Lonely Me, Lonely You: Loneliness and the Longitudinal Course of Relationship Satisfaction

Loneliness was assessed using a single item (“I feel lonely”) answered on a Likert-type rating scale ranging from 1 (not at all) to 5 (absolutely). Single items are among the most frequently used measures of loneliness (Mund et al. 2020; Pinquart and Sörensen 2001). Nevertheless, several authors have raised severe criticism against single-item measures in general (Diamantopoulos et al. 2012) and regarding loneliness in particular (Marangoni and Ickes 1989; Daniel Russell 1982). However, there is much evidence in favor of the appropriateness of single-item measures of loneliness.

First, single-item measures such as the one used here are highly correlated with multi-item questionnaires of loneliness such as the 20-item UCLA Loneliness Scale (Dan Russell et al. 1978) and its shorter forms (Hughes et al. 2004; Iecovich 2013). Second, the results of studies employing single-item measures of loneliness converge very well with studies using multiple items. For instance, the results of two studies on loneliness and mortality (e.g., Patterson and Veenstra 2010; Shiovitz-Ezra and Ayalon 2010) that used even different single-item measures converge very well with each other and with the results of a comprehensive meta-analysis on this topic incorporating a wide variety of loneliness measures (Holt-Lunstad et al. 2015). Likewise, a recent meta-analysis on the stability and change of loneliness found no differences between single- and multi-item measures of loneliness with respect to patterns of mean-level change and rank-order stability (Mund et al. 2020). In a similar vein, a recent study on the sources of inter-individual differences in loneliness also found strong convergence between three different versions of single-item measures employed in nationally representative samples from Germany, Switzerland, and Australia (Mund et al. 2019). Finally, Mund and Neyer (2019) found considerable convergence between a single-item measure of loneliness and an 11-item version of the de Jong Gierveld Loneliness Scale in predicting levels of the Big Five across a time period of 4 years. Third, there is substantial overlap between the nomological networks of single- and multi-item measures of loneliness. For example, von Soest et al. (2020) reported virtually identical correlations with demographic and health-related variables for a single-item measure of loneliness and a 3-item version of the de Jong Gierveld Loneliness Scale. Taken together, we conclude that single-item measures of loneliness demonstrate a considerable amount of convergent validity with multi-item measures. Given all these findings, we conclude that single-item measures of loneliness are valid and appropriate.

In pairfam, loneliness was assessed at measurement waves 1, 4, 5, 7, and 8; for the present analyses, we only used data on loneliness collected at W1 and W8. These two occasions were chosen because we wanted to examine the effects of changes in loneliness on changes in relationship satisfaction. In the analytical framework that we will describe in more detail below, it is only possible to incorporate changes in loneliness as residualized change between two measurement occasions.

At each measurement occasion, relationship satisfaction was assessed using a single item from the Relationship Assessment Scale (Hendrick et al. 1998): “Overall, how satisfied are you with your relationship?” Respondents answered on a Likert-type rating scale ranging from 0 (very dissatisfied) to 10 (very satisfied). This item closely corresponds to other commonly-used single item measures of relationship satisfaction in couple research (e.g., Johnson et al. 2018; Rhoades et al. 2012). As relationship satisfaction was assessed at every wave, we used all available data points to model change in relationship satisfaction over time.

Age and relationship duration (in years) were included as demographic covariates in the analyses. Shyness and depressiveness were included as additional covariates because they most closely resemble neuroticism and extraversion. Neuroticism and extraversion could not be included as covariates because they were not measured in pairfam along with loneliness. However, while depressiveness can be considered a facet of neuroticism (McCrae and Costa 2008), shyness entails aspects of both neuroticism and extraversion (Asendorpf 1990).

Shyness was measured using three items (“I feel inhibited in the presence of others”; “I feel shy in the presence of others”; “I approach other people easily”, reverse coded) answered on Likert-type rating scale ranging from 1 (not at all) to 5 (absolutely). Coefficient \(\omega\), which we used as an indicator of scale reliability (McNeish 2018), amounted to .747 for women (95% confidence interval ranging from .726 to .769) and .733 for men (95% confidence interval ranging from .710 to .756), respectively.

