Distinct social network structures and their cognitive and psychological correlates

Abstract

Social connectedness promotes healthy aging and is associated with lower risk for psychological disorders and cognitive decline. However, little is known about the mechanisms underlying these relationships, and whether different network features are associated with unique health benefits. We used comprehensive data from 386 community-dwelling older adults with and without cognitive impairment to test the relationship between psychological and cognitive function and their personal social networks. Data were collected using a multisite sampling strategy, and included detailed social network interviews and comprehensive measures of episodic memory, executive function, and language. Longitudinal effects were evaluated using a subsample at high-risk for decline, having either at least one copy of APOE ε4 or a current diagnosis of impairment (n = 118; average first follow-up = 2.28 years). Having smaller, affectively closer, and tightly connected social networks (i.e., social bonding) was modestly associated with positive psychological health. However, larger networks that contained weak ties, were more diverse with respect to social roles, and were less dense (i.e., social bridging) demonstrated robust effects on cognition. Longitudinal analyses demonstrated a sustained relationship between social bridging and cognition. Different kinds of social networks, reflecting distinct social mechanisms, may be associated with psychological and cognitive health.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-44571-9.

Introduction

Social connections play a powerful role in human health, well-being, and mortality¹. Meta-analyses suggest that there is a greater than 60% increased likelihood of survival among socially supported and integrated adults – a larger protective effect than that associated with nonsmoking, nondrinking, and exercising². Social connectedness is a modifiable lifestyle factor that could significantly reduce dementia risk³, though the specific aspects of cognition that benefit from connectedness have been underspecified. Social connectedness, especially features like social support and lack of social isolation, has also been associated with greater psychological wellbeing⁴. However, the specific mechanisms by which social connectedness may confer benefits for distinct domains of health and well-being have not been well-characterized or empirically demonstrated, largely because of limitations in the measurement precision of both social connectedness and health outcomes.

Longitudinal and empirical evidence suggests that having affectively close, supportive, homogenous, and densely connected social networks, referred to as social bonding capital in the network literature^5–8, is associated with greater psychological well-being, including less loneliness, lower levels of psychological distress, and fewer symptoms of depression and anxiety^9–12. These findings align with socioemotional selectivity theory¹³ and differential investment of resources¹⁴, which suggest that spending time with close, supportive, and emotionally meaningful social partners is beneficial for older adults’ overall psychological and emotional wellbeing by supporting emotion regulation and effective coping with acute and chronic stress^7,15. Despite these robust findings on psychological well-being, the social network features that relate to cognitive well-being have been less well-characterized. Indeed, empirical evidence on the relationship between key features of social bonding networks (e.g., social support) and cognitive health, including dementia and cognitive decline, have been inconsistent, with some studies finding an association^16,17 and others finding none^18–20. Cognitive health involves distinct neuropsychological processes (e.g., executive function, language, memory) that plausibly benefit from different social inputs than those affecting general or psychological well-being.

While a large majority of research on social connectedness and health has focused on social bonding features, social bridging has received much less theoretical and empirical attention. This has left a critical gap in our ability to disentangle how and why social connectedness affects cognitive health. Specifically, social bridging capital is linked to networks characterized by an expansive set of moderately close, non-kin ties (e.g., friends, coworkers, neighbors) that are more heterogenous and less interconnected^6,8,21. Bridging networks provide access to novel and complex social stimuli and increase cognitive demands, engaging multiple neurocognitive functions^19,22–26. Consistent with this assertion, a small body of prior work, all cross-sectional, has shown that maintaining a large variety of weaker social connections is positively related to global cognitive function, brain volume, and brain reserve^19,22,25.

Although bonding and bridging are widely referenced as theoretical concepts, they are rarely operationalized in a way that allows their independent effects to be directly compared. At least four methodological limitations characterize the existing literature on social connectedness and health. First, social connectedness has been operationalized using a wide variety of measures that likely reflect distinct social processes (e.g., number of relationships or network size^1,27,28, frequency of social contact²⁹, perceived social support³⁰, and participation in social activities^27,31, with little attempt to map them onto mechanistic theories or hypotheses. This heterogeneity across samples and studies has made it difficult to generate consensus about specific social mechanisms and how or why they drive different kinds of health outcomes, particularly in the domain of cognitive health.

Second, most past research has used proxy questions that summarize complex social phenomena using a single measure or scale. These often-unidimensional proxy measures do not capture the richness and complexity of people’s social networks^32,33 and reflect general perceptions rather than precise social measurements of features like network structure or composition^34–36. This prohibits comparative hypothesis testing and provides an incomplete assessment of mechanisms. In contrast, an egocentric social network approach can be used to collect high dimensional data about each of a large and heterogenous set of network members and relationships, providing a wide range of precise measures of the structure, function, and composition of individuals’ social ties across a range of different social spheres³⁷. Indeed, prior work has shown that social networks have multiple correlated structural and compositional features that reflect distinct social mechanisms^6,19,20. These are obscured in studies using single/unidimensional or proxy measures of social connectedness, reifying network theories without isolating which network features are doing the work of improving or preserving health.

