Difference-Making Pathways to Frailty Through Social Factors: A Configurational Analysis

Abstract

Background and Objectives

Social disconnection is highly prevalent in older adults and is associated with frailty. It is unclear which aspects of social disconnection are most associated with frailty, which ones are difference-making, and which combination of social factors are directly linked to frailty.

Research Design and Methods

We conducted a secondary coincidence analysis (CNA) of 1,071 older adults from the Rush Memory and Aging Project (mean age 79.3 ± 7.1; 75.8% female) to identify combinations of social factors that are difference-making for frailty. We included 7 demographic (e.g., age, sex, socioeconomic status) and structural (e.g., social network), functional (e.g., social support, social activity), and quality (e.g., loneliness) aspects of social connection. An established cut score of 0.2 on a frailty index was used to define frailty as the outcome.

Results

CNA produced 46 solution models for the presence of frailty in the data set. The top-scoring model was underfit, leaving a final complex solution path for frailty with the highest fit-robustness score that met the fit parameter cutoffs. We found that the combination of loneliness, low social activity, and older age was present 82% of the time when frailty was present.

Discussion and Implications

The combination of loneliness, social activity, and old age is difference-making for frailty, and supports the inclusion of social factors in frailty prevention and intervention. Further research is needed in diverse data sets to better understand the interrelationships between the 3 aspects of social connection and frailty.

Social connection is an umbrella term for three broad aspects of social relationships that can affect health outcomes: (1) structural (e.g., social networks or living arrangements), (2) functional (e.g., perceived social support, social engagement), and (3) quality (e.g., loneliness; Holt-Lunstad, 2021). In older adults, social disconnection in these three aspects is associated with cognitive and functional decline, dementia, frailty, and death (Holt-Lunstad et al., 2015; Kojima et al., 2022; Perissinotto et al., 2012; Shankar et al., 2011; Wilson et al., 2007). The prevalence of loneliness and social isolation is rising and inadequate social connection is increasingly recognized as a public health concern with poor health outcomes (Our epidemic of loneliness and isolation: The U.S. Surgeon General’s advisory on the healing effects of social connection and community, 2023). The pooled prevalence of loneliness and social isolation among adults aged 65 and older from 15 countries across four continents was 28.6%, and loneliness has increased since the onset of the coronavirus disease 2019 (COVID-19) pandemic (Su et al., 2023). Based on a sample of over 20,000 U.S. middle-aged and older adults, pre-COVID pandemic loneliness prevalence was estimated to be 43.2% (DesHarnais Bruce et al., 2019), representing a large population at risk.

Frailty is a multidimensional construct characterized by decreased physiologic reserves that—if exposed to relatively minor stressors—render an individual vulnerable to disproportionate changes in health status (Rockwood & Mitnitski, 2007). Frailty is associated with falls, institutionalization, disability, and mortality, and is an important factor in loss of independence for older adults (Fried et al., 2001; Rockwood & Mitnitski, 2007).

Theoretical Framework

Prior research suggests that there are causal links between social connection and frailty based on longitudinal studies on social connection and different health outcomes, including frailty and mortality (Holt-Lunstad, 2021; Kojima et al., 2022) via multiple plausible biologic mechanisms including inflammatory response, immune and endocrine dysfunction (Cacioppo et al., 2014; Holt-Lunstad, 2021). The stress-buffering hypothesis is also frequently used to explain how social connection might buffer negative health outcomes. This model was borne out of the observation that individuals with greater levels of material and psychological support have better health compared to individuals with less support (Cohen & Wills, 1985). The stress-buffering hypothesis is useful for understanding how the three aspects of social connection (e.g., social support, relationship quality) can attenuate stress appraisal and response and protect against stress-related poor health outcomes (Cohen & Wills, 1985).

