Larger Increase in Trait Negative Affect is associated with Greater Future Cognitive Decline and Vice Versa Across 23 Years

Abstract

Background:

Trait negative affect (NA) is a central feature of anxiety and depression disorders. Neurocognitive and scar models propose that within-person increase in NA across one period of time relates to decline in cognitive functioning at a future period of time and vice versa. Yet there has been little research on whether within-person change in trait NA across one time-lag precedes and is associated with change in cognition across a future time-lag and vice versa. Due to a growing aging population, such knowledge can inform evidence-based prevention.

Methods:

Participants were 520 dementia-free, community-dwelling adults (mean age = 59.76 years (SD = 8.96), 58.08% females). Trait-level NA (negative emotionality scale), spatial cognition (block design, card rotations), verbal WM (digit span backward), and processing speed (symbol digit modalities) were assessed at five time-points (waves) across 23 years. Bivariate dual latent change score (LCS) approaches were used to adjust for regression to the mean, lagged outcomes, and between-person variability.

Results:

Unique bivariate LCS models showed that within-person increase in trait NA across two sequential waves was related to declines in spatial cognition, verbal WM, and processing speed across the subsequent two waves. Moreover, within-person reductions in spatial cognition, verbal WM, and processing speed across two sequential waves were associated with future increases in trait NA across the subsequent two waves.

Conclusions:

Findings concur with neurobiological and scar theories of psychopathology. Further, results support process-based emotion regulation models that highlight the importance of verbal WM, spatial cognition, and processing speed capacities for downregulating NA.

Every day, many of us rely on our cognitive capacities, such as spatial cognition, processing speed, and verbal working memory (WM), to function healthily, as they are tied to important activities. Such everyday activities include self-regulating, staying on-task, and tracking conversations (Baddeley, 2012). Verbal WM refers to the resource-limited capacity to retain and modify incoming task-salient auditory information in real-time (Baddeley, 2012). Spatial cognition is the capacity to create, maintain, recall, and/or transform well-formed images (Lohman, 1996). Relatedly, processing speed is the ability to fluently perform basic or overlearned cognitive tasks, particularly when high levels of sustained focus are needed (McGrew, 2009). As verbal WM, spatial cognition, and processing speed are important for daily pursuits, their shortcomings have been related to issues with organizing, scholastic achievement, reality testing, personality, physical health, sleep, and psychological well-being (Serper et al., 2014; Willcutt et al., 2013). Moreover, verbal WM, spatial cognition, and processing speed are entwined with basic and higher-order cognitive control processes, such as attention, language, memory, and comprehension (Baddeley, 2003; Etkin, Büchel, & Gross, 2015). Understanding factors linked to verbal WM, spatial cognition, and processing speed problems is thus essential.

In the past 50 years, neurocognitive theories have asserted that verbal WM, spatial cognition, and processing speed depletion might be a correlate of high trait negative affect (NA). Dispositional NA (or neuroticism) is the propensity to experience diverse negative emotions, and is regarded as one of the key personality traits germane to psychopathology, especially anxiety, depression, and trauma-based disorders (Craske, 2003; Watson, Clark, & Stasik, 2011). Specifically, dispositional NA is of central importance to clinical science given its consistent positive cross-sectional and longitudinal relations with depression, anxiety, and non-specific distress (d = 0.48–0.74 across 443313 persons aged 14 to 104 years; Jeronimus, Kotov, Riese, & Ormel, 2016). Attentional control and processing efficiency models assert that heightened NA consumes WM resources and adversely affects processing speed and task accuracy in real-time (Eysenck & Calvo, 1992; Eysenck, Derakshan, Santos, & Calvo, 2007). To date, data accrued from over 600 cross-sectional studies support these theories. Meta-analyses have shown that cognitive functioning problems coincide with diverse emotional disorders, such as major depression (Snyder, 2013), obsessive-compulsive disorder (Snyder, Kaiser, Warren, & Heller, 2015), chronic anxious arousal and worry (Moran, 2016), and schizophrenia (Forbes, Carrick, McIntosh, & Lawrie, 2009). Collectively, a large body of cross-sectional studies have documented mostly small yet significant inverse NA-cognitive functioning relations.

