Head-to-Head Comparison of Social Network Assessments in Stroke Survivors

Abstract

Background and Purpose:

Social networks influence human health and disease through direct biological and indirect psychosocial mechanisms. They have particular importance in neurologic disease because of support, information, and healthy behavior adoption that circulate in networks. Investigations into social networks as determinants of disease risk and health outcomes have historically relied on summary indices of social support, such as the Lubben Social Network Scale–Revised (LSNS-R) or the Stroke Social Network Scale (SSNS). We compared these 2 survey tools to personal network (PERSNET) mapping tool, a novel social network survey that facilitates detailed mapping of social network structure, extraction of quantitative network structural parameters, and characterization of the demographic and health parameters of each network member.

Methods:

In a cohort of inpatient and outpatient stroke survivors, we administered LSNS-R, SSNS, and PERSNET in a randomized order to each patient. We used logistic regression to generate correlation matrices between LSNS-R scores, SSNS scores, and PERSNET’s network structure (eg, size and density) and composition metrics (eg, percent kin in network). We also examined the relationship between LSNS-R-derived risk of social isolation with PERSNET-derived network size.

Results:

We analyzed survey responses for 67 participants and found a significant correlation between LSNS-R, SSNS, and PERSNET-derived indices of network structure. We found no correlation between LSNS-R, SSNS, and PERSNET-derived metrics of network composition. Personal network mapping tool structural and compositional variables were also internally correlated. Social isolation defined by LSNS-R corresponded to a network size of <5.

Conclusions:

Personal network mapping tool is a valid index of social network structure, with a significant correlation to validated indices of perceived social support. Personal network mapping tool also captures a novel range of health behavioral data that have not been well characterized by previous network surveys. Therefore, PERSNET offers a comprehensive social network assessment with visualization capabilities that quantifies the social environment in a valid and unique manner.

Introduction

Social connectivity is a known determinant of health and disease. Social isolation is comparable to smoking, hypertension, and obesity as a risk factor for mortality.¹ Individuals with impoverished social networks have been reported to be 60% more likely to develop dementia compared to those with high levels of social support.² Functional recovery after severe stroke is as much as 65% greater in patients with enriched social networks relative to those with low social support.³ Social isolation has been associated with a higher incidence of traditional cardiovascular risk factors leading to myocardial infarction and stroke as well as increased mortality after cardiovascular events.⁴ The types of relationships that comprise an individual’s network have been shown to influence hospital arrival times for medical emergencies.⁵ Health behaviors have been shown to be socially transmitted, such as predisposition to obesity,⁶ medication adherence,⁷ and exercise.⁸ Understanding patient’s social networks and how they influence disease risk and health outcomes may yield valuable insights for models of human health and reveal modifiable targets for therapeutic intervention.

Social networks are likely to have particular importance in neurologic disease, in which cognitive and physical impairments directly threaten individuals’ ability to engage with family, friends, and colleagues. Social networks also highlight the role of nonprofessional caregivers who are especially important over the course of neurologic illness. Social network characteristics have been shown to contribute to the likelihood of accessing acute stroke care in a timely fashion⁹ as well as outcomes of stroke recovery.^3,10 It has further been shown that stroke survivors are likely to have unmet social support needs in the years following their stroke.¹¹ Understanding the structure and health characteristics of patient’s social network is, therefore, likely to inform predictions of long-term outcomes and may help providers identify vulnerable patients who would benefit from additional social and community supports during the active phase of recovery.

To date, interventions directed at augmenting social support during recovery from illness have not shown positive results.^12-14 Existing social network survey instruments rely on summary indices of social connectivity, which reduce the characteristics of an individual’s network to a single numerical score. These indices do not capture important complexities such as the structure of connections around a particular patient, the composition of health behaviors among the members of a patient’s network that are likely to influence health behaviors or the visual network “sociogram” map of ties. Two social network instruments that have been used to investigate patients with neurologic disease include the Lubben Social Network Scale–Revised (LSNS-R)¹⁵ and the Stroke Social Network Scale (SSNS).¹⁶

