diabetes prevalence and outcomes. Currently, little is known about the structure of depressive symptomology, in terms of the role and importance of specific symptoms and symptom interactions, in people with diabetes. To improve our ability to tailor depression prevention and intervention efforts for people with diabetes, an exploration of specific symptoms and the depressive symptom structure in people with diabetes, and how it differs to those without diabetes, is necessary.
Network psychometrics allows for the underlying symptom structure of mental health conditions, such as depression, to be explored, visualised, and compared between groups (Borsboom & Cramer Reference Borsboom and Cramer2013). Network theory posits that psychological constructs like depression are heterogeneous, dynamic, and complex systems where symptoms are interconnected and influence one another (e.g. ‘fatigue’ and ‘difficulty concentrating’ directly impacting one another) (Cramer et al., Reference Cramer, Van Borkulo, Giltay, Van Der Maas, Kendler, Scheffer and Borsboom2016). Symptom-level interactions are key to conceptualising psychopathology in network theory (Borsboom & Cramer Reference Borsboom and Cramer2013).
A limited number of prior studies have examined depressive symptom networks in the context of diabetes. One previous study examined the depressive symptom networks of individuals with diabetes, heart disease, stroke, and cancer, before and after diagnosis (Airaksinen et al., Reference Airaksinen, Gluschkoff, Kivimäki and Jokela2020). They found that although the mean level of depressive symptoms increased after diagnoses, these changes were not reflected in the depressive network structure. Recent research, also examining depressive symptoms in individuals before and after diagnosis of type 2 diabetes or hypertension, found a significant increase in depressive symptom connectivity (Wan et al., Reference Wan, Feng, Ma, Ma, Wang, Huang, Zhang, Jing, Yang and Yu2022). However, in both studies depressive symptoms were scored using binary values (1 for yes, and 0 for no), and therefore the network models may suffer from a lack of specificity of the strength of depressive symptom interactions. A recent study explored the symptom structure of diabetes distress, depressive, and anxiety symptoms in adults with type 2 diabetes (McInerney et al., Reference Mcinerney, Lindekilde, Nouwen, Schmitz and Deschênes2022) and found regimen and physician-related symptoms to be central to diabetes distress. Feelings of failure were found to be potential bridges (i.e. symptoms of one condition strongly connected to another condition) between diabetes distress and depression, and symptoms related to worry and trouble relaxing to be potential bridges between anxiety and depression, as well as diabetes distress. However, this study was limited to only one time point and did not investigate depressive symptom centrality alone. Furthermore, these prior studies did not directly compare depressive symptom networks in people with and without diabetes. Comparing the depressive symptom interactions and centrality in the networks of those with and without diabetes may lead to a better conceptualisation of the structure of depressive symptomology in people with diabetes.
The present study aimed to identify the depressive symptoms that are most central among those with diabetes and compare depressive symptom networks between those with and without diabetes. Comparisons were made between those with and without diabetes at two time points, using data from the baseline assessment (2009–2011) and the Wave 4 follow-up assessment (2016) of The Irish Longitudinal Study on Ageing (TILDA). We hypothesised that there would be differences in depressive symptom networks between those with and without diabetes at both time points. Specifically, we hypothesised that depressive symptom centrality (i.e. which symptoms are most highly connected) would differ between those with and without diabetes, and that the overall connectivity within the networks would demonstrate heightened connectivity among individuals with diabetes compared to those without.
