Friendship Trajectories and Health across the Lifespan

Abstract

Friends are a vital source of social relations throughout the lifespan and across developmental stages. Our knowledge of how friendships develop over time, especially from childhood through adulthood, is limited. Further, it is now recognized that this specific type of relationship influences health across the life course in unique ways. Using the Convoy Model of Social Relations as a guiding framework, this study charts the multiple and unique trajectories of friendship across adulthood and tests whether these trajectories influence health differentially by age. The sample for the study consisted of 553 adults from the longitudinal Social Relations Study. Respondents ranged in age from 13–77 at Wave 1 (1992), and included only those who reported a best friend in each wave, i.e. Wave 2 (2005) and Wave 3 (2015). Approximately 65% of the respondents were women, and 24.5% were people of color. Latent growth curve analysis identified three trajectories of the presence of friends in one’s network over time, two trajectories of positive friend quality, and three for negative quality. The most consistent findings are associated with positive friend relations over time. Gender was associated with friendship quality where women reported more positive friend relations over time, and increasing positive friend relations predicted better health 23 years later. These findings demonstrate that consistent and increasing positive friendships yield health benefits over time, whereas presence of friends and negative quality do not have an effect. Overall, findings advance understanding of the long term effects of social relations across the lifespan and life course.

INTRODUCTION

Friends are a unique and vital source of social relations throughout the lifespan and across developmental stages. The distinctiveness of the friend relationship has a long and prominent presence in the literature (e.g., Procidano & Heller, 1983). The friend relationship is especially noted for its fundamental role in human development and positive effects on well-being (Blieszner et al., 2019; Holt-Lunstad, 2017; Larsen et al., 2014). Friends help throughout life by connecting individuals beyond their familial relationships and providing sources of emotional support, identity, validation, camaraderie, information, as well as facilitate new ways of thinking and acting (Allan, 1998; Rawlins, 1992). Demographics of family change including lower fertility rates and longer life expectancies have encouraged a research spotlight on friend relationships (Fiori et al., 2020). Further, increasing geographical mobility and changes in the definition of family challenge researchers to break from the traditional assumptions that close social relations are only or predominantly kin relationships (Antonucci et al., 2019; Huxhold, 2019). Hence, attention to how friend relations develop over time is needed, including their effects on health.

Most research on the effects of friendship on health examines this association within a short time frame and/or during specific developmental periods (Asher & Coie, 1990; Bagwell et al., 2005; Hartup & Stevens, 1997; Heinze et al., 2018; Ladd & Troop-Gordon, 2003). Although some researchers have examined the long-term effects of childhood friendship on health in adulthood (Cundiff & Matthews, 2018; Giordano et al., 1998; Narr et al., 2019; Sakyi et al., 2015), our understanding is limited concerning how trajectories of friendship during emerging and early adulthood influence health in middle and older adulthood. This paper uses a unique longitudinal data set to investigate the extent to which friend relations change over 23 years and how these changes influence health among a lifespan sample.

THEORETICAL FRAMEWORK

The Convoy Model of Social Relations, which is grounded in lifespan developmental psychology and life course sociology, provides the guiding framework for the present study (Antonucci et al., 2014; Kahn & Antonucci, 1980). Convoys represent an assembly of social relations, including family and friends, who surround an individual from birth through death. Convoys are considered multi-dimensional and lifelong, subject to both stability and change depending on various personal and situational characteristics, all of which influence health across the lifespan.

The Convoy Model directs us to recognize that friend relations may vary over time. This includes changes in both form, i.e., the extent to which friends are identified as part of one’s convoy or social network, as well as function, i.e., presence of both positive and negative qualities in the friend relationship. Further, friendships earlier in life may hold important implications for those that continue or form anew in later life. Early attachments often set an initial pattern and context for adult relationships (Antonucci et al., 2011), yet growing evidence suggests that early attachment experiences are not destiny in terms of adult relationships (Fraley & Roisman, 2019). Other factors may play a key role that influence the number or quantity of friends one has over time as well as the quality of these relationships. The personal (gender, socio-economic status, race) and situational (marital status) characteristics identified by the Convoy Model may guide the form and function of friendship at various ages.

Beyond accounting for the complexity inherent in the form and function of friend relationships, the systematic study of the multi-dimensional nature of friendship ties in earlier portions of the lifespan may provide critical insights into potential avenues for maximizing health later in life (Chopik, 2018; Huxhold et al., n.d.; Walen & Lachman, 2000). The Convoy Model identifies relationship type as a key avenue for better elucidating the links between social relations and health, showing that not all relationships are equal in their health promoting effects (Ajrouch et al., 2013; Antonucci et al., 2003; Fuller-Iglesias et al., 2013; Webster et al., 2021; Zahodne et al., 2019). Friend relations fluctuate over time and are more likely to require proactive maintenance than family ties (Huxhold et al., 2022; Roberts & Dunbar, 2011). As a result, friend relations (including acquaintances) have been identified as highly influential for health (Fiori et al., 2008; Sandstrom & Dunn, 2014), especially in later years when social networks often become smaller (Ajrouch et al., 2001; Carstensen, 1992; Huxhold et al., 2020; Walen & Lachman, 2014). This paper investigates the ways in which experiences of friend relations during emerging and early adulthood influence health outcomes in middle and older adulthood. Specifically, we explore how this process unfolds similarly or differently at various starting points across the lifespan.

DEVELOPMENTAL PERSPECTIVES ON FRIENDSHIP

The extent to which friends are present in the lives of individuals likely varies throughout the lifespan. Definitions of who is a friend as well as perceptions of friendship quality may reflect developmental trends (Adams et al., 2000). Friendship in childhood and late adulthood has been studied far more than friendship in young and middle adulthood (Blieszner & Ogletree, 2017; Chopik, 2018; Schoeps et al., 2020). Friendship networks begin to expand in adolescence and young adulthood (Blieszner & Roberto, 2004; Wrzus et al., 2013). Though networks tend to get smaller with age (Ajrouch et al. 2001; Carstensen, 1992), it is not clear whether the number of friends identified as close and important also decrease with age. Further, contexts shape friendship experiences across age, such as whether one is male or female, a member of a minoritized group, class privileged, or married (Adams et al., 2000; Allan, 1998; Fiori et al, 2020; Manalel & Antonucci, 2020). Opportunities and expectations related to gender, race, socio-economic and marital status over time may impact shifts in the presence of friends above and beyond age.