The LSNS-R assesses the number and degree of engagement with family and friend contacts within a patient’s network. It is a 12-item scale with subsections dedicated to family ties and friendship ties. Each subsection contains 6 questions that assess network size, frequency of contact, and degree of closeness. It has been used to characterize social networks across a variety of populations, including a prospective cohort study in 13 686 participants that demonstrated a greater risk of incident stroke in individuals with small social network size.¹⁷ The SSNS assesses the number, frequency, and satisfactoriness of a patient’s social contacts among family, friends, and groups. It is a 19-item survey of participants’ social networks over the preceding month that assesses 5 core subdomains: network size, composition (family, close friends, colleagues, neighbors, and groups), frequency of contact, proximity, and subjective satisfaction with the network. The SSNS has been used to assess longitudinal changes in stroke survivors’ social networks in the acute and chronic phases of stroke recovery.^16,18

In the present study, we investigate a newly developed personal network (PERSNET) mapping tool.¹⁹ We developed PERSNET to quantitatively measure the structure and composition of patients’ personal social networks. It is designed to generate a graph theoretic map of a patient’s social network, allowing quantitative structural and compositional variables to be extracted for analysis. Networks are composed of the index patient and their relevant social contacts. Structural metrics captured by PERSNET include network size (the number of network members surrounding an index patient), constraint (the degree of connectivity among network members independent of the index patient), maximum degree (the highest number of ties by a network member), and mean degree (the average number of ties across all network members).²⁰ Constraint is a useful summary metric that encapsulates network size, density, and tie strength, representing the degree to which network members are connected to one another. Smaller and more close-knit networks have greater constraint (Figure 1).²⁰ Perhaps counterintuitively, individuals with more constrained social networks have been shown to be at greater risk of delayed presentation to medical care after an acute stroke, likely owing to restricted information flow and a tendency of patients to negotiate with close network ties for delays in access to care in the setting of a medical emergency.⁹

Figure 1.

Graph theoretic maps depicting hypothetical social networks with high (A) and low (B) constraint.

Compositional metrics captured by PERSNET include demographic and health-related characteristics among the network members, such as sex, race, smoking, alcohol use, and regular exercise (a comprehensive list is provided in Table 1). For example, PERSNET has previously been used among a cohort of patients at risk of multiple sclerosis (MS) to demonstrate a positive association between the percentage of network members deemed to have a negative health influence on the index patient and self-reported MS-related disability,¹⁹ potentially reflecting the influence of health milieu on individual health parameters.

Table 1.

Personal Network Mapping Tool–Derived Network Variables Included in Structural and Compositional Analysis.

Network variable
Structural variables
Network size: the total number of network members within an individual’s social network
Constraint: a measure of the extent to which the members of an individual’s network have ties to one another. Constraint is highest in a close-knit network of individuals who are all highly connected to one another, as in a nuclear family
Maximum degree: the highest number of ties possessed by any network member
Mean degree: the mean number of ties possessed by each network member averaged across the entire network
Compositional variables
Diversity of sex
Diversity of race
Percentage of kin ties
Percentage of ties with infrequent exercise
Percentage of ties with poor diet

In the present study, we administered LSNS-R, SSNS, and PERSNET in random sequence to stroke survivors in order to compare PERSNET-derived structural and compositional metrics against LSNS-R and SSNS summary scores. The rationale is to compare and contrast PERSNET with traditional metrics and to compare the metrics generated within PERSNET as well. Our overall aim is to validate PERSNET against the traditional metrics and to assess whether PERSNET may capture important social network parameters that are not captured by these instruments.

Methods

We stored all data in a secure web-based database hosted by the Research Electronic Data Capture (REDCap).²¹ Requests to access the data set from qualified researchers trained in human subject confidentiality protocols may be sent to Dr Dhand.

Participants

We recruited patients from the Brigham and Women’s Hospital inpatient neurology service and outpatient neurology clinic in accordance with an institutional review board–approved study. Inclusion criteria were diagnosis of ischemic stroke or intraparenchymal hemorrhage at any time point, age > 18, and fluent English speaking. We screened patients with the Short Blessed Test (SBT),²² a brief test of orientation, attention, and language validated for the detection of possible cognitive impairment/dementia. It is scored on a scale of 0 to 28, with points given for response errors. We excluded patients for scores >6. Because of the verbal demands of survey completion, we also screened patients with the NIH Stroke Scale (NIHSS) and excluded patients for scores >1 in the aphasia subsection. Additional exclusion criteria were known diagnosis of dementia or other inability to provide informed consent. We collected responses only from patients themselves, and we did not permit family members’ responses at the beside.