Depressiveness was assessed using a single item (“How often did you feel down and melancholy in the past four weeks?”) answered on a Likert-type rating scale ranging from 1 (never) to 5 (always).¹ Supplemental Tables S2 and S4 show the results for all models without covariates.

We used Actor–Partner Interdependence Models (APIM; Kenny et al. 2006) as depicted in Fig. 1 as the basic tool to examine our research questions. With APIM, it is possible to disentangle actor and partner effects. In the case at hand, actor effects denote to what extent an individual’s loneliness at the first measurement occasion is associated with this individual’s own relationship satisfaction at W8. Partner effects denote to what extent an individual’s loneliness at W1 is capable of predicting his or her partner’s relationship satisfaction at W8 (Kenny et al. 2006).

With the APIM, we examined whether loneliness at W1 is associated with relationship satisfaction at W8 as well as with the development of relationship satisfaction across the 8 years of the study period. For the latter question, the APIM was adapted as shown in Fig. 2. To examine the development of relationship satisfaction across the 8 years of the study period, we employed latent basis growth models (Ram and Grimm 2007). Specifically, we fixed the first slope loading to 0 and the last loading to 1. In this way, the mean of the slope reflects the overall amount of change that has occurred from W1 through W8.

The effects of dyadic similarity in loneliness on later levels and the development of relationship satisfaction were examined using dyadic response surface analysis (DRSA; Schönbrodt et al. 2018; Weidmann et al. 2017), which is a combination of APIM and response surface analysis (see Fig. 3). DRSA extends the canonical APIM by adding higher-order polynomials of the predictor variables to the model. Specifically, in addition to the actor and partner effects included in the APIM, DRSA adds squared actor and partner effects as well as an interaction term between the actor’s and the partner’s levels on the predictor variable (see Fig. 3 and Supplemental Figures S1 and S2; statistical details are provided in the Online Supplement).

The interpretation of results obtained with DRSA is primarily based on response surface parameters, which are derived from the coefficients estimated in the extended APIM (see Fig. 3; Schönbrodt et al. 2018, for more information, see the Online Supplement at https://​osf.​io/​tjuay/​). The response surface parameters are visualized with a response surface plot. There are four response surface parameters for women (\(a_{1f}\), \(a_{2f}\), \(a_{3f}\), \(a_{4f}\)) and men (\(a_{1m}\), \(a_{2m}\), \(a_{3m}\), \(a_{4m}\)), respectively. These parameters characterize two essential features of DRSA: First, the line of congruence (LOC), where values of X (e.g., woman’s loneliness) are identical to Y (e.g., man’s loneliness). As such, the LOC illustrates the effects of congruence (i.e., common scores) on the outcome variable (e.g., relationship satisfaction). The response surface parameter \(a_1\) describes whether the LOC is linear or flat; \(a_2\) contains information on the curvature of the LOC. For example, a significantly positive \(a_2\)-parameter would indicate that the LOC is u-shaped and, thus, that congruence on either high or low levels of the predictor (e.g., loneliness) is particularly beneficial for the outcome (e.g., relationship satisfaction). Second, the line of incongruence (LOIC) illustrates the effects of both partners being incongruent on the predictor variable (i.e., \(X = -Y\); for example, woman scoring high on loneliness, man scoring low). The parameters \(a_3\) and \(a_4\) describe linear and non-linear trends for the LOIC, respectively (Schönbrodt et al. 2018).

All APIM and DRSA were examined in two steps. In the first step, we set up an unconstrained model in which all parameters were estimated freely. In the constrained model, we forced all actor and partner effects to be equal for women and men. Subsequently, we evaluated whether the constrained model fit the data worse than the unconstrained model via a \(\chi ^2\)-difference test. A nonsignificant \(\chi ^2\)-difference test indicates that the regression coefficients can be considered invariant across women and men and that there are no sex differences in the association between loneliness and relationship satisfaction.