Method
Participants
Data were from TILDA, a large nationally representative prospective cohort study of ageing aimed at assessing the health, social, and economic circumstances of the older adult, community-living population in the Republic of Ireland (Kearney et al., Reference Kearney, Cronin, O’regan, Kamiya, Savva, Whelan and Kenny2011). Data collection for Wave 1 was carried out between 2009 and 2011 and 8,175 adults aged 50 and over and 329 younger spouses/partners of participants (N = 8507) were interviewed (TILDA 2019). Wave 4 data collection was carried out between January and December 2016 and included 5,715 TILDA respondents. Detail on sampling, study design, and the TILDA cohort profile has been published elsewhere (Kearney et al., Reference Kearney, Cronin, O’regan, Kamiya, Savva, Whelan and Kenny2011, Whelan & Savva Reference Whelan and Savva2013). Ethical approval for all waves of the TILDA study was obtained from the Trinity College Dublin Faculty of Health Sciences Research Ethics Committee (Kearney et al., Reference Kearney, Cronin, O’regan, Kamiya, Savva, Whelan and Kenny2011). Access to the TILDA dataset was obtained from the Irish Social Science Data Archive.
For the present study, participants were included if they had complete data on depression and diabetes status at baseline (Wave 1). The analysis examining follow-up data (detailed below) was further refined to those with complete data on depression and diabetes status at Wave 4. The final samples for the present study were n = 639 participants with diabetes and n = 7,837 without diabetes at Wave 1, and n = 1,151 diabetes with diabetes and n = 4,531 without diabetes at Wave 4. Between Wave 1 and Wave 4, there were n = 1,010 new cases of diabetes. Due to drop-out and having missing data on depression or diabetes status, n = 498 people with diabetes present in our sample at Wave 1 are not present at Wave 4. As there are five to seven years between data collection at Wave 1 and Wave 4, the sample with diabetes at Wave 4 is predominately represented by people with relatively newly diagnosed diabetes.
Measures
Depressive symptoms were assessed using the Centre for Epidemiological Studies Depression Scale (CES-D) (Radloff Reference Radloff1977). Respondents rate how often they have felt each depressive symptom in the past week on a 4-point Likert scale, from ‘Rarely or None of the Time (Less than 1 Day)’ (score of 0) to ‘Most or all of the time (5–7 days)’ (score of 3) (Briggs et al., Reference Briggs, Carey, O’halloran, Kenny and Kennelly2018, Radloff Reference Radloff1977). While the 20-item CES-D was available at Wave 1, at the Wave 4 follow-up, the 8-item CES-D (CES-D-8), which uses 8 items of the original 20-item scale, was used to screen for depressive symptoms. Therefore, for comparability across waves, we used CES-D-8 at Wave 1 and Wave 4. Previous research compared the CES-D-8 to the CES-D-20 in the TILDA cohort and found the CES-D-8 to be a valid and reliable measure of depressive symptoms in this sample (Briggs et al., Reference Briggs, Carey, O’halloran, Kenny and Kennelly2018). Internal consistency was good for the sample with (α = .83) and without (α = .81) diabetes at Wave 1, acceptable for those with diabetes at Wave 4 (α = .79), and good for those without diabetes at Wave 4 (α = .8).
Diabetes status was determined by a positive response to ‘Do you have diabetes or high blood sugar?’
Statistical analysis
Network analyses were conducted in RStudio (Version 2022.12.0 + 353). The mean, standard deviation, and polychoric correlations were examined for all items. Item informativeness was quantified by examining item standard deviations. In line with previous network analysis research (Marchetti Reference Marchetti2019), an item with a standard deviation 2.5 SDs below the mean standard deviation (MSD) was deemed to be poorly informative. An item redundancy test, from the R package networktools 1.20 (Jones Reference Jones2018), was used to test for overlapping pairs of items. Two items were deemed to be measuring the same construct (i.e. to be redundant) if the polychoric correlations between each of those items and all other items were statistically different in less than 25% of the cases (Jones Reference Jones2018). Networks of depressive symptoms were estimated for people with and without diabetes at Wave 1, respectively. These networks were then compared using the Network Comparison Test (described in detail below). These analyses were then repeated for the Wave 4 data.
