Who Cares What Other People Think? A Longitudinal Investigation of the Role of Autonomy-Connectedness in Self-Esteem Change Trajectories and Instability

The present paper made use of data of the LISS (Longitudinal Internet studies for the Social Sciences) panel administered by CentERdata (Tilburg University, the Netherlands). The LISS panel is a representative sample of the Dutch population aged 16 or older, who complete monthly surveys on social, economic and health related topics. The sample is based on a true probability sample of households drawn from the population register. All respondents provided informed consent. If respondents had no access to a computer with internet, they were provided with one (more information on data collection is available on www.​lissdata.​nl; our analysis code can be accessed through https://​doi.​org/​10.​34894/​4NDGHW).

The LISS dataset contains 12 waves administered between 2008 and 2020 (Fig. 1). Autonomy-connectedness was measured once (in 2009) and was used to predict change trajectories in self-esteem, which had measurements spanning from 2008 through 2020. Self-esteem was usually assessed in May and questionnaires were repeated in June for non-responders. Exceptions include the baseline wave, which was repeated for non-responders in August, and wave 7 and 8, which were administered in November (non-responders in December). In waves 3, 5, 8 and 10, respondents who completed the questionnaires in the preceding year were offered a shorter questionnaire, which did not contain the self-esteem measure, providing lower response rates. Autonomy-connectedness, depressive and anxiety symptoms were administered as an extra questionnaire in November 2009. Demographics were also derived from the 2009 measurement and social desirability was assessed in 2008.

Respondents were included in the dataset if they had completed the 2009 autonomy-connectedness questionnaires and at least one self-esteem measure, providing a total of 5367 respondents (see Table 1). A total of 28,434 (44.15%) self-esteem measurements were completed, out of the total possible 64,404 (5367 respondents × 12 waves). As respondents could not partly complete the self-esteem and autonomy-connectedness questionnaires, they did not have missing values. Missing self-esteem waves were not replaced or imputed, because both statistical approaches (see Statistical analysis) are based on full information maximum likelihood (FIML), which accommodates time-varying covariates and can deal with missing values (e.g., Polit & Gillespie, 2010). Imputing missing values for the waves with few responses would also be invalid, as four waves (3, 5, 8 and 10) had between 83 and 97% observations missing and missing values were systematic (not random) due to the process of data collection in the LISS panel. See https://​doi.​org/​10.​34894/​4NDGHW for the syntax files used for data analysis (Table 1).

All analyses were two-tailed. We used an alpha of 0.001 because statistical power of detecting small true effect sizes still approximates 1 due to the very large sample size, and we wanted to limit the type I error rate. Because of very low completion rates at waves 3, 5, 8 and 10, analyses were carried out using only the remaining 8 waves (colored black in Fig. 1), but all analyses were repeated using all timepoints as sensitivity analyses (see Online Appendix A). Means, standard deviations and correlations of variables under study were computed as descriptive statistics.

Data were analyzed using two complementary statistical techniques answering different research questions. Using linear mixed models (LMM) we tested hypotheses on associations between autonomy-connectedness components and self-esteem in the population sample; we tested individual differences in the linear trajectories of self-esteem over time; and we assessed whether possible differences therein could be explained by autonomy-connectedness. Using latent growth mixture modeling (LGMM) we set out to further explore and illustrate individual differences in self-esteem change trajectories and instability; and we tested whether obtained classes differed with respect to personal characteristics, including autonomy-connectedness. Hence, LMM focused on identifying and explaining linear trajectories of self-esteem over time in the sample as a whole, whereas LGMM built on the LMM by further exploring and illustrating classes with different self-esteem trajectories.