All models were analyzed with the lavaan package (Rosseel 2012) for R (R Core Team 2018). The response surfaces were plotted using the R package RSA (Schönbrodt 2017). APIM and DRSA were run using robust maximum likelihood estimators (MLR); missing data were treated with Full Information Maximum Likelihood (Enders 2010). The models were considered to show a good (reasonable) fit to the data when the ratio of \(\chi ^2\)-to-df was smaller than 2 (3), the Comparative Fit Index (CFI) exceeded 0.97 (0.95), and both the Root Mean Square Error of Approximation (RMSEA) and the Standardized Root Mean Square Residual (SRMR) fell below 0.05 (0.08; Schermelleh-Engel et al. 2003) All analysis scripts and the complete model outputs are available at https://​osf.​io/​tjuay/​.

To determine whether loneliness predicts later levels of relationship satisfaction as well as its development over time, we used the APIM as depicted in Fig. 1. As recommended by Kenny et al. (2006), we first ran a model with no equality constraints on either the actor or the partner effects. We then compared this model to a model with actor and partner effects constrained to be equal. This constrained model did not fit worse than the unconstrained model (\(\Delta \chi ^2 = 0.120\), \(\Delta df = 2\); \(p = .942\)), indicating that there are no differences between women and men with regard to the influence of loneliness on later levels of relationship satisfaction. Furthermore, the constrained model had an excellent fit to the data, as indicated by a CFI of 0.988, an RMSEA of 0.022 (90% CI 0.009–0.034), an SRMR of 0.015, and a \(\chi ^2\)/df ratio of 2.09.

The results, which are displayed in the upper part of Table 2, indicate that individuals who felt lonelier than others at the first measurement occasion were less satisfied than others with their relationship 8 years later (actor effect). Furthermore, the partners of those who felt lonelier than others at the first measurement occasion were likewise less satisfied with their relationship than the partners of less lonely individuals (partner effect). At the first measurement occasion, loneliness was uncorrelated between the partners, whereas the within-couple correlation in relationship satisfaction at W8 was of medium size.

In addition to these key analyses, we ran two sets of supplemental analyses. First, we ran the models for married (\(n = 1394\)) and unmarried (\(n = 933\); 10 couples did not report marital status) couples separately. Second, we split the sample into two groups based on relationship duration at the first measurement occasion. The first group (\(n = 617\) couples) included couples who were together for less than 4 years. The second group (\(n = 649\) couples) consisted of couples who were together for more than 11 years. Regarding marital status, there were no differences in the association between loneliness and later levels of satisfaction. Regarding relationship length, the actor effect of loneliness on later levels of satisfaction was not statistically significant in the group with a relationship length of less than 4 years, whereas the partner effect was significant. In contrast, the actor, but not the partner effect was statistically significant in the group with a relationship duration of more than 11 years. The initial correlation between the partners in loneliness was not significant in the group with the shorter relationship duration, whereas it was significant in the group with a longer relationship history. The correlation in relationship satisfaction tended to be stronger in the group with a longer relationship duration, but it was significantly positive in both groups. Despite these differences in statistical significance, it should be noted that the parameter estimates in both groups did not differ from each other (see Supplemental Figures S4 and S5 and model outputs at https://​osf.​io/​tjuay/​).

In a second set of analyses, we predicted the development of relationship satisfaction over 8 years from loneliness measured at baseline. As in the previous analysis, the model in which actor and partner effects were constrained to be equal for women and men did not fit worse than the unconstrained model (\(\Delta \chi ^2 = 1.614\), \(\Delta df = 2\), \(p = .446\)). Furthermore, the constrained model showed an overall good fit to the data as indicated by a CFI of 0.924, an RMSEA of 0.030 (90% CI 0.028–0.032), an SRMR of 0.048, and a \(chi^2\)/df ratio of 3.08.

The results of this model are displayed in the lower half of Table 2. Taking the general decreasing trend of relationship satisfaction into account, the results indicate that individuals who feel lonelier than others at the first measurement occasion decrease even faster in relationship satisfaction than less lonely individuals. Again, the partner effect reached statistical significance as well and indicate that the relationship satisfaction of partners of individuals feeling lonelier than others decreases faster than the satisfaction of those partnered with less lonely individuals. Loneliness was uncorrelated between the partners at the first measurement occasion. However, the development of relationship satisfaction was highly interdependent among couple members, as indicated by the large correlation between the slopes.

