Most individuals have few close friends, leading to potential isolation after a friend’s death. Do social networks heal to fill the space left by the loss? We conduct such a study of self-healing and resilience in social networks. We compared de-identified, aggregate counts of monthly interactions in approximately 15,000 Facebook networks in which someone had died with similar friendship networks of living Facebook users. As expected, a substantial amount of social interaction was lost with the death of a friend. However, friends of the decedent immediately increased interactions with each other and maintained these added interactions for years after the loss. Through this, the social networks recovered approximately the same number of active connections that had been lost. Interactions between close friends of the decedent peaked immediately after the death and then reached stable levels after a year. Interactions between close friends of the decedent and acquaintances of the decedent stabilized sooner, within a few months. Networks of young adults, ages 18 to 24, were more likely to recover than all other age groups, but unexpected deaths resulted in larger increases in social interactions that did not differ by friends’ ages. Suicides were associated with reduced social-network recovery.
Most individuals have few close friends1,
A friend group could compensate for a loss by strengthening or building new friendships within-network, potentially even returning to similar levels of connectivity and function after a death. Studies on crises document increases in social interactions7,
Recent research on social networks during crises has found that friend groups react in three distinct ways: (1) quickly forming temporary bonds that dissolve, (2) slowly forming longer-term, in-group bonds, or (3) dissolving and never healing. These responses to crises predict a broad range of average responses to a death, covering the spectrum from full connective recovery to collapse. For example, strong connections develop for information sharing and support in the midst of a crisis, but the connections are not long-lasting8. College students have responded to natural disasters by gradually increasing local network clustering — that is, forming relationships with friends-of-friends instead of outsiders10. Some networks never recover from the loss of a central individual: academic collaboration networks are more likely to dissolve after the death of an important member19.
Here, to evaluate the resilience of social networks after a death, we measure how quickly friend networks recover connectivity, how completely they do so, and whether this connective recovery persists beyond the acute grieving period. We used data from the online social network Facebook to measure online social-network connectivity in the months and years before and after a loss. Past research has shown that online social interactions closely correspond to offline interactions20,
To measure social-network responses to loss, we studied the adaptations of 15,129 de-identified social networks on Facebook between 2011 and 2015 in which the central individual died between January 2012 and December 2013, compared with 30,258 similar networks that did not experience a death. We focused on the close friends of the decedent, examining changes in the numbers of interactions by these close friends over a 4-year window, analysing how their communication patterns changed (1) with other close friends of the decedent, (2) with the decedent’s acquaintances (encompassing the close friends’ existing and new friends) and (3) with individuals who did not know the friend who died (again, existing and new friends of the decedent’s close friends). Figure 1 illustrates the effects. We limit our focus to social connections, but we hope that these findings and future research lead to explanations of changes in functional outcomes.
Our focus in all models is on interactions between the decedent’s close friends and others in the decedent’s social network. Interactions are measured as the number of friendships that were ‘activated’ in a social network each month, based on post, comment or photo-tag ties in the network (see Methods). The measure is the total number of interaction partners in a month in each network, weighted by the number of interaction modes (post, comment, photo). We focus on ‘activated’ friendships rather than total volume of communication to avoid biasing findings toward pairs that communicate extremely heavily. We do not distinguish between newly formed and strengthened friendships because we do not know when two people first met. The interactions that we measure online can include conversations between two people who have not talked to each other in many years or former acquaintances who have suddenly become close friends.
We used quasi-Poisson generalized estimating equations to measure changes in the number of interactions between the decedent’s close friends and other members of the decedent’s local social network. These count-models estimated changes in social interactions relative to a pre-death baseline and relative to the control networks, weighted by the control networks’ likelihood of experiencing a death (see Methods). In this control-group set-up, the models estimate the numbers of interactions in the bereaved networks relative to the number of interactions we would have expected without the loss. The difference-in-difference estimates from these models are ratios of the numbers of interactions in the bereaved group compared to the numbers for the control group and pre-death baseline, so that the findings control for pre-existing differences in levels of online activity. Thus, if we see larger increases in activity among younger people, it is not simply because younger people were more active than older people on social media in the first place. The estimates reported in the text and in Fig. 3 are from models that exclude the month of death, as well as the month before the death and the month after it, so that the estimates for longer-term effects are not skewed by peaks in activity immediately around the death and funeral.
