Perceptions of Relationship Quality that Predict Friendship Dissolution During Childhood and Adolescence: Social Support Matters More than Negativity

Abstract

The present study examines perceptions of relationship quality as antecedents of best friendship dissolution. Participants included 230 students in Florida (USA) (54.3% girls; ages 8–13; 39.6% European American, 27.0% Hispanic American, 21.7% African American, 2.6% Asian American) and 496 students in Lithuania (49.0% girls; ages 8–14; 358 boys; nearly all ethnic Lithuanian) attending public schools in Florida (USA) and Lithuania. Reciprocated best friends were identified from friend nominations midway through the school year, at which time each partner described perceptions of social support and negativity in the relationship. Of the 363 reciprocated friend dyads, 26.2% (n=95) were no longer friends 1–3 months later. Dyadic analyses were conducted with initial perceived friend negativity and initial perceived friend social support as predictors of subsequent friendship stability. When considered separately, perceived social support predicted self- and partner reports of friendship stability and perceived negativity predicted self- (but not partner) reports of friendship stability. In each case, lower initial social support and higher initial negativity predicted a greater incidence of friendship dissolution. When considered together, initial perceptions of social support predicted self- and partner reports of friendship dissolution, but perceptions of negativity predicted neither. The findings are consistent with those from married couples, which indicate that the absence of positive interactions is a stronger predictor of relationship instability than is the presence of negative interactions.

Do children’s friendships end with a bang or a whimper? We don’t really know. Not surprisingly, poor quality friendships are less stable than those described as good quality (Hiatt et al., 2015), but quality is a broad description of functioning that encompasses positive and negative dimensions. Global measures of quality cannot determine whether relationships end because the benefits of participation are low or because the costs are high (or both). Meta-analytic findings indicate that positive dimensions are stronger predictors of adult romantic relationship dissolution than are negative ones (Le et al., 2010). However, the extent to which such findings generalize to the friendships of children is not clear because the participants and their relationships differ on several important dimension, including available alternatives, exclusivity, and types of resources exchanged (Delay et al., 2016). The present study examines perceptions of social support and negativity as predictors of best friendship stability over the course of a single semester in a diverse sample of 8–14 year-olds from the USA and Lithuania.

Friend loss is a common occurrence. By some estimates, up to 50% of friendships dissolve over the course of a semester (Meter & Card, 2016). A narrative literature review concluded that rates of friendship dissolution decline across childhood, with a modest uptick during the mid-adolescent years, followed by further decreases (Poulin & Chan, 2010). Best friendships are less likely to dissolve than secondary friendships, regardless of age (Chan & Poulin, 2009). The frequency of friend turnover belies its potentially disruptive nature. Friend loss is benign in most cases, but some children who lose friends (particularly those who become friendless) experience a significant uptick in depressive symptoms (Bukowski et al., 2010).

Both low social support and high negativity have been invoked as antecedents of friendship instability. Social support encompasses positive relationship features such as intimacy, instrumental assistance, and companionship. Negativity encompasses conflict, criticism, and negative affect. The friendship process model (Fehr, 1996) holds that perceived social support is critical to friendship maintenance because support enhances intimacy, which increases relationship satisfaction (Oswald, 2017). In contrast, negativity signals heightened and unresolved conflict indicative of relationship dissatisfaction. There is also merit to considering the cascade model of romantic relationship dissolution (Gottman & Levenson, 2000), which holds that dissolution follows from a persistent unfavorable ratio of positive to negative interactions, which fosters alienation and reduces investment in the relationship. It is worth noting that conceptualizations of friendship change across the transition into adolescence, with an increasing emphasis on loyalty, intimacy, and self-disclosure; some speculate that friendships based on these features should be more stable than those based on features like companionship that tend to define the friendships of children (Laursen & Hartup, 2002).

Few studies have examined relationship quality as a predictor of children’s friendship stability, and those that have tended to combine positive and negative relationship features in ways that suggest both are prerequisites for dissolution. In one instance, high quality (i.e., both partners report high social support and low negativity) adolescent friendships were more stable from one year to the next than low quality (i.e., both partners report low social support and high negativity) friendships (Hiatt et al., 2015). To get a better sense of the relative importance of negativity and social support, we turn to the adult literature. College students retrospectively attributed friendship dissolution to a lack of companionship and affection, not conflict or jealousy (Johnson et al., 2004). The risk of marital dissolution was greater among couples with a relative paucity of positive interactions, both in absolute terms and in proportion to the rate with which negative interactions occur (Gottman & Levenson, 2000).