A network consists of nodes, which represent the variables of interest (i.e. items on the CES-D-8), connected by edges, representing the relationships between nodes (Borsboom & Cramer Reference Borsboom and Cramer2013). In this study, the edges represent partial polychoric correlations, that is the correlation between two CES-D-8 items, when controlling for all other items in the network (Borsboom & Cramer Reference Borsboom and Cramer2013). Edge thickness and colour saturation reflect the edge weight (i.e. the strength of the relationship between two nodes), such that the thicker and more saturated the edge, the higher the absolute weight (Epskamp et al., Reference Epskamp, Cramer, Waldorp, Schmittmann and Borsboom2012). Negative relationships are denoted with a red edge and positive relationships are denoted with a blue edge (Epskamp et al., Reference Epskamp, Cramer, Waldorp, Schmittmann and Borsboom2012).
A Gaussian Graphical Model was estimated, with extended Bayesian Information Criterion model selection (Foygel & Drton Reference Foygel and Drton2010) to examine pairwise associations between nodes, controlling for all other correlations in the network. A polychoric correlation matrix was estimated as input, as the data were ordinal, using the cor_auto function in qgraph (Epskamp et al., Reference Epskamp, Costantini, Haslbeck, Cramer, Epskamp and Rsvgtipsdevice2017) and the R package lavaan (Rosseel Reference Rosseel2012). To estimate many parameters with relatively small datasets, as is needed when estimating psychological networks, a form of regularisation is necessary. To estimate a more interpretable network and to limit the number of spurious edges, the ‘least absolute shrinkage and selection operator’ (LASSO; (Tibshirani Reference Tibshirani1996), a regularisation technique, was used (using the graphical LASSO (glasso; Friedman et al., Reference Friedman, Hastie and Tibshirani2008) with the glasso package (Friedman et al., Reference Friedman, Hastie and Tibshirani2022). To control the balance between a model with more (sensitivity) and fewer (specificity) edges estimated, a hyperparameter must be set. In line with suggestions by Foygel & Drton (Reference Foygel and Drton2010), the hyperparameter was set to 0.5, to estimate a more parsimonious network. The Fructerman-Reingold algorithm was used to visualise the network, which places the nodes with higher centrality at the centre of the network. Further detail on the estimation of regularised partial correlations network can be found elsewhere (Epskamp et al., Reference Epskamp, Borsboom and Fried2018).
Once the networks were estimated and visualised, centrality indices were examined and reported. The centrality index of strength signifies the sum of absolute edge weights directly connecting one node to other nodes in the network (Epskamp et al., Reference Epskamp, Borsboom and Fried2018). The index of one-step expected influence allows for positive edges to outweigh negative, thus, signifies the non-absolute sum of the edges directly connecting one node to others in the network (Robinaugh et al., Reference Robinaugh, Millner and Mcnally2016).
In order to have confidence in the interpretation of network indices, a post hoc bootstrapping framework was used to assess the stability and accuracy of the networks. The non-parametric 2500-boostrapped confidence intervals (CIs) for each edge weight were inspected to assess the variability of edge weights (Epskamp et al., Reference Epskamp, Borsboom and Fried2018). Wide bootstrapped CIs indicate that caution should be exercised in intercepting the strength, but not the presence or direction, of an edge (Epskamp et al., Reference Epskamp, Borsboom and Fried2018). The stability of the networks was also investigated using subset bootstrapping and by computing a correlation stability coefficient (CS-coefficient) (Epskamp et al., Reference Epskamp, Borsboom and Fried2018). A CS-coefficient above 0.25 was considered interpretable and above 0.5 was considered highly stable (Epskamp et al., Reference Epskamp, Borsboom and Fried2018).
To compare networks between people with and without diabetes at each wave, the NetworkComparisonTest package was used (van Borkulo et al., Reference Van Borkulo, Boschloo, Borsboom, Penninx, Waldorp and Schoevers2015). Global strength invariance (i.e. the difference in the absolute sum of network edge weights, that is the overall connectivity of the network) and network invariance (i.e. the possible edge weight differences between networks) were examined.
Results
Sample characteristics are presented in Table 1. The item means and standard deviations, and polychoric correlations between items, were calculated and inspected for each group at both waves. Means and standard deviations for depressive symptoms in each network are reported in Table 2. Polychoric correlations for people with and without diabetes at Wave 1, and for people with and without diabetes at Wave 4, are presented in Supplementary Tables 1–4, respectively.