We first estimate changes in numbers of interactions among close friends of the individuals who died. The purple line in Fig. 2 shows the monthly changes in close-friend interactions before and after the death of a friend. Interactions increase sharply at the death and slowly fade as time goes on (log months from death slope −0.026, 95% confidence interval (CI) −0.032 to −0.020). On average, there were 4.5% (95% CI 3.4–5.7%) more interactions in close-friend networks 9 months after losing a mutual friend than otherwise.
The green line in Fig. 2 shows the monthly changes in interactions from the decedent’s close friends to the decedent’s acquaintances. There were 2.6% (95% CI 1.5%–3.6%) more interactions with acquaintances 2 years after the death than before. These interactions were significantly less likely to fade over time than the close-friend-to-close-friend interactions (slope −0.008, 95% CI −0.015 to −0.001). That is, the networks displayed long-lasting effects, with close friends and acquaintances strengthening and developing new connections that persisted for multiple years and that stabilized at levels similar to the increase in social interactions among close friends.
Finally, the orange line in Fig. 2 shows that there was no overall change in social interactions directed from close friends of the decedent towards individuals who did not know the person who died (P = 0.37), suggesting that interactions formed in the short and longer term were highly localized. In other words, close friends did not increase their online social interactions in general.
Because the increases in close friends’ interactions happen after loss of potential interactions from the death of a friend, we estimate how many interactions were ‘recovered’ through compensation. This comparison between how much social interaction was lost and how much was gained allows us to better interpret the magnitude of potential social support provided after a death. Although we do not know how much online social support individuals need or would like after the death of a close friend, the resulting ‘recovery’ in social connections suggests that friends of the deceased provided compensatory support to each other, reducing potential isolation.
The right side of Fig. 3a shows that in networks of individuals aged 25 or over, the increase in interactions fully compensated or nearly compensated for the loss of the interactions that the deceased individuals had contributed. The grey closed circles in the left side of Fig. 3a estimate the absence of the ego (the decedent) without compensation (for example, the interactions lost after the subject’s death) and the turquoise closed circles in the right side of Fig. 3a estimate interactions with compensation (that is, the lost interactions plus the new/strengthened interactions to other close friends and acquaintances of the decedent). In networks in which the subject was under 25, the surrounding friends actually increased the number of interactions in the local network.
Using estimates from the count-models to simulate the percentage of social interactions recovered among close friends of the decedent and from close friends to acquaintances of the decedent, we estimate that recovery across all age groups was 99% (simulated 95% CI 77% to 126%). When considering only compensation from close friends of the decedent to other close friends, compensation was 78% (simulated 95% CI 63% to 96%). We show estimates for recovery within close-friend groups by age in the Supplementary Information.
In the results described above, we considered the effect of the age of the deceased friend on connective recovery in their social network. To test whether younger or older individuals compensated for the loss of a friend differently, we stratified our estimates based on the ages of close friends. We also distinguished text-based interactions (wall posts and comments) from photo-based interactions (photo tags) to evaluate to what extent recovery might be limited to online interactions and not extend to offline ones. Past research has found that photo tags were more likely to reflect offline interactions22 (see Supplementary Information for a principal component analysis supporting this distinction between text and photos).
In Fig. 3b, we show that increases in social interaction vary by the age of a deceased individual’s friends. We observe smaller compensation effects among older friends: older adults engage in fewer new social interactions with other friends of the decedent. However, we show in Fig. 3d and e that this decline can be explained by the ages of both the ego and their close friends, as well as cause of mortality. When a young person dies unexpectedly (that is, from an unintentional injury; similar to previous work12, we exclude suicides and homicides from the ‘unexpected’ loss category), new interactions are high regardless of close friends’ ages. After the unexpected death of a young person, friends aged 18 to 24 increased mutual friend interactions by 8.7% (95% CI 2.8% to 14.7%), and friends aged 25 to 64 increased mutual friend interactions by 12.9% (95% CI 8.2% to 17.6%). In Fig. 3b, we see no difference in mutual friend interactions by interaction type at young ages, suggesting that support is happening both online (for example by exchanging supportive wall posts and comments) and offline (getting together in person and being photographed). However, older individuals increase interactions with the decedent’s mutual friends through posts and comments without a corresponding increase in photo tags.
