The Relationship between Self-Esteem and Depression when Controlling for Neuroticism

Whereas self-esteem has been considered a viable vulnerability factor for depression, there is another variable that may be a vulnerability factor for both self-esteem and depression, neuroticism. Although defined in various ways, the consensus definition is that neuroticism, at its core, is the propensity to experience negative emotions (Clark & Watson, 1999; Depue & Lenzenweger, 2001; Digman, 1997; Matthews et al., 2003; McCrae & Costa, 1997; Tellegen & Waller, 1997; Widiger, 2009). Neuroticism is believed to encompass many lower-order traits or facets. Most researchers agree anxiety-withdrawal, depression-unhappiness, vulnerability-stress reaction are facets of neuroticism; others would include angry hostility-aggression, impulsivity, inferiority, and dependency should be considered neuroticism as well (Ormel et al., 2013).

A strong relationship with neuroticism has been documented for self-esteem and depression in both cross-sectional and longitudinal studies. For example, Francis and James (1996) showed that emotional stability, a label commonly used for the low end of the neuroticism continuum, is associated with high self-esteem (r = .32) using Eysenck’s two-dimensional model of personality and Rosenberg’s self-esteem scale. Zeigler-Hill and colleagues (2015) found that self-esteem was positively associated with emotional stability using both self-reported scales (r = .45) and perceiver reported scales (r = .35). Also, in a large study of 326,641 participants, Robins et al. (2001) found that emotional stability was positively correlated with self-esteem (r = .50).

Neuroticism has also been shown to be a strong predictor for depression. For example, Kotov, Gamez, Schmidt, and Watson (2010) conducted a meta-analysis of 175 studies and found a strong positive association between neuroticism and depression (d = 1.33). Also, Kendler (2004) conducted a study on over 7500 twins trying to predict the onset of major depression and found an interaction effect between neuroticism and adversity. Individuals with high neuroticism were both at greater overall risk for major depression and were more sensitive to the depressogenic effects of adversity. The relationship between neuroticism and depression goes beyond cross-sectional level of analyses. It has been shown that change in neuroticism was also a predictor of change in depression, such that an increase in neuroticism was associated with an increase in depression (Chow & Roberts, 2014).

Despite significant covariances among the three constructs, they are essentially distinct constructs. For self-esteem and neuroticism, one particular useful way to look at their conceptual distinction is the perspective of core versus surface characteristics (Kandler, Zimmermann, & McAdams, 2014). Neuroticism is considered a core characteristic, which are largely consistent patterns of thoughts, feelings and actions across time and situations. By contrast, self-esteem is considered a surface characteristic, or characteristics that are believed to emerge much later, continue to evolve through lifespan, and are less stable or more environmentally malleable than core characteristics (McAdams & Pals, 2006). According to this point of view, it is believed that self-esteem is the by-product of the interaction of core characteristics, such as neuroticism and environmental influences (McCrae, 2009).

Empirical evidence on self-esteem and neuroticism seems to be in line with this view. First, compared to neuroticism, self-esteem has been found to be a less stable phenotype. Meta-analyses and longitudinal studies on representative samples have revealed that levels of rank-order stability are higher for neuroticism than self-esteem in adulthood (Trzesniewski, Donnellan, & Robins, 2003). Gene-environment interplay studies reveal that whereas neuroticism has been found to be more genetically based, self-esteem is found to be more subject to environmental influences, such as achievements, life stressors and failures (e.g., Kandler, Zimmermann, & McAdams, 2014). Heritability rates for neuroticism have been found to range from 40% to 60% (see Bartels & Boomsma, 2009, for a review), but found to range from 20% to 40% for self-esteem (e.g., Neiss, Sedikides, & Stevensen, 2002; Pedersen, Gatz, Plomin, Nesselroade, & McClearn, 1989). Therefore, it seems that self-esteem is a less stable and more environmentally malleable surface manifestation of personality.

