Age, Loneliness, and Social Media Use in Adults during COVID-19: A Latent Profile Analysis

Loneliness has been linked to morbidity and mortality across the lifespan. Social media could reduce loneliness, though research on the relation between social media and loneliness has been inconclusive. This study used person-centered analyses to elucidate the inconsistencies in the literature and examine the possible role technology barriers played in the relation between social media use and loneliness during the COVID-19 pandemic. Participants (n = 929; M age = 57.58 ± 17.33) responded to a series of online questions covering demographics, loneliness, technology barriers, and social media use (e.g., Facebook, Twitter, etc.) across a range of devices (e.g., computer, smartphone, etc.). A latent profile analysis was conducted to identify distinct profiles of social media use, loneliness patterns, and age. Results yielded five distinct profiles characterized that showed no systematic associations among age, social media use, and loneliness. Demographic characteristics and technology barriers also differed between profiles and were associated with loneliness. In conclusion, person-centered analyses demonstrated distinct groups of older and younger adults that differed on social media use and loneliness and may offer more fruitful insights over variable-centered approaches (e.g., regression/correlation). Technology barriers may be a viable target for reducing loneliness in adults.


Introduction
By 2030, 1 in 5 Americans are projected to be over the age of 65 years old [1]. With a rapidly aging population comes the growing public health concern of increased loneliness, as studies have shown that social network size tends to decrease with age [2] and selfperceived loneliness is strongly associated with health outcomes [3]. Social media offers great potential to decrease loneliness, particularly when mobility is limited due to illness or other factors (e.g., physical limitations, COVID-19 lockdown period). However, existing research on the relation between social media use and loneliness has demonstrated mixed results, possibly because most studies have examined the relation between social media use and loneliness using variable-centered statistics (e.g., regression/correlation) separately in younger or older samples. The aim of this study was to use person-centered analyses to identify meaningful groupings of participants based on their age, social media use, and reported loneliness to help provide clarity regarding the equivocal findings across studies using variable-centered approaches.
Social media may offer an avenue for decreasing loneliness, as previous studies have shown that using social media for communicating with friends and family (as opposed to strangers), regularly self-disclosing thoughts and feelings to peers, and sharing frequent status updates to Facebook friends is associated with reduced loneliness [4][5][6]. Connections formed using social media require minimal effort but can occur frequently and between many different people with great efficiency, which can foster a sense of connectedness among users, leading to reduced loneliness [7]. By contrast, Bonsaksen, et al. (2021a) use of social media to reduce loneliness has not been investigated across the adult lifespan. To address these gaps, this study used person-centered analyses to identify subgroups of adults of various ages who might show different levels of social media use and loneliness during the COVID-19 pandemic when opportunities for socialization outside the home were limited. A person-centered approach could identify subgroups that are similar in age, social media use, and loneliness-a result that is not possible with variable-centered approaches that examine only how variables relate to other variables within the full sample.
Our approach was twofold. First, specific profiles of social media use, loneliness, and age were examined in adults who were 30 years or older during the COVID-19 pandemic. We hypothesized that person-centered analyses would identify specific age groups with distinct profiles on measures of loneliness and social media use. We did not have specific hypotheses about specific features of these subgroups because of the contrasting results reported in the literature to date. However, we anticipated distinct profiles across age groups, with a middle-aged profile of frequent social media users who reported low loneliness [8]. Second, participant subgroups were subsequently examined by their technology barriers, which have been shown to preclude social media use in older adults but have been limitedly studied in younger samples. Comparisons among profiles were conducted to determine whether (and which) technology barriers differed between the profiles. We hypothesized that technology barriers would be negatively associated with social media use. For example, we expected the most technology barriers in groups of older adults who reported loneliness and lower engagement in social media during the COVID-19 pandemic. Demographic features, including, sex, income, education level, race, and ethnicity, were also examined to further characterize the distinct profiles. Results were discussed for their potential to inform strategies to optimize social media features and use for those experiencing loneliness, especially in the context of limited mobility or other external factors that reduce travel outside the home.

Materials and Methods
This observational study used a cross-sectional design and involved the distribution of an online questionnaire. The study was approved by the Temple University Institutional Review Board. All participants provided informed consent prior to participating in the study.