At the same time, neurocognitive and scar theories propose that change across time in verbal WM, spatial cognition, and processing speed might lead to subsequent increases in trait NA over lengthy timescales. For example, scar models, such as the perseverative cognition hypothesis (Ottaviani et al., 2016), posit that increased (vs. lowered) trait NA creates allostatic load in the hypothalamic-pituitary axis (i.e., wear and tear on the biological systems via buildup of inflammatory markers). Such chronic wear-and-tear is associated with subsequent increases in cognitive impairment over prolonged durations (e.g., 9 to 17 years later; Gimson, Schlosser, Huntley, & Marchant, 2018). Further, these scar models argue that increases in NA across time lead to subsequent cognitive decline in systems involved in processing (i.e., processing speed, verbal WM, spatial cognition; Wilson, Begeny, Boyle, Schneider, & Bennett, 2011). Moreover, developmental psychopathology models (Hur, Stockbridge, Fox, & Shackman, 2019) postulate that increases in cognitive dysfunction relate to future increases in NA throughout life. Taken together, change in NA and change in cognitive functioning could be inversely associated across sequential time periods encompassing long durations.

At least five studies have examined these scar and developmental psychopathology theories. First, initial levels of higher neuroticism were connected with initial levels of poorer processing speed, visuospatial WM, and recognition memory in Swedish adult twins (Wetherell, Reynolds, Gatz, & Pedersen, 2002); however, in this study, neuroticism was not associated with 9-year cognitive decline. Another study found that older adults’ heightened anxiety symptoms correlated with decline in verbal memory over 12 years (Gulpers, Oude Voshaar, van Boxtel, Verhey, & Köhler, 2019). Relatedly, higher self-reported depression symptom severity was associated with dementia after nine years (Köhler, van Boxtel, Jolles, & Verhey, 2011). The opposite relation was also observed, such that inhibition, shifting, WM, and other cognitive skills (e.g., verbal reasoning) were related to GAD severity and diagnosis from middle- to late-adulthood (e.g., across nine to 12 years; Zainal & Newman, 2018; Zhang et al., 2015a). Based on this emerging body of longitudinal evidence, it is plausible that elevated NA could be related to subsequent cognitive dysfunction in the long-term.

However, single time-point and two-wave regression studies present a major limitation to the examination of these theories as well as to the psychological sciences and public health fields. It is more crucial to examine how preceding change in cognitive functioning relates to subsequent change in NA (or vice versa) within-persons, across adulthood. Although cross-sectional studies offer ample essential data on average individual differences (between groups) at one time-point, they often do not reveal key within-person (idiographic) change processes across decades (Molenaar, 2004). Noteworthy is the fact that between- and within-person relations among psychological and health variables tend not to coincide in magnitude and direction (Fisher, Medaglia, & Jeronimus, 2018). Uncovering how cognitive functioning and NA evolve within persons over long periods is critical. Many societies worldwide are increasingly facing challenges linked to aging populations, rising life expectancy, and growing prevalence of neurocognitive disorders (e.g., dementia; Chang, Skirbekk, Tyrovolas, Kassebaum, & Dieleman, 2019). Thus, knowledge of factors that are related to the long-term course of cognitive decline or increasing NA within persons can inform evidence-based preventative interventions. Moreover, the aforesaid theories propose that NA-cognition connections occur both between and within persons. Adopting within-person approaches with respect to statistical methods and study design is thus critical to understanding reciprocal idiographic shifts in cognitive functioning and future changes in the propensity to experience emotional distress.

A potent technique that achieves these goals is latent change score (LCS) modeling, a type of structural equation modeling (SEM) (Grimm, An, McArdle, Zonderman, & Resnick, 2012). Compared to conventional non-confirmatory methods (e.g., repeated-measures (M)ANOVA) that are seldom attuned to the intricate error structure of prospective datasets, SEM in general, and LCS models specifically, offer several benefits. SEM can handle missing data optimally by using all available data points (vs. listwise deletion), adjusts for regression to the mean and lagged outcome effects (Falkenström, Finkel, Sandell, Rubel, & Holmqvist, 2017). Moreover, LCS modeling empowers researchers to examine within-person prospective changes by accounting for stable individual differences and reducing measurement error (Hamaker, Kuiper, & Grasman, 2015; Voelkle, 2007). Further, bivariate dual LCS models can test if within-person change in a variable across one time-lag is related to change in another variable at a subsequent time-lag, and thus permits inference regarding temporal precedence of changes (Zainal & Newman, 2019). By testing within-person lead-lag relations, bivariate dual LCS models move us closer toward causality (vs. parallel process latent growth SEM that informs concurrent, between-person relations between slopes of temporal change in unique variables). Bivariate dual LCS modeling can, for example, test if within-person 3-year change in a specific cognitive domain relates to future 3-year change in trait NA across decades (and vice versa), a question of great import to the realms of public health and cognitive sciences.

