1 Introduction

At the same time that global populations are aging [1, 2], household sizes are shrinking and individuals are increasingly living alone [3,4,5]. This raises concerns about social isolation, which has gained the attention of governments and is increasingly viewed as a public health crisis [3, 6, 7]. Social isolation is a concern because it has been shown to cause adverse physical and mental health outcomes [8, 9] and in one study, it was shown to be more predictive of mortality than smoking [10]. Older adults specifically are susceptible to social isolation due to shrinking social networks and mobility impairments [11, 12]. This is problematic due to the strong connection between social isolation and lower health outcomes [9, 13].

Older adult populations have lower rates of adoption of information and communication technologies (ICTs) relative to younger age cohorts, although adoption rates are on the rise [14,15,16]. Given the lower adoption rates, there is concern that they are at more risk of experiencing social isolation given preexisting risk factors [17, 18]. While studies tend to focus on physical social connections [19,20,21], researchers in the biomedical field have also sought to build technologies for older adults and study whether these technologies reduce social isolation and loneliness [22, 23]. The benefit of technology-based interventions is the richness of available data [24,25,26]. More work is needed that leverages time use data to look at social isolation [27].

Given the importance of social connection for health and lower rates of ICT adoption among older adults, this study uses a national-level time use dataset to help answer the following questions:

  1. 1.

    Are there discernable groups of older adults who use technology to socialize a greater proportion of the time (‘tech socializers’), and if so what are these groups’ characteristics?

  2. 2.

    Do the relationships between the amount of time spent socializing and dimensions of health (self-rated well-being, stress, etc.) vary depending on whether it occurred in person or via ICT?

  3. 3.

    Are there seasonal differences between socializing in person or via technology for older adults?

Through answering these questions, this paper situates older adult socializing via technology alongside in-person socializing. It also considers seasonal influences that might drive individuals in northern climates to use one method of socializing (technology) during less hospitable seasons.

2 Literature review

Social isolation is the absence of quantity and quality of social relationships [19, 28]. Sometimes conflated with loneliness, it is important to detangle social isolation from loneliness as they have different conceptualizations and relationships to health [29]. While a person may be socially connected (i.e., having numerous, quality social relationships), they may still feel lonely (i.e., perceiving a deficit in their social network). It has been recommended to include measures of loneliness when studying social isolation to tease out differing effects on health [29], but this information is not often available in federally collected datasets. Instead, social isolation alone is reported, as this concept is more easily quantifiable through variables like the number of social engagements [19, 20, 30]. Traditionally, studies have focused on in-person socializing where the interactions between individuals occur in the same physical space [21, 30,31,32].

Risk factors for social isolation are well documented and often are more prevalent in old age [34]. For example, risk factors include living alone and/or the loss of life partners [6, 11, 35, 36]. Klinenberg suggests that more research is needed to fully realize how much aging and living alone affect health. Recent studies on living alone and socializing present mixed results on the topic [35]. Chai and Margolis leveraged time use data to understand whether older adults living alone are more socially isolated (represented by the time spent socializing either in person or using technology), and ultimately found that older adults living alone spent more time socializing than those who did not [37]. In contrast to this, Fingerman et al.’s study of older adults living alone during the coronavirus pandemic showed that those living alone were less likely to see people in person than those who did not live alone, and they were also less likely to give or receive support (e.g. help getting groceries) [38].

Importantly, social isolation and health have a reciprocal relationship wherein poor health can cause social isolation [13]. For example, an older adult with vision impairments might lose the ability to drive, resulting in less ability to attend social outings. Multiple studies have found links between accessing transportation and social isolation for older adults [11, 39,40,41]. One 2019 scoping literature review found that older adults experience decreased access to activities due to driving cessation and mobility-related disabilities, especially in rural areas where there are gaps in public transportation [11].

In addition to mobility and living alone, individual- and community-level poverty are correlated with social isolation [42, 43]. Community-level poverty relates to how low-income people in areas of concentrated poverty tend to remain low-income longer than peers in relatively high-income areas due to more permanent causes of structural poverty [44]. Studies that operationalize this idea are largely concerned with the related concept of social exclusion, rather than social isolation [42, 43, 45]. Similar to community-level poverty, individual-level poverty can be understood as both a cause and outcome of social isolation. Social isolation makes it harder for individuals to take advantage of economic opportunities and poverty makes it more difficult for individuals to take advantage of social opportunities [46]. Technology may exacerbate these issues, as it can be difficult for older adults to purchase new technologies due to socioeconomic status, which can widen the gap in access to social opportunities [47,48,49].