Still, what aspects of social connection are difference-making and which combination of factors are linked to frailty is less clear. Additionally, the different aspects of social connection are usually studied in isolation. While there is some degree of overlap, structural, functional, and quality aspects of social connection are not highly correlated and confer distinct protection or risk related to health outcomes (Holt-Lunstad, 2021) and must be considered together to understand clinical implications of social disconnection as a whole. The answers to these questions are important for identifying and developing intervention targets for frailty. We aim to address this knowledge gap using a configurational comparative method (CCM) as it uniquely analyzes combinations of sufficient and necessary conditions that lead to an outcome (Baumgartner & Thiem, 2015).

In the current study, we used coincidence analysis (CNA) to identify combinations of social factors that lead to frailty in community-dwelling older adults. One of the great advantages of CNA as an analytic tool is that it is an inherently theory-driven approach through selection of factors to include in the analysis, how to operationalize them, and how to interpret results in terms of what makes biological sense, is based on prior research, and fit-robustness scoring of the models. We considered structural, functional, and quality aspects of social connection that were available in our data set as well as demographic factors that were associated with social connection and frailty in prior research (Fried et al., 2001; Mitnitski et al., 2001; Pollak et al., 2023). Frailty increases with age and is more common in women (Fried et al., 2001). Additionally, prior research suggested differential effects of age and sex on social determinants of health (Nguyen et al., 2020; Pillemer et al., 2019). Low socioeconomic status (SES) also increased the risk for loneliness and frailty (Pollak et al., 2023; Wang & Hulme, 2021) and education was also associated with frailty (Hoogendijk et al., 2014). We incorporated social networks as a measure of structural social connection, social activity and social support as functional measures of social connection, and loneliness as a quality social connection measure (Holt-Lunstad, 2021). Because frailty is defined by reduced physiologic reserve across multiple systems, we expected that frailty would be explained by complex configurations of factors rather than single-factor configurations, a complexity that is supported by CNA (Baumgartner & Thiem, 2015; Rockwood & Mitnitski, 2007). Based on theories that explain mechanism for social factors and health outcomes, the purpose of this analysis was to identify difference-making social factors for frailty across the three dimensions of social relationships (e.g., structural, functional, and quality), and to identify which combination of conditions are difference-making for frailty to identify intervention targets for frailty prevention and treatment in older adults.

Method

Participants

This analysis examined data from 1,071 participants enrolled in the Rush Memory and Aging Project (MAP). The MAP study is a longitudinal, clinical–pathologic study of chronic conditions of aging. Study procedures are described in detail elsewhere (Bennett et al., 2005, 2018). Briefly, participants were recruited from retirement and subsidized housing facilities, church groups, and social services agencies around northeastern Illinois (Bennett et al., 2005). Inclusion criteria were annual clinical evaluation and anatomical gift donation at death (Bennett et al., 2005). There were 2,252 participants in the original data set. For this study, 120 participants with dementia were excluded as these individuals likely have different pathways to frailty and different social characteristics than the general older adult population. Additionally, 1,061 participants with missing data were excluded from the analysis as CNA modeling requires no missing data, which left 1,071 participants included in these analyses. These analyses were restricted to baseline data only to meet current computational limits of CNA. The MAP study was approved by the Rush Medical Center Institutional Review Board. Ethical approval for these secondary analyses was obtained from the Albert Einstein College of Medicine Institutional Review Board.

Study Procedure

A summary of the study procedure and analysis approach is presented in Figure 1. The analysis was conducted in three stages: (1) data transformation, (2) data reduction, and (3) data analysis. The data transformation step is essential to calibrate the variables into factors CNA can process. We calibrated all variables to either so-called crisp-set or fuzzy-set using the QCApro, QCA, and SetMethods packages in R (Baumgartner & Ambuhl, 2020). Crisp-set factors are dichotomous and coded 0/1. Fuzzy-set factors are continuous values ranging from 0 to 1. We calibrated all variables other than sex as fuzzy sets to capture all the information in the continuous or quasi-continuous variables. For all factors, 0 represented full nonmembership in the set and 1 represented full membership in the set. Factors were calibrated with anchors (or cutoffs) based on standardized metrics whenever possible to assign cases as belonging in or out of the set (e.g., frail or not frail). Some factors were calibrated based on membership ratios, which represent the distribution of cases in the data, in this relatively healthy, well-supported sample. Our data transformation procedure and calibration details are thoroughly delineated in Supplementary Section A with in-depth descriptions of included factors.