Likewise, despite significant overlap, neuroticism and depression also differ from each other in fundamental ways. To begin with, while neuroticism is a trait like variable, depression is a state like variable. It is, by definition, a mental disorder, with onsets and episodes. Second, although depressivity, the predisposition to experience depression, is one facet of neuroticism, neuroticism is a much broader construct encompassing many more other facets, such as anxiety-withdrawal, vulnerability-stress reaction, hostility-anger (Ormel et al., 2013). Furthermore, neuroticism has been proposed to be an important vulnerability factor underlying the development of depression and this view has received considerable empirical support (see Klein, Kotov, & Bufferd, 2011 for a review). For example, studies show the rank-order stability of depression tends to be lower than that of neuroticism (Ormel et al., 2013). What is more, it should be noted the prospective association between neuroticism and depression has been well established even across long intervals of multiple years and adjusted for other psychiatric confounding variables. Last but not least, multiple longitudinal studies have found that the lifestyles of high-neuroticism individuals increase the likelihood of stressful experiences, and that these stressors in turn can trigger depression (Hankin, Stone, & Wright, 2010). Collectively, these findings strongly suggest neuroticism plays a very important role, at least partly, in the complex pathways leading to the development of both low self-esteem and high depressive symptoms.

The majority of work showing that low self-esteem prospectively predicts depression has relied on cross-lagged panel models. However, self-esteem and depression are conceptualized and measured as different types of variables. Global self-esteem, as it is usually studied in relationship to depression, is assessed and measured as a trait, denoting the “average tone of self-feeling” that each person carries around (Williams, 1995). It is a person’s long-term, typical, affectively laden self-evaluation (Leary and Baumeister, 1995). By contrast, depression is an affective disorder that is considered to be episodic, with onset as well as remission. Most measures of depression are measures of states and not traits. Typical measures ask people to rate items based on how they have felt over the past two weeks (DSM-5, 2013). Given the different levels of conceptualization and measurement, it is not completely surprising to find the broader and more stable construct (i.e., self-esteem) often out-predicts the less stable construct (i.e., depression).

It is also questionable if cross-lagged panel models are the best way to establish the temporal order of longitudinal relationship among psychological constructs. It has recently been pointed out that standard cross-lagged panel models do not separate within- and between person effects and they assume that each person varies over time around the same mean (Hamaker et al., 2015). In other words, it is assumed there is no time invariant, trait-like individual differences that endure, an assumption that does not hold for most psychological variables (Fraley & Roberts, 2005). For example, it has been shown that there are ample individual differences in both the average level and change for major personality traits (Lüdtke et al., 2011), self-esteem (Wagner et al., 2013) as well as depression (Chow & Roberts, 2014).

Latent growth models are good tools to address this issue. These models separate the stable, between-person component and the within-person, changing aspect of any construct being examined. These models also can be used to estimate individual differences in both the initial levels of a variable (e.g., depression) as well as change over time (Hoffman, 2015). For this reason, latent growth models have been suggested to be useful in studying change in personality traits over time and determining the temporal orders of correlational relationships (Hamaker et al., 2015). They have also been used to study how individual differences in change in personality is related to change of other important outcomes, such as stress (Luo & Roberts, 2015) and mental health (Mu et al., 2016). As noted above, a positive relationship has been found between change in neuroticism and change in depression, such that increases in neuroticism were associated with increases in depression (Chow & Roberts, 2014). In another study, older men who were high in neuroticism at the beginning of the study and who increased in neuroticism over the course of the study experienced a higher risk of mortality than men who began the study low in neuroticism or men who decreased in neuroticism over time (Mroczek & Spiro, 2007). Thus, it is possible that growth in trait-like constructs, such as self-esteem, could be correlated with change in depression over and above antecedent standing (the level in a growth model) and that this is a better way of modeling the interplay of variables over time. Therefore, we will employ a variety of models so as to more thoroughly test the potential confounding of the relation between depression and self-esteem by neuroticism.

Given the aforementioned findings and links among self-esteem, depression, and neuroticism, the current study aimed to examine the relationship between self-esteem and depression while controlling for neuroticism. We employed a data set from a large longitudinal study tracking over 2000 German students in their early 20s that had been used in prior research to replicate the cross-lagged relation between self-esteem and depressive symptoms (Rieger et al., 2016). We investigated the relationship between self-esteem and depression while controlling for neuroticism using both cross-lagged panel models and latent growth models. We hypothesized that the relationship between self-esteem and depression would be reduced after controlling for neuroticism in both cross-sectional and longitudinal analyses.