Third, past research has rarely explicitly empirically examined whether different network features have unique health benefits for cognitive and psychological health. As noted above, most studies focus on a single network feature and relate them to a single, often global, health indicator. This pattern reflects common methodological constraints, including the burden of collecting high dimensional data on social networks and multiple health domains, collinearity among network features, and the analytical complexity of modeling multidimensional networks and outcomes simultaneously³⁷. Consequently, bonding- and bridging-oriented processes are frequently treated as interchangeable or are assumed to operate through similar downstream cognitive or biological pathways, despite strong theoretical reasons to expect that affectively close, dense networks and larger, more diverse networks may differentially shape psychological versus cognitive health. This is problematic because it obscures domain-specific social mechanisms, encourages overly general conclusions about “social connectedness,” and constrains the development of interventions tailored to distinct health outcomes.

Fourth, existing work on social connectedness within the domain of cognitive aging is particularly thin because it lacks specificity regarding the cognitive benefits of network features that may promote resilience to neuropathology and neurological disorders among older adults. That is, empirical studies of the relationship between social networks and cognitive health have focused primarily on global cognitive function, largely overlooking subdomains of cognition, most notably executive function and episodic memory (see¹ for meta-analysis). This is important because episodic memory decline is an early hallmark of Alzheimer’s disease, for example, whereas executive function declines develop as the disease progresses³⁸. Moreover, bridging- and bonding-oriented network features may differentially promote specific cognitive processes, with one study suggesting that larger, more diverse networks predict better episodic memory and social cognition, both cross-sectionally and longitudinally²³, and other work showing that affectively close networks relate to experiencing more positive emotions⁴. Identifying the social network features that relate to unique subdomains of cognition is thus important for characterizing the mechanisms through which social networks predict better cognitive outcomes.

Rather than merely extending prior work, the current study clarifies how distinct social mechanisms that are often conflated in prior research – social bonding and bridging – map onto multiple specific dimensions of cognitive and psychological health. Two primary hypotheses guided this study: First, social bridging, but not social bonding, was expected to be associated with higher cognitive functioning across multiple domains, both cross-sectionally and longitudinally. This prediction is consistent with limited cross-sectional evidence and bridging networks’ association with novel and complex social stimuli and high social cognitive demands^19,22–26. Second, we hypothesized that social bonding, but not social bridging, would be associated with better psychological health, both cross-sectionally and longitudinally, due to bonding networks’ facilitation of emotion regulation and effective coping with acute and chronic stress^5–8,15. Given shared variance between bonding and bridging mechanisms, we added a third exploratory hypothesis, positing that a balanced network typology that maximizes bonding and bridging would be simultaneously beneficial for cognitive and psychological health. For all hypotheses, cross-sectional analyses were conducted using a cognitively diverse sample of older adults. Hypotheses 1 and 2 also included longitudinal analyses focusing on a subsample of adults at high risk for Alzheimer’s disease. This study is unique due to its high dimensional characterization of social networks³⁷ alongside rigorous extensive measures of cognitive and psychological health.

Methods

Participants

All participants provided informed consent prior to completing any part of this study. This protocol was approved by the Institutional Review Board in the IU Human Research Protection Program. Procedures align with international standards for ethical human subjects research as specified in the World Medical Association Declaration of Helsinki.

Data from 386 participants were collected between 2015 and 2022 from two sites. Multisite sampling was employed to enhance sample heterogeneity and to improve the statistical validity and generalizability of this work. As shown in Table 1 and 233 participants (74% White, 58% married, 61% female) were recruited in collaboration with the Indiana Alzheimer’s Disease Research Center. Of these, 62 were formally diagnosed with mild cognitive impairment or Alzheimer’s Disease and related dementia (ADRD). An additional sample of 153 cognitively normal participants (95% White, 54% married, 66% female) was recruited from the community at a second site in southern Indiana. This sample successfully completed a six-item cognitive screener³⁹ on enrollment. Exclusion criteria were aligned with the protocols at the Indiana Alzheimer’s Disease Research Center, including advanced impairment (i.e., less than 10 on the Montreal Cognitive Assessment), being under 55 (n = 9; 2.3% were under 60), family history of dominantly inherited dementia, history of psychiatric disorders (i.e., schizophrenia, substance use), traumatic brain injuries, cancer treatment, and developmental disabilities.

Table 1.

Demographic information by sample.