Figure 1.

Summary of study procedure and analysis approach.

The primary outcome was frailty, based on frailty index scores. The frailty index operationalizes frailty using the deficit accumulation approach that defines frailty as a multidimensional risk state characterized by the quantity of accumulated deficits that reduces the ability of the system to repair damage (Rockwood & Mitnitski, 2007). A detailed description of how the frailty index was constructed is included in Supplementary Section A and variables included in the frailty index are listed in Table 1. Perceived social support was assessed with four items from the Multidimensional Scale of Perceived Social Support (Zimet et al., 1990). Social network size was quantified as the number of community members, family, and friends seen at least once a month. Loneliness was assessed using a modified 5-item De Jong Gierveld Loneliness Scale (De Jong Gierveld & Kamphuis, 1985). Subjective SES status was assessed by the MacArthur Scale of Subjective Social Status (Adler et al., 2000). Late-life social activity was assessed using a 6-item scale inquiring how often participants engaged in common types of social activities during the past year (e.g., going to restaurants, day trips, volunteer work, visit friend’s homes, participate in groups, or attend religious services). Age in years was computed from self-reported date of birth and date of clinical evaluation and sex was assessed by self-report. See Supplementary Section A for detailed descriptions of assessments for each factor as well as calibration details.

Table 1.

Health Variables Included in the Frailty Index

Variable	Original coding	Frailty index coding
1. Leg stand	Continuous	Quartile 1 = 1 Quartile 2 = 0.66 Quartile 3 = 0.33 Quartile 4 = 0
2. Pinch strength	Continuous	Quartile 1 = 1 Quartile 2 = 0.66 Quartile 3 = 0.33 Quartile 4 = 0
3. Finger tap	Continuous	Quartile 1 = 1 Quartile 2 = 0.66 Quartile 3 = 0.33 Quartile 4 = 0
4. Purdue pegboard	Continuous	Quartile 1 = 1 Quartile 2 = 0.66 Quartile 3 = 0.33 Quartile 4 = 0
5. Grip strength	Continuous	Females: 0–15 = 1; 16–20 = 0.5; >20 = 0 Males: 0–25 = 1; 26–32 = 0.5; >32 = 0
6. Body mass index	Continuous	1 if BMI < 18.5; OR ≥ 30
7. Memory problems	Dichotomous	0 = no; 1 = yes
8. Hypertension	Dichotomous	0 = no; 1 = yes
9. Heart conditions	Dichotomous	0 = no; 1 = yes
10. Cancer	Dichotomous	0 = no; 1 = yes
11. Congestive heart failure	Dichotomous	0 = no; 1 = yes
12. Stroke	Dichotomous	0 = no; 1 = yes
13. Parkinson’s disease	1 = highly probable; 2 = probable; 3 = possible; 4 = not present	0 = no (possible, not present); 1 = yes (highly probable and probable)
14. Diabetes	Dichotomous	0 = no; 1 = yes
15. Help dressing	0 = no help; 1 = help or unable	0 = no help; 1 = help or unable
16. Help walking across small room	0 = no help; 1 = help or unable	0 = no help; 1 = help or unable
17. Help toileting	0 = no help; 1 = help or unable	0 = no help; 1 = help or unable
18. Help eating	0 = no help; 1 = help or unable	0 = no help; 1 = help or unable
19. Help bathing	0 = no help; 1 = help or unable	0 = no help; 1 = help or unable
20. Help transferring	0 = no help; 1 = help or unable	0 = no help; 1 = help or unable