Alongside the rising use of ICT and social media have come studies seeking to understand if they might serve as possible remedies to social isolation [23, 50]. Chen and Schulz's systematic review of ICT’s ability to reduce social isolation among older adults revealed that ICT has the potential to be an effective tool for improving social connectedness [23]. One 2013 study of older adults in assisted and independent living communities found a negative relationship between internet usage and social isolation [51], indicating that ICT has the potential to be as effective at limiting social isolation as in-person interventions. ICT has also been shown to be effective at supporting information sharing, physical activity, and mental health for older adults with cognitive impairments and their caregivers [25, 52, 53]. Because of the passive data collection associated with ICT use, studies have leveraged these data to study mental health and isolation using quantitative and qualitative analyses [52, 54, 55]. Researchers using passively collected ICT data are sometimes limited by its quality and bias but if sourced from piloted applications or websites, these data can provide more robust information about how individuals use ICT platforms [52, 56].

Our objective is to examine older adults’ time spent socializing in person or using technology to understand which characteristics demonstrated in the literature relate to socializing mode in later life. Regarding RQ1, we have two hypotheses. First, we hypothesize that most adults will not primarily socialize using technology (i.e., be a ‘tech socializer’) given the lower rates of technology adoption in the population, but there will be a discernable group who spend more time using technology to socialize more than time spent socializing in person (Hypothesis 1). Because technology use is often a solo activity, we also predict tech socializers will spend significantly more time alone (Hypothesis 2). Given the vast literature demonstrating health and socializing are related, for RQ2 we hypothesize that health and well-being measures will be associated with socializing more, regardless of whether it occurs in person or using technology (Hypothesis 3). For RQ3, because winter conditions influence Canadians’ and older adults’ mobility, we suspect time spent socializing via technology will be significantly associated with seasonal changes (Hypothesis 4).

3 Research design

3.1 Dataset

We use Statistics Canada’s 2015 General Social Survey (GSS) Cycle 29 time-use data, which was collected from April 2015 to April 2016 and released for public use in 2017. Eligible individuals were non-institutionalized residents of Canadian provinces and at least 15 years old. Participants completed a general survey, which included modules on demographic and household characteristics, perceptions of time, health and well-being, and employment. In addition, each participant supplied a retrospective, single-day time-use diary that details the activities, locations, and other contextual information (e.g., with whom the activity occurred). More information about how the GSS time-use data are collected is available on the Statistics Canada website [57]. The resulting episodic data provide a sequence of activities and social interactions for a given participant.

As this study is concerned with the habits of older adults, only respondents aged 65 and older are included. While aging is a continuous process, and attention on those reaching retirement is called for, 65 is the cutoff to reduce the number of respondents who are still members of the workforce.

3.2 Methods of analysis

3.2.1 Independent variable selection

Independent variables contextualize each participant’s sociodemographic situation and perception of time use. These include sex of respondent (1 = female, 0 = male), age group (0 = 65–74, 1 = 75 +), low-income (1 = low-income, 0 = not low-income), marriage status (represented by a series of dummy variables: married; widowed; single), population center indicator (represented by categories: urban; rural), perception of having extra time (1 = every day, 2 = a few times a week, 3 = about once a week, 4 = about once a month, 5 = less than once a month, 6 = never), the season of the survey (represented by a series of dummy variables: spring; summer; winter; autumn), and day of the survey (0 = weekday, 1 = weekend). Marriage status, season, and survey day variables are recoded to reduce the number of categories to aid with model interpretation (for original categories see Statistics Canada, [58]). We generate the low-income variable using the household income group, household size, and population centre indicator variables from the GSS, in addition to information from Statistics Canada on the Low-Income Cut-Off in 2015.

The population centre indicator, which we collapse into rural and urban for use in the model, is a GSS-reported variable characterizing a survey participant’s residential location as falling into a Census Metropolitan Area (CMA) or Census Agglomeration (CA). If they do not live in one of these areas, or live in a ‘small population centre,’ the participant’s population centre indicator is coded as ‘rural.’ The GSS codes participants who live in Prince Edward Island (P.E.I) separately, regardless of whether they live in a CMA/CA or not. Due to the small populations in P.E.I.’s CMAs, we treat these participants as ‘rural.’ The Low-Income Cut-Off data provided by Statistics Canada supply income thresholds based on a household’s size and location. For example, a household of three people in an urban population center (population at least 100,000) is considered low-income if their income is less than $40,000 [59, 60].