The data come from a large, ongoing longitudinal German study (Transformation of the Secondary School System and Academic Careers; TOSCA; for a detailed overview see Trautwein, Neumann, Nagy, Lüdtke, & Maaz, 2010). The TOSCA study currently encompasses six time points. Data for self-esteem, depression and neuroticism are available for three waves. T1 is 2 years after graduation from high school (February to May, 2004). Participants completed an extensive questionnaire taking about 2 hours in exchange for a financial reward of 10 Euros. The second (T2) and third (T3) assessment took place from February to May, 2006 and from February to May, 2008, respectively. Again, participants completed an extensive questionnaire taking about 2 hours in exchange for a financial reward of 10 Euros.

Given this is a very large panel study that examined a very wide range of variables for an extended period of time, the current dataset has allowed many important questions regarding personality to be examined and research to be published. Among others, the most relevant ones have examined neuroticism (Lüdtke et al., 2011), self-esteem (Wagner et al., 2013), as well as self-esteem and depression (Rieger et al., 2016). However, it should be noted that the current analyses regarding self-esteem, depression and neuroticism have never been previously reported.

Data of the interested constructs (i.e., self-esteem, depression, and neuroticism) were available for n = 2,318 (64.0% female) individuals at T1, n = 1,912 (64% female) individuals at T2, n = 1,871 (63% female) individuals at T3. The sample size of the pooled data set is N = 2,512. Mean age of participants was M = 21.5 years (SD = .8) at T1, M = 23.4 years (SD = .6) at T2, and M = 25.4 years (SD = .7) at T3.

For attrition analyses, we compared continuers, who completed all three time points, with dropouts, who participated only in the first wave. There were no significant differences on the study variables: self-esteem (d = –.00, p = .99), depression (d = –.01, p = .94), and neuroticism (d = –.01, p = .84). However, continuers were more likely to be female (OR = .76, p = .02), had better grade point averages (d = .20, p = .001) and performed better on a reasoning ability test (Heller & Perleth, 2000; d = 0.13, p = .03). Overall, the differences between continuers and dropouts were small (|ds| ≤ .20).

Self-esteem. Self-esteem was measured at the trait level. The Rosenberg Self-esteem Scale (RSE; Rosenberg, 1965) was used to assess self-esteem: three items were administered: (a) “At times, I think I am no good at all.” (b) “All in all, I am inclined to feel that I am a failure.” and (c) “I wish I could have more respect for myself.” These items were translated into German. Participants were asked to rate these items using a likert-type 4-point scale ranging from 1 (“not at all”) to 4 (“totally”). Internal consistency was good across all three waves (α = .84 at T1, .84 at T2, .86 at T3).

Depression. Depressive symptoms were assessed with the 15-item German version (“Allgemeine Depressionsskala”; ADS-K; Hautzinger & Bailer, 1993) of the Center for Epidemiologic Studies Depression Scale (CES-D; Radloff, 1977). A sample item was “I felt lonely.” Participants were asked to rate how often they have felt this way during the last week, using a 4-point likert-type scale (0 = “rarely or none of the time”, 1 = “sometimes”, 2 = “frequently”, 3 = “most of the time”). Internal consistency was good across three waves (α = .90 at T1, .91 at T2, .91 at T3).

Neuroticism. Neuroticism was measured using the German version (Borkenau & Ostendorf, 1993) of the NEO Five-factor Inventory (NEO-FFI; Costa & McCrae, 1992). Extensive work on the German translation has demonstrated the instrument’s high reliability, validity, and comparability with the English original (e.g., Borkenau & Ostendorf, 1993). A sample item was “I often feel tense and jittery.” The items were rated on a 4-point scale (1 = “strongly disagree”, 4 = “strongly agree”). In-depth psychometric analyses of the 4-point response format show that this format has some advantages over a 5-point scale (Lüdtke, Trautwein, Nagy, & Koller, 2004). Internal consistency was good across three waves (α = .87 at T1, .88 at T2, .89 at T3).

All models were estimated in the framework of longitudinal confirmatory factor analyses using Mplus 7.3 (Muthén & Muthén, 1998–2012). Two-sided statistical tests were performed at a level of significance of 5%. However, due to the observational character of our study, we rely on effect sizes and confidence intervals in addition to p-values (Groot, 2014).