In addition to friend networks, relationship quality with friends may also reflect normative developmental trends. The effects of age on relationship quality have been documented to show that older age may lead to a heightened ability to maximize positive aspects of friendship (Birditt et al., 2005; Luong et al., 2011). In particular, emotional closeness with friends appears to increase throughout adulthood (Carstensen, 1992). Further, negative aspects of the friendship tie do not increase, but rather tend to remain stable over time (Akiyama et al., 2003; Birditt et al., 2009; Bruine de Bruin et al., 2020). Though longitudinal research suggests there is more continuity than change in later life friendships, gender may impact friendship trajectories. Field (1999) found that in old age, men displayed more change than women in terms of not only decreased number of friends, but also less desire for closeness compared to women, a difference that was not evident in earlier years. Nevertheless, it appears that social skills evolve with life experience, and hence the likelihood of positive relationship quality with friends increasing as one grows older is high (Blieszner & Roberto, 2004). Using age as a proxy for development, we examine the likelihood of having friends over time as well as the quality of friend relationships, accounting for gender as well as other critical personal and situational characteristics such as socioeconomic status, race and marital status. Next, we consider the increasing evidence that changes in the friendship tie effects health over time.

FRIEND RELATIONSHIPS INFLUENCE ON HEALTH OVER TIME

The positive impact of having friends appears to garner long-lasting effects over time. Indeed, less close non-family ties generally offer adults a more effective avenue for promoting health and well-being over time than close family ties (Huxhold et al., 2020). For instance, Almquist (2011) found that childhood friendships are linked to middle age health disparities; those without friends reported worse self-rated health over 35 years later. Yet, most research has focused on earlier parts of the life course, and often examines shorter time frames. Narr and colleagues (2019) studied friendship among middle adolescents and found that close and strong friendships predicted increases in self-worth and decreases in anxiety and depressive symptoms by early adulthood. Further, having no friends in childhood predicted a higher likelihood of exhibiting internalizing and externalizing problems in early adulthood (Sakyi et al., 2015). Health benefits of having friends has been documented at older ages as well (Fiori et al., 2006; 2007). We investigate whether there are long-term effects on health in later life from earlier friendship ties starting in adolescence.

Beyond the importance of having friends, the benefits of friendship quality on health outcomes have also been documented (Efeoglu & Sen, 2022; Holt-Lunstad et al., 2007). Though positive friend relations yield health benefits (Blieszner et al., 2019; Holt-Lunstad, 2017; Larsen et al., 2014), negative aspects of friendship may produce distinct health benefits as well (Antonucci et al., 2010). Moreover, the health benefits of friendship quality may vary by personal and situational characteristics such as gender and marital status (Antonucci et al., 2001), race (Mouzon, 2014) or socioeconomic status (Almquist, 2012). We investigate whether changes in friendship form and function matter more for younger, middle aged or older adults’ health over and above the effects of gender, race, socioeconomic status and marital status.

STUDY GOALS

Our study goals were derived from the Convoy Model and examined the influence of personal and situational characteristics on friendship and health outcomes over time. The first goal was to chart the multiple and unique trajectories of friendship across early, middle and late adulthood. For this exploratory goal, we hypothesized that age would predict the likelihood of trajectory classification concerning the proportion of friends in one’s network, as well as the quality of the friendship tie. Our second goal was to test whether the trajectories of friendship influence self-rated health and depressive symptoms differentially by age. We hypothesized that older adults with trajectories of high proportions of friends, as well as high positive and low negative relationships quality with a best friend would report better self-rated health and less depressive symptoms than younger adults. We further investigate the additional personal and situational characteristics of gender, racial/ethnic group membership, education, and marital status.

METHODS

Sample

Data for this study come from three waves of the Social Relations Study collected by the Survey Research Center at the University of Michigan. The first wave of this regionally representative sample (N=1,703) was collected in 1992 from those aged 8–93, with a 72% response rate. Wave 2, collected in 2005, consisted of 1,076 of the original respondents and Wave 3, collected in 2015, included 727. At Wave 3 six hundred thirty-three (37%) were deceased, 54 incapacitated (3%), and 319 (19%) were coded as drop-out attrition (final refusal by informant, incarcerated, lost-unable to track, unavailable for entire study period, and unable to complete interview). The sample for the current study consisted of 553 adults who ranged in age from 13–77 at Wave 1, and who reported a best friend, though not necessarily the same best friend, in each wave. To explore data attrition from young adulthood to later adulthood, a series of attrition analyses were performed (see Appendix Table 1). We compared the demographic characteristics among three groups: 1) respondents excluded from analysis (n = 954) due to dropouts in Wave 2 and/or Wave 3; 2) those included in the proportion friends analyses only (n = 297), but did not identify a best friend; and 3) those used in all analyses (n = 256). Two differences were found: people excluded from the final analyses were older in Wave 1, and those who did not identify a best friend across three waves were more likely to be married.

Measures

Age was measured as a continuous variable, calculated from birth date.

Friendship.

Proportion friend in network was assessed in all three waves using the hierarchical mapping technique (Antonucci, 1986). Respondents were shown a diagram containing three concentric circles with the word YOU in the middle. They were asked to nominate people in their lives based on varying levels of closeness. Respondents were then asked to identify their relationship with each of the first 10 people named in their network age 13 or older. The proportion of those identified as a friend in each respondent’s network ranged from 0–100%.