In the supplemental analyses, no differences in the parameter estimates emerged for married versus unmarried couples. However, the initial correlation between partners in loneliness was significant in the couples with a relationship duration of more than 11 years, whereas it was not significant for the couples with shorter relationships (i.e., \(< 4\) years). However, despite this difference in statistical significance, the parameters did not significantly differ from each other (see Supplemental Figures S6 and S7 and model outputs at https://​osf.​io/​tjuay/​).

In this analysis, we examined the effects of dyadic similarity in loneliness at the first measurement occasion on each partners’ level of relationship satisfaction 8 years later. As with the ordinary APIM, we first examined whether the relevant paths of the model could be set equal for men and women. The results of the comparison between the constrained and the unconstrained model indicated differences in how similarity in loneliness affects relationship satisfaction of women and men (\(\Delta \chi ^2 = 15.58\), \(\Delta df = 5\), \(p < .001\)). We, thus, report the results for the unconstrained model, which is saturated. The results are displayed in the upper part of Table 3 and in Fig. 4a, b. The path coefficients, from which the surface parameters were derived, are reported in Supplemental Table S3.

For women, relationship satisfaction was higher only when she and her partner were uniformly low on loneliness, which is reflected by the significantly negative \(a_1\)-parameter (see upper part of Table 3). When couples were congruently high on loneliness, women were somewhat more satisfied than women who were congruent with their partners on medium scores of loneliness, as reflected by the significantly positive \(a_2\)-parameter (see also Fig. 4a). Incongruence in loneliness was associated with low relationship satisfaction across the entire spectrum of the scale (i.e., nonsignificant parameters \(a_3\) and \(a_4\); see also Fig. 4a). Thus, women are less satisfied prospectively when either partner scored higher on loneliness than the other, indicating an absence of buffering or compensatory effects. Overall, the pattern observed for women resembles a couple-oriented picture (Schönbrodt et al. 2018).

For men, in contrast, loneliness predicted later satisfaction in an actor-driven way. This means that the female partner’s loneliness did not affect the male partner’s relationship satisfaction (see Fig. 4b). This pattern is indicated by only the surface parameters \(a_1\) and \(a_3\) being significant. Thus, irrespective of the woman’s degree of loneliness, men were more satisfied with their relationships when they scored low on loneliness. Men having higher levels of loneliness than others were less satisfied with their relationships 8 years later, again irrespective of their partner’s loneliness (see Fig. 4b).

In the supplemental analyses on married versus unmarried couples and on couples with shorter (i.e., \(<4\) years) versus longer (\(>11\) years) relationship duration, we found several differences regarding the effects of similarity in loneliness and later levels of satisfaction see Supplemental Figures S10 and S11 and model outputs at https://​osf.​io/​tjuay/​). Specifically, differences between women and men only occurred in unmarried couples in these analyses. Furthermore, the parameter \(a_1\) was somewhat less negative for unmarried women than for their partners (i.e., the decline in satisfaction with increasing levels of loneliness was less strong for unmarried women). The parameter \(a_2\) was not significant for unmarried men, but for their spouses. The parameter \(a_3\), in contrast, was negative for unmarried men but not statistically significant for their female partners. Finally, the parameter \(a_4\) was significantly positive for unmarried men only. Despite the differences in statistical significance, however, the parameters did not differ significantly from each other. Several more differences occurred for shorter- versus longer-term relationships. In shorter-term relationships, the parameter \(a_1\) was significantly negative for men, but not for women. In longer-term relationships, the pattern was reversed, that is, \(a_1\) was significantly negative for women, but not for men. The parameter \(a_2\) was significantly positive for women in both shorter- and longer-term relationships, but not for men. The parameter \(a_3\) was statistically significant only for men in shorter-term relationships. For women, it tended to be positive (i.e., incongruence tended to have a positive association with later satisfaction), but did not reach statistical significance. The difference in the \(a_3\)-parameter between women and men in shorter-term relationships was itself statistically significant. The parameter \(a_4\) was significantly positive for men in shorter-term relationships, but not for their partners.