The statistical procedure encompassed roughly three steps: First, to determine whether the three constructs should be modeled separately or as indicators of a common factor, we tested a series of models: 1) the one-factor model vs. the two-factor model of self-esteem and neuroticism; 2) the one-factor model vs. the two-factor model of depression and neuroticism; 3) the one-factor model vs. the two-factor model of self-esteem and depression; and 4) the one-factor model vs. the three-factor model of self-esteem, depression, and neuroticism. Second, to properly interpret latent variable change in longitudinal models, at least strong measurement invariance has to be established (Meredith, 1993; Meredith & Teresi, 2006). Thus, we specified a latent state model with imposed strong measurement invariance (same loadings and intercepts for each indicator over time) for all constructs within one model. This model served as our baseline model and we derived the means, standard deviations as well as latent correlations between all three constructs from it. Third, to investigate the prospective relationship between self-esteem and depression, we estimated a cross-lagged panel model (Model 1) and thereby reproduced the results from Rieger et al. (2016). Following this, we specified a cross-lagged panel model controlling for neuroticism at each time point (Model 2, see Figure 1). Fourth, to study interindividual difference in change over time we specified latent growth curve models. In a first step, we estimated three univariate latent growth models for each construct separately (Model 3a, 3b, 3c). Following this, we estimated a dual latent growth model for self-esteem and depression (Model 4). In a last step, we constructed a tri-variate latent growth model, to examine the relationship between self-esteem and depression while controlling for both the initial level as well as change in neuroticism over time (Model 5, see Figure 2).¹

Missing data. To deal with missing values, we used full-information maximum likelihood estimation, as this procedure has been shown to produce less biased and more reliable results compared with the more conventional methods (e.g., listwise or pairwise deletion; Allison, 2003; Graham, 2009).

Model fit criteria and parceling strategy. Following the recommendations of Hu and Bentler (1998, 1999), we employed the comparative fit index (CFI), Tucker-Lewis-Index (TLI), and the root-mean-square error of approximation (RMSEA) to assess model fit. Self-esteem has three items and therefore was modeled on the item level. For each of the depression and neuroticism scales, we randomly aggregated the items into three parcels. Item parcels have been shown to produce more reliable latent variables than individual items when used as indicators (Little, Cunningham, Shahar, & Widaman, 2002; but see Marsh, Lüdtke, Nagengast, Morin & von Davier, 2013).

Table 1 shows the means and standard deviations of the measures used in the current study. The latent correlations are depicted in Table 2.

We first tested and compared a one-factor model vs. a two-factor model of self-esteem and neuroticism using data at Time 1 to determine whether these constructs were distinguishable. In the one-factor model, all indicators loaded on one common factor, separately for each time point; in the two-factor model, two latent factors (self-esteem and neuroticism) were specified. The two-factor model (χ² (8) = 45.20, CFI = .99, TLI = .99, RMSEA = .05, SRMR = .02) fit significantly better than the one-factor model (χ² (9) = 709.09, CFI = .87, TLI = .81, RMSEA = .19, SRMR = .06). Likewise, for depression and neuroticism, the two-factor model (χ² (8) = 83.14, CFI = .99, TLI = .98, RMSEA = .06, SRMR = .04) fit significantly better than the one-factor model (χ² (9) = 2085.30, CFI = .71, TLI = .52, RMSEA = .32, SRMR = .10). The same is true for self-esteem and depression, with the two-factor model (χ² (8) = 3.87, CFI = 1.00, TLI = .99, RMSEA = .04, SRMR = .01) fitting significantly better than the one-factor model (χ² (9) = 150.77, CFI = .78, TLI = .64, RMSEA = .26, SRMR = .11). Lastly, we also tested the one-factor model vs. the three-factor model of all the three constructs. As predicted, the three-factor model (χ² (24) = 6.31, CFI = .99, TLI = .98, RMSEA = .05, SRMR = .02) fit significantly better than the one-factor model (χ² (24) = 150.82, CFI = .78, TLI = .64, RMSEA = .26, SRMR = .11). Because the one-factor and the two/three-factor models are non-nested, we did not conduct formal test of difference in fit; however, the fit indices clearly indicate the favorability of the two-factor and the three-factor models. Similarly, previous research also found the two-factor model fit better for depression and self-esteem than the one-factor model (Orth et al., 2008). Thus, in the following analyses, the three constructs were modeled separately rather than as indicators of a common factor.