Descriptive statistics for the sample
Mean scores on the centre for epidemiological studies depression 8-item (CES-D-8) scale for each item. Scores ranged from zero to three
Item informativeness was examined and no item was found to be poorly informative (i.e. 2.5 SD below the MSD: Wave 1 people with diabetes MSD = 0.44 ± 0.78; Wave 1 people without diabetes MSD = 0.37 ± 0.71; Wave 4 people without diabetes MSD = 0.43 ± 0.73; and Wave 4 people without diabetesMSD = 0.4 ± 0.7). The redundancy test suggested no reductions (i.e. no item was found to be measuring the same construct as another). Therefore, all items were included in the analysis.
The networks of CES-D-8 items for people with and without diabetes at Wave 1 and Wave 4 are displayed in Figure 1. Node strength is presented in Figure 2. Expected influence is presented in Supplementary Figure 1. For those with diabetes at wave 1 and 4, respectively, the estimated network had 75% (21/28) non-zero edges, and 86% (24/28) non-zero edges. For those without diabetes at wave 1 and 4, respectively, the estimated network had 86% (24/28) non-zero edges, and 82% (23/28) non-zero edges. These indicate highly connected depressive symptom networks overall.
Network of depressive symptoms for people with (i) and without (ii) diabetes at wave 1 and with (iii) and without (iv) diabetes at wave 4. Line thickness and colour denote the strength and direction, respectively, of polychoric correlations between symptoms. Blue lines indicate positive correlations, and red indicate negative. The thicker the line, the stronger the correlation between two symptoms. Dep1 indicates I felt depressed; Dep2, I felt that everything I did was an effort; Dep3, my sleep was restless; Dep4, I was happy; Dep5, I felt lonely; Dep6, I enjoyed life; Dep7, I felt sad; Dep8, I could not get going. For those with diabetes at wave 1 (i), the node highest in node strength was depressed (Dep1), followed by everything’s an effort (Dep2) and couldn’t get going (Dep8). The order of the three nodes highest in expected influence was depressed (Dep1), couldn’t get going (Dep8) and everything’s an effort (Dep2). For those without diabetes at wave 1 (ii), the nodes highest in node strength were: depressed (Dep1), everything’s an effort (Dep2) and sad (Dep7). The nodes highest in expected influence were: depressed (Dep1), everything’s an effort (Dep2), and sad (Dep7). For those with diabetes at wave 4 (iii), the nodes highest in strength were sad (Dep7), not enjoying life (Dep6), and everything’s an effort (Dep2). The nodes highest in expected influence were depressed (Dep1), sad (Dep 7), and everything’s an effort (Dep2). For those without diabetes at wave 4 (iv), the nodes highest in node strength were depressed (Dep1), not enjoying life (Dep6), and everything’s an effort (Dep2). The nodes highest in expected influence were depressed (Dep1), sad (Dep7), and everything’s an effort (Dep2).
Depressive symptom strength scores for people with (i) and without diabetes (ii) at wave 1, as well as for people with (iii) and without diabetes (iv) at wave 4. Dep1 indicates I felt depressed; Dep2, I felt that everything I did was an effort; Dep3, my sleep was restless; Dep4, I was happy; Dep5, I felt lonely; Dep6, I enjoyed life; Dep7, I felt sad; Dep8, I could not get going.
Network stability among those with diabetes
At wave 1, the bootstrapped CIs around the estimated edge weights were inspected and some were found to overlap, suggesting that edge weight order should be interpreted with caution. However, five edges were significantly different to most other edges in the network (Fig. 3). Case-dropping subset bootstrapping indicated that node strength (CS[cor = 0.7] = 0.516) and expected influence (CS[cor = 0.7] = 0.673) were above the threshold of 0.5 to be considered highly stable. At wave 4, while many estimated edge weights are overlapping and therefore, their order should be interpreted with caution, three edge weights were significantly different from almost all others, namely, Not happy (Dep4) to Not enjoying life (Dep6), Everything’s an effort (Dep2) to Couldn’t get going (Dep8), and Felt lonely (Dep5) to Sad (Dep7). Node strength (CS[cor = 0.7] = 0.59) and expected influence (CS[cor = 0.7] = 0.75) were above the cut-off to be considered stable (Fig. 3).