To date, at least four studies have examined the dynamic, within-person, prospective links between NA and cognition or related constructs using LCS analyses. Within persons, better caregiver-rated emotion regulation of preschoolers was connected with future 1-year improvement in cognitive functioning (Blankson et al., 2013); despite that, the opposite effect was not observed, and the two-time-point study hindered tests of change-to-future change relations. Relatedly, in Australian adolescents, 1-hour increase in worry vulnerability dovetailed with subsequent 1-hour reduction in performance on WM-based algebraic tests (and conversely) thrice within a day (Trezise & Reeve, 2016); however, it remains unknown if findings extrapolate to longer time-lags as proposed by scar theories. Another study indicated that possibility by showing that across two years, higher initial anxiety coincided with depleted future cognition in older adults (Tetzner & Schuth, 2016); however, its two-time point design does not speak to whether change in verbal WM preceded and related to future change in trait NA (and vice versa). Relatedly, another study observed that 3-year rise in anxiety proneness was associated with steeper future 3- to 6-year declines in processing speed and attention in older adults (Petkus, Reynolds, Wetherell, Kremen, & Gatz, 2017); however, whether findings extend to broader trait NA across more than two decades starting from middle-to-late adulthood continues unanswered. This is an important empirical question. Although on average, trait NA tends to decline from middle-to-late adulthood (Terracciano, McCrae, Brant, & Costa, 2005), there is large variability in individual change trajectories (Mroczek & Spiro, 2003).

Building on this growing yet nascent body of evidence, the current study aimed to answer the following unexplored question: In dementia-free middle-aged adults, how does within-person change across one time lag in three specific cognitive facets (spatial cognition, verbal WM, processing speed) relate to future change in NA (and vice versa) across a sequential time lag using 5 assessment points spanning 23 years? The current study adds to prior literature in a number of ways. First, high trait NA cuts across various psychological disorders (Böhnke, Lutz, & Delgadillo, 2014). Thus, shedding light on how within-person change in trait NA relates to future change in cognition can offer insights into the development of neuropsychiatric disorders in late adulthood. Moreover, high trait NA in cognitively-intact older adults is common (e.g., Stanley & Novy, 2000). Also, although theories suggest that cognitive functioning facets like WM are necessary for optimal emotion regulation (Teper, Segal, & Inzlicht, 2013), seldom have such theories been tested using LCS methods. In the context of aging, sequential intervals across 23 years offer unique perspectives. Last, emotion regulation theories which posit inverse NA-cognition relations (e.g., Schmeichel & Tang, 2015) emphasize that such relations involve NA at the trait (or individual difference) level. This means that adults who are more prone to experiencing poorly managed spirals of anxiety or depressive symptoms are likely to be more susceptible to future cognitive decline over time. Accordingly, we hypothesized that across 23 years and 4 sequential time lags, within-person increased NA across a preceding time lag would be significantly related to reductions in cognitive functioning facets – spatial cognition (Hypothesis 1), verbal WM (Hypothesis 2), and processing speed (and vice versa; Hypothesis 3) – over the subsequent time lag.

Method

Participants.

This was a secondary analysis of the Swedish Adoption/Twin Study of Aging (SATSA; Finkel & Pedersen, 2004) publicly available dataset (https://www.icpsr.umich.edu/icpsrweb/NACDA/studies/3843). Ethics approval was obtained from the Karolinska Institutet, and all participants provided informed consent. The current study focused on the 560 participants who consented to complete both self-reports (including the trait NA measure used herein) and in-person cognitive testing. Of these, we removed 40 participants diagnosed with dementia at baseline to examine our hypotheses in a relatively cognitively intact sample. Selected participants (n = 520) averaged 59.76 years of age (SD = 8.96, range = 40.74 – 84.28), 58.08% were female, and 59.04% had tertiary education at baseline.

Procedures.

In the present study, we examined the phenotypic relations between cognition (verbal WM, spatial cognition, processing speed) and NA as opposed to genetic heritability. The SATSA researchers administered in-person the following measures of trait-level NA, spatial cognition, verbal WM, and processing speed at all 5 waves of data collection herein.