In Fig. 3c, we show that photo-based interactions with people who did not share the loss of a friend (that is, strangers to the decedent) decrease among older people. At ages 55 to 64, photo-based interactions with individuals who did not know the friend who died were only 93% (point-wise 95% CI 88% to 99%) of, or approximately 7% lower than, their expected level. We did not observe this effect among younger people, suggesting that a decrease in photo-based interactions may be specific to older populations.
Finally, in Fig. 3d, we describe variation in these effects by cause of death. In the panel, cause-specific estimates are from separate models, each of which were stratified by the age of death as in Fig. 3a. The red dashed line at 1.0 is the average change in interactions after a death for all causes other than the one shown, and the estimates are changes in interactions above or below that average increase in interactions after a death for the specific cause. Some causes of death are linked to stronger changes in close friends’ social interactions; close friends of individuals who die suddenly and unexpectedly, such as from an unintentional injury, interact more with each other. Friends of suicide victims, however, are less likely than friends of people who died of other causes to strengthen or form new interactions with the decedent’s other friends. Friends of people who die from drug overdoses exhibit a similar pattern, although not at a statistically significant level. We show in the Supplementary Information that other causes of death associated with substance abuse26 and stigma27 — deaths from liver disease and sexually transmitted diseases — were also associated with reduced recovery.
This large-scale, longitudinal study documents social-network adaptations to structural trauma. We found that, on average, social networks fully recovered the volume of social interactions lost from a death. Healing occurred through connective recovery, and friends were more closely connected to each other years after the shared loss. These effects were highly localized and, on the whole, the increase in interactions did not extend beyond the immediate social network.
Although we observed some age variation in close friends’ adaptations to a death, these differences seemed to be limited to close friends aged 18–24 and, after age 24, could be explained by cause of mortality and the age of the person who died. In other words, the youngest people in these networks — conceivably those with the most fluid lives and ties — tended to contribute a disproportionate amount to connective recovery, but individuals of all ages adapted and greatly contributed to recovery when a young person died unexpectedly from an unintentional injury. These differences were not due to different overall levels of online activity by age. The difference-in-difference design and the multiplicative model, which estimated ratios of activity rather than absolute changes, aided comparisons of effect sizes between younger and older groups with different baseline levels of activity.
We leave many questions unanswered. It is likely that networks do not always adapt to a loss and network-level recovery might not translate into recovery at the individual level; notably, we were not able to evaluate both connective recovery and the subjective experience of loss here. Evidence on recovery among widows, for example, suggests that social support from friends does not often compensate for the loss of a spouse28 and that recovery may depend on a shared loss, as widowed individuals fare better when they live close to others who have experienced the death of a spouse29. Furthermore, the apparently smooth trajectories of network recovery seen here might correspond to noisier oscillations in recovery at the individual level30. Even with full connective recovery, the networks might have a changed ‘personality’ and function differently from before. The strengthened and newly active friendships did not replace the deceased; although levels of connectivity were the same, the networks restructured to accomplish it.
Although we establish that online social networks recovered lost social connections after a death, it is difficult to further evaluate whether the effect sizes identified here are large or small. As a reference, the subjects sent an average of 13 wall posts and messages and were tagged in 3 photos during the 6-month baseline period January to June 201124. For comparison, another large-scale Facebook study demonstrated that receiving approximately twice as many comments from close friends, an increase of 100% compared with the average number received in the study, was linked to increases in perceived social support comparable to a major life event, such as having a baby31. Here, after accounting for baseline differences in activity between the two studies, interactions during the month of death increased by roughly 60% among young adults ages 18 to 24, and 30% among all adults, while long-term interaction increased by approximately 5%. Further, these increases in interaction occurred for over 20 (median 27) close friends in the bereaved networks. This rough comparison suggests that even small increases in online social network activity, such as the effects seen here, have the potential to be meaningful. However, as we note above, we are not able to measure the subjective experience of loss and recovery at the individual level.
Because we do not have data on offline interactions, we cannot say for certain that social support online reflected increased offline interactions. Although past studies have documented similarities between online and offline interactions20,
A practical implication of these findings is that the typical social response to loss seems to happen faster than psychological diagnosis periods. Whereas psychological maladaptation to a loss is diagnosed at 14 months after a death35, the findings here suggest that social interventions might take place substantially earlier. Although we do not know whether this increased connectivity translates to immediate perceptions of closeness, it suggests that the precursors to new or strengthened close friendships manifest immediately after a loss. The potential for age differences in recovery merits further study, particularly in the context of modern, aging populations.