The ICC of 0.70 indicated that most variation in self-esteem scores stemmed from differences between people rather than from intra-individual changes in self-esteem (variation within their own scores). Incorporating the ARH1 covariance structure improved model fit, χ²(8) = 1,113.56, p < 0.001. Adding a random effect of time to the latter model further improved the fit, χ²(2) = 334.61, p < 0.001, showing that there were individual differences in (linear) trajectories of self-esteem over time. In the latter model, error variances generally decreased over time, from values equal to 0.318 at wave 1 (2008) to 0.186 at wave 12 (2020). The estimated autocorrelation was 0.143 (Wald z = 11.65, p < 0.001). The variances of random intercept and slope were 0.657 (z = 37.68, p < 0.001) and 0.287 (z = 15.22, p < 0.001), respectively, and were negatively correlated (r = − 0.156, Wald z = − 4.66, p < 0.001). The negative correlation means that respondents with higher baseline self-esteem tended to have lower (less positive or more negative) slopes, indicative of a lower increase or a higher decrease of self-esteem over time, compared to people with lower baseline self-esteem.

Table 3 shows the fixed-effect estimates, variances, intercept and slope and their covariance, and fit indices of models 1–4. On average, self-esteem decreased slightly from 2008 to 2020 (− 0.079 points per 10 years, d = -0.081, small effect; Table 3, Model 1), suggesting a largely stable average trajectory on population level. Upon adding the autonomy-connectedness components to Model 2, model fit improved compared to Model 1, χ²(3) = 1766.20, p < 0.001, Table 3. Partly in line with the hypotheses, Self-awareness and Capacity for managing new situations had positive associations with self-esteem (Table 3, Model 2), whereas Sensitivity to others appeared to be unrelated to self-esteem. Thus, individuals with low Self-awareness and Capacity for managing new situations, seem to have poorer self-esteem, whereas for Sensitivity to others, the association seems more complex.

Adding the three interactions of autonomy-connectedness with time again improved model fit (χ²(3) = 24.16, p < 0.001; Table 3, comparing Model 3 to Model 2), but Model 3’s BIC suggested worse fit than Model 2. As Model 3 indicated that Self-awareness and Capacity for managing new situations did not predict self-esteem trajectory over time, we fitted our final Model 4 with all three main effects and only the interaction of time with Sensitivity to others. Model 4 showed improved model fit compared to Model 2 (χ²(1) = 16.37, p < 0.001; BIC = 52,843.36), with a positive interaction effect (B = 0.10, SE = 0.026, t(3,240.44) = 4.051, p < 0.001). Simple slope analyses showed that self-esteem did not change over time for people high in Sensitivity to others (for + 1SD: slope = − 0.025, SE = 0.019, t(3,280.92) = − 1.28, p = 0.201), whereas self-esteem decreased for people with average (slope = − 0.080, SE = 0.014, t(3,249.37) = -5.80, p < 0.001), and low Sensitivity to others (for -1SD: slope = − 0.13, SE = 0.019, t(3,209.17) =  − 7.00, p < 0.001). In summary, Self-awareness and Capacity for managing new situations were unrelated to self-esteem development over time for the sample population as a whole. Individuals with high Sensitivity to others reported less decrease in self-esteem over time than those with lower Sensitivity to others.

Sensitivity analyses of the mixed models analyses are reported in Online Appendix A. Using all 12 waves instead of 8 did not alter the results (Online Appendix A). Adding demographic characteristics age, sex, net income of the household and education level to Model 4 improved model fit (− 2LL = 48,502.95, χ²(8) = 4,156.72, p < 0.001; BIC = 48,766.30). Age was positively associated with self-esteem, B = 0.0045, SE = 0.00064, p < 0.001, suggesting that older individuals had higher self-esteem. Older individuals however also showed more decline in self-esteem over time (time x age interaction: B = − 0.0041, SE = 0.00092, p < 0.001). Having completed primary education only was associated with lower self-esteem than having a university degree (B = − 0.23, SE = 0.051, p < 0.001; Online Appendix A). However, addition of these demographic characteristics did not change the findings on autonomy-connectedness.