Negative affect was measured with the 4-item trait-level negative emotionality-distress scale (Buss & Plomin, 1984) (i.e., “I frequently get distressed.”; “Everyday events make me troubled and worried”; “I get emotionally upset easily.”; and “I often feel frustrated.”). Items were endorsed on a 5-point Likert scale (1 = not characteristic of me to 5 = completely characteristic of me). In this study, between-person internal consistency (α) was acceptable at .73 on average across all waves. The scale items have also demonstrated acceptable two-week retest-reliability (r = .53–.63) and construct validity (Naerde, Roysamb, & Tambs, 2004). Within-person α across waves (.71) was also good herein.¹

Verbal WM was assessed using the backward digit span test (DST) (Finkel & Pedersen, 2004). Participants repeated backward three to nine-digit sequences of increasing length. Total score was derived by adding the highest number sequences correctly recalled. Whereas DST forward tests attention, DST backward taps into attention and WM processes reliably (Pedersen, Plomin, Nesselroade, & McClearn, 1992; Wechsler, 2008). The DST has good between-person α (.93) and within-person α (.89) herein, and two-week retest-reliability (r = .83) (Wechsler, 2008). It also has good convergent and discriminant validity (Pomplun & Custer, 2005).

Spatial cognition was indexed as scores on the card rotations test (CRT) and Koh’s block design test (BDT) (Pedersen et al., 1992). For the CRT, participants were presented with a target design and four options for each item. Their task was to accurately choose the item that would form the target upon mental rotation. The BDT mirrored the Wechsler’s Adult Intelligence Scale (Wechsler, 1997) such that participants used blocks to produce seven forms of design. For both tasks, scores were based on completion time (in seconds) for each design. These spatial cognition measures have excellent between-person α (.91) and within-person α (.83) in this study, 33-day retest-reliability (r = .90) (Brand, Pieterse, & Frost, 1986), and strong convergent and discriminant validity (Finkel, Reynolds, McArdle, & Pedersen, 2007).

Processing speed was assessed with the symbol digit modalities test (SDMT) (Pedersen et al., 1992). Respondents verbally stated numbers that matched unique symbols within a limited time. SDMT scores (digits accurately linked to each symbol) can range from 0 to 100. In this study, this two-part SDMT had good between-person α (.96) and within-person α (.91). The SDMT also has good two-week retest reliability (r =.74) (Hinton-Bayre & Geffen, 2005), convergent validity with other processing speed measures, and discriminant validity with assessments of other constructs (Benedict et al., 2017).

Data Analyses.

The original study contained six waves of assessment: 1984 (Time 1 [T1]); 1987 (Time 2 [T2]); 1990 (Time 3 [T3]); 1993 (Time 4 [T4]), 2004 (Time 5 [T5]); and 2007 (Time 6 [T6]). However, because of funding deficits during the course of data collection, in-person testing was largely absent at T4 resulting in 5 available data points for each individual and 4 sequential time lags with each lag spanning anywhere from 3 to 14 years. For the current study, we tested whether within-person change across each time lag (t) was related to change across the sequential lag (t+1).

To test the fit of our measurement models, we used SEM with the lavaan (Rosseel, 2012) R package and used practical fit indices and heuristic cut-offs: confirmatory fit index (CFI; Bentler, 1990), Tucker-Lewis Index (TLI; Tucker & Lewis, 1973), root mean square error of approximation (RMSEA; Steiger, 1990), and standardized root mean square residual (SRMR; Hu & Bentler, 1999). As our continuous data were clustered within twins, we used the lavaan.survey (Oberski, 2014) package to account for nesting of data within twins. We added the twin identification number as a cluster variable (each person was assigned a distinct twin number), and selected maximum likelihood (ML) with robust sandwich estimators. In total, 3.85% of our data were missing completely at random (MCAR), based on the MCAR test (Little, 1988) (χ²(df = 10,865) = 120.47, p = 1.000). Full information ML (FIML) was used to handle missing data as this method uses all available data (vs. listwise deletion) to estimate model parameters (Graham, 2009). Moreover, FIML was a suitable approach as the data was missing at random (Dominicus, Ripatti, Pedersen, & Palmgren, 2008).