To answer our first research question (prospective relationship between self-esteem and depression), we first reproduced the results of Rieger et al. (2016) by constructing a regular cross-lagged panel model with freely structural coefficients (Model 1). In cross-lagged models, a latent variable at Time 2 is predicted by the same variable at Time 1 (the autoregressor) and the other latent variable at Time 1. The cross-lagged paths indicate the relation of one variable to the other, after controlling for the stability of the same variables over time (Finkel, 1995).

Model 1 showed a good fit to the data, χ² (122) = 387.35, CFI = .99, TLI = .98, RMSEA = .03, and SRMR = .04 (see also Table 3). Consistent with Rieger et al. (2016), the cross-lagged paths from self-esteem to depression were both statistically significant for T1 to T2 and T2 to T3 (β = –.22, SE = .04 and β = –.26, SE = .04, ps < .001). By contrast, the cross-lagged paths from depression to self-esteem were both nonsignificant (β = –.05, SE = .03, p = .09 and β = –.01, SE = .04, p = .76). The stability coefficients of self-esteem were β = .72, SE = .03 and β = .74, SE = .03 (ps < .001) and β = .34, SE = .04 and β = .32, SE = .04 (ps < .001) for depression.²

We next tested a second structural cross-lagged model (Model 2, see Figure 1), that is identical to model 1, except that we controlled for neuroticism at each time point. The model also showed a good fit to the data, χ² (301) = 1107.50, CFI = .98, TLI = .97, RMSEA = .03, and SRMR = .03 (see Table 3). The cross-sectional paths from neuroticism to self-esteem at the three time points were β = –.78, SE = .01, β = –.62, SE = .03, and β = –.66, SE = .03 (ps < .001); and the cross-sectional paths from neuroticism to depression were β = .66, SE = .02, β = .62, SE = .03 and β = .68, SE = .03 (all ps < .001), respectively. When controlling for neuroticism, the cross-lagged paths from self-esteem to depression switched from the negative direction to the positive direction. Self-esteem at T1 positively predicted depression at T2 (β = .12, SE = .04, p < .001) and self-esteem at T2 positively predicted depression at T3 (β = .16, SE = .03, p < .001). Likewise, the cross-lagged paths from depression to self-esteem were in the positive direction for both time points as well, with depression at T1 predicted self-esteem at T2 (β = .08, SE = .03, p < .01) and depression at T2 predicted self-esteem at T3 (β = .08, SE = .03, p = .004). The stability coefficients of neuroticism were r = .77, SE = .01 and r = .77, SE = .01 (ps < .001). Compared to Model 1 in which neuroticism was not controlled for, the stability coefficients of self-esteem dropped from .72 and .74 to .37 and .28 (ps < .001) and those of depression dropped from .34 and .32 to .21 and .21 (ps < .001). Given the high correlation of neuroticism with both self-esteem and depression, our interpretation of the switch from a negative relation in Model 1 (self-esteem to depression: –.22 & –.26; depression to self-esteem: –.05 & –.01) to a positive relation in Model 2 (self-esteem to depression: .12 & .16; depression to self-esteem: .08 & .08) is consistent with an artifact that results from controlling for all of the valid variance that explains the relation between self-esteem and depression.

Given the criticisms of cross-lagged regression models, we also tested the relations of all three variables using latent growth models. To assess the magnitude of interindividual differences in intraindividual change in neuroticism, self-esteem and depression, we constructed three univariate latent models with multiple indicators (Model 3a, 3b, 3c). All the three models showed excellent fit to the data (CFI and TLI > .95 and RMSEA and SRMR < .05; see Table 3). The latent intercepts represent the initial level of a personality trait at T1 and the variance of the latent intercepts indicates the amount of reliable individual differences at T1. The mean of the latent slope factors indicates the rate of change across the 2-year period and the variance of the slope indicates the amount of reliable individual differences in change. Addressing our primary question, the statistically significant variance components revealed that all three factors showed significant interindividual differences in both the initial level and the intraindividual change over time. Specifically, the variance of initial level is .20, SE = .01, p < .001 for neuroticism, .28, SE = .02, p < .001 for self-esteem, and .14, SE = .02, p < .001 for depression. Likewise, the variance of change was .02, SE = .01, p < .001 for neuroticism, .04, SE = .01, p < .001 for self-esteem, and .02, SE = .01, p = .007 for depression. In terms of the change direction, on average, neuroticism and depression declined over the 2-year period (m = –.03, SE = .01 and m = –.05, SE = .01, ps < .001), and self-esteem was found to increase across time (m = .07, SE = .01, p < .001). Taken together, all variables exhibited significant inter-individual difference in change over time.

