Age Differences in the Effects of Self-Esteem

Participants

A convenience sample of community-residing adults (N = 128; M age = 56.7 years; SD = 18.7; range = 24.2–90.2; 46% women) participated in a 7-day daily-diary study. Participants were recruited for a follow-up of a previous larger study (Diehl & Hay, 2007). Because the study focused on the effects of daily stress in healthy community-residing adults, participants were screened out for any major sensory impairments, concurrent depression, or history of severe mental illness. These eligibility criteria were established during a screening interview conducted by phone. Of the original sample of 239 participants, 132 (55%) were successfully recruited for the current study. Four participants were excluded from analyses because they failed to return the baseline although they completed the daily diaries. The overall compliance rate was 98%, with 91% participants completing all 7 days’ diary, 7% missing one diary, and 2% missing two to four diaries.

Most participants identified themselves as Caucasian (91%; n = 116), 9% (n = 11) as Black, and 1% (n = 1) as American Indian or Alaska Native. The majority of participants were married (58.6%), 21.1% were single, 12.5% were divorced, 7.0% were widowed, and 0.8% were separated. Participants’ annual household income ranged from less than $5,000 to over $150,000, with a mean of $59,100. Most participants (85%) had a bachelor’s degree or higher. Participants described themselves as having good vision, M = 4.70; SD = 1.05, and hearing, M = 4.73; SD = 1.02, (1 = very poor ... 6 = very good), and being satisfied with their lives, M = 4.68, SD = .77 (1 = extremely unhappy ... 6 = extremely happy).

Table 1 presents additional baseline information of the three age groups: young adults (n = 33; 24–39 years), middle-aged adults (n = 34; 40–59 years), and older adults (n = 61; 60 and older). At baseline, the three age groups differed significantly from each other in neuroticism, F(2, 125) = 6.36, p < .01, negative affect (NA), F(2, 125) = 6.51, p < .01, and physical symptoms, F(2, 125) = 4.44, p < .05. In contrast, the three age groups did not differ from each other in life satisfaction, subjective health, self-esteem, and positive affect (PA), ps > .05. Results of Spearman’s rank correlation test showed that income, ρ = –.01, and education, ρ = .08, were not significantly associated with self-esteem ps > .05.

Table 1 Descriptive statistics at baseline by age group

	Young adults (n = 33)	Middle-aged adults (n = 34)	Older adults (n = 61)	Correlations
Mean				1	2	3	4	5	6	7
Note. Standard deviations are in parentheses.***p < .001.
1. Age	29.90 (3.95)	54.65 (5.15)	72.37 (8. 46)	–
2. Life satisfaction	4.73 (.63)	4.41 (.82)	4.80 (.79)	.03	–
3. Neuroticism	18.30 (8.83)	17.03 (9.11)	12.16 (8.66)	–.30***	–.34***	–
4. Subjective health	4.82 (1.13)	5.00 (1.02)	5.23 (.84)	.12	.30***	–.17	–
5. Self-esteem	22.85 (4.62)	22.85 (4.32)	24.62 (4.14)	.16	.28***	–.71***	.12	–
6. Positive affect	33.39 (7.08)	34.47 (6.50)	35.62 (5.70)	.11	.35***	–.25***	.11	.34***	–
7. Negative affect	19.70 (6.62)	18.68 (6.65)	15.54 (4.77)	–.29***	–.16	.58***	–.11	–.50***	–.27***	–
8. Physical symptoms	18.39 (4.51)	16.29 (3.35)	16.08 (3.46)	–.18	–.07	.23	–.31***	–.09	–.08	.21

Table 1 Descriptive statistics at baseline by age group

View as image HTML

Procedure

Participants were recruited for a 6-year follow-up study. Participants were sent a baseline questionnaire and paper diaries in the mail. Trained research assistants gave instructions to the participants over the phone on how to complete the baseline questionnaire and paper diaries. Participants were instructed to first complete the baseline questionnaire. They then started the daily diaries the next day. Participants received daily reminders to complete the diaries before they went to bed, via phone calls, emails, or texts. To ensure compliance with the protocol, participants used prepaid envelopes to return the completed baseline questionnaire and each completed diary the next day. Overall compliance was 98%. Participants provided an average of 6.9 diaries (SD = .51; range = 3–7). The final sample consisted of 128 individuals who returned at least six completed diaries. Given our analytic strategy (see the section below: Statistical Analysis: Multivariate Marginal Modeling) only days with complete data on all measures of interest were included in the analyses. Thus, the present analyses were based on 878 days of data (an average of 6.86 days of data per person).