Bootstrapped edge weight confidence intervals for people with (i) and without diabetes (ii) at wave 1, as well as for people with (iii) and without diabetes (iv) at wave 4. Dep1 indicates I felt depressed; Dep2, I felt that everything I did was an effort; Dep3, my sleep was restless; Dep4, I was happy; Dep5, I felt lonely; Dep6, I enjoyed life; Dep7, I felt sad; Dep8, I could not get going.
Network stability among those without diabetes
At wave 1, inspection of the bootstrapped CIs around the estimated edge weights indicated that, while some were overlapping, the edge between Not happy (Dep4) and Not enjoying life (Dep 6) was significantly different to all other edges (Fig. 3). The network was highly stable, with the CS coefficients for strength and expected influence being at the highest level tested (CS[cor = 0.7] = 0.75) and therefore, above the cut-off to be considered interpretable (0.5). At wave 4, similarly, while many bootstrapped CIs overlap (Fig. 3), some edge weights were significantly different from almost all other edge weights, namely, Not happy (Dep4) to Not enjoying life (Dep6); Everything’s an effort (Dep2) to Couldn’t get going (Dep8); Felt lonely (Dep5) to Sad (Dep7); Depressed (Dep1) to Everything’s an effort (Dep2); and Depressed (Dep1) to Sad (Dep7). The network was highly stable, with the CS coefficients for strength and expected influence being at the highest level tested (CS[cor = 0.7] = 0.75) and therefore, well above the cut-off to be considered interpretable (0.5; Fig. 3).
Network comparison
First, the networks were investigated to determine if they differed in terms of connectivity (i.e. if one network is more strongly connected than the other). At Wave 1, results from the global strength invariance test indicated that the difference in global strength between the depression networks for people with and without diabetes was not significant (S = 0.07, p = .39), where S refers to the difference in global strength between the networks. Next, the network invariance test was used to investigate whether the edges differed between networks. Results suggested that there was at least one edge that differed between the networks (M = 0.19, p = .03). Therefore, we performed exploratory post hoc testing of all the edges in the network to determine which edges differed between the networks. There were two edges that differed significantly between the networks at Wave 1, between Felt lonely (Dep5) and Couldn’t get going (Dep8) (E = 0.19, p = 0.00) and between Not enjoying life (Dep6) and Sad (Dep7) (E = 0.19, p = 0.00), where E refers to the value of the edge weight difference between the networks. The edge between Felt lonely (Dep5) and Couldn’t get going (Dep8) and the edge between Not enjoying life (Dep6) and Sad (Dep7) were significantly stronger for those with diabetes than those without diabetes.
The network comparison test indicated that the Wave 4 networks were not different between people with and without diabetes in terms of global strength (S = 0.01, p = .8) and edge weights (M = 0.12, p = .01).
Discussion
The present study estimated and compared networks of depressive symptoms in a cohort of adults with and without diabetes at each of two time points, approximately five to seven years apart. This study provides a more detailed understanding of depressive symptom interactions in older Irish adults with and without type 1 or type 2 diabetes. To our knowledge, this is the first study to investigate the network structure of CES-D-8 items using the 4-point response scale for people with diabetes and compare them to people without diabetes. Understanding the differences in depressive symptomology between people with and without diabetes may allow for more tailored, and thus effective, clinical treatment. We identified the most highly connected and influential symptoms in the depressive networks, and differences in the strength of symptom relationships between people with and without diabetes. Across the networks estimated, four symptoms emerged with consistently high node strength in all networks, that is, for people with and without diabetes and at all time points. These were feeling depressed, feeling like everything’s an effort, feeling sad, and couldn’t get going. Similar results were found when we examined the expected influence of each item. That these symptoms were consistently the most strong and influential may suggest that they are core components of depressive symptomology as measured by the CES-D-8. For people without diabetes, two of the nodes highest in node strength were the same at both time points (feeling depressed and feeling like everything’s an effort) while for people with diabetes, feeling like everything’s an effort was the only one to remain in the three most central items at Wave 4.