As a final sensitivity analysis, depressive and anxiety symptoms (measured in 2009) and their interactions with time were added to Model 4. This model again showed improved fit, − 2LL = 47,868.18, χ²(2) = 634.77, p < 0.001; BIC = 48,151.78, with both depressive (B = − 0.039, SE = 0.0023, p < 0.001) and anxiety (B = -0.016, SE = 0.0025, p < 0.001) symptoms being negatively associated with self-esteem. Main results on effects of autonomy-connectedness on Self-esteem were unaffected, that is, Self-awareness (B = 0.42, SE = 0.016, p < 0.001) and Capacity for managing new situations (B = 0.13, SE = 0.014, p < 0.001) still predicted self-esteem, and there was a positive interaction effect between Sensitivity to others and time (B = 0.10, SE = 0.027, p < 0.001). No interactions between depressive symptoms and time, and anxiety symptoms and time were found (see Online Appendix A).

The random effects in the mixed models analyses (see Table 3) suggest individual differences in baseline self-esteem as well as change trajectories, which were partly explained by effects of autonomy-connectedness, demographics, and depressive and anxiety symptoms. These individuals differences and self-esteem instability were further explored using LGMM.

Table 4 summarizes the model fit indices for the 1 to 10 class solutions for LGMM, with the 10-class solution providing the best fit. The 7-class solution was selected because classes 8 through 10 did not provide additional substantive information compared to classes 1 through 7 and yielded very small classes. More specifically, the 8-class solution resulted in a class that contained less than 1% of all cases, and the 9- and 10-class solutions further split up the high and stable self-esteem classes, not contributing to an enhanced understanding of individual differences in self-esteem trajectories.

Figure 2 displays self-esteem trajectories of 50 randomly selected respondents per latent class. Table 5 displays the class parameters and autonomy-connectedness, clinical correlates and demographic characteristics across latent classes, including 99.9% confidence intervals of class averages on these variables and their (Pseudo) R² values. Note that class membership had a large effect on Self-awareness, Capacity for managing new situations, anxiety- and depressive symptoms, (R² ≥ 0.16), medium effects on Sensitivity to others and age (R² between 0.06-0.08), and hardly an effect on income (R² < 0.01). As income hardly differed across classes, we focus on the other variables for interpreting the classes.

Classes A and G (lightest grey, together 28%) had the highest baseline self-esteem (6.37 and 6.87, respectively), were most stable (ICCs of 0.77 and 0.84), and showed no (decreasing or increasing) trend. Both classes were characterized by higher than average autonomy (i.e., higher Self-awareness, d_A = 0.62 and d_G = 0.86; lower Sensitivity to others, d_A = -0.27 and d_G = -0.49; higher Capacity for managing new situations, d_A = 0.48 and d_G = 0.78), as well as less depressive (d_A = − 0.41 and d_G = -0.45) and anxiety symptoms (d_A = -0.39 and d_G = − 0.45). People in these classes were relatively older (d_A = 0.15 and d_G = 0.062) and well-educated. Neither sex was overrepresented in these classes.

Classes B and C (light grey, together 38%) had average to above average baseline self-esteem (5.93 and 5.51, respectively), but were less stable (ICCs of 0.65 and 0.73), as Class B showed a small negative trend over time (b = − 0.072), and Class C had a small positive trend (b = 0.085). Their autonomy-connectedness was close to average (i.e., Self-awareness, d_B = 0.085 and d_C = 0.027; Sensitivity to others, d_B = − 0.010 and d_C = 0.039; Capacity for managing new situations, d_B = -0.0073 and d_C = − 0.048), with below average depressive (d_B = -0.23 and d_C = − 0.26) and anxiety (d_B = − 0.23 and d_C = − 0.14) symptoms. Individuals in Class B were older than in other classes (d_B = 0.36), whereas individuals in Class C were younger (d_C = − 0.11). Neither sex was overrepresented in these classes.