Measures

Measures administered in the baseline assessed a variety of sociodemographic and personal information, including positive affect (PA) and negative affect (NA), self-esteem, physical symptoms, and selected individual difference variables, such as neuroticism (McCrae & Costa, 2003) and psychological well-being (Ryff, 1995). Measures in the daily diaries assessed PA and NA, self-esteem, and physical symptoms on a day-to-day basis. Additional measures in the daily diaries included perceived control (Eizenman, Nesselroade, Featherman, & Rowe, 1997) and a checklist of self-attributes. Below we described the measures pertaining to the research questions of the current study.

Statistical Analysis: Multivariate Marginal Modeling

To examine the associations between affect, physical symptoms, and self-esteem, the typical multilevel modeling analysis would entail a univariate approach. However, the univariate multilevel modeling approach ignores the multivariate response association and does not yield valid statistical inferences (Asar & İlk, 2013). The current study moved the typical univariate analyses one step further and used a multivariate marginal modeling approach.

Models for longitudinal data may be grouped into (1) single-level and (2) multilevel models (Asar & İlk, 2013). In contrast to the multivariate multilevel approach (Raudenbush & Bryk, 2002), which is multilevel, the multivariate marginal approach is single level. For multilevel models, both the regression parameters, random effects, and dependencies are estimated simultaneously. The computation of this group of models leads to statistical difficulties such as convergence problems and computational burden (Asar & İlk, 2013). For marginal models, however, the dependencies are modeled separately and treated as nuisance parameters. Thus, multivariate marginal models allow the interpretation of the regression parameters on the mean response, but not the random effects and dependencies. We applied the multivariate marginal approach in this study because we are primarily interested in the regression parameters, rather than the random effects and dependencies (Asar & İlk, 2013).

Analyses were performed using the R package “mmm” (Asar, 2014; Asar & İlk, 2013). The basic multivariate marginal model can be presented in the following equations.

Y_itj = β_0j + β_1jSelf-Esteem_it + β_{2jPhysical Symptoms}

This model indicates two different models for Y_itj = PA (j = 1) and Y_itj = NA (j = 2) for participant i at time t. Thus, for each assessment at time t, there are two affect variables, namely, PA and NA within participants. In the single-level model, the independent variables entered may be time-invariant or time-varying. In subsequent models, covariates including age, sex, neuroticism, subjective health, and interaction terms were added.

The multivariate marginal modeling approach used generalized estimating equations (GEEs) for parameter estimation (Asar & İlk, 2013; Liang & Zeger, 1986). GEEs use the robust sandwich estimator to estimate the variance-covariance matrix. The robust sandwich variance estimates are insensitive to incorrect choices of the association structures. In contrast, the naive variance estimates are sensitive to the misspecification of the association structures (Asar & İlk, 2013). Interested readers may consult Asar’s work (Asar, 2014; Asar & İlk, 2013) and the corresponding web resources (https://cran.r-project.org/web/packages/mmm/index.html) for details of the theory and application of multivariate marginal modeling.

Descriptive Statistics

Table 1 presents selected baseline information of the three age groups. For the measures administered in the diaries, the intraclass correlation (ICC) for PA was .69 and .43 for NA. The ICCs for self-esteem was .60 and .61 for physical symptoms. The ICCs indicated that PA, NA, self-esteem, and physical symptoms showed substantial within-person variability in the study period. Across individuals and days, mean PA was 31.92 (SD = 8.14; range = 10–50). Mean NA was 13.70 (SD = 4.73; range = 10–37). Mean self-esteem was 24.51 (SD = 5.05; range = 0–30). Mean physical symptoms was 14.21, SD = 3.48, range = 11–38.