21. Difficulty rising from chair	0 = normal; 1 = slow; 2 = >1 attempt, no arms; 3 = uses arms; 4 = uses arms, >1 attempt; 5 = unable to arise without help	0 = no help; 0.5 = slow, no arms; 1 = uses arms or unable
22. Help with finances	0 = no help; 1 = help or unable	0 = no help; 1 = help or unable
23. Using the telephone	0 = no help; 1 = help or unable	0 = no help; 1 = help or unable
24. Help with meal preparation	0 = no help; 1 = help or unable	0 = no help; 1 = help or unable
25. Help taking medications	0 = no help; 1 = help or unable	0 = no help; 1 = help or unable
26. Heavy housekeeping	0 = no help; 1 = help or unable	0 = no help; 1 = help or unable
27. Light housekeeping	0 = no help; 1 = help or unable	0 = no help; 1 = help or unable
28. Help shopping	0 = no help; 1 = help or unable	0 = no help; 1 = help or unable
29. Traveling within community	0 = no help; 1 = help or unable	0 = no help; 1 = help or unable
30. Help going up/down stairs	0 = no help; 1 = help or unable	0 = no help; 1 = help or unable
31. Walking half a mile without help	0 = no help; 1 = help or unable	0 = no help; 1 = help or unable
32. Taking care of home	0 = no help; 1 = help or unable	0 = no help; 1 = help or unable
33. Sleep item	Dichotomous	1 = yes
34. Could not get going	Dichotomous	1 = yes
35. Felt depressed	Dichotomous	1 = yes
36. Felt happy	Dichotomous	1 = no
37. Felt everything was an effort	Dichotomous	1 = yes
38. Falls yes = 1 no = 0	0 = 0 falls; 1 = one or more falls	1 = yes
39. Cognition: Mini-Mental State Examination	Continuous	<10 = 1; 10–17 = 0.75; 18–20 = 0.5; 21–23 = 0.25; ≥24 = 0
40. Physical activity	Continuous	Quartile 1 = 1 Quartile 2 = 0.66 Quartile 3 = 0.33 Quartile 4 = 0

Notes: BMI = body mass index; OR = odds ratio.

Configurational Analysis

CCMs are a family of configurational analytic methods, including qualitative comparative analysis and CNA (Baumgartner & Ambuhl, 2020). CCMs have been applied in political and social science research since the 1980s. More recently, CCMs have been used in implementation science and healthcare-related research (Hickman et al., 2020; Petrik et al., 2020). CCMs, and CNA applied in this analysis, are designed to investigate different research questions, and uncover different properties of causal structures compared to traditional regression techniques and thus complement traditional methods (Baumgartner & Ambuhl, 2020; Thiem et al., 2016). CNA is both a mathematical and theoretical approach to identify combinations of conditions that are sufficient, necessary, and nonredundant for an outcome of interest. The configurational approach contrasts with traditional regression-based methods in that it (1) employs Boolean algebra versus linear algebra, (2) finds solutions at the case level rather than the variable level, (3) searches for necessary and sufficient conditions for an outcome rather than marginal/incremental effects between variables in a sample, (4) operates under the regularity framework compared to the probability framework in regression, and (5) CNA can model causal complexity (conjunctions of conditions) and equifinality (multiple paths to the same outcome; Baumgartner & Ambuhl, 2020; Baumgartner & Falk, 2018; Baumgartner & Thiem, 2015).