Next, as a prerequisite to using bivariate dual LCS models, we tested for equivalence of all measures across each and every wave of assessments (i.e., four-item NA, one-item spatial cognition², one-item verbal WM, and one-item processing speed measures). Longitudinal measurement invariance (LMI) determines the degree to which the assessments were measured along the same scale at all time periods, based on their equivalence of factor loadings (λs), intercepts (τs), and residual variances (εs) (Millsap & Yun-Tein, 2004). LMI was tested in the following order: (1) configural (equal factor structures across waves); (2) metric (λs fixed equal, τs and εs varying freely across waves); (3) scalar (λs and τs fixed equal, but εs varying freely across waves); and (4) strict (λs, τs, and εs fixed equal across waves). At each step, change in practical fit indices were used to assess measurement equivalence (Meade, Johnson, & Braddy, 2008). Change in practical fit indices (ΔCFI < −.010, ΔTLI < −.015, ΔRMSEA > +.015, ΔSRMR > +.030) from the unconstrained to constrained model suggest that the unconstrained model is preferable. Table 1 establishes this sequence of steps to establish LMI.

Table 1.

Longitudinal measurement invariance of the four-factor model of NA and cognition

	χ2	df	p	CFI	TLI	RMSEA	SRMR
Four-factor model of NA and cognition (spatial WM, verbal WM, processing speed)
Time 1 (1986–1988)	42.417	16	< .001	.972	.951	.064	.060
Time 2 (1989–1991)	24.852	16	.072	.993	.989	.037	.037
Time 3 (1992–1994)	17.911	16	.329	.999	.998	.017	.037
Time 4 (2002–2004)	29.408	16	.021	.986	.976	.057	.059
Time 5 (2005–2007)	17.265	16	.369	.998	.997	.020	.053
Level of measurement invariance
Configural (varying λ, τ, ε across time)	35.888	80	1.000	1.000	1.072	.000	.052
Metric (equal λ, varying τ, ε across time)	50.738	96	1.000	1.000	1.062	.000	.060
Scalar (equal λ, τ, varying ε across time)	66.125	112	1.000	1.000	1.054	.000	.068
Strict (equal λ, τ, ε across time)	100.452	136	.990	1.000	1.034	.000	.086
Tests of measurement invariance
Configural invariance vs. metric invariance	Δχ2(df = 16) = 14.850, p = .464, ΔCFI = .000, ΔTLI = −.010, ΔRMSEA = .000, ΔSRMR = +.008
Metric invariance vs. scalar invariance	Δχ2(df = 16) = 15.387, p = .504, ΔCFI = .000, ΔTLI = −.008, ΔRMSEA = .000, ΔSRMR = +.008
Scalar invariance vs. strict invariance	Δχ2(df = 24) = 34.327, p = .921, ΔCFI = .000, ΔTLI = −.0194, ΔRMSEA = .000, ΔSRMR = .018

Note. NA = negative affectivity; WM = working memory; CFI = confirmatory fit index; RMSEA = root mean square error of approximation; λs = factor loadings; τs = intercepts; εs = residual variances.

Subsequently, to separate between- and within-person effects, establish temporal precedence, and adjust for regression to the mean and lagged outcome effects, we used bivariate dual LCS models. As bivariate dual LCS models approximate causality by combining cross-lagged panel and parallel process latent growth SEM, they determine if within-person change in a variable at a time-lag relates to change in another variable at a later time-lag (Grimm & Ram, 2018). LCS thus models true, within-person latent change of a variable across two adjacent occasions while minimizing measurement error (McArdle, 2009). The prospective trajectory of a variable can thus be modeled in terms of its initial status and latent score difference between each occasion. Equation (1) computes within-person change in LCS models:

$$\Delta {\text{X}}_{\left[t\right]}={\alpha }_{\text{x}}{}^{*}{\text{X}}_s+{\beta }_{\text{X}}{}^{*}{\text{X}}_{\left[t-1\right]}$$ (1)

where ΔX_[t] signifies the latent change in variable X at occasion t, α_X signifies the between-person constant change parameter (usually with equality constraints imposed across occasions) connected to the latent slope (X_S), and β_X signifies the within-person self-feedback loops (same construct relating to its future change). Equation (1) reflects the dual LCS model, wherein the constant change parameter (α_X) and proportional effect (β_X) models within-person trajectories of change as the buildup of changes between measurements up to that time. Based on the fact that permitting the coupling and proportional change parameters to vary freely increased parameter estimates’ standard errors, we fixed them to be equal to guide interpretation and retain parsimony in all our models; an approach aligned with recommended practices (see Figure S4 which depicts a bivariate dual LCS model in McArdle, 2009). Moreover, unstandardized regression coefficients and 95% confidence intervals [CIs] are presented in this paper. To facilitate interpretation of parameter estimates, Cohen’s d effect sizes were computed using the following formula for effect sizes: d = t√(2(1 – r)/N) where r = √(t²/(t² + df)) (Dunlap, Cortina, Vaslow, & Burke, 1996; Dunst, Hamby, & Trivette, 2004), where t refers to the t-statistic of the parameter estimate, N the sample size, and df the degrees of freedom of the model. Cohen’s d of 0.20, 0.50, and 0.80 denoted small, moderate, and large effects, respectively.