Participants
Participants were required to be 30 years or older, fluent in English, live in the US or Canada, and have access to a computer, tablet, or smartphone device to complete an online questionnaire. A total of 946 participants responded to the survey. After eliminating participants who failed to pass a manipulation check embedded within the technology barrier questions ("Please select the right-most option [i.e., Strongly agree]"; n = 17), the final sample consisted of 929 participants ranging from 30-98 years old. Data were missing for the social media variable (n = 5), race (n = 4), ethnicity (n = 12), and income (n = 31). Participant characteristics are summarized in Table 1 and show that the sample included individuals between the ages of 30 to 98 (median = 58.7), the majority of whom were White women. Approximately half of the sample was college educated, with a variety of education levels represented. The sample also contained a range of yearly income, with 34.1% earning less than $35 K per year, 43.1% earning between $35-99 K per year, and 22.8% earning more than $100 K per year.

Procedures
Data were collected online from September 2021 to February 2021 using the Qualtrics web-based survey and recruitment services. Qualtrics relies on an actively managed, double-opt-in market research panel to recruit participants based on designated inclusion/exclusion parameters. We sought an even distribution of adult participants across six 10-year age bands ranging from 30 to 90+ years old. Qualtrics determines compensation for the participants they recruit based on survey length and other factors and the type of reward varies (e.g., cash, gift cards, airline miles, redeemable points, etc.). Our survey included a series of questionnaires that took approximately 15-20 min to complete and assessed demographic information, social media use, loneliness, technology use, and technology barriers.

Indicator Variables Included in the Latent Profile Analysis
Social Media Use-Social media use was estimated using questions about use of the following digital devices: smartphone, tablet, and personal computer/laptop. For each device that the participant owned and/or used, they were asked to indicate whether they use the device to view or engage in social media at least once per week. Responses collapsed as follows: 0 = no weekly social media use, 1 = uses social media on one device weekly, 2 = uses social media on two devices weekly, 3 = uses social media on three devices weekly. Higher scores were interpreted as reflecting greater social media use and engagement. Frequency of overall device usage was also factored into the overall social media score to estimate the amount of time spent on social media. Participants were asked how frequently they used each device (smartphone, tablet, and computer/laptop). The final social media use score ranged from 0-20, with a score of 0 indicating that the participant did not use social media on any device and a score of 20 indicating that the participant used social media on all of their electronic devices several times per day.
Loneliness-Loneliness was evaluated using the Revised UCLA Loneliness Scale [26] which includes 20 items scored from 1 to 5 (1 = strongly disagree, 2 = disagree, 3 = neutral, 4 = agree, 5 = strongly agree). Some questions were reversed scored. After standardizing all responses, the total score ranges between 20 and 80, with higher scores indicating more loneliness. A median split was used to dichotomize the loneliness variable, with scores under 36 characterized as "not lonely" and scores greater than 36 characterized as "lonely".
Age-Participants were asked to provide their age in years.

Predictor Variables
Technology Barriers-technology barrier questions consisted of a 25-item survey using a 5-point Likert scale ranging from 1 ('strongly disagree') to 5 ('strongly agree'). Items covered a variety of barriers, such as cost, complexity, limited knowledge, privacy concerns, medical/physical obstacles, and others (see Supplemental Materials Table S1). Technology barrier questions were taken from multiple sources, including the technostress questionnaire [27], the Older Adults' Computer Technology Attitudes Scale (OACTAS) [28], and a focus group on technology barriers in older adult population [29]. A total sum score was computed, ranging from 12 to 125 with higher scores reflecting more technology barriers. Based on the result of a principal component analyses (see Supplemental Materials), items were grouped to measure specific technology barriers (i.e., time/money concerns, limited knowledge, sensory/motor difficulties, and privacy concerns). Because a different number of items was included in each sub-component score, average scores were computed for each component that ranged from 1-5, with higher scores indicating more barriers.
Demographics-In addition to reporting age (indicator carriable described above), participants were asked to indicate their race, ethnicity, sex, education level, and annual income from a series of options.

Statistical Analyses
Statistical analyses were performed with MPlus Version 5.2 (Los Angeles, CA, USA) [30] and SPSS version 25 (IBM SPSS Statistics for Windows, Armonk, NY, USA). Descriptive analyses were performed for all study variables. Means and standard deviations were used for all variables that were normally distributed. Median and range were used for variables that did not meet normality assumptions. A factor analysis of technology barriers was performed to determine areas where respondents grouped together. Four factors were identified: knowledge, time and money, sensory motor, and privacy. Once items were grouped into a factor, a composite score was created by adding all items with the same factor together. Descriptive statistics of technology barriers are shown in Table 1. Mean averages of each factor within the 5 distinct profiles is shown in Figure 1.