Classes D and F (darker grey, together 22%) were characterized by highest instability (ICCs 0.46 and 0.51). Whereas Class D was characterized by low baseline self-esteem (4.63) that increased over time (b = 0.13), Class F had average self-esteem at baseline (5.64), which decreased over time (b = − 0.20). Both classes were characterized by some autonomy deficits (measured in 2009), which were more severe in Class D than in Class F (i.e., Self-awareness, d_D = − 0.50 and d_F = − 0.27; Sensitivity to others, d_D = 0.26 and d_F = − 0.012; Capacity for managing new situations, d_D = − 0.33 and d_F = -0.18). Both classes had above average depressive (d_D = 0.40 and d_F = 0.42) and anxiety (d_D = 0.40 and d_F = 0.35) symptoms. Classes D and F were younger (d_D = -0.43 and d_F = − 0.41) than other classes, and had lower education. Females were more likely to belong to Class D (62%) than males.

Class E (darkest grey, 11%) was characterized by low baseline self-esteem (4.27) with a neutral trend and high stability (ICC = 0.75). Individuals in this class had marked autonomy-connectedness deficits (Self-awareness d_E = -1.06, Sensitivity to others d_E = 0.45, Capacity for managing new situations d_E = − 0.72) and the highest depressive (d_E = 0.99) and anxiety (d_E = 0.88) symptoms. Individuals in Class E were of average age (d_E = − 0.092), and had lower education than other classes. Females were more likely to belong to Class E (63%) than males.

Sensitivity analyses with time not being restricted in its functional form did not suggest a non-linear relation (see Online Appendix A). Appendix. Fig. 2 in Appendix A displays the results of the sensitivity analysis including all available waves in the LGMM. As can be seen, except for Class C, class compositions are identical. For Class C, there are minor differences with a somewhat higher intercept and smaller slope. We do not judge these differences to be impactful enough to diverge from our initial solution.

As an additional analysis suggested during the review process, we also assessed class differences in socially desirable response tendencies, to explore characteristics of the ceiling class (G). Social desirability was highest in the more ‘confident’ classes (d_A = 0.27, d_B = 0.15, d_G = 0.25, d_F = 0.012) and lowest in less ‘confident’ classes (d_C = − 0.017, d_D = − 0.28, d_E = − 0.44), revealing a medium association between self-esteem and providing social desirable responses (R² = 0.056).

In line with the hypotheses, healthy autonomy-connectedness was positively associated with trait self-esteem, as evidenced by correlations between the autonomy-connectedness components and self-esteem, and differences in autonomy-connectedness across classes. Classes characterized by high and stable self-esteem (i.e., A and G) displayed remarkably high self-awareness compared to the other classes, in line with Blatt’s (2008) focus on the importance of self-definition for psychological well-being. The results suggest that a healthy capacity for self-governance seems to correspond with favorable self-esteem trajectories and good mental health. Our findings are also consistent with earlier studies on autonomy-connectedness, showing autonomy deficits in individuals with depression and anxiety symptoms (e.g., Bekker & Belt, 2006; Bekker & Croon, 2010; Kunst et al., 2019; Rutten et al., 2016). The results are less consistent with studies using Beck et al.’s (1983) original ‘autonomy’ subscale (e.g. Bieling et al., 2000), probably due to different operationalizations of autonomy (i.e., ‘self-governance’ in the present study).

The directionality of the association between autonomy-connectedness and self-esteem remains to be elucidated in future research, as we found no clear indications of, for instance, healthy autonomy-connectedness clearly predicting recovering self-esteem trajectories. Thus, both vulnerability and scar mechanisms still seem conceivable based on theory. Individuals with low self-esteem have been described to have global, negative self-beliefs, which can elicit fears and corresponding maladaptive behavioral patterns (e.g., avoidance and pleasing; Fennell, 1997). Low self-esteem could contribute to autonomy deficits when people consistently fail to act upon their authentic needs because of these patterns. Inversely, autonomy deficits could also contribute to lowered self-esteem, because effective self-governance seems important to shape one’s life in a way that benefits (social) need fulfillment, hampering self-esteem (Leary, 2012). Self-neglectful maladaptive behavior can also create and confirm negative self-beliefs (e.g., ‘if my needs are never met, apparently I am not worthwhile’; Fennell, 1997).