We chose CNA for this analysis because we were looking to identify combinations of conditions that lead to frailty, a question that CNA is uniquely suited to determine (Baumgartner & Ambuhl, 2020). One of the main advantages of CNA is the identification of complex configurations of conditions that, when jointly present, directly link to an outcome (Baumgartner & Falk, 2018). CNA can also recover multiple alternative paths to the same outcome, where if one path is blocked, the outcome might still occur through an alternative path (Baumgartner & Falk, 2018). Another important advantage of CNA as an analytic tool is that it is both a mathematical and theoretical approach in terms of decisions regarding what factors to include in the analysis, how to operationalize them, and how to interpret results. Cohort data are inherently noisy due to unmeasured factors, rendering it unlikely that strictly necessary or sufficient conditions exist in the data (Baumgartner & Ambuhl, 2020). Consistency and coverage measures are fit parameters used to set acceptable thresholds for necessary and sufficient conditions to still extract causal information from the data (Baumgartner & Ambuhl, 2020). The CNA algorithm uses a bottoms-up approach to find configurations that meet the consistency/coverage thresholds, starting with a single-factor value, then proceeds with two-, then three-factor conjunctions, and so on (Baumgartner & Ambuhl, 2020). By considering single factors first, the algorithm ensures no redundant conditions are included in model outputs. While it is accepted practice to use a consistency/coverage cutoff of 0.8 for model fit (Hickman et al., 2020; Yakovchenko et al., 2020), we used the minimum accepted threshold for consistency/coverage of 0.75 given that our data, while large and diverse, derive from real-world, noisy cohort data.

Data Analysis

Stata (StataCorp LLC, College Station, TX) version 17.0 was used to assess participant demographics and relevant social and health characteristics. CCM analyses were conducted using the R packages “cna” and R Studio software applications (version 4.3.0; Baumgartner & Ambuhl, 2023). We applied cna models to examine potential causal paths between seven social factors and frailty. Participant demographics known to be associated with frailty and all social factors available in the data set that were linked to frailty based on prior studies were considered for inclusion in the analysis. The minimally sufficient condition (msc) function within the R package “cna,” also described in detail elsewhere (Yakovchenko et al., 2020), was applied to identify configurations of conditions with the strongest connection to frailty for purposes of data reduction. For a detailed description of data reduction procedures, see Supplementary Section B.

The eight factors identified by the msc routine were included in the final model. Frailty was modeled as the outcome. We used the frscore function to uncover solution paths with the best model fit (Parkkinen & Baumgartner, 2021). We chose the best models out of the 46 by balancing the following criteria: (1) the compliance of the model solutions with theoretical knowledge, (2) consistency and coverage values, and (3) the fit-robustness of the model. We describe our model selection process and detailed justification of the selection of the final two models in Supplementary Section C, with an accompanying table describing the 46 models.

Results

Baseline Characteristics

Baseline characteristics of study participants are summarized in Table 2. The mean age was 79.3 ± 7.1 years. The sample was mostly female (75.8%), mostly Caucasian (94.8%), highly educated, and relatively healthy, without disability, cognitive impairment, or depressive symptoms. Approximately 40% of participants were married. Participants were generally well-supported with a median of six contacts in their social networks (interquartile range 3, 10). The mean loneliness score was 2.2 out of a range of 5.

Table 2.

Characteristics of Participants (n = 1,071)

Age (years), mean, SD	79.3 ± 7.1
Female, % (n)	75.8 (812)
Race/ethnicity, % (n)
Caucasian	94.8 (1,015)
African American	4.1 (44)
Other	1.2 (12)
Education (years), mean, SD	15.5 ± 3.1
Social factors
Social network size, median (IQR)	6 (3,10)
Social support,a median (IQR), range 1–4	4.5 (4, 5)
Loneliness,b mean, SD, range 1–5	2.2 ± 0.6
Marital status, % (n)	39.7 (767)
Cognitive function
Global cognitive function summary, mean, SD	0.23 ± 0.5
Physical function, median (IQR)
ADL	0 (0, 0)
IADL	1 (0, 2)
Physical activity (hours/week)	2.9 (1.1, 5)
Frailty index, mean, SD	0.16 ± 0.1
Psychological factors, median (IQR)
Depressive symptoms (CES-D),c range 1–10	0 (0, 1)
Comorbidities, mean, SD
Self-reported conditionsd	1.4 ± 1.1

Notes: ADL = activities of daily living; IADL = instrumental activities of daily living; IQR = interquartile range; SD = standard deviation.