Variable	Cross-sectional analysis			Longitudinal
	Cognitively normal			ADRC pilot sample
	Community (n = 153)	IADRC (n = 171)	MCI or ADRD (n = 62)	High ADRD risk (n = 118)
Age	74.0 (7.02)	71.0 (6.78)	75.4 (6.27)	71.2 (6.27)
Living alone	30 (20%)	30 (18%)	13 (21%)	17 (14%)
Functional limitations
None	130 (85%)	122 (71%)	12 (19%)	–
1 or more	23 (15%)	49 (29%)	50 (81%)	–
Gender
Men	52 (34%)	54 (32%)	37 (60%)	51 (27%)
Women	101 (66%)	117 (68%)	25 (40%)	67 (73%)
Race
Native American	–	1 (0.5%)	–	–
Pacific Islander	–	1 (0.5%)	–	–
Asian	2 (1%)	1 (0.5%)	–	1 (1%)
Black	1 (1%)	49 (29%)	7 (11%)	16 (13%)
White	146 (95%)	118 (69%)	54 (87%)	101 (86%)
Other	4 (3%)	1 (0.5%)	1 (2%)	–
Relationship status
Never married	8 (5%)	3 (2%)	8 (13%)	21 (18%)
Not reported	8 (5%)	–	5 (8%)	26 (22%)
Married	83 (54%)	93 (54%)	42 (68%)	71 (60%)
Divorced	36 (24%)	47 (27%)	3 (5%)	–
Separated	–	2 (1%)	–	–
Widowed	17 (12%)	26 (15%)	2 (3%)	–
Years of education
> High school	–	2 (1%)	–	1 (1)
High school	9 (6%)	13 (8%)	8 (13%)	9 (8%)
Some college	23 (15%)	31 (18%)	17 (27%)	26 (22%)
College graduate	46 (30%)	45 (26%)	12 (19%)	23 (19%)
Advanced degree	67 (44%)	80 (47%)	25 (40%)	59 (50%)
Employment status
Retired/not working	124 (81%)	140 (82%)	53 (85%)
Part-time	23 (15%)	11 (6%)	4 (7%)
Full-time	6 (4%)	20 (12%)	–
Not reported	–	–	5 (8%)

IADRC = Indiana Alzheimer’s disease research center. MCI = Mild cognitive impairment. ADRD = Alzheimer’s disease and related dementias. The longitudinal sample does not have data on functional limitations and only specifies marital status as either married or not.

To assess longitudinal associations between social networks and health, we leveraged longitudinal data from a subset (n = 118) of the Indiana Alzheimer’s Disease Research Center sample who were at high-risk for ADRD and had completed at least one longitudinal follow-up. These individuals all had at least one allele of APOE ε4 and/or a diagnosis of mild to moderate impairment at their initial visit. Official diagnoses of mild cognitive impairment and Alzheimer’s disease for Alzheimer’s Disease Research Center sample were provided via a consensus panel including neurologists, neuropsychologists, psychiatrists, clinical trainees, and other staff using standardized diagnostic criteria. We focused on this sample for longitudinal analysis because it ensured a high probability of observing measurable changes in cognitive function over a relatively short follow-up period. Participants averaged between 2 and 3 follow-up visits with a majority having at least two (78%), some with three (34%) and a smaller number with four (16%). The high attrition rate was due in part to the COVID-19 pandemic occurring in the middle of the study, temporarily pausing data collection. The average time between visits was 1.28 years (SD = 0.64), with a range from six months to over four years.

Materials

Social network interview

Personal social networks were collected using an expanded version of the PhenX Social Network Battery⁴⁰, developed by a panel of researchers from the National Human Genome Research Institute to provide high-quality, standard measures for research. Trained interviewers gathered the names of individuals (i.e., alters) that participants had regular contact with in the past six months to discuss important and/or health matters³⁷. These name generators targeted people whom the participant actively sought out for advice and/or people who hassled and/or encouraged health behavior change to maximize the amount and types of interactions captured. The interview ended here for all longitudinal cases and subsequent analyses; see the supplemental materials for a list of name generators. For a more robust assessment of social bridging in the full sample, additional name generators asked about relationships in a variety of social settings (i.e., family, weekly activities, neighbors, work, volunteering) to elicit ties beyond a “core” network. Interviews usually took between 45 min and 1.5 h.

Participants could name an unlimited number of people to their networks, with the total (i.e., network size) in the sample ranging from 0 (six subjects reported no ties) to 54. After names were generated, participants answered follow-up questions about each alter, including the frequency of contact (1 = hardly ever; 2 = occasionally; 3 = often), perceived closeness (1 = not very; 2 = sort of; 3 = very), relationship strength (1 to 10, with 10 being high), whether the person provided support (i.e., listening, emotional, advice, daily chores, financial; 1 = Yes; 0 = No), and the relationship type (e.g., friend, family, coworker, medical professional). Measures of network structure were gathered by first asking whether alters know each other (0 = No) and, if so, how close they are (1 = not very; 2 = sort of; 3 = very). For all measures, aggregated outcomes were created by averaging across alters in a participant’s network. Other alter-level characteristics (i.e., age, gender, race) were collected but were not relevant to the analyses.