Interaction Effect of Self-Esteem and Physical Symptoms

Table 2 presents the results of three nested models. Model 1 tested the effects of self-esteem (within-person) and physical symptoms on PA and NA simultaneously. For PA, self-esteem was associated with a higher level of PA, β₁₁ = .54, robust z = 3.99. The association between PA and physical symptoms was not significant, β₂₁ = –.20, robust z = –1.34. For NA, self-esteem was associated with a lower level of NA, β₁₂ = –.38, robust z = –3.57. Physical symptoms were associated with a higher level of NA, β₂₂ = .15, robust z = 2.03. To address the first hypothesis, Model 2 tested the Self-Esteem × Physical Symptoms two-way interaction on PA and NA. For both PA and NA, the Self-Esteem × Physical Symptoms effect was not significant, β₃₁ = –.03, robust z = –.78, and β₃₂ = .003, robust z = .15.

Table 2 Results of multivariate multilevel modeling

Parameter		Model 1	Model 2	Model 3
		Unstandardized Estimate	Unstandardized Estimate	Unstandardized Estimate
		(Robust SE)	(Robust SE)	(Robust SE)
			Fixed effects
Note. All time-varying covariates are within-person centered. All time invariant covariates are grand-mean centered. *Robust z > 1.98.
PA
Time-varying variables
	Intercept	32.30 (.67)*	32.25 (.66)*	32.26 (.66)*
	Self-esteem	.54 (.14)*	.55 (.13)*	.56 (.11)*
	Physical symptoms	–.20 (.15)	–.22 (.15)	–.08 (.13)
	Self-Esteem × Physical Symptoms		–.03 (.04)	–.07 (.03)*
	Self-Esteem × Age			–.01 (.01)
	Physical Symptoms × Age			.01 (.01)
	Self-Esteem × Physical Symptoms × Age			–.003 (.002)
Time invariant variables
	Age	.06 (.03)*	.06 (.03)*	.06 (.03)*
	Health	1.49 (.46)*	1.49 (.46)*	1.50 (.45)*
	Neuroticism	.12 (.07)	.12 (.07)	.13 (.07)
	Sex	–.74 (.90)	–.70 (.90)	–.71 (.90)
	Self-esteem (between-person)	.74 (.15)*	.76 (.15)*	.76 (.15)*
	Physical symptoms (between-person)	–.04 (.17)	–.06 (.17)	–.06 (.17)
NA
Time-varying variables
	Intercept	13.32 (.32)*	13.32 (.33)*	13.28 (.32)*
	Self-esteem	–.38 (.11)*	–.38 (.10)*	–.39 (.08)*
	Physical symptoms	.15 (.07)*	.15 (.07)*	.13 (.10)
	Self-Esteem × Physical Symptoms		.003 (.02)	.03 (.03)
	Self-Esteem × Age			.11 (.004)*
	Physical Symptoms × Age			.01 (.004)
	Self-Esteem × Physical Symptoms × Age			.004 (.002)*
Time invariant variables
	Age	–.02 (.01)	–.02 (.01)	–.02 (.01)
	Health	–.12 (.32)	–.12 (.31)	–.15 (.31)
	Neuroticism	.12 (.04)*	.12 (.04)*	.12 (.04)*
	Sex	.57 (.56)	.57 (.56)	.62 (.56)
	Self-esteem (between-person)	–.15 (.08)	–.15 (.08)	–.15 (.08)*
	Physical symptoms (between-person)	.28 (.11)*	.28 (.10)*	.26 (.10)*