The present study compares perceptions of social support and negativity as predictors of best friendship stability during the late primary and early middle school years. To determine the degree to which individual perceptions of quality shape self and partner reports of friendship stability, social support and negativity will be examined as separate predictors of dissolution across a 1–3 month period during a single academic semester. The prospect of amplification (i.e., dissolution risk increases only when both high negativity and low social support are present) will also be considered. Replication is a strength of the study, with cross-national contrasts of participants drawn from public schools in Florida (USA) and Lithuania. Comparisons of primary and middle school students will test competing hypotheses that suggest, on the one hand, heightened instability during early adolescent transitions, and, on the other hand, heightened stability arising from more mature conceptualizations of friendship.

Method

Participants

USA.

Participants included 122 (65 girls, 57 boys) primary school (4^th-5^th grade; range=8–10 years old) students and 108 (60 girls, 48 boys) middle school (6^th-7^th grade; range=11–13 years old) students attending a public school required to represent Florida public school students in terms of ethnicity and family income. Primary school students attended classes with the same peers and teacher throughout the day. Middle school students attended classes with different teachers and a varying subset of peers drawn from the same pod (i.e., four homeroom classes). School records indicated that 39.6% were European-American, 27.0% were Hispanic-American, 21.7% were African-American, 2.6% were Asian-American, and 9.1% were mixed-race or another race.

Lithuania.

Participants included 120 (60 girls, 60 boys) primary school (4^th grade; range=9–10 years old) students and 376 (183 girls, 193 boys) middle school (5^th-7^th grade; range=10–14 years old) students in public schools in Lithuania. Students attended all classes with the same peers. Teachers rotated between classes in middle school. Nearly all students were ethnic Lithuanian.

Intraclass correlations (ρ) examined the proportion of variance in predictor and outcome variables attributable to grouping structure (i.e., classrooms). Low within-classroom ICCs (ρ=0.07–0.08) revealed minimal variation between classrooms (Hox, 1998), suggesting that nestedness did not bias results.

Procedure

Written parent consent and child assent were required for participation. Trained research assistants administered surveys to students on computer tablets in a quiet school setting. The study was not preregistered. Study materials are not publicly available, but can be obtained on reasonable request.

USA.

A total of 653 students from 28 classes in two schools were invited to participate (M=70.2% classroom participation, SD=13.0; range=48–96%). The same surveys were administered twice during an academic year: (1) January 2022 and (2) February-March 2022. The study was approved by the university Institutional Review Board (USA #135501-16).

Lithuania.

A total of 1154 students from 45 classes in 7 schools were invited to participate (M=63.1% classroom participation, SD=12.1; range=40–87%). The same surveys were administered twice during an academic year: (1) February 2022 and (2) May 2022. Measures were translated from English to Lithuanian by a bilingual team of research assistants, then back-translated by a separate team. Differences were resolved by discussion. The study was approved by the university ethics committee (Lithuania #6/−2020).

Measures

Friendship.

At both time points, participants identified, on a class/pod roster, up to two rank-ordered best friends (i.e., “Who is your best friend? and Who is your next best friend?”) and up to three (Lithuania) or five (USA) additional rank-ordered friends (i.e., Who are your other friends?”). Primary school rosters contained an average of 15 students in USA (SD=2.80, range=11–20) and 23 students in Lithuania (SD=1.56, range=19–25); middle school rosters contained an average of 61 students in USA (SD=5.06, range=55–67) and 26 students in Lithuania (SD=3.25, range=14–30). Participants could also identify, rank, and describe friends not on the roster (i.e., My friend is not listed”); top-ranked friendships with nonparticipants were not included in the present study, necessitating the use of second-ranked friendships (USA n=17, Lithuania n=4). Reciprocated friends were dyads in which both partners nominated one another as friends. Reciprocated best friends were dyads in which at least one partner in a reciprocated friend dyad nominated the other as a best friend. Best friendship dissolution occurred when one or both members of a Time 1 reciprocated best friend dyad failed to nominate the other as a best friend or friend at Time 2.

Friendship Quality.

At Time 1, participants completed a short version of the Network of Relationships Inventory (Furman & Buhrmester, 1985), describing their first and second-ranked friendships. Items were rated on a scale ranging from 1 (never true) to 5 (always true). Social support (M=4.09, SD=0.08) was measured with 5 items (e.g., “NAME OF FRIEND and I really care about each other”). Negativity (M=2.10, SD=0.82) was measured with 4 items (e.g., “NAME OF FRIEND and I argue with each other”). Internal reliability was acceptable (social support α=.83-.85, negativity α=.80-.83).