Next, parallel to what we examined using cross-lagged panel models, we examined the relationship between self-esteem and depression using a dual latent growth model (Model 4). Model 4 showed a very good fit to the data, χ² (122) = 354.56, CFI = .99, TLI = .99, RMSEA = .03 and SRMR = .03. We examined the associations between self-esteem and depression by focusing on both the initial levels as well changes of the two constructs. Correlations among the latent intercepts reflect associations among the initial levels of the two variables at T1. We found the levels of self-esteem and depression were highly correlated (r = –.78, SE = .04, p < .001). Correlations among the latent slopes reflect associations between the changes of the two constructs across time. Like the pattern we observed regarding initial levels, change in self-esteem was found to be significantly negatively associated with change in depression (r = –.82, SE = .15, p < .001).³

Last, we constructed the same dual latent growth model but controlled for both the initial level as well as change in neuroticism over time (Model 5, Figure 2). Again, the model showed a good fit to the data (CFI and TLI > .95, RMSEA and SRMR < .05; see Table 3). The association between neuroticism and self-esteem or depression was extremely high: for initial level, β = –.83, SE = .02 and β = .81, SE = .03 respectively, ps < .001, and for change, β = –.82, SE = .08 and β = .81, SE = .12 respectively, ps < .001. When controlling for neuroticism, the association between initial levels of self-esteem and depression dropped from –.78 (Model 4) to –.35⁴ (Model 5), SE = .10 (p = .001), and the association between change in self-esteem and depression dropped from –.82 (Model4) to –.37 (Model5), SE = .49, p = .45. The magnitude of the relationship between the intercept of self-esteem and the slope of depression and vice versa was similar to that of Model 4 (r = .41, SE = .26, p = .12; r = .28, SE = .19, p = .16).

The current research sought to address a set of fundamental questions regarding the relationship between self-esteem and depression. Two theories have been proposed to explain this relationship: the vulnerability model and the scar model. Although a growing body of research has supported the vulnerability model by finding a prospective relationship from low self-esteem to depression (e.g., Orth et al., 2008), two sets of observations raise further questions about this conclusion. First, another variable, neuroticism, has been shown to be strongly related to both self-esteem and depression. Such findings raise the possibility that the relationship between self-esteem and depression may be accounted for by their respective overlap with neuroticism. However, to date, no studies have explicitly tested this hypothesis. Second, the strongest evidence supporting the vulnerability model has come from research employing cross-lagged panel models. However, these cross-lagged panel models have recently been called into question and have been shown to provide biased estimates of the relation between variables like self-esteem and depression over time (Hamaker et al., 2015). To address the aforementioned problems in the past research, the present study examined the relationship between self-esteem and depression while controlling for neuroticism using a variety of modeling techniques. Specifically, we hypothesized that the relationship between self-esteem and depression would be significantly reduced after controlling for neuroticism.

We first sought to reproduce the basic findings of the cross-lag panel regression analyses and determine what effect controlling for neuroticism in these models would have. Like prior research with this sample, the cross-lagged panel regression analyses showed that self-esteem prospectively predicted depression and not the reverse when neuroticism was not incorporated into the model. When neuroticism was controlled for, the prospective relationships from self-esteem to depression or that from depression to self-esteem were not only reduced, but unexpectedly reversed to a significantly positive coefficient, suggesting the models may be mis-specified. In fact, Hamaker et al. (2015) showed that cross-lagged panel models sometimes may reveal reciprocal effects that do not exist. They further demonstrated that such problems often result from the inability of cross-lagged models to adequately separate the within-person and the between-person level when the constructs contain time-invariant, trait-like individual differences. Therefore, when using the original cross-lag panel model while controlling for neuroticism we went one step further than the methodological fix and specified the most likely confound. Consistent with this idea, controlling for the effect of neuroticism not only reduced the relationship between self-esteem and depression, but reversed it. The latter pattern most likely resulted from both the misspecification implicit in the cross-lagged panel model and the importance of neuroticism to the relation of self-esteem and depression.