Afterwards, we conducted bivariate dual LCS models to test the change-to-future change dynamic relations between NA and cognition (spatial cognition, verbal WM, or processing speed). In bivariate dual LCS models, within-person change is captured in the equation below:

$$\Delta {\text{X}}_{\left[t\right]}={\alpha }_{\text{x}}{}^{*}{\text{X}}_s+{\beta }_{\text{X}}{}^{*}\Delta {\text{X}}_{\left[t-1\right]}+{\delta }_{\text{Y}}{}^{*}\Delta {\text{Y}}_{\left[t-1\right]}$$ (2)

Equation (2) expands the prior equation (1) by including a within-person change-to-future change coupling parameter (δ_Y) that signifies the impact of change (Δ) in one variable on future change in another variable. We included the coupling parameters (δ_Y) for both outcomes (Δcognition and ΔNA) to permit the test of bi-directional, within-person, temporal dependencies in change of two variables. The δ_Y values are the most important estimates in these bivariate dual LCS models as they indicate the impact of within-person change in one variable on future change in the other variable, accounting for autoregressive self-feedback processes (β_X). Figures 1 to 3 show each bivariate dual LCS model (i.e., how latent and observed variables relate to each other). Also, Figures S1 to S3 in the Supplementary Materials show the coupling effects.

Figure 1. Bivariate dual LCS models of NA and visual-spatial processing.

Note. SPT = visual-spatial processing; LCS = latent change score; NA = negative affect; T1 = time 1; T2 = time 2; T3 = time 3; T4 = time 4; T5 = time 5; λ = factor loadings; β = autoregressive feedback loop of within-person change in a construct predicting for future within-person change in itself; Δ denotes change in a construct. Of most interest to this study is the coupling parameter (δ_Y) that indicates the effect of prior within-person change in NA severity correlating with future within-person change in visual-spatial processing (and vice versa) while adjusting for other parameters in the model.

Figure 3. Bivariate dual LCS models of NA and processing speed.

Note. PS = processing speed; LCS = latent change score; NA = negative affect; T1 = time 1; T2 = time 2; T3 = time 3; T4 = time 4; T5 = time 5; λ = factor loadings; β = autoregressive feedback loop of within-person change in a construct predicting for future within-person change in itself; Δ denotes change in a construct. Of most interest to this study is the coupling parameter (δ_Y) that indicates the effect of prior within-person change in NA severity correlating with future within-person change in processing speed (and vice versa) while adjusting for other parameters in the model.

Power analysis.

Using the RAMpath R package (Zhang, Hamagami, Grimm, & McArdle, 2015b), we conducted a power analysis based on a priori Monte Carlo simulations that paralleled study conditions, a recommended approach (Green & MacLeod, 2016). Power was based on an effect size of d = 0.20 for the coupling effect (change-to-future change NA-cognition relation, and vice versa). Following 1,000 replications per condition, we found 95.48% to 98.65% power to identify significant within-person coupling effects.

Results

Attrition.

Across the 5 waves, the average missing data rate was 3.85%. Attrition across waves was not predicted by age (odds ratio [OR] 95% CI = −0.01–0.10), gender (OR 95% CI = −0.80–0.90), or education level (OR 95% CI = −0.47–1.22). Attrition status also showed negligible-to-small correlations with key study variables of interest (see Table S1 of Supplementary Materials).

Longitudinal measurement equivalence.

Table 1 shows that across the 5 waves of data collection, the four-factor measurement model of cognition (spatial cognition, verbal WM, processing speed) and NA showed invariant factor structure (configural invariance), item loadings (λs) (metric invariance), item intercepts (τs) (scalar invariance), and residual errors (θs) (strict invariance). Comparison of latent scores across time is thus meaningful as the measures have been found to produce equal and stable measurement properties.

Univariate LCS models.

Table 2 demonstrates that in the entire sample, normative between-person reductions in spatial cognition ability (β = −13.92, 95% CI [−19.09, −8.76], d = −0.10), verbal WM (β = −3.19, 95% CI [−5.33, −1.04], d = −0.11), and processing speed (β = −12.01, 95% CI [−16.67, −7.35], d = −0.11) were observed across 23 years. However, no between-person changes occurred for NA (β = 0.67, 95% CI [−2.10, 3.43], d = 0.03). Further, between persons, initial status was substantially negatively related to latent change for each of these constructs with small effect sizes (ds = −0.11 to −0.02).