In the next step, we investigated the effects of dyadic similarity in loneliness at W1 on the development of relationship satisfaction across the 8 years of the study. The constrained model did not fit worse than the unconstrained model (\(\Delta \chi ^2 = 6.4939\), \(\Delta df = 5\), \(p = .261\)), indicating that there were no differences between women and men regarding the effects of loneliness on the development of relationship satisfaction. The constrained model had an acceptable fit to the data (\(\chi ^2 / df = 2.99\); CFI \(= 0.952\); RMSEA \(= 0.029\); SRMR \(= 0.045\)).

The resulting response surface parameters are displayed in the middle portion of Table 3 and are visualized in Fig. 4c. The path coefficients from which the surface parameters were derived are displayed in Supplemental Table S3. As indicated by the significant surface parameter \(a_1\), relationship satisfaction slightly increased during the study period if both partners were congruently low on loneliness at W1. As indicated by the significant surface parameter \(a_2\), this trend was curvilinear so that relationship satisfaction decreased more slowly when both partners were congruently high on loneliness at W1, compared to congruence around average scores of loneliness (see Fig. 4c). The significant parameter \(a_3\) further indicates an effect of incongruence: Relationship satisfaction decreased fastest when the actor had high levels on loneliness while the partner had low levels (see also Fig. 4c).

In the supplemental analyses, the parameter \(a_3\) was significantly negative for unmarried couples only. That is, incongruence in loneliness between partners tended to have a negative association with changes in satisfaction over time for unmarried, but not for married couples. For couples with a relationship duration of less than 4 years, the parameter \(a_1\) tended to be less negative than for couples being together for more than 11 years. In contrast, the parameter \(a_2\) was significantly positive for individuals in longer-term relationships, indicating an upward bend of the LOC. No differences occurred for \(a_3\) and \(a_4\). Despite these differences in statistical significance, it should be noted that the surface parameters did not significantly differ between the groups (see Supplemental Figures S12 and S13 as well as the model outputs at https://​osf.​io/​tjuay/​).

In the final analysis, we examined whether dyadic similarity in the development of loneliness across 8 years was associated with the development of relationship satisfaction. When comparing the constrained and the unconstrained models, the \(\chi ^2\)-difference test indicated no difference in the effects between men and women (\(\Delta \chi ^2 = 10.208\), \(\Delta df = 5\), \(p = .07\)). The constrained model had a good fit to the data as indicated by a ratio of \(\chi ^2\) to df of 2.45, a CFI of 0.948, an RMSEA of 0.025 and an SRMR amounting to 0.037. The resulting surface parameters are displayed in the bottom portion of Table 3 and are visualized in Fig. 4d. The path coefficients from which the surface parameters were derived are displayed in Supplemental Table S3.

As indicated by a significantly negative \(a_1\)-parameter, relationship satisfaction decreased slowest over time when both partners showed relative decreases in loneliness over time. Furthermore, the parameters \(a_3\) and \(a_4\) reached statistical significance. This indicates that incongruent constellations in the development of loneliness were associated with the development of relationship satisfaction. Specifically, relationship satisfaction decreased slower when at least one partner decreased in loneliness (i.e., \(a_4\)-parameter). This trend was slightly shifted in a way that the least decreases in satisfaction were found when Partner A (the actor) strongly decreased in loneliness (i.e., \(a_3\)-parameter; see Fig. 4d).

In the supplemental analyses, several differences occurred regarding the surface parameters for the groups. Regarding marital status, gender differences occurred for unmarried but not for married couples. The strongest difference occurred for the parameter \(a_3\), which was significantly negative for unmarried men but not for their partners. That is, incongruence in relative changes in loneliness appear to be negatively associated with the development of relationship satisfaction for unmarried men, but not unmarried women. For married couples, \(a_3\) was significantly negative. Regarding the parameter \(a_{4}\), there was a further group difference. The parameter was positive for married couples but not statistically significant for unmarried women and men. That is, the LOIC is bend upwards for married couples but shows no curvature for unmarried couples. With regard to relationship duration, gender differences in the surface parameters were found only for couples being together for less than 4 years. As found before, the parameter \(a_3\) was significantly negative for men in such shorter-term relationships but not for their partners. Despite these differences in statistical significance, however, the parameters did not differ significantly between the groups (for details, see Supplemental Figures S14 and S15 as well as the model outputs at https://​osf.​io/​tjuay/​).