"lonely."
Age-Participants were asked to provide their age in years.

Predictor Variables
Technology Barriers-technology barrier questions consisted of a 25-item surv ing a 5-point Likert scale ranging from 1 ( strongly disagree') to 5 ( strongly agree') covered a variety of barriers, such as cost, complexity, limited knowledge, privac cerns, medical/physical obstacles, and others (see Supplemental Materials Table S1) nology barrier questions were taken from multiple sources, including the techno questionnaire [27], the Older Adults' Computer Technology Attitudes Scale (OA [28], and a focus group on technology barriers in older adult population [29]. A tot score was computed, ranging from 12 to 125 with higher scores reflecting more techn barriers. Based on the result of a principal component analyses (see Supplemental rials), items were grouped to measure specific technology barriers (i.e., time/mone cerns, limited knowledge, sensory/motor difficulties, and privacy concerns). Becaus ferent number of items was included in each sub-component score, average score computed for each component that ranged from 1-5, with higher scores indicating barriers.
Demographics-In addition to reporting age (indicator carriable described a participants were asked to indicate their race, ethnicity, sex, education level, and a income from a series of options.

Statistical Analyses
Statistical analyses were performed with MPlus Version 5.2 (Los Angeles, CA [30] and SPSS version 25 (IBM SPSS Statistics for Windows, Armonk, NY, USA). D tive analyses were performed for all study variables. Means and standard deviation used for all variables that were normally distributed. Median and range were us variables that did not meet normality assumptions. A factor analysis of technology ers was performed to determine areas where respondents grouped together. Four were identified: knowledge, time and money, sensory motor, and privacy. Once were grouped into a factor, a composite score was created by adding all items w same factor together. Descriptive statistics of technology barriers are shown in T Mean averages of each factor within the 5 distinct profiles is shown in Figure 1  Latent profile analysis (LPA) was used to empirically identify participant groups based on age, social media use, and loneliness scores. LPA was used because of several advantages over other person-centered techniques such as cluster analysis and k-means clustering. LPA uses a stepwise procedure and a variety of fit indices to determine whether the addition of classes improves the fit to the data [31]. LPA can accommodate predictor and outcome variables of differing scales and variances, and normally distributed data are not required. LPA assumes that latent profiles are homogenous and that relationships between variables are due to the underlying latent profile; outcomes include model-based, probabilistic classification scores. Mplus default settings were used for the LPA analyses: variances were held equal across classes and covariances among indicators were fixed at zero.
Although there are no "gold standard" guidelines regarding the number of participants and power for a proposed LPA, a variety of statistical indices and conceptual considerations are taken into account to determine which model fits the data best [32,33]. LPA models are fit in a series of steps, starting with a one-profile (independence) model. The number of profiles then is increased one at a time until there is no additional improvement to the fit of the model [32,33]. Statistical fit indices are examined, including the Akaike information criterion (AIC) [34], Bayesian information criterion (BIC) [35], and sample-size adjusted (ABIC) [36]. Lower scores on these indices suggest a better model fit. We also considered the bootstrap likelihood ratio test (BLRT), which examines the fit of the model with k profiles to the fit with k − 1 profiles [32,33]. Entropy, an index of how well participants fit into distinct profiles that differ from each other, values greater than 0.80 indicate good separation between the identified groups [37]. The size of the smallest profile was also considered; profiles that are too small (e.g., <5-10% of the sample) suggest overfitting of the model to the data, which could limit generalizability. Last, theoretical implications and consistency with conceptual models were also considered.
Following identification of profiles, between-group differences were examined using Chi-square tests to determine whether profiles significantly differed on demographic variables. A Kruskal-Wallis test was conducted to evaluate differences among the 5 profiles on their total number of technology barriers. Post-hoc tests were then conducted to test pairwise comparisons of total technology barriers between profiles. All analyses were determined to be significant when p < 0.05.

Psychometric Properties of the Study Questionnaires
Cronbach's α and McDonald's ω were computed for both the UCLA Loneliness Scale (α = 0.94; ω = 0.94) and the technology barriers total score (α = 0.95; ω = 0.95) and indicated good internal consistency. Internal consistency was not expected for the items that comprise the social media use questionnaire; therefore, Cronbach's α was not computed for this questionnaire.
Bivariate correlations among the primary measures are reported in Supplementary Materials (See Table S2) and supported the validity of the scales. For example, as expected, there were significant, negative associations between the social media use total score and both age and the technology barriers total score. That is, higher social media use scores were associated with younger age and fewer technology barriers. There was no significant bivariate relation between social media use and the UCLA Loneliness score; however, the UCLA Loneliness score was significantly associated with age and technology barriers, such that greater loneliness was associated with older age and more technology barriers.