High self-esteem on the other hand likely facilitates behaviors corresponding with self-governance (e.g., assertiveness, self-expression; Rutten et al., 2016), which may allow for (social) needs fulfillment and psychological wellbeing, further boosting self-esteem (Leary, 2012). Indications exist that improved interpersonal functioning can, in turn, buffer against even the effects of depression on self-esteem (e.g., Shahar & Davidson, 2003). Studies including multiple autonomy-connectedness and self-esteem measurements may disentangle the interplay between autonomy-connectedness and self-esteem in the development and maintenance of depression and anxiety symptoms.

Partly in line with our hypotheses, sensitivity to others was weakly and positively related to self-esteem in the zero order correlations, and unrelated to self-esteem in the mixed models analyses. For any given individual, being highly sensitive to others’ needs and wishes seems to say little about their self-esteem. At a glance, this finding seems contradictory to earlier studies showing associations between depression and anxiety symptoms on one hand and sociotropy on the other hand (e.g. Bieling et al., 2000; Grondin et al., 2011); as well as anaclitic formulations of psychopathology (Blatt, 2008). However, depression and anxiety symptoms could be most pronounced in individuals with high levels of maladaptive others-directedness, whereas the autonomy-connectedness scale seems to measure adaptive as well as maladaptive aspects of connectedness (Kunst et al., 2021). Similarly, earlier studies on interpersonal dependence showed smaller associations between psychopathology and relatively adaptive ‘connectedness’, and larger relations with maladaptive ‘neediness’ (Dunkley et al., 2006).

Related to this, sensitivity to others seemed to ‘buffer’ against the effects of age on declining self-esteem. Across the lifespan, self-esteem gradually decreases from age 50–60 and on (Orth et al., 2018), possibly corresponding with increasing social isolation and employment changes as we age (Orth et al., 2010). Individuals with high sensitivity to others may be more motivated to maintain interpersonal connectedness throughout their lives, thereby keeping a social network to support their sense of relational value (Leary, 2012). All findings on autonomy-connectedness were maintained when controlling for age, education, sex, depressive and anxiety symptoms, and hence are unlikely confounded by these characteristics. However, it should be noted that the buffering effects of sensitivity to others on self-esteem decline were small, and research including measures on life events and quality of social networks seems warranted to gain more insight into these processes.

Whereas mixed models showed that individual differences existed, LGMM allowed for a more detailed exploration and illustration of these differences. The identified classes are in line with earlier research showing weak and negative associations between trait self-esteem and self-esteem instability (Okada, 2010), as classes characterized by high self-esteem were also relatively stable. Additionally, the existence of a ‘low and stable’ class could explain why the association between trait self-esteem and self-esteem instability is merely small, thereby also illustrating the added value of LGMM. Depressive and anxiety symptoms corresponded with the overall self-esteem levels, with most symptoms in the ‘low and stable’ class (E) and less symptoms as classes had higher and more stable self-esteem.