Table 2 Results of multivariate multilevel modeling

View as image HTML

Age Differences in Interaction Effect of Self-Esteem and Physical Symptoms

Finally, to address the second hypothesis, Model 3 tested the age difference in the interaction effect of self-esteem (within-person) on PA and NA, by including the Self-Esteem × Physical Symptoms × Age three-way interaction. Although no goodness-of-fit indices were available from the mmm package, Model 3 was used for further interpretation for three reasons: (1) our theoretical interest to address the buffering effect of self-esteem, (2) statistically significant Self-Esteem × Physical Symptoms × Age interaction on NA, and (3) interaction terms should not be dropped simply because the interaction effects do not reach statistical significance. For point (3) in particular, an interaction term is included in the model to test the relevant hypothesis. Dropping nonsignificant interaction terms may potentially lead to missing important conditional relationship between variables (Brambor, Clark, & Golder, 2006). We noted that the Self-Esteem × Physical Symptoms effect on PA was significant in Model 3 but not in Model 2. However, effects of lower order terms are almost certain to change with the inclusion of a higher order interaction term (Brambor et al., 2006). According to Brambor et al. (2006) and Friedrich (1982), the change in the effects of lower order terms with the inclusion of a higher-order interaction should not be interpreted as a sign of multicollinearity. In addition, results from multivariate marginal modeling reported here were essentially identical to results using univariate multilevel modeling (Chui & Diehl, 2014, November). The selection of Model 3 for interpretation was further supported based on the goodness-of-fit indices available in results from univariate multilevel modeling.

In Model 3, the Self-Esteem × Physical Symptoms effect was significant for PA, β₃₁ = –.07, robust z = –2.22, but not significant for NA, β₃₂ = .03, robust z = 1.23. In contrast, the Self-Esteem × Physical Symptoms × Age effect was not significant for PA, β₆₁ = –.003, robust z = –1.54, but significant for NA, β₆₂ = .004, robust z = 2.34. Results of Model 3 are graphically presented in Figures 1 and 2 for PA and NA, respectively.

We followed the procedures and recommendations to interpret interaction effects (Aiken & West, 1991; Friedrich, 1982). Figure 1 graphically represents the two-way interaction of self-esteem and physical symptoms on PA. When self-esteem was higher (solid black line), PA was higher compared to when self-esteem was lower (dotted gray line), across all levels of physical symptoms. In addition, the effect of self-esteem on PA depended on the level of physical symptoms. Specifically, the effect of self-esteem on PA was higher when the level of physical symptoms was lower. Thus, the Self-Esteem × Age interaction effect was in the opposite direction as hypothesized, such that, if the buffering effect of self-esteem was present, the effect of self-esteem on PA would be higher at the higher end of physical symptoms (Cohen & Wills, 1985). Figure 2 illustrates the three-way interaction of self-esteem, physical symptoms, and age on NA. Across all three age groups, when self-esteem was higher (solid black line), NA was lower compared to when self-esteem was lower (dotted gray line). In addition, the association between physical symptoms and NA depended jointly on self-esteem and age. The effect of self-esteem on NA was lower when both the levels of physical symptoms and age were higher. In particular, in older adults (Panel c), the effect of self-esteem on NA was lower at the higher end of physical symptoms, such that when the level of physical symptoms was at +2SD, the level of NA was the same regardless of the level of self-esteem. As shown from Figures 2(a) to 2(c), the effect of self-esteem was reduced from young to old age when the level of physical symptoms was higher.

Affect and Health in Everyday Life

Consistent with findings of past cross-sectional, longitudinal, and intensive repeated-measures studies, the daily physical symptoms of adults were associated with negative emotional experiences, including depression and negative affect (Murrell et al., 1991; Zautra, 2003). For example, in an earlier study by Conner et al. (2006), physical pain was associated with negative emotional experiences in individuals with rheumatoid arthritis. Our results advance those reported in intensive repeated-measures studies, such that the application of the multivariate marginal approach (Asar & İlk, 2013) enabled the simultaneous examination of PA and NA in their associations with physical symptoms. Although PA and NA are conceptualized as relatively independent dimensions (Watson et al., 1988), ignoring the multivariate responses within person may not yield valid statistical inferences because of the inaccurate estimates of standard errors (Asar & İlk, 2013; Zeger, Irizarry, & Peng, 2006). In addition, the univariate examination of either PA or NA fails to provide a more complete picture of the complexity of emotional experience in everyday life (Hay & Diehl, 2011). Thus, the findings reported here contribute to the existing literature by examining the associations between PA, NA, physical symptoms, and self-esteem simultaneously.