To better estimate the static and dynamic relations between self-esteem, depression, and neuroticism over time, we modeled these variables using latent growth models. These models specify intercept and growth parameters, and can still be extended to include lagged relations from intercepts to growth parameters. Using these better specified models, we found that neuroticism accounted for most, if not all, of the association between both the level and change of self-esteem and depression. Specifically, in these growth models, the link between overall level in self-esteem and overall level in depression dropped by more than a half from –.60 to –.24 and the link between change in self-esteem and change in depression dropped from –.36 to –.01 when controlling for neuroticism. Meanwhile, both the level and change of neuroticism was highly correlated with those of self-esteem and depression. In addition, the stability coefficients of neuroticism were much higher (average .78) compared to those of self-esteem (average .33) or depression (average .21). Put together, these findings further suggest neuroticism serves as a confounding variable for both self-esteem and depression. That is neuroticism is a confound factor that, in part, explains the relation between the two variables and fully accounts for any dynamic relation between self-esteem and depression.

Our findings imply that neuroticism may be the cause of self-esteem and depression. These findings are also consistent with the abundance of evidence in the clinical literature, which shows that neuroticism predicts most forms of psychopathology, such as depression, anxiety, psychological distress, and substance abuse, to name a few (Kotov, Gamez, Schdmidt & Watson, 2010; Mu, Luo, Nickel & Roberts, 2016). The broad associations observed between neuroticism and various forms of psychopathology have not only led people to theorize neuroticism as a trait vulnerability factor underpinning the risk of developing many forms of psychiatric disorders (see a review for Klein, Kotov, Bufferd, 2011), but also as a higher order factor accounting for the high levels of diagnostic overlap and comorbidity among the wide range of psychopathology (Krueger & Markon, 2006; Watson, 2005). It appears that neuroticism also plays this role for the overlap between self-esteem and depression.

One finding of note was that even after controlling for neuroticism, the initial levels of self-esteem and depression were still significantly correlated (–.24). This suggests there is something left over between the self-esteem and depression, even after neuroticism was controlled for. One possibility is that there is something common to both self-esteem and depression, yet is not captured by neuroticism. This postulation is in line with the findings that refute the common factor model, in which self-esteem and depression are assumed to tap the same construct that overlaps highly with neuroticism (Orth et al., 2008). The inability of neuroticism to fully account for the overlap between self-esteem and depression cautions concluding that self-esteem, depression or neuroticism are indistinguishable constructs.

Whereas the levels of self-esteem and depression is still significantly correlated controlling for neuroticism, the correlation coefficients between changes of self-esteem and depression dropped to almost zero when neuroticism was controlled for. One limitation of the growth modeling approach is that we did not measure these constructs often enough to get an optimal index of change. Thus, it is still possible that some small portion of self-esteem and depression are dynamically related over time. We suspect that better data will be needed to adequately test the relation among dynamic components of neuroticism, depression, and self-esteem. For example, more thorough and continuous assessments of self-esteem and depression, rather than assessment at several years’ interval, would be necessary to provide more reliable estimates of change. Future research should endeavor to conduct deeper assessments of the constructs of interest more often to address such limitations.

What implications do these findings have for the vulnerability and scar models? Numerous studies and have been devoted to exploring this question, many of which involve very well-designed and rigorous longitudinal studies. Despite the accumulating evidence leaning towards the vulnerability model, our finding suggests that neither of these two models is adequate to address this question, at least for the age group examined in the current study (21–25), because they both omit an important confounding variable, neuroticism. Indeed, the relationship between self-esteem and depression disappeared or was even reversed when neuroticism was taken into account. Our finding suggests that future research should switch focus to the role of neuroticism in the development of self-esteem and depression. Does it represent some broad liability factor? What is the genetic or neural underpinning of this liability process?