Table 2.

Univariate latent change score models of each variable

		NA				Spatial Cognition
Within-person proportional change (β)		β	[95% CI]	ES d		β	[95% CI]	ES d
		−0.04	[−0.27, 0.18]	−0.02		0.34***	[0.192, 0.491]	0.27
Initial status
Mean		12.32***	[12.10, 12.54]	2.29		36.68***	[35.62, 37.75]	0.51
Variance		3.95***	[3.23, 4.67]	0.46		127.40***	[111.70, 143.11]	0.03
Constant change (α)
Mean		0.67	[−2.10, 3.43]	0.03		−13.92***	[−19.09, −8.76]	−0.10
Variance		0.12	[−0.07, 0.31]	0.08		14.54***	[1.93, 27.15]	0.04
Correlation between initial status and α		−0.11	[−0.76, 0.55]	−0.02		−41.97***	[−62.09, −21.84]	−0.03
Residual error
${\sigma }_e^2$		2.72***	[2.35, 3.08]	0.70		29.44***	[26.61, 32.27]	0.19
Model fit
χ2		62.308				63.031
df		13				13
p		< .001				< .001
CFI		.925				.969
TLI		.942				.976
RMSEA		.086				.088
SRMR		.090				.051
		Verbal WM				Processing Speed
Within-person proportional change (β)		β	[95% CI]	ES d		β	[95% CI]	ES d
		0.73***	[0.21, 1.25]	0.11		0.23***	[0.11, 0.35]	0.23
Initial status
Mean		4.26***	[4.15, 4.37]	0.16		41.76***	[40.81, 42.72]	0.56
Variance		0.95***	[0.75, 1.15]	0.15		108.22***	[92.55, 123.88]	0.03
Constant change (α)
Mean		−3.19***	[−5.33, −1.04]	−0.11		−12.01***	[−16.67, −7.35]	−0.11
Variance		0.53	[−0.20, 1.26]	0.08		6.08***	[0.21, 11.96]	0.07
Correlation between initial status and α		−0.71***	[−1.21, −0.21]	−0.11		−23.20***	[−37.59, −8.80]	−0.04
Residual error
${\sigma }_e^2$		0.94***	[0.82, 1.06]	0.15		27.20***	[23.11, 31.29]	0.13
Model fit
χ2		31.779				75.714
df		13				13
p		.003				< .001
CFI		.954				.941
TLI		.965				.942
RMSEA		.053				.097
SRMR		.097				.069

Note.

^*p < .05;

^**p < .01;

^***p < .001;

NA = negative affectivity; VSP = visual-spatial processing; df = degrees of freedom; CFI = confirmatory fit index; RMSEA = root mean square error of approximation.

Spatial Processing and NA.

Table 3 shows that the bivariate dual LCS model testing the mutual reciprocal influences of change in spatial cognition (ΔSpatial Processing) relating to future change in NA (ΔNA) (and conversely) had good fit based on practical fit indices. Within persons, an increase in NA across each preceding time lag (t) was significantly related to a decrease in spatial cognition across the sequential time lag (t + 1), with a small-to-moderate effect (β = −0.20, 95% CI [−0.28, −0.12], d = −0.29). Also, within-person depletion in spatial cognition across t was substantially related to rise in NA across t + 1, with small effect (β = −4.03, 95% CI [−6.12, −1.94], d = −0.16). Findings thus supported Hypothesis 1.

Table 3.