Self-Esteem as a Resilience Factor across the Adult Lifespan

In accord with past research, our findings provided partial support for the role of self-esteem as a resilience factor in the face of daily physical symptoms (Baumeister et al., 2003; Greenberg et al., 1992). Self-esteem has been associated with various indicators of life success, including better academic and occupational achievement, higher income, better physical health, and more happiness (Leary & Baumeister, 2000; Orth et al., 2012; Trautwein, Ludtke, Koller, & Baumert, 2006). Self-esteem was also shown to buffer the negative effect of death-related existential threats in young adults (Routledge et al., 2010). Most studies, however, focused on self-esteem as a trait-like individual difference variable (Kernis, 2005) and its associations with indicators of life success. Although between-person associations do not necessarily translate into within-person associations (Hamaker, 2012), recent findings supported that the time-varying aspect of self-esteem may have implications for emotional well-being (Paradise & Kernis, 2002; Sowislo et al., 2014). Instead of the degree of stability, the present study moved one step further and examined the buffering effect of within-person self-esteem in the everyday context. Our findings suggest that within-person self-esteem does not buffer the effect of physical symptoms on daily PA or NA. Although within-person self-esteem showed an overall effect on daily PA and NA, its effect on PA was reduced when the level of physical symptoms was higher. In addition, the effect of within-person self-esteem on daily NA became negligible when the level of physical symptoms was high in older adults. Our findings suggest that the effect of within-person self-esteem diminishes in old age when age-related limitations in resources and challenges to well-being become more personally relevant (Orth et al., 2010; Wagner et al., 2013). Similarly, results from longitudinal studies have shown that the level of depressive symptoms increased and the use of coping strategies decreased in old age (Chui, Gerstorf, Hoppmann, & Luszcz, 2015; Diehl et al., 2014). Thus, self-esteem may serve as a psychological resource to cope with minor stress earlier in life. In contrast, age-related vulnerabilities such as chronic health conditions may render the implementation of coping and self-regulation strategies more difficult if not impossible (Charles, 2010).

Self-Esteem in Old Age

Our results suggest that there is no universal generalization of the benefits of self-esteem, in relation to chronological age, that can be made regarding whether self-esteem confers greater resilience on adults (Diehl, Hay, & Chui, 2012). Despite the plethora of research on the benefits of high self-esteem, recent studies examined the vulnerabilities associated with high self-esteem (Crocker & Park, 2004). In particular, the different routes in the pursuit of high self-esteem were examined. When motivated by self-validation goals, individuals react to threats to the self in ways that undermine learning, relatedness, autonomy and self-regulation. Consistent with Crocker and Park (2004), we suggest that the costs associated with the pursuit of self-esteem may differ in different individuals and under different circumstances. In particular, the threat to the self due to a high level of chronic physical symptoms may not be easily dealt with by a boost of temporary self-esteem (Schimel et al., 2008).

Limitations and Outlook

Our findings should be considered in light of several limitations. First, our sample was relatively healthy, high functioning, and community-dwelling. The sample was also predominantly of European-American descent. Research suggests that individuals of more disadvantaged socioeconomic backgrounds, i.e., lower income and educational background, tend to experience more stress and poorer physical health (Gallo, Bogart, Vranceanu, & Matthews, 2005; Gruenewald et al., 2012). Thus, our findings may underestimate the association between affect and physical symptoms in a more diverse and less healthy population.