Some cautions regarding this study should be considered. We did not test our research question in any other datasets besides this one, which limits the generalizability of our results. For example, one limitation is the generalizability of results with participants from Germany to the United States or other cultures. Although there are certainly cultural differences between Germany and other countries, to date, no major differences have been documented on change of personality traits (Ludtke et al., 2011) or associations between self-esteem and depression (Rieger et al., 2016). Furthermore, research on cross-national comparisons have found the rates of depression are similar across countries (Weissman et al., 1996). Nevertheless, it is unclear to what degree cultural influences might affect the relations among neuroticism, self-esteem and depression. Future research should test and see if the findings of the current study can be generalized to other diverse samples.

Another caveat involves the generalizability of our results to other age groups. The TOSCA sample consists of students in young adulthood, a critical period in personality development marked by confluence of multiple developmental tasks (Arnett, 2000) and dramatic increase in multiple personality traits in putatively positive directions (Roberts, Walton & Viechtbauer, 2006). However, given that other age groups have shown differential change trajectories of personality traits (Roberts, Walton & Viechtbauer, 2006; Schwaba & Bleidorn, 2018) and depression (Hankin et al., 1998), the associations observed among neuroticism, self-esteem and depression could be affected by developmental challenges specific to other age groups. Future research should examine the associations among self-esteem, depression and neuroticism as well as their continuity and change in other life stages, such as adolescence or senior adulthood.

It should be noted that our measurement of personality traits and psychological functioning were all based on self-report data. Self-report measures reflect mostly the individual’s own perspective of one’s personality, behaviors and mood, and can be possibly confounded by individual differences in social desirability, response styles and level of insight. Future studies should assess personality, mood using other approaches and perspectives, such as observer ratings. Another issue of employing the same methodology (i.e., self-report) to assess all the three constructs is common method variances (Podsakoff et al., 2003), which could have inflated the observed associations among the constructs of interest. Indeed, in our studies, the absolute values of cross-sectional correlations among self-esteem, depression and neuroticism were quite high, ranging from .61 to .83. Future studies should employ multiple methods in measuring self-esteem, depression and neuroticism, to obtain more comprehensive estimates of associations among the three thus reducing the common method variance.

Relatedly, we measured self-esteem using three items from the original Rosenberg Self-esteem Scale, given the current project is part of large longitudinal panel study and it was difficult to include all items for each scale. Past research has shown that self-esteem can be measured adequately with only one item (Robins et al., 2001). Also, the internal consistency and test-retest reliability of the three-item version were similar to those of its full scale (Rosenberg, 1979). Nevertheless, the content validity of our three-item version has not been formally tested and future research should explore if our findings can be replicated when more thorough measurement of self-esteem is employed.

In conclusion, the present study significantly extends prior research on self-esteem and depression by controlling for neuroticism, and examining not only the concurrent, but also the dynamic relationships among the three variables. Our results suggest that neuroticism is a confound variable that, in part, explains the relation between self-esteem and depression and fully accounts for any dynamic relation between the two variables. It is clear from our results that the relationship between self-esteem and depression may not be meaningful when neuroticism is taken into consideration.

All participant data and analysis scripts are available at the following link: https://osf.io/3rw7x/.

DOI: https://doi.org/10.17605/OSF.IO/3RW7X

This work was supported by a grant to Ulrich Trautwein from the Ministry of Science, Research and the Arts of Baden-Württemberg (Az: 33-7532.20/735).

The authors have no competing interests to declare.

• Contributed to conception and design: Wenting Mu, Brent Roberts, Ulrich Trautwein

• Contributed to acquisition of data: Ulrich Trautwein

• Contributed to analysis and interpretation of data: Wenting Mu, Jing Luo, Sven Rieger, Brent Roberts

• Drafted and/or revised the article: Wenting Mu, Brent Roberts, Sven Rieger

• Approved the submitted version for publication: Brent Roberts, Ulrich Trautwein

Sven Rieger was a doctoral student at the LEAD Graduate School & Research Network [GSC1028] at the time of the study, who was funded by the Excellence Initiative of the German federal and state governments.

The author(s) of this paper chose the Open Review option, and Streamlined Review option, and the peer review comments are available at: http://doi.org/10.1525/collabra.204.pr