Quantifying social bonding and bridging

The operationalization of social bridging and bonding follows well-established theory^{5–8,15,21,34}. Because measures derived from egocentric social network data are often highly correlated (e.g., network size negatively correlates with density, average tie strength positively correlates with average contact), a latent variable approach⁴¹ was used to estimate bridging and bonding, following previous work^19,23,24. Latent variable modeling adjusts for collinearity and measurement error inherent in egocentric network data³⁷. Variables in each model were justified using existing theory and past evidence; however, these latent constructs were estimated de novo to ensure an accurate reflection of the covariance structure and measurement properties of the current sample. Structural equation modeling was used to reduce measurement error, allow multiple observed indicators to contribute to each construct, and accommodate potential heterogeneity in how theoretically defined variables manifest. This approach is particularly appropriate given the multisite sample and allows the latent variables to reflect sample-specific expressions of social network characteristics.

A latent variable was estimated to represent social bridging using six structure-focused variables used in prior work (see⁴¹: network size, effective size, diversity, density, sole bridge status, and minimum tie strength. All measures were calculated and entered independently into the latent model. Effective size identifies the number of groups of nonredundant ties (i.e., ties that are not connected) by subtracting network size from the mean number of ties that each alter had to all other alters. Diversity assessed the presence of multiple types of social relationships⁴² using the number of social roles that a participant reports in their network (e.g., spouse, child, friend, coworker, neighbor) which ranged from 0 to 9. Density was measured by taking the average closeness of ties between alters (ranging from 0 to 3) with scores of 0 indicating that no network members knew each other. Sole bridge status was a binary indicator of having an unconnected or isolated alter (i.e., someone that no one else in the network knows) with 1 = sole bridge present. Minimum tie strength ranged from 1 to 10 based on the weakest tie.

The latent estimation of social bonding used four network variables focused on relationship quality – contact frequency, relationship strength, closeness, and support – following theoretical precedent^5–8,15 and the distinct operationalizations implemented in past work^27–31. For each alter, participants reported on frequency of contact ranging from 1 to 3 (corresponding to “hardly ever,” “occasionally,” and “often”), closeness ranging from 1 to 3 (corresponding to “not very close”, “sort of close” and “very close”), and tie strength using a sliding scale from 1 to 10 (i.e., no response categories). We then calculated the average for each of these variables across all alters. Support was measured using the number of distinct functions provided by each alter (i.e., listening, caring, providing advice, help with chores, loaning money; potential range of 0–5), which was averaged across all alters.

A confirmatory factor analyses confirmed that a single factor was needed for each construct. Fit statistics for each suggested excellent model fit (model vs. saturated Χ² is insignificant at 0.05 level, CFI and TLI > 0.95, RMSEA < 0.08). All factor loadings can be found in supplemental Fig. 1. The six participants (1.5% of the sample) reporting no alters were imputed using the lowest possible score after constructing the latent models, following precedent⁴¹. Omitting these cases did not impact the significance of any effects. Importantly, these latent variables were normally distributed (see supplemental figures), but significantly negatively correlated (r = − .48, p < .001).

Cognitive function

Cognition was measured using two approaches. First, we collected the Montreal Cognitive Assessment (MoCA)⁴³ as a global cognition measure; possible scores ranged here from 0 to 30. Second, we collected a comprehensive assessment of six cognitive domains – attention, processing speed, executive function, episodic memory, language, and visuospatial ability – using the Uniform Data Set (UDS) 3.0, which is administered in all NIH-funded Alzheimer’s Disease Research Centers⁴⁴. In addition to providing a rigorous assessment of a wide range of cognitive abilities, this measure also includes sample norms and validated latent factor structures drawing on thousands of cases from across the US⁴⁵. Details on the UDS administration are presented in supplemental materials. All cognitive data were standardized using z-score transformations prior to any analyses (i.e., calculating composites, inferential tests), consistent with past work²⁴.

Mental health

All participants completed the 15-item Geriatric Depression Scale⁴⁶ and four items from the anxiety subscale of the Hopkins Symptom Checklist⁴⁷. Given its strong correlation to depression and cognitive decline⁴⁸, loneliness was collected using the 3-item UCLA loneliness scale⁴⁹. Two positive outcomes were also collected: a single-item measure of happiness ranging from 1 (not at all) to 10 (very much so) and the eight-item measure of perceived global social support from the Multidimensional Scale of Perceived Support⁵⁰. Participants in the longitudinal analysis did not complete measures of loneliness or social support. All measures were standardized via z-score transformations prior to analyses.