^aAssessed by the Multidimensional Scale of Perceived Social Support.

^bAssessed by the De Jong Gierveld Loneliness Scale.

^cAssessed by the Center for Epidemiological Studies—Depression scale.

^dIncluded conditions: hypertension, diabetes, heart disease, cancer, thyroid disease, head injury, and stroke.

Solution Pathway to Frailty

The analysis revealed Model 4 as the most plausible solution path for the outcome frailty because this model met all four criteria for model selection in that (1) was a supermodel of the top-most scoring model (e.g., included loneliness as a difference-maker for frailty), (2) fit at least as well as the top-scoring model, (3) theoretically consistent, and (4) met the consistency/coverage threshold of 0.75, which is standard practice for CNA models based on other publications (Hickman et al., 2020; Yakovchenko et al., 2020). Model 4 showed loneliness, low social activity, and older age lead to frailty. When frailty was present, this combination of factors was present 75% of the time. Additionally, 82% of frailty in this sample was explained by this combination of conditions.

To enhance robustness of our analysis, we considered Model 7 as a plausible alternative path for the outcome frailty as a model that met criteria 1–3 but not 4, because the coverage for Model 7 did not meet the 0.75 cutoff. Model 7 showed low perceived SES, loneliness, and older age lead to frailty. When frailty was present, this combination of factors was present 81% of the time and 72% of frailty in this sample was explained by this combination of conditions. The two models agree with each other based on loneliness and age but differ by one factor; one includes low perceived SES in the solution pathway, and one includes low social activity in the solution pathway. While Model 7 represents a plausible, alternative path for substantive reasons, we chose Model 4 as the better model for frailty because it was the higher scoring, better fitting model. We discuss further sensitivity analyses in Supplementary Section D.

Discussion

The main finding of this study was the identification of a specific combination of functional and quality aspects of social relationships that are difference-makers for frailty. The final solution path to frailty included the following conjunction: loneliness (quality), low social activity (functional), and older age, which was present 82% of the time when frailty was present in the sample. This is the first known application of CNA to explore difference-making pathways between social factors and frailty. CNA is an ideal analysis tool to investigate the complex geriatric syndrome of frailty that likely has multiple potential solution paths that can be variably influenced by different combinations of factors.

The difference-making solution path for frailty included the combination of loneliness, low social activity, and older age. An association between frailty and age is well-established (Fried et al., 2001; Mitnitski et al., 2001). Studies using various frailty measures found frailty increased with age and frailty can be a proxy measure for aging (Fried et al., 2001; Mitnitski et al., 2001). We show that age is a difference-maker for frailty, and therefore emphasizes where to target research and healthcare resources for the planning and investment of frailty prevention and intervention.