Friend relationship quality was measured in all three waves to assess both positive and negative aspects with a best friend. Positive quality was indicated by creating an index that averaged the responses (1=disagree; 5=agree) to the following five items: “I feel my friend supports me”, “I can share my very private feelings and concerns with my friend.”, “I enjoy being with my friend.”, “I feel my friend encourages me in whatever I do.”, “I feel that my friend believes in me” (alpha=.77). (Revelle’s ω_t=.77; ω_h=.63 [McNeish, 2018]). Negative quality was indicated by creating an index that averages the responses (1=disagree; 5=agree) to the following two items: “My friend gets on my nerves” “My friend makes too many demands” (ω_t =.55). The relationship quality measures have been used across multiple national and international studies (Akiyama et al., 2003; Birditt & Antonucci, 2007; Birditt et al, 2016) and have been shown to have consistent and expected associations with health and wellbeing (Cohen, 2004; Ha, 2010; Rook, 2015).

Health was assessed at Wave 3, controlling for baseline. Self-rated health measured how respondents rated their health at the present time on a 5-point scale ranging from very poor=1 to excellent=5. Depressive symptomatology was measured using the 20-item Center for Epidemiological Studies Depression (CES-D) scale (Radloff, 1977). Respondents reported the experience of depressive symptoms in the past week on a 4-point scale ranging from 1 (rarely/none of the time) to 4 (most of the time). Item scores were summed to create a total composite score with higher values indicating greater depressive symptoms (α = .85).

Personal and Situational Characteristics.

The following characteristics that have been found to be linked to both friend relations and health were included in this study: Gender, coded as male=0; female=1. Race/ethnicity was assessed with a self-report item and coded so that White=1 and people of color=0. Education was measured with a self-report item worded as “What is the highest grade of school or year of college you have completed?” Education level ranged from 0–17+ years. Marital status was assessed through the question, “Are you currently married or living with a partner, widowed, divorced, separated, or have you never married?” and coded 1 = married/cohabiting; 0 = other.

Missing data.

We used multiple imputation to address missing data. We used the MICE package in R to generate 50 datasets (Buuren & Groothuis-Oudshoorn, 2011). While 50 is larger than current standards, which suggests that one dataset for each percentage of missing data is sufficient, we observed that the group trajectory analyses showed greater variation across datasets compared to typical regression models. We therefore used more datasets than usual to create a more stable approximate of the posterior distribution. We used predictive mean matching for trajectory variables and linear regression for all remaining variables.

Analysis Strategy.

Analyses involved two approaches to address our research questions. First, to examine research question 1, Latent Class Growth Analysis (LCGA) was used to identify discrete subgroups of friendship trajectories over time (Nagin, 2009). We used LCGA based on previous research that identified theoretically grounded and empirically meaningful subgroups of relationship quality trajectories (Birditt & Antonucci, 2007) and patterns of social network change (Litwin et al., 2020). We extend this previous work by focusing specifically on the friend relationship. Our LCGA model uses observations across all three waves as indicators of latent subgroups that were not directly observed. We aimed to identify discrete subgroups of friendship trajectories to illustrate heterogeneity in the developmental nature of friend relationships. Notably, LCGM has the advantage of being able to identify latent classes according to the longitudinal change pattern of friend relationships, and at the same time, can identify factors related to the classification of the subgroups (Lee & Chung, 2021). For each measure of friendship, we estimated quadratic trajectories for 2 to 5-class solutions, meaning each class had its own intercept, linear change, and quadratic change terms. Because only three time points were observed, this corresponds to freely estimating the average of each trajectory at each time point. We allowed the residual variation around each trajectory to vary by class. To avoid computation problems which occur when the residual variation approaches 0 and the likelihood approaches infinity, we constrained the residual variation to be larger than a particular small value which we subjectively selected. We used the Akaike information criterion (AIC) and Bayesian information criterion (BIC) to inform model selection. In two of the three cases, models continued to improve with the addition of more classes, though by extracting smaller and smaller subsets of respondents. For these reasons, we used subjective criteria to determine our models of choice, as recommended by Nagin, (2009, p.74). We also calculated scaled entropy, a measure of the classification accuracy of sample respondents into classes where a value of 1 indicates no uncertainty in classification and 0 indicates that the model classifies the sample no more precisely than random chance.

Once we determined the number of classes, we predicted class membership with age as well as the personal and situational characteristic data using the automated three-step approach in Mplus (R3STEP; Vermunt, 2010). The R3STEP procedure allows uncertainty in classification to influence the predictive model while ensuring that the trajectory model converges to the same trajectories as in the unconditional model. For these analyses, we centered respondent age at 40. We also centered years of education at 12.

To examine research question 2, we conducted multivariate regression analysis using R to determine links between identified classes and health outcomes in Wave 3, controlling for age as well as personal and situational characteristics. To conduct these models, first we exported from Mplus the class assignment probabilities for each of our three friendship measures (proportion friends in the network, best friend positive relationship quality, and best friend negative relationship quality). Next, we imported these data into R and merged these data with other study variables. We then conducted a total of six main effect regression models, one model predicting self-rated health and one predicting depressive symptoms for each of the three friendship variables. We included age and the same personal and situational characteristic variables as in the R3STEP models and added the probability of class assignment as a predictor as well as an age × probability interaction, with age and education centered as before. We also centered the class probability at 0.5. When we estimated more than two classes, we used the classification probability for one of the trajectories that had the highest or lowest mean value on the particular friendship scale. This study was not pre-registered. The code behind this analysis has been made publicly available at the APA’s repository on the Open Science Framework (OSF); see https://doi.org/10.17605/OSF.IO/Y7K4W.

RESULTS

We present first a descriptive analysis of the sample, including a correlation matrix to allow for replication. This is followed by findings that address each hypothesis. Table 1 provides the means and sample distribution of study variables examined for those who participated in all three waves of data collection.

Table 1.