Transparency and openness

Sample size determinations were done a priori in G*Power 3.1⁵¹. For cross-sectional regression models, a minimum sample size of 58 is needed to detect a single, significant coefficient with a small effect size (R²_p = 0.10, effect size f² = 0.11) with α = 0.05 and β = 0.80. Lowering the Type I error rate (α = 0.01) and increasing Type II errors (β = 0.95) yielded a minimum of 145. Thus, adequate power was present for these tests. For longitudinal modeling, power demands vary based on the exact data structure⁵². The online tool, GLIMMPSE⁵³, was used to confirm that the reduced longitudinal sample met power demands. Sufficient power (β > 0.80) was present using the Hotelling-Lawley Trace and conditional power method with α = 0.05 and sample size of 118, even if variability tripled. Data exclusions and analytic processes are detailed here with all materials available through the main and supplemental files. Cognitive tests can be found on the National Alzheimer’s Coordinating Center website. All data and code will be shared upon request to the first author.

Analytic strategy

For all hypotheses, analytic models were constructed using Stata SE 17.0 with covariates including age, gender, education, clinical diagnosis, and site, when relevant. To control for bidirectional relationships between psychological and cognitive health in older adults^12,18, depression symptoms were controlled when predicting cognitive outcomes; MoCA scores were controlled when predicting psychological outcomes. Functional limitations were controlled in both analyses using the Functional Activities Questionnaire⁵⁴. Scores were binarized such that individuals who indicated no limitations (n = 264) had a score of 0 and those who indicated any limitations had a score of 1 (n = 122). Marital status was included as a covariate (1 = married, 0 = not married; alternative coding separating divorced, widowed, etc. did not affect results) after an independent samples t-test showed that social bonding was higher in networks of people who were married (M = 0.073, SE = 0.09) versus those who were not (M = − 0.128, SE = 0.06), t(384) = 2.11, two-tailed p = .05. Similarly, employment status was included (0 = retired or not working, 1 = working full- or part-time) since people still working had more social bridging in their networks (M = 0.235, SE = 0.11) versus those who were not (M = − 0.05, SD = 0.06), t(384) = 2.11, two-tailed p = .035. To address missingness across demographic variables, multiple imputations were used via full information maximum likelihood in structural equation modeling⁵⁵.

For Hypothesis 1 and 2, fixed effects regressions were used to test longitudinal effects, specifically whether changes in social bridging (Hypothesis 1) or bonding (Hypothesis 2) predicted changes in health outcomes in a sample at high-risk for ADRD. These models treated all variables as repeated measures to uncover within-person associations across time. While removing time-invariant heterogeneity, fixed effects models are robust against unbalanced observations by controlling subject-level variance when estimating standard errors for predictors⁵². A key requirement for accurate estimation is the presence of change in the dependent variable, in this case cognition. Moreover, these models provide arguably the most conservative estimate of intraindividual change over time. Given the limited sample size, this approach is more appropriate than those requiring many repeated observations from a large number of individuals (e.g., dynamic structural equation modeling).

For Hypothesis 3, exploratory k-means cluster analysis⁵⁶ was used to determine whether networks that contained elements of both social bridging and bonding were associated with better outcomes overall. Solutions were computed for up to 20 groups with the optimal solution being detected by scree plots of within-group sum of squares, effect size (i.e., η²), and proportional reduction of error. The best solution was used as a predictor to determine how network typology was associated with cognitive and psychological health.

Results

Hypothesis 1: bridging, but not bonding, is positively associated with cognition over time

Supplemental Table 2 presents coefficients for the following models testing relationships between social bridging and cognitive function in the full sample. As shown in Fig. 1, social bridging was significantly associated with higher MoCA scores (β = 0.13, p = .03), executive function (β = 0.28, p = .002), memory (β = 0.11, p = .023), and language (β = 0.16, p = .002). These models explained 33% of variance in MoCA scores, 30% in executive function, 16% in episodic memory, and 38% in language. Associations between bridging and attention, speed of processing, and visuospatial function were not statistically significant (all ps > 0.10). Curiously, social bonding was negatively associated with executive function (β = − 0.14, p = .019) and entirely unrelated to all other cognitive outcomes (all ps > 0.10).

Fig. 1.

Associations for bridging and bonding with cognitive and psychological outcomes in the full sample analyses. Bold points indicate a significant effect. Error bars are 95% CIs.