In addition to the above-listed controls, the exposure variables in this study are a series of health and well-being variables. The health variables included are self-rated health and self-rated mental health, which are captured using a Likert scale (1 = excellent, 2 = very good, 3 = good, 4 = fair, 5 = poor). We also include variables representing self-rated stress (1 = not at all stressful, 2 = not very stressful, 3 = a bit stressful, 4 = quite a bit stressful, 5 = extremely stressful), subjective well-being (1–10, very dissatisfied-very satisfied). However, self-rated health, mental health, and stress variables are recoded as binary variables to aid in interpretation. Excellent, very good, and good responses to questions about health and mental health are considered to be good health/mental health, while fair and poor responses are considered bad. Likewise, those who responded to the self-rated stress question with ‘a bit stressful’, ‘quite a bit stressful’, and ‘extremely stressful’ are classified as being stressed. In addition to these health variables, disability status (1 = presence of disability, 0 = no reported disability) is also included.

3.2.2 Outcome variables

The GSS reports two useful variables for socializing: duration socializing or communicating in person and duration socializing or communicating using technology. These duration variables are generated by Statistics Canada using respondents’ single day time use diaries. Any amount of time an individual reported socializing or communicating either in person or using technology (phone, email, video conferencing, etc.) as their primary activity is included. Using the episode file for each participant, Statistics Canada generates these duration variables by aggregating time spent doing activities as they relate to socializing in person or using technology. We use k-means clustering to group older adults by their socializing behaviors represented by these two variables. To our knowledge, k-means clustering does not allow for the incorporation of survey weights. This is the only analysis performed with the un-weighted survey data. The number of clusters is selected by analyzing the within groups sum of squares across different numbers of clusters to find the lowest possible number of clusters while also reducing the within groups sum of squares; ultimately, we use three clusters. After we generated the clusters, we qualitatively labeled the categories ‘tech socializers,’ ‘common socializers,’ and ‘in-person socializers,’ and used them to build a series of regression models. Descriptive analyses for these categories are provided in Results. ANOVA tests to determine significant differences between clusters were not possible due to the complex nature of the survey data; in place of ANOVA, we use a bivariate general linearized model to estimate the relationship between cluster assignment and time spent alone. We also tested the sensitivity of the k-means clusters by performing a second cluster analysis using occurrences of socializing or communicating in person/using technology, rather than the duration of each activity. As the effect is limited with 97 percent of records being assigned the same category, we report these results alongside the duration clusters in Appendix A.

3.2.3 Regression analysis

Using the clusters obtained from the k-means analysis, we perform two regression models to estimate relationships between the explanatory variables and the mode of socialization (in person, tech, common) assigned to the participant. Due to the categorical structure of the outcome variables, we use a multinomial logistic regression. For all models, the ‘common socializer’ category is the reference category, since these participants show no extreme tendency towards in-person socializing or tech socializing and in general, socialize less overall.

Ultimately, we use two models. Model 2 uses all independent variables, including health, mental health, stress, subjective well-being, low-income status, sex, age group, partner status, visible minority status, disability status, population center indicator, survey day, having extra time, household size, and season, while Model 1 omits survey season. More information about the variables included in the models is reported in Results.

4 Results

This section provides descriptive information for the older adult population included in the study, then moves into the results of the population cluster analysis and regression results. Regression model results are presented as outlined in Regression Analysis, and all results use the GSS-provided probability weights unless specified.

4.1 Descriptive statistics and population cluster analysis

Table 1 provides weighted sample characteristics. In general, the sample is majority female (54%), white (92%), and predominately living in a two-person household or having a long-term partner (58% and 65%, respectively). Nearly half of respondents reported having a disability (49%), and they predominately rated their health as good (82%). While not reported on the table, most participants lived in Ontario (39%), Quebec (25%), or British Columbia (14%), and 80 percent lived in a large urban population centre, which includes all Census Metropolitan Areas and Census Agglomerations except those in P.E.I.

Table 1 Descriptive statistics of older adult sample from GSS
Full size table

We employed a k-means algorithm to segment the older adult population into three categories characterizing socializing behavior (called ‘mode of socializing’), which are used as the outcome during regression analysis. Table 2 provides descriptive information about the clusters. Tech socializers are individuals who, on average, socialized in-person four times less than they socialized using technology. In-person socializers socialized physically 90 times more, on average, than they used technology to socialize. Common socializers tended to spend more time socializing in person, but the difference was much smaller than the in-person category (seven times more, on average) and socialized less overall, on average. According to cluster totals accounting for participant weights, tech socializers had, on average, the highest amount of time spent alone while in-person socializers had the lowest. A bivariate general linearized model between mode of socializing and time spent alone revealed mode of socializing (reference: in-person socializer) is significantly associated with time spent alone (βcommon = 90.26, p < 0.001; βtech = 252.54, p < 0.001).