In contrast with our hypotheses, autonomy-connectedness deficits did not necessarily predict unstable self-esteem trajectories. Although the most unstable Classes D and F indeed displayed some autonomy deficits (small to medium effects), the most severe autonomy deficits (medium to large effects) were found in the ‘low and stable’ Class E. Our expectation was based on presumed high sensitivity to external evaluations in people with poor capacity for self-governance. However, if such a sensitivity exists, it does not seem visible when assessing trait self-esteem yearly. Momentary fluctuations in state self-esteem were beyond the scope of this study and might be interesting to study in laboratory or experience sampling studies in relation to autonomy-connectedness. Assessed on the long-term, however, severe autonomy deficits seem to correspond with alarmingly low self-esteem scores, high depressive and anxiety symptoms (all large effects), that are unlikely to improve without intervention, possibly due to reciprocal effects of autonomy and self-esteem. Therapies targeting self-esteem (Staring et al., 2016) or autonomy-connectedness (Bekker et al., 2016) might therefore be recommended for this group. An alternative interpretation is that this class is characterized by severe psychopathology or socio-occupational impairments, which collectively contribute to poor self-esteem, autonomy-connectedness and mental health. The directionality of the obtained associations remains to be clarified in future research.

The present study extended previous LGMM studies by describing not only mental health, but also potential underlying psychological characteristics of self-esteem classes in terms of autonomy-connectedness deficits. We identified more diverse classes based on self-esteem trajectories than in previous studies, possibly due to the nationally representative sample including a wider age range (16–91). In our sample, age was positively associated with self-esteem, and older individuals also showed more decline in self-esteem over time, consistent with meta-analytic research on self-esteem across the lifespan (Orth et al., 2018). The unstable classes (D and F), accordingly, included younger individuals than most other classes. However, it should be noted that we did not find classes specific to certain age groups (e.g., adolescents, older individuals), suggesting that the obtained class solutions were not merely reflective of lifespan effects. Nevertheless, our sample contained older individuals than previous studies (ages 12 through 30; Birkeland et al., 2012; Morin et al., 2013; Mund & Neyer, 2016; Oshri et al., 2017; and 16 through 40; Kiviruusu et al., 2016), possibly explaining the identification of a ‘decreasing’ (F) and ‘low and stable’ (E) class. Class G, the ‘ceiling class’, was also a noteworthy variant on the ‘high and stable’ class found previously in LGMM studies. Their high scores did not seem to stem from socially desirable responding tendencies, as social desirability was comparable in other ‘confident’ classes (A and B). This class therefore seems highly content and resilient.

The different latent classes we found compared to earlier studies can also partly be explained by differences in statistical approaches. For instance, we did not obtain the ‘recovering and stabilizing’ classes, that Mund and Neyer (2016) and Morin et al. (2013) did identify. In their studies, non-linear slopes were permitted, thereby allowing trajectories that initially increase and then reach a plateau. Even though our data overall did not support the use of non-linear slopes, Class C (‘high and increasing’) did resemble a ‘stabilizing’ class when non-linear slopes were allowed (see Online Appendix A). Additionally, whereas our, Mund and Neyer (2016) and Morin et al. (2013)’s studies involved error variance in the process of estimating latent classes, other research based the latent classes purely on intercept and slope trajectories (Birkeland et al., 2012; Kiviruusu et al., 2016; Oshri et al., 2017). In these studies, most respondents were classified as ‘high and stable’: 87.4% in Oshri et al. (2017) and 87.1% in Birkeland et al. (2012); compared to 28% in our sample, 29% in Mund and Neyer (2016) and 13.5% in (Morin et al., 2013). Incorporating error variances in class estimation thus allows for the identification of more heterogeneous classes with respect to stability in self-esteem development.

The present study suggests that a relatively large section of the population (11%) suffers from persistent self-esteem and autonomy-connectedness problems. In some of these individuals, self-esteem issues likely preceded the development of their anxiety and depressive symptoms. Prominent social anxiety disorder (SAD) models, for instance, propose that socially anxious individuals have baseline negative self-representations that are activated during conversations, triggering a cascade of cognitive (e.g., overestimation of threat and social cost, self-focused attention) and behavioral (e.g., avoidance, safety behavior) processes that maintain SAD (Hofmann, 2007; Rapee & Heimberg, 1997; Spence & Rapee, 2016). Similarly, low self-esteem tends to precede the development of depressive symptoms (Sowislo & Orth, 2013).