Sample Characteristics (N = 553)

	Waves 1–3
	1992	2005	2015
Age (M, SD)	40.0 (13.4)	53.6 (13.4)	63.6 (13.4)
[Range]	[13 – 77]	[26 – 90]	[36 – 100]
Female (%)	64.9%	-	-
White (%)	75.5%	-	-
Education (M, SD)	13.4 (2.4)	14.0 (2.1)	14.1 (2.2)
Married/living with a partner (%)	63.8%	70.3%	64.4%
Proportion of friends in the network (M, SD)	19% (23%)	16% (21%)	17% (22%)
Positive quality [1–5] (M, SD)	4.78 (0.34)	4.83 (0.33)	4.82 (0.35)
Negative quality [1–5] (M, SD)	1.92 (1.06)	1.54 (0.85)	1.48 (0.82)

The average age of the study sample at Wave 1 (1992) was 40 years old and ranged from 13 to 77 years. At Wave 2 (2005), the average age was 53.6 years and ranged from 26 to 90 years old. The average age at Wave 3 (2015) was 63 years and ranged between 36 and 100 years old. At all three waves, the standard deviation of the respondents’ average age was 13.4. At Wave 1 approximately 65.0% of the respondents were women, and 24.5% were people of color. Average years of education increased over time from 13.4 (SD = 2.4) at Wave 1 to 14.1 (SD = 2.2) at Wave 3. Marital status also changed over time. At Wave 1 63.8% of the sample were married or living with a partner, the percentage increased to 70.3% at Wave 2 and decreased to 64.4% at Wave 3. Finally, friend relations changed slightly from 1992 to 2015. The percentage of friends in the respondents’ networks decreased: at Wave 1 19% of respondents’ networks were comprised of friends; the percentage decreased to 16% at Wave 2 and increased to 17% at Wave 3. Positive friend quality increased over time from 4.78 (SD = .34) at Wave 1 to 4.82 (SD = .35) at Wave 3. Negative friend quality decreased over time from 1.92 (SD = 1.06) at Wave 1 to 1.48 (SD = .82) at Wave 3.

The correlation matrix (Table 2) indicates a statistically significant association between the personal and situational characteristics at Wave 1 and proportion of friends in the network at Wave 1. Further, personal and situational characteristics at Wave 1 were associated with best friend relationship quality across three waves. Significant associations were observed between proportion friends at Wave 2 and self-rated health at Wave 3; and positive quality of best friendship at Wave 1 and 3 with self-rated health at Wave 1 and 3, and depressive symptoms at Wave 3.

Table 2.

Descriptive Statistics and Intercorrelations between Study Variables

Study Variables	1	2	3	4	5	6	7	8	9	10	11	12	13	14
1. Female	--
2. Age (W1)	−0.06	--
3. White	−0.12**	0.10*	--
4. Married/living with partner (W1)	−0.06	0.22***	0.24***	--
5. Education, years (W1)	−0.11**	0.15***	0.15***	0.20***	--
6. Proportion friends (W1)	−0.09*	−0.10*	0.09*	−0.32***	0.09*	--
7. Proportion friends (W2)	0.04	0.02	0.10*	−0.11*	0.14***	0.40***	--
8. Proportion friends (W3)	0.05	0.02	0.12**	−0.10*	0.17***	0.41***	0.59***	--
9. Positive quality (W1)	0.27***	0.03	0.01	0.03	0.00	0.03	0.04	0.01	--
10. Positive quality (W2)	0.31***	−0.10	−0.01	−0.07	−0.07	−0.09	0.04	−0.02	0.25***	--
11. Positive quality (W3)	0.21***	−0.01	−0.04	−0.03	0.01	0.02	0.06	−0.01	0.19**	0.35***	--
12. Negative quality (W1)	0.03	0.16***	0.18***	0.14**	0.17***	−0.02	0.06	0.09	0.24***	0.12*	0.03	--
13. Negative quality (W2)	0.01	−0.01	0.10	0.00	0.01	−0.03	0.03	0.02	0.16**	0.28***	0.17**	0.32***	--
14. Negative quality (W3)	0.10	0.04	0.05	−0.05	0.11*	0.00	0.04	0.04	0.10	0.05	0.30***	0.15*	0.35***	--
15. Self-rated health (W1)	−0.07	−0.10*	0.13**	0.07	0.16***	0.07	0.06	0.07	0.01	0.03	0.03	0.04	0.01	0.0
16. Self-rated health (W3)	−0.03	−0.06	0.07	0.10*	0.17***	−0.03	0.10*	0.05	0.10*	0.08	0.12*	0.06	0.08	0.0
17. CES-D (W1)	0.15***	−0.17***	−0.10*	−0.17***	−0.15***	0.05	−0.02	0.00	−0.07	0.03	−0.03	−0.08	−0.12*	−0.0
18. CES-D (W3)	0.01	0.05	−0.09*	−0.16***	−0.25***	0.00	−0.05	−0.04	−0.05	−0.02	−0.15**	−0.03	−0.04	−0.0

Note:

^*p<0.05

^**p<0.01

^***p<0.001

Proportion and Quality of Friendship Change Over Time

We identified trajectories of friend relations over time. We first report on how the proportion of friends in one’s social network changed over time, followed by a depiction of the extent to which positive and negative aspects of relationship quality with best friend change. Multiple and distinct trajectories emerged.

Proportion of Friends in Network

Results of the Latent Class Growth Analysis (LCGA) for the presence of friends in one’s network showed that model fit improved with each additional class (see Table 3). We selected the three-class solution because the fourth class was comparatively small (8%) and its inclusion had very little impact on the other trajectories. Review of the intercept and slope, shown in Appendix Table 2, was used to identify and name each class. Class 1 - identified as the sample reporting a Higher and Varied Proportion of Friends (25%) - is characterized by both a high intercept and high residual variation (I=42%, SD=25) compared to the other classes. This group had more and greater variability in the reported proportion of friends over the three time points. Class 2 - identified as having a Moderate level and Declining Proportion of Friends - (45%) - is characterized by both a moderate intercept and a declining residual variation (I=17%, SD=15) compared to the other classes. Class 3 - identified as having a Low Level and Stable Proportion of Friends (31%) - is characterized by both a low intercept and a low level of residual variation (I=1%, SD=5) compared to the other classes. Figure 1 displays the mean trajectory as well as 50% and 95% residual variation—that is, the range we can expect to find 50% and 95% of observations around the mean. The substantial variability of Class 1 indicates that members vary substantially across time, while the low variability of Class 3 indicates that members stay close to the mean trajectory. These differences appear to suggest fundamentally different developmental trajectories.