To evaluate whether these effects were persistent and offered protection against decline over time, fixed effects models tested within-person association between change in social bridging and change in cognitive function using the high-risk subsample (see supplemental Table 3 for results summaries). Controlling clinical status and age, social bridging was associated with within-person increases in episodic memory (β = 0.08, p = .004) and language (β = 0.06, p = .022), explaining 25% and 9% of within-person variance alongside 59% and 32% of between-variance.

Sensitivity analyses using regression-based residuals for both cross-sectional and longitudinal models (described in the supplemental materials and presented in supplemental Tables 4 and 5) confirmed the robustness of these effects. Including bonding simultaneously in these models did not change the effects of bridging in direction or magnitude. Along with including regression-based residuals, mixed effects variance components models with observations over time nested in individuals were used to conduct longitudinal sensitivity analyses. All longitudinal effects of social bridging on cognition were replicated using the more liberal estimation procedures with mixed effects models (see supplemental Table 6).

Hypothesis 2: bonding, but not bridging, is associated with psychological health

As in the previous analyses, several models were constructed to examine the relationship between social bonding and psychological health (see Fig. 1). See supplemental Table 7 for full results and supplemental Table 8 for longitudinal results. Social bonding was associated with lower depression (β = − 0.11, p = .037) and loneliness (β = − 0.15, p = .018) along with higher happiness (β = 0.16, p = .006). Associations were not significant for anxiety or perceived social support. When conducting fixed effects analyses, none of these effects persisted over time (all ps > 0.10) with very small within-subjects variance being explained (all R² < 0.04). However, social bonding was predictive of within-person change in MoCA scores (β = 0.15, p = .015). Cross-sectional analyses for social bridging found that it was associated with higher perceived social support (β = 0.13, p = .033), but no other psychological outcomes. Fixed effects showed no relationship between social bridging and mental health over time.

Sensitivity analyses for both cross-sectional and longitudinal models (described in the supplemental materials and presented in supplemental Tables 4 and 5) confirmed the robustness of these effects. First, including bridging simultaneously did not change the effects of bonding in direction or magnitude, although bonding also predicted perceived social support (β = 0.19) (see supplemental Table 9). Though the longitudinal fixed effects models did not change when bridging and bonding were entered together, the more liberal mixed effects model revealed that bonding was no longer predictive of MoCA scores over time (β = − 0.004, p = .93), and predicted lower depression (β = − 0.13, p = .029) and higher happiness (β = 0.30, p = .001) (see supplemental Table 6).

As an additional robustness analysis for Hypothesis 1 and 2, we compared coefficients for bridging (Hypothesis 1) and bonding (Hypothesis 2). To achieve this, postestimation comparisons were across each independent models using structural equation modeling and controlling for within-subjects clustering of standard errors. As shown in Fig. 1 via the confidence intervals for each point estimate, bridging was a stronger predictor for MoCA, executive functioning, episodic memory, and language (all ps < 0.01) whereas bonding was stronger for happiness (p = .03).

Figure 2 displays two hypothetical sociograms to visualize social networks that are related to cognition. Participants themselves are omitted to focus on the interaction patterns of the network to better reflect social bridging and bonding. Executive function, episodic memory, and language were directly related to social bridging; consequently, the sociograms were constructed to reflect social network structures of individuals who were high-performing (+ 1 SD) or low-performing (-1 SD) in at least one of these domains and three domains overall. After identifying 21 low- and 7 high-performing individuals that satisfied these criteria, average values were computed across key social network indicators: size, diversity, and density. Scores were mapped to a representative, “nearest neighbor” network to generate an approximation of size, network member types, and interconnectedness. Smaller, interconnected, and kin-focused networks (i.e., low social bridging; see left panel) were associated with poor cognition. Larger, loosely connected networks with more non-kin members (i.e., high social bridging; see right panel) may offer critical cognitive protection.

Fig. 2.

Two fictional networks associated with scoring + 1 SD on 3 or more cognitive outcomes. Ties are color-coded via relationship type. Participants are omitted.

Hypothesis 3: network typologies and associations with cognitive and psychological health

A k-means analysis identified specific network typologies, and a three-cluster solution best captured the spread of bridging and bonding without overfitting the data, explaining 62% and 69% of variance, respectively. Figure 3 displays a scatterplot of bridging and bonding with cases displayed based on their respective cluster. While bridging-focused networks (n = 104) were larger, less dense, and had more non-kin members compared to bonding-focused networks (n = 107), balanced networks were the most common (n = 175) being moderately sized and somewhat densely connected. See supplemental Table 11 for item-level averages for indicators of bridging and bonding across clusters. Using these details, it is important to note that balanced networks had similar levels of latent bridging compared to the bridging-focused networks (M_Balanced = 0.645, SD = 0.30; M_{Bridging−Focused} = 0.675, SD = 0.555), and differed from the bridging-focused networks mainly in their level of bonding (M_Balanced = 0.181, SD = 0.42; M_{Bridging−Focused} = -1.219, SD = 0.599). However, the level of latent bonding in the balanced networks was relatively low compared to the bonding-focused networks (M_{Bonding−Focused} = 0.889, SD = 0.624).