Latent Profile Analysis
LPA models were conducted using z-transformed scores for age, loneliness, and social media use. The one-profile model was a fit first, followed by models with two, three, four, five, and six profiles. As shown in Table 2, multiple model indicators suggested that the five-profile model was the best fit for the data. First, the VLMRLRT and BLRT p-values were not statistically significant for the six-profile model. Thus, when considering only the first five models, the AIC, BIC, and ABIC are smallest for the five-profile model. The fiveprofile model also had good entropy, and the profile with the fewest participants included at least ten percent of the sample (n = 99). Finally, the five-profiles were interpretable and conceptually coherent. Therefore, the five-profile model was identified as the final model for interpretation. Mean z-scores are shown in Figure 2 to enable comparisons across profiles and variables. Additionally, mean raw scores (and standard deviations) for age, loneliness, and social media use across each of the five profiles, along with their characterizations, are included in Table 3. UCLA Loneliness score was significantly associated with age and technology barriers, such that greater loneliness was associated with older age and more technology barriers.

Latent Profile Analysis
LPA models were conducted using z-transformed scores for age, loneliness, and social media use. The one-profile model was a fit first, followed by models with two, three, four, five, and six profiles. As shown in Table 2, multiple model indicators suggested that the five-profile model was the best fit for the data. First, the VLMRLRT and BLRT p-values were not statistically significant for the six-profile model. Thus, when considering only the first five models, the AIC, BIC, and ABIC are smallest for the five-profile model. The five-profile model also had good entropy, and the profile with the fewest participants included at least ten percent of the sample (n = 99). Finally, the five-profiles were interpretable and conceptually coherent. Therefore, the five-profile model was identified as the final model for interpretation. Mean z-scores are shown in Figure 2 to enable comparisons across profiles and variables. Additionally, mean raw scores (and standard deviations) for age, loneliness, and social media use across each of the five profiles, along with their characterizations, are included in Table 3.

Demographic Differences among the Profiles
Omnibus chi-square tests were performed to examine demographic differences among the profiles. To control for type I error, only omnibus tests with p-values < 0.01 were considered meaningful and further investigated with post-hoc chi-square tests. Results of the omnibus tests are shown in Table 4. As seen in Table 4, post-hoc tests showed that the profiles statistically differed on several demographic variables with a medium effect size for all comparisons, except ethnicity (small effect). Results can be summarized as follows: (1) profile 5 included significantly fewer men; (2) profile 2 included more participants with greater education; (3) profiles 1 and 2 included more participants with higher annual incomes, whereas profile 4 included more participants with lower annual incomes; (4) profile 5 included more participants that identified as races other than White. Table 4. Demographic information and between-group differences for profiles 1-5.

Technology Barriers
A Kruskal-Wallis test was conducted to evaluate differences among the five profiles on their total number of technology barriers (see Table 4). Post-hoc pairwise comparisons showed that profile 2 had the lowest total technology barriers score compared to all other profiles. Profile 5 also had a significantly lower total technology barriers score than profiles 1, 3, and 4. There were no other significant differences between the profiles. The results of post-hoc between group statistical comparisons are included in Supplemental Materials (Table S3).
Mean scores for each of the technology barriers component scores were examined qualitatively. As shown in Figure 1, privacy concerns and limited knowledge were consistently reported as greater barriers than time/money and sensory/motor limitations. Knowledge barriers were higher in the older profiles (e.g., profiles 3 and 4).