Potential Confounding Variables

To isolate the contributions of relationship quality to friendship stability, friendship rank (reports of relationship order) and friendship duration (reports of friendship initiation) were included as Time 1 covariates.

Plan of Analysis

A total of 929 participants (Florida n=327, Lithuania n=602) were involved in at least one reciprocated friendship at Time 1, resulting in 1,177 reciprocated friend dyads (Florida n=545, Lithuania n=632). Friendship quality data were only available for 2 top-ranked friends. As a consequence, 323 (Florida n=228, Lithuania n=95) friend dyads were omitted because neither partner nominated the other as a best friend at Time 1. Participation at both time points was a precondition for inclusion. As a result, 32 (Florida n=27, Lithuania n=5) friend dyads were omitted because one or both partners were absent at Time 2. To prevent unequal individual contributions to the data, participants were restricted to a single Time 1 reciprocated best friendship. Highest ranked friendships were prioritized. As a result, 157 students (Florida n=58, Lithuania n=99) were omitted because their best friends participated in other reciprocated best friend dyads. The final sample included 726 students who were involved in 363 reciprocated best friend dyads (Florida n=57 female, n=47 male, n=11 other; Lithuania n=117 female, n=122 male, n=9 other). Supplemental Table 1 describes best friendship rankings at each time point.

Actor Partner Interdependence Models (APIM; Kenny et al., 2006) for indistinguishable dyads were conducted, in a structural equation modeling framework using Mplus v8.3 (Muthén & Muthén, 2017). The APIM partitions the variance shared across partners on the same variables from variance that describes associations within and between partners. Within-dyad equality constraints, reflecting the interchangeable nature of partners, were imposed on within-individual dissolution paths, between-individual dissolution paths, within-individual correlations, and between-individual correlations, as well as on the means, intercepts, variances, and residual variances of friendship quality scores. Model fit was assessed following recommendations made by Olsen and Kenny (2006). A nonsignificant χ² indicates that the model fits the data; the RMSEA should be <.08; and the TLI should exceed .95 (Hu & Bentler, 1999).

Three sets of analyses were conducted. The first consisted of two path analyses, conducted separately for social support and negativity (see Supplemental Figure 1). Actor paths describe the association from one partner’s reports of relationship quality (i.e., social support or negativity) at Time 1 to the same partner’s report of friendship stability (presence or absence of best friend or friend nomination) at Time 2. Partner paths describe the association from one partner’s reports of relationship quality at Time 1 to the other partner’s report of friendship stability at Time 2.

The second consisted of a single path analysis that included both social support and negativity at Time 1 as predictors of self- and partner reports of friendship dissolution at Time 2 (see Supplemental Figure 2). Wald tests compared social support→friendship dissolution paths to negativity→friendship dissolution paths to identify differences in their strength of association.

The third included interaction terms (social support × negativity) to test for amplification effects (see Supplemental Figure S3). Actor-actor interaction effects (i.e., each partner’s perception of social support crossed with the same partner’s perception of negativity predicting self-reports of friendship dissolution) and partner-partner interaction effects (i.e., each partner’s perception of social support crossed with the same partner’s perception of negativity predicting partner reports of friendship dissolution) were added to the model (Garcia et al., 2015).

Finally, four sets of supplemental analyses were conducted. First, multiple group contrasts examined differences in patterns of associations as a function of age (primary school and middle school) and location (USA and Lithuania). Wald tests of parameter constraints compared groups on actor and partner paths. Second, all analyses were rerun to include each partner’s reports of friendship duration and friendship rank as separate Time 1 covariates. Third, analyses were rerun with the 439 dyads excluded from the primary analyses because one or both partners participated in other reciprocated friendships. Fourth, multiple group contrasts examined differences in patterns of association between top ranked best friendships and lower ranked best friendships.

Power.

There was adequate power for the main analyses. For actor and partner effects, the APIMPowerR calculator (Ackerman & Kenny, 2016) indicated that a minimum of 360 dyads were required to detect small (B=.10) actor and partner effects. For multiple group contrasts, there was adequate power to detect large (e.g., B₁=.05, B₂=.55; Δ ≥ .50) differences between groups on direct effects paths, but not small (e.g., B₁=.10, B₂=.35; Δ = .10 to .23) or medium (e.g., B₁=.10, B₂=.40; Δ = .24 to .30) differences.