Table 3.

Latent Class Growth Analysis Model Fit

#Classes	AIC	AICC	BIC	aBIC	Entropy
Proportion friends
2	13869	13869	13908	13879	0.91
3	13660	13661	13721	13676	0.77
4	13565	13566	13647	13586	0.80
5	13518	13520	13621	13545	0.82
Positive quality
2	65	65	97	68	0.85
3	52	54	102	57	0.89
4	46	49	113	53	0.92
5	49	54	134	58	0.94
Negative quality
2	1470	1471	1502	1473	0.94
3	1290	1292	1340	1295	0.94
4	1231	1235	1299	1239	0.95
5	1175	1180	1260	1184	0.95

Figure 1.

Latent Class Growth Analysis: Proportion Friends

Positive Friendship Quality

Results of the LCGA for positive friend quality relations suggest an AIC that yielded the four-class solution while the BIC suggest a two-class solution (see Table 3). We selected the two-class solution because the three- and four-class solutions added comparatively small classes (4%–6%). The first class – High and Rising Positive Quality (86%) - is characterized by high positive relationship quality, increasing slightly across time, and low residual variation (I=4.9; S=0.01; SD=0.25) compared to the second class. The second class - Moderately High and Stable Positive Quality (14%) - is characterized by a lower intercept, no change across time, and high residual variation (I=4.5; SD=0.53) compared to the first class. Figure 2 displays each trajectory’s mean across time, indicates residual variation, and suggests somewhat different developmental trajectories of positive friendship quality.

Figure 2.

Latent Class Growth Analysis: Positive Friend Quality

Negative Friendship Quality

Results of the LCGA for negative friend quality relations showed that the model fit improved with the inclusion of four and five classes. The four- and five-class solutions extracted comparatively small classes (6%–7%), so we selected the three-class solution (see Table 3). Class 1 - Low and Stable Negative Quality (38%) - is characterized by a low intercept, no change across time, and low residual variation (I=1.1; SD=0.25) compared to the other classes. Class 2 - Moderate and Declining Negative Quality (17%) - is characterized by a higher intercept, significant linear (negative) and quadratic (positive) slopes, and a low residual variation (I=2.5; S=−0.17; Q=0.005; SD=0.25) compared to other classes. Class 3 - Moderate and Stable Negative Quality (45%) -- is characterized by a similarly high intercept as Class 2, but no change across time, and larger residual variation (I=2.3; SD=1.02). Figure 3 displays each trajectory’s mean across time and residual variation. The combination of linear and quadratic effects for Class 2 is associated with a decline in negative relationship quality from Wave 1 to 2 and a similar level in Wave 3 as Wave 2.

Figure 3.

Latent Class Growth Analysis: Negative Friend Quality

Age Predicting Friendship Trajectories

Table 4 reports the results of the R3STEP models which predicted class membership. Our exploratory hypothesis that age would predict the likelihood of trajectory classification was not supported. Results also showed that gender was not significantly associated with the proportion of friends class trajectories. However, race, education and marital status were. Respondents who were White had lower odds of being in Class 3 (Low and Stable Proportion Friends; b=−1.63, SE = 0.51, p<0.01) compared to Class 1 (Higher and Varied Proportion Friends). Respondents with more years of education had lower odds of being in Class 2 (Moderate and Declining Proportion Friends; b = −0.21, SE = 0.07, p <0.01) and Class 3 (Low and Stable Proportion Friends; b = −0.32, SE = 0.07, p < 0.001) than Class 1 (Higher and Varied Proportion Friends). Respondents who were married or living with a partner had greater odds of being in Class 2 (Moderate and Declining Proportion Friends; b = 1.02, SE = 0.34, p <0.01) or Class 3 (Low and Stable Proportion Friends; b = 2.15, SE = 0.37, p <0.001) compared to Class 1 (Higher and Varied Proportion Friends). In other words, at all ages (and both genders), respondents who were White, highly educated and unmarried were more likely to have a high average proportion of friends in their social networks across time than people of color, less-educated, and married/cohabitating respondents.

Table 4.

Personal Characteristics Predicting Proportion Friends Class Membership

	Proportion friends		Positive quality	Negative quality
Predictor	Class 2	Class 3	Class 2	Class 2	Class 3
Intercept	0.60** (0.18)	0.23 (0.16)	−1.80*** (0.22)	−0.78*** (0.19)	0.17 (0.18)
Age W1	−0.02 (0.12)	0.13 (0.12)	−0.28 (0.19)	−0.28 (0.16)	−0.04 (0.11)
Education W1	−0.21** (0.07)	−0.32*** (0.07)	0.00 (0.11)	−0.03 (0.08)	−0.12 (0.07)
Female	−0.29 (0.34)	−0.28 (0.34)	−1.39* (0.55)	−0.21 (0.4)	−0.22 (0.32)
Married W1	1.02** (0.34)	2.15*** (0.37)	0.14 (0.62)	0.16 (0.4)	0.5 (0.32)
White	−0.92 (0.52)	−1.63** (0.51)	−0.23 (0.69)	−0.3 (0.49)	−0.54 (0.36)

Notes:

^*p < 0.05

^**p < 0.01

^***p < 0.001

Though age was not significantly associated with positive or negative quality friend trajectory classes, there was an effect of gender predicting positive relationship quality. Female respondents had lower odds of being in Class 2 (“High and Stable Positive Quality”; b = −1.39, SE = 0.55, p < 0.05) than class 1 (“High and Rising Positive Quality”); that is, they are less likely to have low positive relationship quality with their best friend over time. These findings clearly indicate that while age was not associated with different friendship developmental trajectories, other personal and situational characteristics were.

Friendship Trajectories and Health

Table 5 reports the results from the models testing whether trajectories of friendship influence Wave 3 self-rated health and depressive symptoms differentially by age. Our hypothesis that both the presence and quality of relationship with best friends would affect self-rated health and depressive symptomology in later adulthood was partially supported.