Loneliness alone was identified as a difference-maker for frailty (Supplementary Section C, Table S1), but we concluded that this was an underfit model and it was theoretically more likely that several factors were at play to bring about the outcome. Loneliness, in combination with low social activity and older age, was the most plausible solution path for frailty. Loneliness was associated with frailty cross-sectionally and longitudinally in prior studies; however, mechanisms for the association were unclear (Kojima et al., 2022). Our results suggest that loneliness is a difference-making factor for frailty. We note this with an important caveat. We were unable to include depression as a factor in this analysis due to skewness of the sample. Loneliness and depressive symptoms often co-occur, but are considered distinct constructs and are assessed using different measurement tools (Cacioppo et al., 2006). Furthermore, the influences of loneliness and depressive symptoms were reciprocal in older adults, and both were associated with frailty in prior studies, but the directionality of the relationship was unclear (Cacioppo et al., 2006; Kojima et al., 2022; Lambert Van As et al., 2022). Similarly to its effects on frailty, loneliness increased the risk of worsening depression and hindered recovery from depression (Lambert Van As et al., 2022). Depression was additionally suggested as a potential mechanism through which loneliness influences morbidity and mortality (Cacioppo et al., 2006). Thus, we cannot rule out that the addition of depression into our model would change the difference-making output and we cannot draw any conclusions regarding depression and whether it is important for frailty based on these results. Because the sample had very low depressive symptoms, which is what precluded inclusion of this factor in our analysis, this factor was held constant in the data and is therefore not a source of confounding here. Additionally, loneliness was established as a predictor for frailty in several longitudinal studies, which provides support for loneliness as an important part of the causal chain (Kojima et al., 2022). Loneliness was also associated with a range of poor health outcomes that might explain some of the mechanisms behind the association between loneliness and frailty including inflammation, poor self-regulation of health behaviors such as smoking and physical activity, increased stress, and decreased ability to cope, to name a few (Pollak et al., 2023; Shankar et al., 2011; Uchino et al., 1996). Loneliness is also associated with sympathetic nervous system activation which perpetuates the inflammatory response and immunologic dysfunction, both of which are linked to a host of poor health outcomes (Cacioppo et al., 2014; Cole et al., 2015). What is made clear from our results is that loneliness is a potentially important modifiable risk factor for frailty that may not be difference-making alone but becomes operative when combined with low social activity and older age and should be considered when planning for frailty interventions. The role of depressive symptoms and depression remains to be investigated in other, more diverse samples.

Social activity was also identified as an important factor for frailty in combination with loneliness and older age. This finding supports our hypothesis that a combination of social factors comprises the causal path to frailty rather than any one factor alone. This finding also supports prior research emphasizing social connection as a multifactorial risk factor for poor health outcomes as well as the U.S. Surgeon General’s call for inclusion of all three aspects of social relationships in epidemiological and intervention research of social connection (Holt-Lunstad, 2021; Our epidemic of loneliness and isolation: The U.S. Surgeon General’s advisory on the healing effects of social connection and community, 2023). Social activity represents social engagement, or activities conducted in the company of others. Given our findings that social activity is difference-making for frailty, more research is needed on social interventions as part of frailty prevention and amelioration. While no gold standard exists, there is a growing body of research on interventions to address social disconnection. Which intervention is most effective and when and how much to best intervene to reduce health risks, however, are still unclear. A systematic review on loneliness interventions for older adults included interventions such as social support interventions (e.g., counseling, education), physical activity interventions (e.g., recreational or fitness activities), social technologies (e.g., companion robots, telephone befriending, internet use), or other social activities (e.g., singing or horticultural therapy; Poscia et al., 2018). Results suggested multimodal interventions (e.g., educational, cognitive, and social support) as the most effective for loneliness, which is supported by our results (Poscia et al., 2018). We showed that it is not loneliness alone that affects frailty but loneliness in combination with social activity and older age. Thus, interventions for frailty prevention must address not just loneliness, but also other structural and functional aspects of social connection. A systematic review of qualitative studies revealed that autonomy in designing the intervention as well as level of participation in the intervention, new social connections, and a sense of belonging yielded reduction in loneliness (Noone & Yang, 2021). The results of this review also supported multidimensional interventions for frailty that address structural, functional, and quality aspects of social relationships. Further research is needed on social interventions that include frailty as the outcome of interest that also target older adults and address both loneliness and social activity to prevent frailty.

Value of Solutions Despite Ambiguity

Not all cases of frailty in the sample were explained by this analysis but 82% of frailty cases were explained by our final model. Additionally, 75% of the time when loneliness, low social activity, and older age were present in combination, frailty was present. We do not argue that this solution path explains all possible paths to frailty, as we will discuss in the limitations section. However, in an area of study where little was previously known, this analysis significantly narrows the realm of possibilities and sheds further light on difference-making social factors for frailty and, therefore, intervention targets and areas for future research. Importantly, the complex solution formula revealed by this analysis includes three specific conditions in combination on one solution path. Based on the theoretical underpinnings of CNA, all three conditions are difference-makers for frailty in combination and if one of the conditions was missing, frailty would not occur along that causal path (Baumgartner & Thiem, 2015). These findings support a multipronged approach to frailty prevention and intervention.