We note two possible explanations for the quick and nearly complete recovery of social interactions that we observed. First, the compensation effects might be driven by a lower bound on individuals’ level of social connection. Because individuals might have a carrying capacity in their social activity36, we might expect them to be driven to replace lost friendships more quickly than they are driven to establish friendships in general. However, we rule out a Facebook-specific compensation effect by showing that we do not observe a similar compensatory increase in social interaction among friends of a living individual who simply deactivated their account on Facebook (see Supplementary Information).
Second, compensatory social interactions could result from bonding during crisis. The extent of recovery observed here would imply that grief responses tend to produce a level of increased social interaction that compensates for the loss of a single individual.
Finally, recovery dynamics here did not correspond to the hypothesized ‘five stages of loss’37. Instead, they were similar to patterns seen in resilient psychological responses to grief and trauma17. These responses to loss mathematically resemble responses to shock in small-scale biological networks. As a quantitative analogue for the patterns of social recovery that we observed, we highlight in Supplementary Fig. 5 that the dynamics here resemble patterns observed in synaptic potentiation — the set of processes that underlie learning and memory in the brain38.
We hope that these findings spur greater interest in how social networks adapt to trauma and crisis. Better understanding of social-network adaptations could help us to identify why social networks succeed or fail in recovery, and how such failures might be prevented. The findings here, we believe, are an important first step in this direction.
To conduct this study, we used Facebook data as well as public vital records from the State of California. Our study protocol was approved by four bodies: the Institutional Review Board at the University of California, San Diego (UCSD); the State of California Committee for the Protection of Human Subjects; the Vital Statistics Advisory Committee at the California Department of Public Health; and Facebook’s internal review group. The UCSD institutional review board approved a waiver of informed consent for analysis of existing data. We have created an aggregate dataset that preserves data privacy.
The analysis is restricted to the social networks of Facebook users in California who met basic, pre-analysis criteria: they had a ‘real’ first and last name, birthdate between 1945 and 1989 (see Supplementary Information), and at least two ‘close friends’ (defined below). A total of 12,689,047 profiles fitted the eligibility criteria. Once we had identified the eligible population, we matched profile information (first name or nickname, last name and date of birth) to California Department of Public Health vital records for 2012 and 2013 to ascertain whether the individual was still living, and if not, his or her cause of death. In 15,129 cases, the vital records indicated that the person died between January 2012 and December 2013. To preserve privacy, after automatically matching to public records, all analyses were performed on de-identified, aggregate data. All data were observational; no one’s experience on the site was different from usual.
The focus of the study is on the close friends of deceased individuals: how their friendship connections and communication patterns changed after the death of a friend (referred to as the subject or the decedent throughout). We characterized types of friends of the subject based on their communication during the period January to June 2011. ‘Close friends’ were defined as people who communicated with the subject using Facebook comments, posts or photo tags, or who appeared in a photo with the subject during this 6-month window. We use the term ‘close friends’ loosely to represent individuals who interacted with the subject; this is likely to include the subject’s closest confidants20,22 and other less important communication partners.
We contrast these close friends with ‘acquaintances’, Facebook friends who did not communicate with the subject during during January to June 2011, and ‘strangers’, individuals who were not Facebook friends with the subject and did not communicate with or appear in any photos with the subject. Within the analysis sample (see Methods), the median number of close friends was 27 (25–75th percentiles: 10–69) and the median number of acquaintances (Facebook friends excluding close friends) was 64 (30–138). These numbers are lower than those for all Facebook users, but social connections and social media activity are typically lower in older populations. All Facebook users not in the close friend and acquaintance groups were counted as strangers (that is, any Facebook user at two degrees of separation from the decedent). For computational reasons, we measured wall posts, comments and photo tags of close friends who were also based in California, but recipients of interactions were not constrained to California.
We then counted how many different people the subject’s close friends communicated with each month who were (1) other close friends of the decedent, (2) acquaintances of the decedent and (3) strangers to the decedent. We separately counted text wall posts, comments and photo tags, counting the number of people towards whom each of the subject’s close friends directed each of those actions during the month. To avoid counting a small number of individuals extremely heavily in our outcome measure, we did not count multiple interactions of the same type between the same two people (although including them did not alter our results).