Second, the present study cannot tease apart the cause-and-effect associations between affect, physical symptoms, and self-esteem in everyday life. Instead of viewing self-esteem as a psychological resource of resilience (Leary & Baumeister, 2000), some studies view self-esteem as the outcome of good health, better education, and more positive personality profiles (Wagner, Lang, Neyer, & Wagner, 2014). We cannot ascertain whether PA and NA reported were reactions to physical symptoms or other intervening events. However, in an evaluation of the prospective reciprocal relation between self-esteem and depression, results from a meta-analysis of longitudinal studies suggested that self-esteem predicts depression (Sowislo & Orth, 2013). In particular, the effect of self-esteem on depression (β = –.16) was significantly stronger than the effect of depression on self-esteem (β = –.08). Future research might include a lead/lag effect of self-esteem to examine the reciprocal relation between multiple domains in everyday life using the intense repeated-measure design (Stawski et al., 2015).

Third, recent studies applied statistical models such as the latent trait-state model to quantify the extent to which self-esteem is a trait or a state (Anusic & Schimmack, 2016; Hank, 2015; Wagner et al., 2015). Using repeated measures of different time-scales spanning from days to years, these studies examined the between-person and within-person variances in self-esteem. In general, trait self-esteem pertains to the between-person variance. In contrast, state self-esteem pertains to within-person variance. It is not clear, however, whether trait self-esteem measured using the original self-esteem scale (Rosenberg, 1989) may or may not be comparable to the measure of trait self-esteem derived from the between-person variances in repeated measures of self-esteem. Although self-esteem was measured using the original Rosenberg’s self-esteem scale at baseline, the conceptualization and estimation of trait and state self-esteem were not the focus of the present study. Future studies should examine the similarities and differences in trait self-esteem measured at baseline, compared to trait self-esteem derived from the between-person variance in repeated measures.

Fourth, in this study age was a between-person variable. Thus, the effects of cohort and age differences are confounded (Li & Schmiedek, 2002). We cannot ascertain whether the age difference in the effect of self-esteem stems from age-related changes or cohort differences. Past studies found evidence both for and against cohort differences in self-esteem (Gentile, Twenge, & Campbell, 2010; Trzesniewski et al., 2003; Twenge & Campbell, 2001). However, less is known about the cohort differences in the buffering effect of self-esteem (Orth, Robins, & Meier, 2009). Future studies should adopt multiple time-scales as different metrics of time to disentangle the effects of cohort and age in the buffering effect of self-esteem (Gerstorf, Hoppmann, & Ram, 2014; Li & Schmiedek, 2002).

Fifth, although statistical methods, such as the region of significance test, are available to evaluate interaction effects in univariate multilevel modeling (Preacher, Curran, & Bauer, 2006), no equivalent computational tools are available for multivariate marginal modeling. The interaction effects reported in the present study cannot be evaluated using the Johnson-Neyman technique or confidence bands. Future developments of statistical methods and implementations in software are needed to enable better interpretation of two-way and three-way interaction effects in multivariate marginal modeling.

The last limitation affects all intensive repeated-measures designs in real life: The strength of an intensive repeated-measure design is that the phenomenon of interest is examined as it unfolds in the natural environment; thus, ecological validity is enhanced. However, unlike laboratory-based studies, our study cannot control for all other factors that may impact the affective experience in everyday life. Despite these limitations, this study offers insights for future research. Our findings support the assertions that interrelationships across multiple domains should be considered in the research of affective experiences in everyday life (Ong & Zautra, 2015). In addition to physical symptoms, research should also examine multiple types of stressors, as they may be associated with different outcomes. For instance, among stressors of different domains, social stressors were particularly relevant for physical symptoms (Hay & Diehl, 2010).

Conclusion

This study showed that the effect of self-esteem on the association between physical symptoms and affect differed between PA and NA, and in different age groups. In contrast to our initial hypotheses, self-esteem did not act as a buffer on daily PA or NA, although self-esteem showed an overall effect on daily affect. The self-esteem literature has provided a plethora of evidence on the important role of self-esteem as a psychological resource of resilience. Our findings attest to the value of the examination of the role of self-esteem as a resilience factor in the lifespan development of emotional well-being. These findings converge with results from the aging literature that late life entails increasing vulnerabilities that challenge individuals’ well-being (Baltes & Smith, 2003). Future research may substantiate our findings by examining both vulnerability and resilience that self-esteem may bring in coping with the varying challenges across the adult lifespan.