Table 5.

Friendship Trajectories Predicting Health

	Proportion friends (3-class)		Positive quality (2-class)		Negative quality (3-class)
Predictor	CES-D W3	SRH W3	CES-D W3	SRH W3	CES-D W3	SRH W3
Probability class 1	−0.31 (0.98)	0.08 (0.11)	−4.28* (1.65)	0.63*** (0.18)	0.81 (1.05)	0.19 (0.12)
Probability class 1 × Age W1	0.27 (0.74)	−0.05 (0.08)	2.35 (1.20)	−0.09 (0.13)	−0.60 (0.77)	−0.05 (0.09)
Intercept	23.95*** (1.55)	2.20*** (0.22)	26.43*** (2.20)	1.93*** (0.30)	25.46*** (2.18)	2.11*** (0.30)
Age W1	0.99** (0.34)	−0.06 (0.04)	0.18 (0.57)	0.00 (0.06; 0.00)	1.00** (0.38; 0.02)	−0.05 (0.04)
CES-D/SRH W1	0.25*** (0.04)	0.05** (0.02)	0.19*** (0.05)	0.04 (0.02; 0.01)	0.20*** (0.05)	0.03 (0.03)
Education W1	−0.78*** (0.15)	0.01 (0.08)	−0.68** (0.22)	0.05 (0.12; 0.00)	−0.68** (0.23)	0.12 (0.12)
Female	−0.86 (0.74)	0.33*** (0.05)	−0.93 (1.09)	0.37*** (0.07	−1.44 (1.09)	0.38*** (0.07)
Married W1	−1.93* (0.79)	0.15 (0.09)	−1.35 (1.06)	−0.06 (0.12)	−1.16 (1.08)	−0.04 (0.12)
White	−0.52 (0.84)	0.01 (0.09)	−1.16 (1.23)	0.00 (0.14)	−1.62 (1.25)	−0.02 (0.14)
Adjusted R-squared	0.14	0.11	0.13	0.15	0.11	0.12
$f^2(R_{\mathit{\text{model}}}^2,R_{\mathit{\text{change}}}^2)$	0.00 (0.15, 0.00)	0.00 (0.13, 0.00)	0.04 (0.13, 0.03)	0.04 (0.14, 0.03)	0.01 (0.13, 0.01)	0.01 (0.14, 0.01)

Notes: Cells: Coefficient (SE).

^*p < 0.05

^**p < 0.01

^***p < 0.001

f² calculated by comparing each of the models presented to the same model but excluding Probability of class 1 and its interaction with age.

Results indicated that neither the main effects for class membership trajectory nor the trajectory-by-age interactions were significant for proportion friends and negative relationship quality models. By contrast, the main effect of class 1 membership trajectory for the positive quality trajectories was significant when predicting both CES-D and self-rated health. Specifically, being in Class 1, the High and Rising Positive Quality trajectory, was associated with a 4.28 (SE = 1.65; p < 0.05, f² = 0.04) lower score on the CES-D scale and 0.63 (SE = 0.18; p < 0.001, f² = 0.04) points higher on the five point self-rated health scale. The effect size of being in Class 1 on depressive symptoms and health were small based on Cohen’s convention (Cohen, 1988, p 413–414). There were, however, no age effects. These findings do not support our hypothesis that older adults who report trajectories of high proportions of friends, as well as high positive and low negative relationships quality with best friend trajectories would report fewer depressive symptoms and better self-rated health than younger adults. Instead, it appears that regardless of age, positive friend quality yields health benefits.

DISCUSSION

This study uses lifespan and life course developmental perspectives (Fuller-Iglesias et al., 2010) to examine trajectories of friendship over 23 years among people who were 13 to 77 years of age at Wave 1 and reported a best friend at each wave. Findings identify multiple unique friendship trajectories for the proportion of friends in one’s total network, and the quality, both positive and negative, of friend relations over time. Of special interest is the finding that age is not associated with these trajectories, suggesting that developmental drivers of friend relations involve more than simple age. Finally, and perhaps most importantly, increasing positive friendship trajectories are associated with both depressive symptomatology and self-rated health at every point in the lifespan, while negative trajectories and quantity of friends over time are not. The empirical demonstration of long-term effects of positive quality illustrates the importance of cultivating and maintaining such ties over the life course. In the paragraphs below, and using the Convoy Model framework, we discuss the dynamic but also stable nature of the friendship tie, and its influence on health.

Multiple Friendship Trajectories

The current findings identified friendship trajectories among the same individuals over twenty-three years who at each time point reported having a best friend. We examined a regional sample of those aged 13 and older to ensure the ability to identify developmental elements of friendship across a wide age range. Hence, our sample is not age restricted as is often the case in other studies of friendship (Huxhold, 2019; Pinquart & Sorensen, 2001; Waters et al., 2000; Wrzus et al., 2017). These findings suggest that friendships do fall into identifiable trajectories, supporting the Convoy Model tenet asserting that social relations develop over time, are dynamic, influenced by personal and situational characteristics resulting in some changes but at the same time some stability (Antonucci et al., 2014). Of key significance is that neither proportion of friends in the network nor friend quality changed with age. It is generally understood that friend networks decrease with age (Wruz et al., 2013) and that negative aspects of the friendship tie tend to remain stable over time (Akiyama et al., 2003; Birditt et al., 2009; Bruine de Bruin et al., 2020). Huxhold et al., (2022) suggest here may be historical effects at play, which could explain the discrepancy between our findings and those in the literature. A cohort analysis of convoys showed that the proportion of family was smaller in 2005 compared to 1980 (Antonucci et al., 2019), thereby documenting evidence of change, specifically that family is less likely to comprise ones’ network in later years. Such changes may reflect demographic transitions (e.g., smaller families), and lead to the greater presence of friends across the life course (Fiori et al., 2020). For instance, adults without siblings report less frequent social activities with relatives compared to those who have siblings (Trent & Spitze, 2011), suggesting that friends may increasingly be present in the networks of those with fewer family members. Demographic changes including lower fertility and longer life expectancy may lead to the situation where friends are deemed close and important at every age.