Fig. 3.

Scatterplot of bridging and bonding, coded by k-means cluster assignment. Centroids for each cluster are noted by filled circles.

Next. this three-cluster solution was used as a predictor, as with the original tests of Hypothesis 1 and 2. Balanced networks were tested against bridging-focused networks for cognitive and bonding-focused networks for psychological outcomes, (see supplemental Tables 11 and 12, respectively). We did not find support for Hypothesis 3 that balanced networks would be associated with positive health in both domains. Balanced-networks were not associated with better cognition than bridging-focused networks, and instead were associated with worse episodic memory, (β = − 0.17, p = .04). No differences in psychological outcomes were revealed when comparing balanced and bonding-focused networks, (all ps > 0.05). Therefore, these network types reflect the broad influences of bridging and bonding, and more importantly, balanced networks were not associated with cognitive health and psychological health as hypothesized (Fig. 4).

Fig. 4.

A visualization of networks across the three clusters. Network characteristics were derived using cluster centroids. Thicker lines indicate greater mean tie strength.

Discussion

This study sought to clarify how distinct features of personal social networks relate to cognitive and psychological well-being in later life by directly comparing theoretically motivated bonding and bridging mechanisms across multiple health domains. Prior research has established that social connectedness broadly benefits health^{1,2,9,16,19,25}; however, it has often relied on global health or general cognitive outcomes, assessed social ties cross-sectionally, and examined single, unidimensional, or proxy indicators of social connectedness. This has provided limited insight into whether different network mechanisms confer distinct health benefits and whether such associations persist over time. By leveraging high-dimensional network data, latent network constructs, and both cross-sectional and longitudinal analyses, the present study addressed these limitations and tested whether bonding- and bridging-oriented networks are differentially associated with domain-specific cognitive function and psychological well-being.

Consistent with Hypothesis 1, these data suggest that social bridging-focused networks were associated with significant cognitive benefits for older adults. Specifically, higher social bridging was associated with better memory, executive function, and language, all abilities prototypically impacted by ADRD^38,57. Longitudinal analyses support these findings, demonstrating that reductions in bridging were associated with more significant cognitive decline, particularly in episodic memory, among people at high risk. Moreover, it is important to highlight that the cognitive correlates of social bridging were highly domain-sensitive – a novel contribution of this study made possible by the robustness of the UDS cognitive battery. Though our findings also suggest that social bonding was cross-sectionally associated with better mental health, as predicted, these findings did not hold longitudinally in more conservative fixed-effects models, providing only partial support for Hypothesis 2. Critically, social bridging was not significantly associated with psychological well-being, nor was social bonding predictive of cognitive health outcomes.

Bridging-focused networks may associate with better cognitive function because being embedded in larger, more heterogenous, and less densely interconnected social worlds exposes people to novel or diverse social stimuli that promotes brain and cognitive resilience^22,25. Relationships and interactions in bridging-focused networks likely require high levels of effort, particularly social cognitive skills, which may help build cognitive reserve^19,20. This may explain the robust longitudinal relationship between changes in social bridging and cognitive decline uncovered here and identified cross-sectionally in some past work^19,22–26. Moreover, existing neuroimaging research corroborates associations between social bridging and bonding and divergent neural correlates of cognitive and psychological outcomes. Specifically, greater thickness in the frontal lobe and higher volume in the insula, neural regions that support lucidity and cognition, have been associated with bridging whereas social bonding has been related to greater volume and thickness in emotion regulation regions²⁵. Taken together with past studies, the current findings provide direction for investigating the protective effects of social bridging on core neuropsychological functions typically degraded in ADRD.

Our findings suggest that social bonding might not provide sufficient stimulation for building cognitive reserve. However, bonding-focused networks were associated with higher perceived social support, less loneliness, and less depression, consistent with aging theories and empirical evidence about the psychological benefits of strong, supportive, and interconnected relationships^4,5,7,12,14. Though speculative, one reason why bonding networks might not have been associated with cognitive health is that interacting with intimate and familiar social partners may require little social cognitive effort, especially for repetitive, day-to-day social exchanges⁵⁸. Indeed, navigating interactions with close bonding ties (e.g., marital or romantic partners, adult children) may engage less social cognition, including theory of mind, face memory, and emotion recognition, compared to exchanges with a diverse set of relatively less familiar interaction partners, such as friends, neighbors, and members of social clubs⁵⁹.