The symptoms of restless sleep, not feeling happy, and feeling lonely were consistently amongst the symptoms lowest in strength and expected influence. In particular, restless sleep was the node lowest in strength and expected influence, across all networks. This suggests that restless sleep may be a peripheral symptom of depression, as measured by the CES-D-8. According to network theory, the most central symptoms trigger and maintain the depressive network (Borsboom Reference Borsboom2017). Therefore, targeting poor sleep alone through clinical interventions may not significantly decrease the connectivity or strength of the overall depressive network compared to targeting other, stronger, and more influential symptoms. These findings contradict recent research on non-pharmacological sleep interventions, which showed their effectiveness in reducing depression severity, especially in clinical populations (Gee et al., Reference Gee, Orchard, Clarke, Joy, Clarke and Reynolds2019). However, there were notable differences as the present study was observational and focused on specific populations.
The connectivity of the networks (i.e. how strongly or densely connected the networks are) were also compared between people with and without diabetes at each wave in the present study. The network comparison test determined that the networks were not significantly different between people with and without diabetes in terms of connectivity, at either wave. However, for people with diabetes, the depressive network at Wave 4 was more highly connected (86%; 24/28 non-zero edges) than the depressive network at Wave 1 (75%; 21/28 non-zero edges). Those with diabetes at Wave 4 were predominately (87.74%) people with relatively new (within the last 5–7 years) diagnoses of diabetes. This is in line with recent research by Wan et al. (Reference Wan, Feng, Ma, Ma, Wang, Huang, Zhang, Jing, Yang and Yu2022), which found a statistically significant increase in depressive symptom connectivity in the 2 years following diagnosis. In contrast, Airaksinen et al. (Reference Airaksinen, Gluschkoff, Kivimäki and Jokela2020) found the connectivity of depressive symptoms remained unchanged before and after diagnosis (Airaksinen et al., Reference Airaksinen, Gluschkoff, Kivimäki and Jokela2020). However, Airaksinen et al. (Reference Airaksinen, Gluschkoff, Kivimäki and Jokela2020) used a binary measure of depressive symptoms, which may have impacted the sensitivity of their findings. The network theory of psychopathology proposes that a more densely connected network is indicative of stronger reinforcing feedback loops between symptoms, important to the progression and maintenance of psychopathology (Borsboom Reference Borsboom2017, van Borkulo et al., Reference Van Borkulo, Boschloo, Borsboom, Penninx, Waldorp and Schoevers2015). Our sample was between five and seven years older at Wave 4 than at Wave 1 and predominately people with newly diagnosed diabetes. Therefore, the increased connectivity between these time points could suggest an increase in the reactivity and self-reinforcing nature of the depressive profile for people with diabetes as they age, or in the aftermath of diagnosis. There is also evidence that the density of a network’s connectivity may render it more responsive to treatment (Esfahlani et al., Reference Esfahlani, Sayama, Visser and Strauss2017, Peralta et al., Reference Peralta, Gil-Berrozpe, Librero, Sánchez-Torres and Cuesta2020). Still, more research is necessary to test and confirm the clinical implications of network indices.