Missing Data.

Little’s MCAR test indicated that friendship quality data were missing completely at random, χ²(47)=31.73, p=.957. Missing item-level data (social support=5.2%; negativity=2.7%) were imputed with an EM algorithm with 25 iterations. Participants completed friendship quality data for two top-ranked best friends only. A total of 111 (15.3%) students had wave-level missing data on perceived social support and negativity because their partners were 3^rd (n=44), 4^th (n=31), 5^th (n=28), or 6^th-7^th (USA only n=8) ranked friends. Full information maximum-likelihood estimation (FIML) was applied to wave-level missing data.

Results

Preliminary Analyses

Concurrent intraclass correlations (ICC) established nonindependence between friends on perceived social support and perceived negativity, a precondition for dyadic analyses. There were statistically significant (p<.001) within-dyad ICC (interpreted as r²) for social support (ICC=.28) and negativity (ICC=.26). Concurrent interclass (within individual) correlations, conducted with one randomly selected member of each dyad to avoid bias arising from nonindependent reports (i.e., two friends describing the same relationships), indicated that negativity was inversely associated with social support (r=−.22).

Separate 2 (location) × 2 (age) ANOVAs were conducted with one randomly selected member from each friend dyad. Perceived social support and perceived negativity were the dependent variables. There were no statistically significant main effects for social support. For negativity, there was a statistically significant main effect for age, F(1, 574)=8.50, p=.004. Negativity (d=0.26) was higher in middle school than in primary school (d=0.34).

Of the initial 363 Time 1 reciprocated best friend dyads, 268 (73.8%) remained reciprocated friends 1–3 months later at Time 2. Separate 2 × 2 chi-square contrasts indicated that rates of best friendship dissolution did not differ as a function of location or age χ²(1)=0.03–1.71, p=.19-.87.

Perceived Social Support and Perceived Negativity as Predictors of Friendship Dissolution: Results from Separate Relationship Quality Models

Two separate models were conducted with either Time 1 perceived social support or Time 1 perceived negativity as predictors of Time 2 best friendship dissolution. Figure 1 depicts the results. Both models fit the data: χ²(2)=0.37–0.43, p=.81-.83, RMSEA=.00, CFI>.999, TLI>.999.

Figure 1. Perceived Social Support and Perceived Negativity as Predictors of Adolescent Best Friendship Dissolution: Results from Separate Friendship Quality Indistinguishable Dyad APIM Analyses.

Note. N=363 dyads. Standardized beta weights reported. Results for perceived social support are presented on the left of the slash; results for perceived negativity are presented on the right. Time 1=January in Florida, February in Lithuania; Time 2=February/March in Florida, May in Lithuania. *p<.05, **p<.01.

Perceived social support.

Statistically significant actor and partner effects emerged such that lower initial perceptions of social support were associated with higher rates of subsequent self- and partner reported best friendship dissolution.

Perceived negativity.

Statistically significant actor effects emerged such that greater initial perceptions of negativity were associated with higher rates of subsequent self- (but not partner) reported best friendship dissolution.

Follow-up multiple group contrasts.

There were no statistically significant differences on actor or partner effects between (a) primary and middle school dyads, and (b) USA and Lithuania dyads (social support Wald=1.00–3.58, p=.06-.32; negativity Wald=0.25–1.30, p=.26-.62).

Perceived Social Support and Perceived Negativity as Predictors of Friendship Dissolution: Results from the Combined Relationship Quality Model

Figure 2 presents results from analyses that examined longitudinal associations from Time 1 perceived social support and Time 1 perceived negativity to Time 2 best friendship dissolution. The model fit the data, χ²(2)=1.35, p=.51, RMSEA=.00, CFI>.999, TLI>.999.

Figure 2. Perceived Social Support and Perceived Negativity as Predictors of Adolescent Best Friendship Dissolution: Results from the Combined Friendship Quality Indistinguishable Dyad APIM Analysis.

Note. N=363 dyads. Standardized beta weights reported. Time 1= January in Florida, February in Lithuania; Time 2=February/March in Florida, May in Lithuania. *p<.05, **p<.01.

For perceived social support, statistically significant actor and partner effects emerged, such that lower initial perceptions of social support were associated with higher rates of subsequent self- and partner reported best friendship dissolution. For perceived negativity, neither actor nor partner effects were statistically significant. Comparisons indicated that social support was a stronger predictor of best friendship dissolution than negativity (actor Wald=20.31, p<.001; partner Wald=4.73, p=.03).