Table 2 Descriptive statistics of clusters resulting from k-means clustering
Full size table

4.2 Regression results

Results for all models are provided in Table 3. Coefficients have been exponentiated resulting in an odds ratio, which aids in the interpretation of results. Reference categories for explanatory variables are provided in the table.

Table 3 Odds ratios result from multinomial logistic regression
Full size table

Model 2 includes a seasonal variable and has a lower AIC score relative to Model 1, indicating the model has better fit. Given Model 2 is a more complex model yet has a lower AIC score, its results are focused on below.

4.2.1 Health and well-being

Neither in-person or tech socializers have a significantly different relationship to health, stress, or subjective well-being, compared to common socializers. However, presence of a disability is significant to both categories, versus common socializers. The odds of being an in-person socializer versus common socializer for those without a disability is 1.405 times that of those with a disability in Model 2 (p < 0.01). To the contrary, the odds of being a tech socializer versus a common socializer for those without a disability is 0.719 times that of a person with a disability (p = 0.02). In other words, those without a disability have a higher likelihood of being an in-person socializer versus common socializer and a lower likelihood of being a tech socializer versus common socializer.

4.2.2 Socio-demographics

Sex, partner status, and age have significant relationships with in-person socializers and tech socializers. The odds of being an in-person socializer versus a common socializer for females is 1.227 times that of males in Model 2 (p = 0.04); this odds ratio increases to 2.073 for tech socializers versus common socializer (p < 0.01). This indicates that females have a higher likelihood of being in-person socializers versus common socializers when compared to males, and this is especially true for tech socializers.

For partner status, being widowed (versus being in a long-term relationship) has a significant, positive relationship to being an in-person socializer at the 0.10 level (OR = 1.264). For tech socializers, the relationship with being widowed or single is stronger and more significant. The odds that a widowed or single person is a tech socializer versus a common socializer is 1.665 and 1.803, respectively (p < 0.01). As with in-person socializers, this indicates those who are single or widowed have a higher likelihood of being a tech socializer versus a common socializer.

The odds of a person in the older age group (75 +) being a tech socializer are 0.774 times that of a person in the lower age group (65–74), meaning that those in the older group have a lower likelihood of being a tech socializer than those in the younger group (p = 0.08). This relationship is weak and is not observed for the in-person socializers. Being low-income and living in a rural location did not have significant relationships to being an in-person socializer or tech socializer versus common socializer.

4.2.3 Time and season

Being surveyed on the weekend is strongly associated with being an in-person socializer versus common socializer (OR = 2.506, p < 0.01). Because survey day corresponds to time use data collection, it is logical that these individuals were often spending more time socializing in person. Having extra time is not statistically significant in either outcome category, nor is being surveyed on the weekend with being a tech socializer versus common socializer.

Spring, compared to winter, does not have a significant relationship to being an in-person socializer or tech socializer versus common socializer. However, being surveyed in the summer or autumn months exhibits a significant relationship with being an in-person socializer versus common socializer. An individual is 1.622 times more likely to be an in-person socializer versus common socializer in summer compared to individuals surveyed in winter (p < 0.01). In other words, summertime is associated with higher odds of a person being an in-person socializer versus common socializer. A less significant relationship is observed for autumn, with individuals being 1.305 times more likely to be in-person socializers versus common socializers (p = 0.05). There are no seasonal relationships for tech socializers.

5 Discussion

The results of the k-means clustering and regression analyses provide some evidence supporting the hypotheses generated prior to analysis and answer the research questions outlined in the Introduction. First, the results support Hypothesis 1, which speculated that most adults would not be technology socializers, but there would be a discernable group primarily using technology to socialize. During the k-means clustering, most older adult participants were classified as common socializers or in-person socializers (n = 3649 and 596, respectively), while the fewest were classified as tech socializers (n = 308)Footnote 1. This suggests that while most older adults socialize in person versus using technology, some older adults are ‘tech-savvy’ or tech-oriented. These older adults could be key to helping other older adults adopt new technologies, which has been suggested elsewhere [61].

Hypothesis 2, which stated that technology socializers will spend significantly more time alone, is supported by our findings. Tech socializers had the highest average time spent alone (µtech = 828.59 min, 95% CI [758.84, 898.34]), while in-person and common socializers had lower averages (µin-person = 576.05 min, 95% CI [534.54, 617.56]; µcommon = 666.31 min, 95% CI   [645.95, 686.67]). These differences between clusters were confirmed using a bivariate GLM. Studies of social isolation that solely focus on physical outings or activities with companions may be missing a crucial element of social interaction. Given tech socializers did not have negative relationships with health, it can be preliminarily suggested that technology socializers may benefit emotionally from virtually mediated social interactions as much as in-person socializers benefit from physical socializing. Further research on this topic is called for to investigate health and wellness trade-offs, if any.