Furthermore, our analyses identified two edges, or relationships between symptoms, that were significantly different between the networks for people with and without diabetes at Wave 1. The relationship between ‘felt lonely’ and ‘couldn’t get going’ was significantly stronger for people with diabetes than for those without. Perceived loneliness has been shown to be significantly predictive of higher HbA1c in people with self-reported diabetes, after controlling for demographics, depression, and number of chronic illnesses (Huang et al., Reference Huang, Cho, Kuo and García2022). Diabetes self-management behaviours appear to be impacted by levels of social support (Koetsenruijter et al., Reference Koetsenruijter, Van Eikelenboom, Van Lieshout, Vassilev, Lionis, Todorova, Portillo, Foss, Serrano Gil, Roukova, Angelaki, Mujika, Knutsen, Rogers and Wensing2016, Strom & Egede Reference Strom and Egede2012). Our findings suggest that the link between feeling lonely and the somatic symptom of not being able to ‘get going’ might be an important pathway for both clinicians and researchers to explore to better understand the mechanisms underlying the relationship between loneliness and diabetes management. Results of the network comparison test also indicated that the relationship between ‘not enjoying life’ and ’Sad’ was significantly stronger for people with diabetes than those without at Wave 1. These findings highlight the capability of network analyses to provide unique insights into the structure of mental health conditions and associations between symptoms, which may lead to more personalised clinical care.
Strengths and limitations
To our knowledge, this was the first study to compare depressive networks between people with and without diabetes at two time points. Our findings highlight differences in individual symptom relationships between those with and without diabetes. As the TILDA dataset is a nationally representative prospective study of adults over 50 living in Ireland, the findings should be generalisable to the population of Ireland. By taking a symptom-level approach, this study highlights individual symptoms and symptom relationships that could be targeted as part of a personalised approach to mental health treatment in the future.
The study has several limitations. First, the findings may not be generalisable beyond the predominantly White and older adults living in Ireland who were included in the TILDA dataset. Second, the data collection took place between 2009 and 2016, and changes in diabetes technology and healthcare services since then may affect the applicability of the results to the present day. Third, confounding factors were not accounted for in the network analysis; only the depressive symptom structure was modelled. It is possible that connections between nodes and node strength can be explained by factors not modelled in the network. While our focus was on the crude structure of depressive symptoms in people with diabetes, future research should use datasets with much larger numbers of people with diabetes, and with diabetes specific data, to consider factors such as complications, diabetes duration, HbA1c levels, and demographics. Fourth, while there was a significant difference in edge weights between people with and without diabetes at Wave 1, this difference did not persist at Wave 4, possibly due to sample size limitations. Additionally, the study did not focus on diagnosed clinical depression. In a study using the network comparison test to compare depressive networks in pregnant women, there was a difference in global strength (connectivity) but not in network structure between women with and without depression (Santos et al., Reference Santos, Fried, Asafu-Adjei and Ruiz2017). Future research should further explore differences between clinical and non-clinical populations. At present, our findings should be interpreted in relation to depressive symptomology in the general, non-clinical population. Finally, although we used polychoric correlations to preserve information on data severity and order, there is currently no established ‘gold standard’ for handling ordinal psychological data in network analysis (Epskamp Reference Epskamp2017).
Conclusion
The present study highlights the symptoms which may be central to depressive symptomology in adults over 50 living in Ireland with and without diabetes. Findings suggest that specific symptom relationships, namely, between ‘felt lonely’ and ‘couldn’t get going’ and between ‘not enjoying life’ and ’sad’, were significantly stronger for people with diabetes than for those without. As such, the present paper identifies key areas of symptom-level inquiry for researchers and clinicians working with older adults in Ireland with diabetes. Furthermore, clinicians could consider utilising questionnaires, such as the CES-D-8, as a tool to supplement a full clinical interview and gain a more in-depth understanding of each individual’s unique disease burden.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/ipm.2024.10.
Acknowledgements
Data from The Irish Longitudinal study on Ageing (TILDA) were accessed via the Irish Social Science Data Archive (ISSDA) – www.ucd.ie/issda.
Financial support
TILDA is co-funded by the Government of Ireland through the Office of the Minister for Health and Children, by Atlantic Philanthropies, and by Irish Life. Access to TILDA data can be requested via the Irish Social Science Data Archive.
Competing interests
None.
Ethical standards
The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committee on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008.
Open access