Additional insight about benefits of social bridging and its relationship to social bonding emerged from our analysis of network typologies in conjunction with our third exploratory hypothesis. Specifically, the k-means cluster analysis showed that a majority of networks were classified as balanced with both bridging and some levels of bonding, but with no apparent associations to cognitive and psychological health, showing only partial support of exploratory Hypothesis 3. This contrasts key predictions of socioemotional selectivity theory¹³ regarding the preponderance of bonding-focused relationships, but more importantly, such networks did not offer a “best of both worlds” scenario for psychological and cognitive health outcomes. It must be noted, however, that there were no networks that possessed exceptionally high bridging and bonding (i.e., above 1 on each dimension). Given that bridging and bonding are negatively correlated (see Fig. 3), it is likely that simultaneously maintaining networks with high levels of both may be difficult, constituting a tradeoff between more bridging- or bonding-dominant social organization strategies. A constellation of weak peripheral ties and a tight nucleus of very close ties is theoretically possible (e.g., with a strong core-periphery structure), but it is unclear whether truly balanced networks might be created and maintained or potentially optimized through social interventions.

An important direction for future research is determining whether, and to what extent, social bridging can be meaningfully increased through intervention to support cognitive health. Although technology-mediated interactions such as video chats have been shown to improve mental health, it remains unclear whether they provide sufficient cognitive and social complexity to promote cognitive resilience, despite evidence from at least one small randomized clinical trial suggesting benefits for cognitive resilience⁶⁰. Other structured social activities (e.g., book clubs) have also been shown to offer cognitive benefits for older adults⁶¹. However, it is unknown whether these activities work by increasing social bridging or instead primarily reinforce existing close ties or provide other non-social forms of cognitive stimulation. Together, these open questions highlight the need for interventions that explicitly target bridging-oriented network features and experimentally test social bridging as a mechanism for maintaining cognitive health in later life.

Constraints on Generalizability

These findings are limited by the relatively short period of observation. While the longitudinal results provide evidence consistent with a directional association and causal interpretation, they were collected after participants reached older adulthood. It is possible that these late-life networks reflect reverse causation playing out over the life course, such that those with greater cognitive abilities and resources were able to form and maintain a more complex and expansive set of social ties. Relatedly, people with bonding-focused networks in older adulthood might have already pruned social ties or experienced social contraction in response to diminished cognitive functioning or other unobserved time-varying characteristics. Perhaps the most plausible explanation is bidirectionality, whereby social networks and cognition are interconnected with experiences in early life that promote more engagement over the life course, in turn, building cognitive reserve and resilience that support network maintenance.

To further explore the issue of bidirectionality in the relationship between social networks and health, additional research should investigate the social factors that influence bridging and bonding and therefore precede cognitive decline. For example, these data showed that individuals who were retired had lower social bridging than those who were still working. Retirement may lead to a pruning of work-related ties and a reduction in social bridging and has also been associated with declines in cognitive function⁶². Likewise, some research has found that occupational complexity in working with people – an exposure that precedes cognitive decline and which promotes social cognitive skills analogous to bridging networks – is linked to better cognitive outcomes and lower risk for dementia²⁴. These results do not solve the question of reverse causality, but they offer insight into future longitudinal research that can interrogate issues of causality over the life course.

Lastly, it should be noted that the sample was relatively well-educated and largely White. Based on US Census estimates⁶³, both samples were reasonably representative of the adult population over the age of 60 in the sampled area. Compared to the entire US population of older adults, however, these findings may have limited generalizability to racial and ethnic minorities in the US, those from lower socioeconomic backgrounds, and other geographic regions. Nevertheless, these findings provide a firm foundation for replication with larger and more representative samples.

Conclusion

In summary, these findings indicate that social connectedness is not a single, uniform construct in its relationship to health. Rather, we found that social bridging is robustly associated with cognitive functioning and cognitive change, whereas social bonding is more closely linked to a range of dimensions of psychological well-being. By demonstrating that bridging-oriented networks are uniquely associated with cognitive resilience in the context of cognitive aging, this study highlights social bridging as a promising modifiable target for delaying the progression of ADRD, while reaffirming the importance of social bonding for emotional health. In future work, understanding how social bridging ties can be cultivated, sustained, and scaled up through intervention will be critical for developing social policies and programmatic initiatives that promote cognitive resilience across the life course.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Conceptualization: LJH, ACK, BLP. Funding acquisition: ACK, BLP, LGA. Project administration: ACK, BLP. Formal analysis: LJH, SP. Visualization: LJH. Supervision: ACK, BLP. Writing – original draft: LJH, MEC, SP. Writing – review & editing: LJH, ACK, MEC, SP, LGA, BLP.

Data availability

Data are not yet publicly available. The research team is currently preparing a curated dataset to be released in accordance with funding regulations on data sharing. This work will be linked to that public file when it becomes available.

Declarations

Competing interests

The authors declare no competing interests.