Strengths and Limitations

Cohort data are inherently noisy and can create error signals during the analysis where causality is attributed to a factor where none exists. CNA, however, is built to minimize false positives and reach nonredundant configurations based on a bottoms-up approach to identify only difference-making factors (Baumgartner & Thiem, 2015). Our large, well-characterized sample is an asset in terms of identifying all possible difference-making solutions with the given set of factors. Our theory and data-driven approach to factor selection and reduction via use of the msc routine lends further confidence in our findings. We also utilized multiple goodness-of-fit tests to ensure robustness of our model (Parkkinen & Baumgartner, 2021). Our results are limited by potential unmeasured confounders. This includes known relevant factors such as race, depression, and comorbidities. Due to skewness of the sample, we were not able to assess race as a factor. Depression and comorbidities were not included as factors as they are components of the frailty index score and would therefore likely show up in the causal path to frailty due to diagnostic circularity and collinearity. Furthermore, we cannot make any claims regarding causal irrelevance of factors that were not identified as causal in this data set; merely, these factors were not causal in this sample and merit further study. A related limitation is that our sample was mostly Caucasian, mostly female, well-educated, relatively healthy, and cognitively and functionally intact, which limits the generalizability of our findings. Our cohort is also comprised of individuals with a mean age 79.3 (SD 7.1) and we used a 75-year cutoff for membership in the set for age due to skewness of our sample. It is possible there are differential effects with different age cutoffs that bear future study. We excluded approximately 50% of cases due to missing data, which introduces a concern for bias. However, we believe our data are reasonably diverse that supports the robustness of our findings. We additionally present sensitivity analyses (Supplementary Section D), where we test the robustness of our results. We additionally cannot make causal inferences or rule out reverse causality given the cross-sectional and observational nature of our data. It is theoretically plausible that frailty may cause loneliness due to physical limitation, need for tangible support, or other factors and is an area for future research. In these analyses, we highlight difference-making factors for frailty that may represent candidate causes to be further investigated using other methods. Finally, while prior literature points to gender differences in frailty (Fried et al., 2001), we were unable to investigate this further with CNA for this analysis due to the nature of the case-level algorithm under which CNA operates and this remains to be further investigated using more traditional methods.

Conclusion

The configurations of loneliness, social activity, and age difference-making factors for frailty. Frailty interventions should include a combination of structural, functional, and quality aspects of social connection. Further research is needed in diverse data sets to better understand the relationships of structural, functional, and quality social factors and frailty.

Funding

This work was supported by NIH/National Center for Advancing Translational Science Einstein-Montefiore CTSA (KL2 TR002558); National Institute on Aging (R01AG062659-01A1, R01AG044007-01A1, R01AG036921, R01AG017917).

Conflict of Interest

None.

Data Availability

This study was not preregistered and materials and data are not available as these are secondary analyses and we have not completed original work with the data. Data can be requested through https://www.radc.rush.edu/requests.htm. Reproducibility code is included in Supplementary Section E.

Author Contributions

Chava Pollak (Conceptualization [equal], Data curation [lead], Formal analysis [lead], Funding acquisition [equal], Investigation [lead], Methodology [lead], Validation [equal], Writing—original draft [lead], Writing—review & editing [equal]), Joe Verghese (Conceptualization [equal], Funding acquisition [equal], Methodology [equal], Supervision [equal], Validation [equal], Writing—original draft [equal], Writing—review & editing [equal]), and Helena Blumen (Conceptualization [equal], Funding acquisition [equal], Supervision [equal], Validation [equal], Writing—original draft [supporting], Writing—review & editing [equal]).