Data Collection

We administered in-person PERSNET, LSNS-R, and SSNS surveys in both inpatient and outpatient settings (all 3 survey instruments are available for review in Supplemental Materials). The SSNS (19 items, 5-minute collection time), LSNS-R (12 items, 5-minute collection time), and PERSNET (∼48 items with adaptation to responses, 20-minute collection time) surveys were administered in a systematic interview style.

The PERSNET questionnaire contains 3 sections: name generator, network inter-relator, and name interpreter. The name generator asks 3 questions paraphrased as follows: “Who do you typically discuss important matters with?” “Who do you often socialize with?” and “Who provides support for your health needs?” The number of names is not capped, allowing network size to be measured without a ceiling, but the subsequent survey questions are limited to the first 10 persons in the network for feasibility of survey administration and statistical analysis. Participants answer questions that assess the strength of connection between each pair of individuals within the network (no tie, weak tie, and strong tie). Finally, in the name interpreter section, participants answer questions about the demographics and health habits of each network member. The instrument is included in Supplement 1.

We randomized the order of the 3 surveys to avoid order effects on participants’ responses. The average duration of each encounter, including SBT and NIHSS screenings, informed consent, and survey administration, was 40 minutes.

Data Analysis

Using the R statistical package, we imported data from REDCap. We generated summary scores for SSNS (scores reported on 0-100 scale) and LSNS-R (scores reported on 0-60 scale) for each patient. From each patient’s PERSNET survey responses, we extracted structural and compositional network indices and generated a network map.

To compare the instruments, we used Spearman correlation to assess the relation of SSNS score, LSNS-R score, and PERSNET metrics. We then used regression analysis to confirm the relation between variables. The PERSNET structural parameters that we studied were total network size, constraint, maximum degree, and mean degree. The PERSNET compositional parameters that we studied were diversity of sex, diversity of race, percentage of contacts who are direct kin, percentage of ties who do not exercise regularly, and percentage of ties who eat a poor diet. Sex was coded as female, male, or other. Race was coded black, white, or other. Direct kinship was coded as a binary yes/no variable. Infrequent exercise and poor diet were coded as yes/no variables and expressed as a percentage of the participant’s network members.

Results

Participants

We enrolled 70 patients who provided informed consent. We did not include 3 patients because 2 failed to complete all 3 surveys and 1 was excluded for memory impairment during the interview. We summarize patient demographics in Table 2.

Table 2.

Demographic Composition of the Patient Cohort.

	Overall
Number of enrolled patients	67
Sex (%)
Female	27 (40.3)
Male	40 (59.7)
Age (mean [SD])	65.36 (13.24)
Race (%)
American Indian	1 (1.5)
Asian	4 (6.0)
Black	8 (11.9)
White	50 (74.6)
Other/unknown	3 (4.5)
Mixed	1 (1.5)
Ethnicity (%)
Hispanic or Latino	8 (12.1)
Not Hispanic or Latino	58 (87.9)
Education (%)
High school grad	14 (20.9)
Some college	16 (23.9)
Associate degree	5 (7.5)
Bachelor’s degree	16 (23.9)
Graduate degree	16 (23.9)
Employment (%)
Employed for wages	22 (32.8)
Self-employed	3 (4.5)
Out of work and looking for work	2 (3.0)
Out of work but not currently looking for work	1 (1.5)
Retired	31 (46.3)
Unable to work	8 (11.9)
Net worth (%)
Less than $5000	4 (6.2)
$5000-$49 000	11 (17.2)
$50 000-$169 000	10 (15.6)
$170 000-$490 000	12 (18.8)
More than $500 000	27 (42.2)
Married (%)	42 (62.7)
Live alone (%)	9 (13.4)
Household number (mean [SD])	2.71 (1.23)
Patient type (%)
Inpatient	43 (64.2)
Outpatient	24 (35.8)
Short Blessed Test (mean [SD])	1.84 (2.00)
NIHSS (mean [SD])	1.66 (2.93)
Stroke type (%)
Ischemic	52 (77.6)
Hemorrhagic	15 (22.4)
Stroke area (%)
Cortical	29 (43.3)
Subcortical	25 (37.3)
Both cortical and subcortical	2 (3.0)
Brainstem	9 (13.4)
Unknown	2 (3.0)
Stroke side (%)
Left	22 (39.3)
Right	30 (53.6)
Both	4 (7.1)

Abbreviation: NIHSS, NIH Stroke Scale.