For each action type (wall post, comment, photo tag) and recipient type (close friend, acquaintance, stranger), we summed these monthly social interactions for all close friends in the networks. To combine the three action types of differing scales without making assumptions about their importance, we then used the geometric average of the wall post, comment and photo-tag edge sums in each network with an adjustment to account for zeroes. Per network in the analysis sample, there was a cumulative median of 113 of these monthly interactions between close friends and other close friends of the decedent (25–75th percentiles: 16–463), 87 (22–261) between close friends and acquaintances of the decedent, and 4,049 (1,117–12,041) between close friends and strangers to the decedent. Supplementary Fig. 6 displays the distributions of counts of close friends and acquaintances, as well as interactions between close friends and the decedents’ other close friends, acquaintances and strangers.
To ensure age and gender covariate balance in our analyses, we compared the deceased individuals to a stratified random sample of non-deceased individuals. This comparison sample contained two networks matched on age, gender and name validation (see Supplementary Information) for each network that had experienced a death. These comparison networks were randomly paired, given same age, gender and name validation, to networks in which the central individual died. The comparison networks were assigned counterfactual dates of ‘death’ from the paired networks. There were 30,258 social networks in this comparison sample, referred to as the ‘control’ group to be consistent with other studies, and 15,129 networks in which the central individual died, referred to here as the ‘treatment’ group. In total, there were 2,020,493 close friends and acquaintances in this sample, and 771,034 who experienced the death of a friend.
To further reduce confounding and ensure parallel trends in the treatment and control groups, especially unmeasured confounders related to social values, culture and socioeconomic status, we used stabilized inverse-probability regression weights. The propensity scores were estimated using a penalized regression on subject and friend characteristics (counts of subject Facebook activity, counts of close-friend Facebook activity, Facebook-friend self-reported education, self-reported marital status, whether they used a smartphone, and a set of page‘like’ based latent social characteristics, which we describe in the Supplementary Information). This propensity score method was previously validated using an experimental baseline (D. Eckles & E. Bakshy, manuscript in preparation).
We used quasi-Poisson generalized estimating equations with independent working correlation to measure changes in the number of interactions between the decedent’s close friends and other members of the decedent’s local social network. We used a diagnostic plot of the relationship of variance to mean in our data to choose quasi-Poisson over negative binomial and chose these models over Poisson owing to over-dispersion in the data. The mean and variance of the treatment and control groups did not differ before the deaths. In the quasi-Poisson models, the treatment estimate was the difference-in-difference interaction between (1) whether the network included a deceased individual, and (2) whether the time period was before or after the death. The standard errors were clustered at the ego network level. We included controls for interactions among close friends during the 6-month period January to June 2011 to account for differences in network clustering at baseline, along with a control for Facebook activity outside the local network (interactions with strangers) in models that measured interactions within the local social network (close-friend interactions with other close friends and acquaintances). This online sociality control slightly attenuated the effect sizes, but helped to account for changes in overall Facebook activity over time.
To estimate the number of communication interactions ‘lost’ by the death of the central subjects, we added close friends’ communications sent to the central subjects for the control group only. This allowed us to estimate the potential interactions lost in the treated networks compared with the control networks. We do not include wall posts, comments and photo tags to the deceased individual’s account in these models. All statistical tests are two-sided. We did not adjust for multiple comparisons in Fig. 3. The tests were conducted for causes of death that were unexpected (unintentional injury)12 and for causes that past works have found to be strongly associated with low levels of social support (drug overdose and suicide)24,39,
For each of the month-by-month figures, we ran the same models, substituting a continuous variable (months from death — included as fixed effects) for the binary (pre-/post-death) variable. This paired sampling and model set-up is very similar to the coarsened exact-matching approach used in a previous study19.
We have created an aggregate dataset that preserves data privacy, and we will make this dataset available to researchers who request it from the corresponding author.
We will make replication code available to researchers who request it from the corresponding author.
How to cite this article: Hobbs, W. R. & Burke, M. K. Connective recovery in social networks after the death of a friend. Nat. Hum. Behav. 1, 0092 (2017).
Publisher’s note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
We thank J. Fowler, N. Christakis, C. Marlow, L. Adamic, D. Ferrante, A. Bejar, P. Fleming, W. Nevius and M. Jackman for their support on this project.
Supplementary Notes, Supplementary Methods, Supplementary Figures 1–8, Supplementary Tables 1–12, Supplementary References.