Neither the actor-actor interaction effect (B=−.04, p=.49) nor the partner-partner interaction effect (B=−.04, p=.46) were statistically significant, indicating that there were no amplification effects.

Follow-up multiple group contrasts.

There were no statistically significant differences on actor or partner effects between (a) primary and middle school dyads, and (b) USA and Lithuania dyads (social support Wald=1.22–3.46, p=.06-.27; negativity Wald=0.44–1.06, p=.30-.51).

Supplemental Analyses.

The same pattern of statistically significant results emerged when analyses were rerun with potential confounding variables (i.e., friendship length and friendship duration) as Time 1 covariates (see Supplemental Table S2). The same pattern of statistically significant results emerged when analyses were rerun to include dyads omitted from the primary analyses because one or both partners were in other friendships; multiple group contrasts failed to reveal differences between groups on actor or partner effects (see Supplemental Table S3). Multiple group contrasts failed to reveal differences between top ranked and lower best friendships on actor or partner effects (see Supplemental Table S4).

Discussion

Low social support was a better predictor of best friendship dissolution than high negativity. Indeed, social support was inversely related to both self and partner reports of friend stability, suggesting that one friend responded to the other friend’s lack of enthusiasm about the affiliation. Perceived negativity, in contrast, predicted only self-reports of friend stability, an association that disappeared after accounting for variance shared with perceived social support. To answer the question posed at the outset, children’s friendships appear to end with a whimper (i.e., declining positive engagement), not a bang (i.e., escalating conflict).

As friendships ascend the ladder of important interpersonal relationships during childhood and early adolescence, they are increasingly characterized by positive features: Helping, sharing, loyalty, and intimacy (McDougall & Hymel, 2007). Close friends depend on one another for exclusive interpersonal rewards (Laursen & Hartup, 2002). Should rewards dissipate or should others become more reliable providers, the need to maintain the investment declines. Friends proffer support in ways that other relationships cannot. These benefits are so important that their absence imperils the affiliation. College students cite decreased companionship and positive affect as the most common reasons for dissolving a friendship (Johnson et al., 2004).The absence of positive interactions is also a strong indicator of divorce (Gottman & Levenson, 2000).

The finding that negativity does not uniquely predict friendship dissolution may strike some as counterintuitive. But it is worth noting that conflict between friends is commonplace; disagreements with friends are more frequent than with other peers (Laursen & Adams, 2018). What matters (in terms of relationship impact) is how disagreements are resolved (Laursen & Hafen, 2010). Friends are noteworthy for the ability to settle differences without damaging affiliations or interrupting ongoing interactions. Disruptive, coercive conflicts signal a lack of relationship closeness (Shulman & Laursen, 2002). Most measures of relationship negativity overlook the nuances of conflict resolution. Successful conflict mitigation is perhaps better captured by supportive provisions of friendship, such as communication and respect, as opposed to low levels of conflict and criticism. It follows that parents, teachers, and caregivers concerned about the social development of children should focus not on the presence of disputes between friends but on the manner in which they are resolved, for these behaviors may (paradoxically) offer clues to positive relationship features and the long-term stability of the affiliation.

There are limitations that warrant mention. First, data were collected twice, 1–3 months apart, during the middle of the academic year. It is premature to assume that the same relationship quality antecedents apply to new friendships at the start of the school year or to well-established friendships that spans many months or years. Second, we focused on friends in the same class. We can only assume that similar processes are at play in cross-classroom and out-of-school friendships, as well as between those without consent to participate. Third, our conclusions are limited to first and second best friends; lesser friends may be more sensitive to perceived negativity. Finally, our sample was underpowered for tests of age and geographic differences; conclusions about their absence should be made with caution.

In conclusion, social support is the foundation of best friendship stability. Perceived support predicts not only self-perceptions of relationship continuity, but also partner perceptions. Perceptions of negativity did not meaningfully add to the prediction of best friendship stability, independent of the variance they share with perceptions of social support. Of course, disagreements are detrimental to some friendships, but we suspect that few dyads whose interactions are plagued by toxic negativity become best friends. Even those whose friendships end with a bang may not have suffered from chronic negativity, but may instead have found that a dearth of positive experiences spawned a relationship ending dispute.

Public Significance Statement.

What is the public health significance of this article?—

The detrimental consequences of friend loss are well documented, but the characteristics that forecast it are not clear. Results from this study indicate that the absence of social support is more of a threat to friendship continuity than the presence of negativity.