Further, the DIRe-Model (Huxhold et al., 2022) states that in order to understand changes in social relationships a number of age-related developments need to be taken into account simultaneously. For example, the capacities for engaging with friends tend to decline with age, because of decreases in health and mobility. At the same time, social skills, that are helpful for maintaining social relationships may increase with life experience. Overall, the antagonistic effects of both developments on the quality of friendship relationships may cancel each other out. Furthermore, a recent longitudinal study showed friendship quality relationships remains relatively stable across middle adulthood and old age and declines marginally in very old age (Böger & Huxhold, 2018). Though previous research found negative aspects of the friendship tie do not increase, but rather tend to remain stable over time (Akiyama et al., 2003; Birditt et al., 2009; Bruine de Bruin et al., 2020), the dynamics of the friend tie may begin to take on new dimensions given the historical changes noted above, leading to various positive and negative quality trajectories as friends continue to be present, close and important in later life. In sum, friend trajectories did not vary by age, yet changes in the friend tie were uniquely influenced by other key personal and situational characteristics over time as identified by the Convoy Model.

The personal characteristics of gender, socio-economic status, and race as well as the situational characteristic of marital status appear to distinctively guide the form and function of friendship within social networks. Namely, marital status, race and education influenced the extent to which friends are present within the network. In general, the findings suggest that Whites and those with higher levels of education have more and varied friendships as indicated by their higher likelihood of being in Class 1 (Higher and Varied Proportion of Friends), than in either class 2 (Moderate and Declining Proportion of Friends) or 3 (Low and Stable Proportion of Friends). It may be that being part of a privileged group encourages branching out beyond family. For example, being White means being part of the majority, which minimizes the need for a strong racial or ethnic identity and therefore facilitates opportunities to connect with diverse and numerous others (Hedegard, 2018). Experiences of racism, discrimination and unfair treatment motivate the significance of including race as a personal characteristic in the Convoy Model. Such experiences may discourage forging ties beyond family and a very few trusted close friends (Ajrouch et al., 2001). Conversely, it may be that cultural definitions of friendship lead to an undercount of who is labeled a friend among racial-ethnic minorities. Because people of color are more likely to have fictive or chosen kin relationships (Chatters et al., 1994; Taylor et al., 2022), the relationship may not be captured as a friend per se, but instead identified as a family member. Further, staying in school longer provides access to more resources, for example, opportunities to meet new people, make friends and take part in social opportunities (e.g., joining clubs, new classes, attending organized functions) than one may find in the workplace. Pursuit of higher education may encourage moving away from family and home, which in turn encourages developing more friendships. On the other hand, married or partnered people exhibited a slightly different pattern, perhaps reflecting the likelihood that a partner or spouse served as a best friend. For instance, it has been suggested that being married implies friends are less important for preventing loneliness (Pinquart & Sorenson, 2001). Partnered people were more likely to be in Class 2 (Moderate and Declining Proportion of Friends) or Class 3 (Low and Stable Proportion of Friends) than in Class 1 (Higher and Varied Proportion of Friends). One might conclude that among those who are not married, being in Class 1 with High and Varied friendship is perhaps more advantageous. Given that they do not have a spouse or partner available, they may have more time to maintain friendships (Kalmijn, 2003). These findings are especially important when considered with current overall population demographic changes indicating that fewer people are marrying, or choose to marry later, and that more people are experiencing divorce when they do marry. The developmental trajectories of friendship may, therefore, become even more dynamic. The Convoy Model guided the identification of various personal and situational characteristics to examine. In particular, the dynamics of having friends over time is not influenced by age, but instead appears to be more sensitive to the personal characteristics of race and education and the situational characteristic of whether one is married or not.

Contrary to findings with respect to overall friendship network trajectories, there was a gender effect in positive, and not negative, relationship quality. Female respondents were more like to be in Class 1 (High and Rising Positive Quality) than in Class 2 (High and Stable Positive Quality), drawing attention to the intriguing and poorly understood issue of gender differences in social relations. In many ways, these findings are consistent with studies that have used the Convoy Model to examine gender differences in social relations more generally (Ajrouch et al., 2005; Antonucci & Akiyama, 1987; Pearce et al., 2021). The current findings add depth to our understanding of gender and friend relations. Contrary to Field (1999), who found that in old age men displayed more change than women in the friendship tie, we found that transitions in the middle part of life yield more change for women than men. Further, the likelihood of positive relationship quality with friends increasing as one grows older (Blieszner & Roberto, 2004) appears to be particularly true for women. More recently, Ermer and Kanter (2020) reported that women showed more fluctuation in friend support quality than men over time. The complexity of women’s social relations is illustrated in this study, as girls and women appear to be more invested in, and likely more influenced by friend relations starting in early adulthood given a greater tendency to self-disclose and share intimate details of their lives with friends (Youniss & Haynie, 1992; Sherman et al., 2000).

In sum, these findings compel the field to consider developmental elements beyond age, such as race, gender, education and marital status. Examining those factors in tandem provides a more nuanced portrait of change and stability in social relations. Further, these findings suggest that unique situational characteristics arising from socio-historical shifts in population and family demography should be included in the Convoy Model framework (Antonucci et al., 2017; Antonucci et al., 2019). Next, we turn to a consideration of the findings concerning how friendship trajectories are associated with depressive symptomatology and self-rated health.

Friendship Trajectories and Health

The final set of analyses examined whether and how overall friendship network and quality trajectories influence health depending on age. The Convoy Model suggests that multiple dimensions of social relations have unique effects on health (Antonucci et al., 2014). Only positive friend quality was significantly related to health, and, in fact, to both health measures, depressive symptomatology and self-rated health. Importantly, regardless of age, positive relationship quality is associated with health. Neither the percentage of friends in network nor negative quality friendship trajectories have a long-term effect on health outcomes 23 years later. These findings challenge that of Uchino and colleagues (2012), who found that ambivalent friend ties were negatively associated with physical health. This discrepancy may be due to different way of measuring quality, as Uchino and colleagues assessed support quality as a count of positive support with multiple friends whereas in this study, we assessed support quality (e.g., support, share private feelings) of a best friend. Hence, the degree of positivity may vary more given the specific relationship, e.g., best friend, under study. Notably, if negativity were high in the best friend tie, that relationship would likely be terminated. Importantly, our findings suggest that one can recover from deficits or from less-than-optimal trajectories as time and developmental circumstances change.