Network Structural Analysis

Comparison of LSNS-R and SSNS to PERSNET-derived structural variables: Scores on LSNS-R and SSNS were significantly correlated with multiple structural network parameters derived from PERSNET (Figure 2). Specifically, LSNS-R was directly correlated with total network size (Spearman ρ = 0.45, P = .0002), mean degree (Spearman ρ = 0.36, P = .003), and maximum degree (Spearman ρ = 0.36, P = .002) and inversely correlated with PERSNET constraint (Spearman ρ = −0.32, P < .05). Stroke Social Network Scale was directly correlated with a mean degree (Spearman ρ = 0.40, P = .001), and maximum degree (Spearman ρ = 0.30, P = .013), and inversely correlated with PERSNET constraint (Spearman ρ = −0.25, P = .040). Stroke Social Network Scale and LSNS-R scores were significantly correlated with one another (Spearman ρ = 0.61, P = .0001).

Figure 2.

Correlation of personal network (PERSNET) mapping tool structural (panel A) and compositional (panel B) variables with LSNS-R and SSNS total scores. Spearman ρ statistic for each correlation is represented on the Y-axis and individual personal network (PERSNET) mapping tool–derived network parameters are indicated on the X-axis. LSNS-R indicates Lubben Social Network Scale–Revised; SNSS, Stroke Social Network Scale.

Internal comparison of PERSNET-derived structural variables: Total network size was inversely correlated with constraint (Spearman ρ = −0.77, P = .0001) and directly correlated with mean degree (Spearman ρ = 0.58, P = .0001) and max degree (Spearman ρ = 0.79, P = .0001). Constraint was inversely correlated with a mean degree (Spearman’s ρ −0.57, P = .0001) and maximum degree (Spearman’s ρ = −0.87, P = .0001). The mean and maximum degree were directly correlated (Spearman ρ = 0.76, P = .0001).

Network Compositional Analysis

Scores on SNSS and LSNS-R were not significantly correlated with PERSNET-derived compositional variables, including diversity of sex, diversity of race, percentage of ties who are kin, who abuse alcohol, smoke, who do not regularly exercise, who have health problems, who provide support to the ego, who live far distances, who have frequency of contact, who have short length of relationship to the patient, and who have a negative overall health influence on the ego (Figure 3). There were modest correlations of the compositional variables against each other including: percentage of ties who do not exercise regularly and percentage of ties with a bad diet (Spearman ρ = 0.39, P < .01); percentage of ties who do not exercise regularly and diversity of sex (Spearman ρ = −0.36, P < .01). Comparison of the metrics using regression analysis confirmed these results (Supplement Table 1).

Figure 3.

Graph theoretic maps of each participant’s social network derived from personal network (PERSNET) mapping tool responses. Each map is labeled to the bottom right with the participant number. The index patient is represented by a black dot, and each of the remaining network members are represented by an open dot. Red lines represent strong ties and blue lines represent weak ties.

Cutoff for Social Isolation

We determined an appropriate cutoff score at which network size patients may be socially isolated by using the LSNS-R definition of social isolation. We used an LSNS-R cutoff score of 24 based on prior studies.¹⁵ The participants with an LSNS-R score of <24 had a mean social network size of 4 (SD = 2.4). Those who had a score of >24 had a mean social network size of 7.7 (SD = 4.1). From these distributions, we determined that a network size <5 was a cutoff for being at-risk of social isolation. Further study of the specificity and sensitivity of this cutoff is needed.