Our findings therefore suggest that negative self-beliefs could be a valuable additional target in psychotherapy for SAD and major depression. Clinicians may evaluate clients' self-esteem histories to determine the potential benefits of adding self-esteem (e.g. Fennell, 1997; Staring et al., 2016) or autonomy-connectedness (Kunst et al., 2022; Maas et al., 2019) enhancing interventions. In cognitive behavioral therapy for SAD, for instance, challenging thoughts about social performance could be accompanied by challenging more global self-beliefs. Cognitive behavioral therapy for SAD typically also involves teaching clients to focus less on ‘what other people think’ during conversations (i.e., decreasing self-focused attention; Spence & Rapee, 2016). Given the large association between self-awareness and self-esteem, a worthwhile venue might be to coach clients with SAD to reflect more on their own interests, needs, preferences and opinions, and to support clients in learning to act accordingly; the main aims of autonomy enhancing treatment (Maas et al., 2019). Developing a stronger sense of self and identity might help clients feel more confident during social interactions and help reduce social anxiety.

The present study was subject to a number of limitations. First, due to the data collection method of the LISS panel the timing of measurement sometimes varied across participants (e.g., T1 was completed either in May or August of 2008). Moreover, four waves had many missing self-esteem measurements and for unclear reasons, and the Rosenberg Self-Esteem Scale used an adapted 7-point scale instead of the original 4-point scale. Second, due to the single autonomy-connectedness measurement, we could not assess changes over time and its associations with self-esteem on timepoints other than 2009. This may be important, as we found that the autonomy-connectedness scores across latent classes were most strongly associated with baseline self-esteem in that class, instead of the increasing or decreasing trend within the class, possibly because of the proximity of the 2009 autonomy-connectedness measure to the first self-esteem measure (in 2008). Theoretically, autonomy-connectedness is viewed as a relatively stable psychological characteristic, stemming from secure attachment experiences (Bekker & van Assen, 2008). Nevertheless, autonomy deficits seem susceptible to positive therapeutic influences (e.g., see Kunst et al., 2022). Its test–retest stability and potential changes across the lifespan remain to be investigated.

Furthermore, we chose to analyze data of all respondents (ages 16–91) without creating subgroups according to life phase, because splitting the sample would reduce the statistical power to detect small to medium meaningful patterns in self-esteem trajectories in the population. We were also unable to use ‘age’ instead of ‘wave’ as a time indicator because of convergence issues. Additionally, our focus was on the role of autonomy-connectedness in self-esteem trajectories, and by incorporating age in the analyses we assessed whether age is associated with self-esteem and its development. Further exploration of the effects of age on self-esteem was beyond the scope of the present study aims. Similarly, we did not take into account potential measurement variance of the self-esteem concept. The RSES factor structure appears largely invariant across nations (e.g., Alessandri et al., 2015; Schmitt & Allik, 2005) as well as across several groups such as clinical versus non-clinical groups (Salerno et al., 2017). Nevertheless, changes in self-esteem could partly reflect differences in interpreting questions and this should be assessed in future research.

Whereas the present study focused on identifying and describing latent classes in terms of autonomy-connectedness, it was beyond of its scope to develop and test models predicting class membership as accurately as possible for specific individuals. Clinically, it would be valuable to predict likely developmental pathways of self-esteem and mental health, based on characteristics that can be targeted in psychotherapy. We regard our study as a first step towards creating optimal prediction models involving more social- and personality variables than those available in our study, to improve the accuracy of the predictions in multinomial regression models. For clinical practice, it would also be interesting to test whether different interventions are required depending on pre-treatment symptom trajectories. We recommend directly comparing effectiveness of autonomy enhancing treatment (Kunst et al., 2022; Maas et al., 2019) to self-esteem boosting therapies (e.g., Staring et al., 2016) to shed light on causal and working mechanisms.