Given that the present study examined friendship trajectories over an extended time period, one might argue that the findings are both developmentally meaningful and encouraging. It is noteworthy that when all dimensions of social relations are examined simultaneously (structure, positive and negative quality), there is one dimension that emerges as key over time. Findings underscore the Convoy Model assumption that supportive relationships across the life course are important. By highlighting a specific relationship type, that of friend, we explicitly document that the security of positive friend ties over time clearly bestows a positive health benefit. This outcome parallels findings concerning the benefits of secure attachments over time (Cassidy & Shaver, 2002; Heinz et al., 2018; Mikulincer & Shaver, 2007; Waters et al., 2000). Future research may consider identifying mechanisms through which positive friendship quality effects health, such as emotions (Boger & Huxhold, 2018; Uchino & Rook, 2020), lower stress and higher efficacy (Antonucci & Jackson 1987; Huxhold et al., 2022).

The Convoy Model identifies relationship type and various dimensions of social relations as a key avenue for better elucidating the links between social relations and health. These findings show that for the friend tie, not all dimensions of social relations are equal in their health promoting effects. Importantly it conveys that there are no long-term disadvantages to the negative aspects of friendship, but there are cumulative health benefits from the positive.

Future Directions

While the findings reported here offer unique insights into friendships over time, several limitations should be recognized. Concerning both positive and negative relationship quality, participants were asked only about one best friend at each time point and the best friend was not necessarily the same person over time. Inclusion of a broader array of friendships would provide more complete information about the form and function of friendships. In addition, the available data about those friendships were limited. Specifically, we investigated the presence of friends over time by measuring the proportion of friends in one’s convoy. This, we felt, was a reasonable choice, as having five friends among twenty family members is quite different from having five friends in a convoy of six. Yet, there is no denying that the absolute number of friends may be important. Further, our findings suggest that friends may play a different role when the respondent is married versus not-married. A more in-depth longitudinal examination of friendships among those who are married and unmarried would be useful (see Antonucci et al., 2001; Birditt & Antonucci, 2007; Ermer & Kanter, 2020). Finally, while we are able to examine friendships over an extended period of time, it is a limitation that we are limited to three data points, i.e., 1992; 2005 and 2015. More frequent assessments might provide greater specificity concerning change and stability over time in this relationship.

Summary and Conclusions

This study provides support for the Convoy Model by demonstrating that personal and situational characteristics influence friendship trajectories as well as the ways in which earlier experiences of friendship relations influence later life health outcomes at different starting points across the life course. Our findings offer a unique contribution to the developmental literature on friendships across the lifespan. Three distinct friendship patterns were identified as were different positive and negative friendship trajectories. It is noteworthy that only positive friend relationship quality had a long-term, 23-year association with health, and this effect was evident across age. Developmentally, it is striking how far reaching the effects of positive friendship quality are. Also remarkable is the fact that this is the only long-term effect, i.e., neither proportion of friend patterns nor negative relationship quality had the same far-reaching effects. These findings have both theoretical and practical importance. Theoretically, these findings lend support for the Convoy Model by demonstrating that a convoy of positive friendships yields long-term health benefits. Practically, the findings highlight the remarkable long-term effects of positive relationship quality on health. Both contribute to our understanding of social relations and their long term effects, potentially paving the way for intervention programs designed to maximize health and well-being across the lifespan and over the life course.

Public Health Significance.

This study suggests that there are distinct trajectories of friendship beginning in young and mid-adulthood indicated by proportion of friends in one’s network and relationship quality. Identifying how personal and situational characteristics influence multiple dimensions of the friend relationship over time provides key information to guide social support-based interventions aimed at enriching developmental outcomes and maximizing well-being across the lifespan and life course.

Appendix Table 1.

Sample Attrition

	<3 waves	<3 identify best friends	3 best friends
	M(SD)	M(SD)	M(SD)
% Female	55	60	70
% White	68	73	79
Age, W1	54 (20)	42 (13)	39 (14)
Years of education, W1	12.04 (2.81)	13.42 (2.43)	13.4 (2.25)
% Married, W1	46	68	59
CES-D, W1	31.36 (9.86)	30.19 (9.08)	29.53 (9.6)
Self-rated health, W1	3.67 (1.05)	4.15 (0.86)	4.13 (0.83)
Sample size	945	297	256

Appendix Table 2.

Latent Class Growth Analysis Trajectories

Class	Class probability	Intercept	Linear slope	Quadratic slope	SD
Proportion friends
1	25%	42.41*** (3.22)	−0.77 (0.5)	0.03 (0.02)	25.42*** (7.05)
2	45%	17.47*** (1.55)	−0.48 (0.25)	0.01 (0.01)	14.97*** (4.54)
3	31%	1.12*** (0.27)	−0.05 (0.06)	0.00 (0.00)	5.001 (0.00)
Positive quality
1	86%	4.88*** (0.02)	0.01** (0.00)	0.00 (0.00)	0.251 (0.00)
2	14%	4.47*** (0.08)	0.00 (0.02)	0.00 (0.00)	0.53 (0.16)
Negative quality
1	38%	1.06*** (0.02)	0.00 (0.00)	0.00 (0.00)	0.251 (0.00)
2	17%	2.47*** (0.05)	−0.17*** (0.01)	0.005*** (0.000)	0.251 (0.00)
3	45%	2.27*** (0.11)	0.00 (0.02)	0.00 (0.00)	1.02*** (0.25)

Notes:

^*p<0.05

^**p<0.01

^***p<0.001.

¹Standard deviation at boundary constraint