Discussion
Person-centered analyses showed five profiles of adults that differed in age, social media use, and loneliness. Profiles that reported high levels of loneliness (profiles 1 and 4) also reported average social media use, and profiles that did not report feeling lonely (profiles 2, 3, and 5) showed all levels of social media use (frequent, average, infrequent). When considering age, among the younger adults in their late 30s (profiles 1 and 2), more frequent social media use was associated with lower levels of loneliness, but this pattern was not observed for older participants. For example, we identified a subgroup of lonely older adults in their late 60's (post-retirement age) who used social media more than a subgroup of the oldest adults in their 80's that reported low levels of loneliness. Thus, our findings, along with the equivocal findings reported in the literature to date suggest personcentered analyses are highly suitable for identifying meaningful profiles of participants with unique patterns of social media use, age, and loneliness. Variable-centered analyses and any approach that assumes linear relations among social media use, age, and loneliness will likely continue to yield conflicting results and obscure meaningful patterns among subgroups of people, particularly among adults from different age groups who may be in very different life-stages and social roles with very different demands.
Demographic features, including, sex, income, education level, race, and ethnicity and technology barriers were examined to further characterize the distinct profiles. The profiles with the highest loneliness did not possess similar demographic characteristics, but they were comparable in terms of technology barriers. Profiles 1 and 4 (participants with more loneliness), as well as Profile 3 (oldest participants), reported more technology barriers than people in the other profiles. Among both the middle-aged and older adult age groups (i.e., people in their late 30s and late 60s), participants who reported more technology barriers also reported greater loneliness. When considering specific technology barriers, our results showed that limited knowledge and privacy concerns were most relevant for participants in every profile. These specific barriers might be targeted to reduce loneliness. Although speculative, it is possible that removing technology barriers may offer people who feel lonely ways to use technology to (1) engage with people through avenues that do not involve social media (i.e., learning about community events, etc.) or (2) access non-social activities that are stimulating and indirectly reduce loneliness.
As stated, our results show no clear and consistent relation between frequent social media use and decreased loneliness among all adults. However, our measure of social media use was general and did not specify specific platforms or social media use behaviors, which may have specific relations with loneliness. For example, Facebook usage was associated with an increase in relationship tie-strength more than face-to-face communication in a sample of 3649 Facebook users. These effects were seen in both direct forms of Facebook use (posting on walls, messaging with friends and family members, commenting on posts) and indirect forms (reading a friend's status updates) [38]. However, Phu and Gow (2019) found that more persistent usage of Facebook was associated with greater loneliness, and a larger number of friends on Facebook predicted higher levels of loneliness in younger adults in their 20's [39]. With respect to Instagram, using the app to browse and interact with content posted by friends and peers was related to lower levels of loneliness, whereas posting content without interacting with posts made by others (i.e., broadcasting) was related to higher levels of loneliness [40]. These findings suggest that social media use is multi-faceted and nuanced, with particular platforms and specific behaviors differentially contributing to social connectedness and loneliness. Though our social media use measure did not assess the use of specific platforms or specific social media behaviors, future work should do so using person-centered analyses.
We acknowledge several limitations and highlight strengths of our study. First, the sample was racially and ethnically homogeneous and comprised mostly of Non-Hispanic White individuals, which limits the generalizability of our results. Additionally, our measure of social media use was developed for this study and has not been extensively studied. However, results from bivariate correlations support the validity of the measure as it was predictably (and significantly) associated with age and technology barriers (see Supplementary Materials). Additionally, asking questions about barriers to technology use through an online questionnaire may have resulted in a biased sample. Respondents in this study had to have some familiarity with technology in order to complete the questionnaire. It is possible this study did not capture extremely lonely adults who do not use social media because they have barriers that prevent them from even accessing or using any technology at all. These limitations notwithstanding, our study also had numerous strengths that served to address gaps in the extant literature. For example, our use of person-centered analyses is novel and conceptually advantageous to map the complexities among social media use, loneliness, and age. Additionally, we recruited a large sample that provided more than sufficient power to detect effects in all of our study analyses. Finally, to our knowledge, this study is one of few that examine technology barriers in a sample that includes people in late young adulthood and middle age (i.e., 30-65 years old). As such, our novel findings indicate that technology barriers can impact adults of all ages and can influence subjective experiences of loneliness regardless of social media use.

Conclusions
In conclusion, person-centered analysis proved to be useful in identifying subgroups of adults with different patterns of social media usage and loneliness during the COVID-19 pandemic. Across the five profiles, social media use was not clearly and directly linked to loneliness. Though, among profiles of younger middle-aged adults in their late 30s and early 40s (profiles 1 and 2), more frequent social media use was associated with lower levels of loneliness, but this pattern was not observed for older participants. Additionally, higher technology barriers were associated with higher levels of loneliness across the five profiles. Results support the importance of addressing technology barriers, particularly privacy concerns and limited technology knowledge across all ages, which may work to alleviate technology barriers as well as loneliness. Future studies should incorporate person-centered analyses to examine specific aspects of social media use (frequency, type of social media platform) and their effects on loneliness across the lifespan.