Finally, we find little evidence supporting Hypothesis 3, which stated that health and well-being measures would be associated with socializing more, regardless of whether in-person or using technology. All health and well-being variables, except for disability status, did not have a significant relationship to in-person or tech socializers versus common socializers. Disability status was strongly related to socializing both in-person and via technology, although the relationships were in opposition. Those without a disability had higher odds of being an in-person socializer versus common socializer, while they had lower odds of being a tech socializer versus common socializer (compared to those who do have a disability). This suggests that older adults with disabilities may be more apt to use technology to socialize if they are unable to do so in person, while older adults who can socialize in person will do so and may be less likely to use technology. The senior technology adoption model (STAM) might lend insight into why this is the case—older adults who have disabilities may perceive a benefit to using technology to socialize in place of in-person socializing [47]. Given the GSS reported more fine-grained disability status variables (e.g., presence of a mobility disability), future work should focus more closely on types of disabilities to fully realize the relationships between disability and socializing.

Hypothesis 4 posited that time spent socializing via technology would be significantly associated with seasonal changes. However, regression results revealed an entirely different relationship. Tech socializers did not have any significant relationships to seasonal changes, while older adults had higher odds of being an in-person socializer in the summer and autumn months, relative to the winter. This suggests that in-person socializing is dependent on the time of year, as older adults might be less able to move around freely in the wintertime due to snow and other weather-related mobility hazards. A study in Alberta found that older adults do indeed change their physical activities in the wintertime, which could explain the significantly higher odds of socializing in person in the summer [62]. Since technology socializers did not have seasonal relationships, it stems to reason that older adults may benefit from transitioning to technology in the winter to maintain the same level of social participation (i.e., become tech socializers). Future work should collect time-use information from the same individuals in summer and winter so seasonal transitions between technology socializing and in-person socializing may be further explored.

A few limitations should be noted. First, the data used for this study has a complex design which limited the integrity of some analyses. For example, survey weights could not be used during the generation of the mode of socializing clusters, which ultimately formed the dependent variable for regression analysis. Some variations on the clustering analysis were performed to validate those generated, and similar clusters emerged each time. Additionally, our analysis conflates socializing with the mode of socializing. It might be useful to separate time spent socializing from the mode to understand how individuals proportion their socializing time. Other cluster techniques which can account for survey weights and for how individuals proportion their socializing activities warrant further exploration.

Historically marginalized groups (e.g., low-income, racial & ethnic minorities, people with disabilities) may have differing levels of technology use [48]. The data in the GSS (not accounting for survey weights) are largely white, affluent individuals. While the regression analysis did attempt to account for being low-income, this variable was generated using coarse information about income and location provided in the public use GSS main file, which could have impacted our results. Given the historically low response rate for the 2015 cycle of the GSS time-use survey, data on marginalized groups may be less robust than in years prior.

No geographic analysis was performed in this study, which could have affected the results. Built environments have been shown to influence older adults’ social participation and ability to be mobile [11, 31]. While a location indicator was included, this was a binary that could not account for the nuanced differences across space that enable or disable individuals to socially participate. Future research should focus energy on obtaining access to private use GSS microdata so built environment analysis can be performed.

Lastly, some work has shown that social connectedness or social belonging can be detrimental to an individual’s health [63]. This study operates on the assumption that social interaction is purely for the benefit of the individual. Without more information on the type of socializing in which individuals participate or the dynamics of these relationships (e.g., bridging, bonding social capital), it is hard to model the relationships between socializing and health. Given the information available in the GSS, work on this issue might be best served using an alternative dataset.

6 Conclusion

This study revealed that there are discernable groups of older adults who use technology to socialize a greater amount than they socialize in person. Tech socializers spend the most time alone, which would flag them as socially isolated in some studies, while in-person socializers spend the least amount of time alone. Individuals with poor health did not have higher odds of being tech socializers, but those with a disability did. The opposite was observed for in-person socializers, as those without a disability had higher odds of being in-person socializers. Results of the seasonal models indicate that individuals have higher odds of being an in-person socializer in summer and autumn versus winter, but these trends did not appear for tech socializers, suggesting tech socializing is less sensitive to changes in weather. Future work should break down the relationship between being a tech or in-person socializer, having a disability, and seasonality to further specify those relationships. While there are some limitations, this study has added to a body of literature on social isolation and has suggested that time spent socializing in different contexts should be considered to better estimate relationships.