Discussion

In this comparison of a novel personal social network mapping instrument to 2 traditional social network surveys, we demonstrate a high degree of correlation between the structural network parameters captured by PERSNET and the summary scores generated by SSNS and LSNS-R. This indicates that PERSNET is a valid tool for assessing the degree of social support conferred by an individual’s social network. We further demonstrate a high degree of correlation among the various structural network data derived from PERSNET, suggesting that this instrument has a high degree of internal validity. We found no correlation between the SSNS or LSNS-R scores and PERSNET network compositional parameters, indicating that the traditional social network indices do not capture data regarding the demographics and health habits of individuals in the network.

The greatest association was between the SSNS and LSNS-R scores and PERSNET network constraint in an inverse direction. As previously discussed, more highly constrained networks are smaller and have stronger ties among members, resulting in less overall diversity in social contacts across individual members of the network. It has been proposed that constrained social networks benefit less from the free flow of information and behaviors that predominates in more open networks and that this may underlie a range of adverse health outcomes including delayed presentation to care during medical emergencies.⁹ It is plausible that greater PERSNET-derived network constraint predicts lower scores on SSNS and LSNS-R, both of which account for the number and range of contacts within an individual’s network.

The compositional analysis of numerous demographic parameters and health behaviors revealed no correlation between PERSNET and SSNS or LSNS-R. This is because the latter 2 surveys focus primarily on the patient’s perceived social supports rather than on the qualities of the individual members that comprise the patient’s network. Therefore, PERSNET captures important socially derived determinants of health that are not captured by standard survey instruments. Because individual health behaviors are known to be influenced by prevailing trends within an individual’s social network,²³ PERSNET may prove useful not only identifying patients with high-risk network structures but also those with high-risk network composition characterized by maladaptive health habits. The identification of harmful health behaviors in an individual’s social network could be used to facilitate behavioral modifications to mitigate the effect of those behaviors on individual patients.

One of the values of PERSNET over the other 2 indices studied is the ability to more precisely characterize and visualize network structure, moving beyond global summary statistics to map the nodes and ties within an individual’s social network, and from that map to extract parameters that are amenable to individual and group-level analysis. Such precision in the clinical setting may help identify individuals with high-risk social networks (ie, highly constrained networks with few nodes) that could benefit from augmented social supports. As further investigations proceed into the role of social networks in health and disease, analytical tools that provide such quantifiable metrics and statistical rigor are likely to be of enormous value. While PERSNET has to date primarily been used as a research tool, we are hopeful that, moving forward, it will generate clinically useful information for patients recovering from stroke and a variety of other neurologic disorders that entail prolonged recoveries and debilitating deficits. While more work is needed to pilot and implement specific interventions, we believe that the identification of patients at high risk of social isolation will serve as an opportunity to buttress social supports and thereby confer a therapeutic benefit to the recovering patient.

Our study was primarily designed to compare the performance of the 3 surveys tested at a single time point in a specified population of stroke survivors. Because we lacked longitudinal follow-up, we are unable to compare how each instrument may capture relevant changes in social networks over time or to establish the validity of these instruments in predicting long-term clinical or network-related outcomes. Because we elicited survey responses from patients themselves, our sample was limited to those patients who were sufficiently neurologically intact to engage the roughly 40 minutes of questions required to complete the screening process and social network surveys. This is reflected in the low mean NIHSS of 1.6 in our cohort. It is unknown whether close family surrogates may serve as a valid proxy for patients who are unable to engage in survey completion. In the present study, we included only individuals able to complete self-report surveys in order to generate the most authentic assessment of each participant’s perceived social network. A subsequent study, however, assessing the correlation between patient responses to PERSNET and family members answering on their behalf would be informative, potentially expanding the range of eligible patients to those with a greater degree of neurologic impairment and helping to determine whether stroke severity may be associated with an individual’s social network parameters.

In conclusion, PERSNET is an easily administered survey tool with external validity as an index of perceived social supports and internal validity as an assay of social network structural parameters. The capacity for precise social network mapping as well as the demographic and behavioral data that PERSNET captures on each member of a patient’s social network represent novel additions to the armamentarium of existing social network survey tools. Data derived from PERSNET may help patients, families, physicians, and social workers optimize social supports for individuals recovering from illnesses such as stroke and are likely to allow for a greater degree of quantitative rigor in future investigations of social networks in human health and disease.