Bivariate latent difference score models of spatial cognition and NA

Parameter estimate	Spatial Cognition and NA
Within-person bivariate change-to-change coupling effects (δ)	ΔSpatial CognitionT – 1 ➔ ΔNAT				ΔNAT – 1 ➔ ΔSpatial CognitionT
	β	[95% CI]	d		β	[95% CI]	d
	−4.03***	[−6.12, −1.94]	−0.16		−0.20***	[−0.28, −0.12]	−0.29
Within-person proportional change (β)	ΔSpatial CognitionT– 1 ➔ ΔSpatial CognitionT				ΔNAT – 1 ➔ ΔNAT
	0.11	[−0.07, 0.30]	0.07		−0.05	[−0.23, 0.14]	−0.03
	Spatial Cognition				NA
Initial status	β	[95% CI]	d		β	[95% CI]	d
Mean	36.50***	[35.43, 37.56]	0.51		12.35***	[12.13, 12.57]	2.25
Variance	126.39***	[110.24, 142.53]	0.29		4.08***	[3.30, 4.80]	0.46
Constant change (α)
Mean	−5.42	[−11.96, 1.13]	−0.06		0.56	[−1.80, 2.92]	0.03
Variance	7.16**	[2.33, 11.28]	0.09		0.54***	[0.28, 0.80]	0.24
Residual error
${\sigma }_e^2$	28.51***	[25.60, 31.43]	0.18		2.88***	[2.56, 3.20]	0.84
Model fit
χ2	138.89
df	40
p	< .001
CFI	.957
TLI	.952
RMSEA	.071
SRMR	.083

Note.

^*p < .05;

^**p < .01;

^***p < .001;

NA = negative affectivity; CI = confidence interval; df = degrees of freedom; CFI = confirmatory fit index; RMSEA = root mean square error of approximation. All estimates have accounted for nesting within twins.

Verbal WM and NA.

Table 4 shows the bivariate dual LCS model assessing bi-directional influences of change in verbal WM (ΔVerbal WM) and ΔNA. Based on practical fit indices, the model had acceptable fit. Within persons, increased NA across t was significantly related to reduction in verbal WM across t+1, with small effect (β = −2.98, 95% CI [−3.97, −1.99], d = −0.28). Likewise, within-person decrease in verbal WM across t was substantially associated with t+1 rise in NA, with small effect (β = −0.57, 95% CI [−0.78, −0.36], d = −0.32). Thus, the findings offered support for Hypothesis 2.

Table 4.

Bivariate latent difference score models of verbal WM and NA

Parameter estimate	Verbal WM and NA
Within-person bivariate change-to-change coupling effects (δ)	ΔVWMt – 1 ➔ ΔNAt				ΔNAt – 1 ➔ ΔVWMt
	β	[95% CI]	d		β	[95% CI]	d
	−0.57***	[−0.78, −0.36]	−0.32		−2.98***	[−3.97,−1.99]	−0.28
Within-person proportional change (β)	VWMt – 1 ➔ ΔVWMt				NAt – 1 ➔ ΔNAt
	0.42	[−0.11, 0.95]	0.09		0.03	[−0.20, 0.26]	0.02
	Verbal WM				NA
Initial status	β	[95% CI]	d		β	[95% CI]	d
Mean	4.21***	[4.10, 4.32]	2.96		12.34***	[12.12, 12.56]	2.29
Variance	1.01***	[0.78, 1.24]	0.49		4.08***	[3.35, 4.81]	0.46
Constant change (α)
Mean	−1.81	[−3.99, 0.38]	−0.04		−0.27	[−3.14, 2.59]	−0.01
Variance	0.31	[−0.15, 0.77]	0.08		0.71***	[0.43, 0.99]	0.29
Residual error
${\sigma }_e^2$	0.92***	[0.81, 1.04]	0.89		2.81***	[2.50, 3.12]	0.86
Model fit
χ2	110.86
df	40
p	< .001
CFI	.938
TLI	.931
RMSEA	.062
SRMR	.086

Note.

^*p < .05;

^**p < .01;

^***p < .001;

WM = working memory; NA = negative affectivity; CI = confidence interval; df = degrees of freedom; CFI = confirmatory fit index; RMSEA = root mean square error of approximation. All estimates have accounted for nesting within twins.

Processing speed and NA.

Table 5 demonstrates that the bivariate dual LCS model examining the reciprocal impacts of change in processing speed (ΔProcessing Speed) and ΔNA. The practical fit indices suggest that the model offered satisfactory fit to the data. Within-person increase in NA across t was significantly related to future t+1 decrease in processing speed, with moderate effect (β = −0.19, 95% CI [−0.24, −0.15], d = −0.48). Similarly, within-person reduction in processing speed across t was significantly related to rise in NA across t+1, with small effect (β = −4.41, 95% CI [−6.16, −2.65], d = −0.20). The data thus substantiated Hypothesis 3. Table S2 in the Supplementary Materials show between-person descriptive statistics of study variables and additional within-person SEM models that also fit the data well.³

Table 5.

Bivariate latent difference score models of processing speed and NA

Parameter estimate	Processing Speed and NA
Within-person bivariate change-to-change coupling effects (δ)	Δ Processing Speedt – 1 ➔ ΔNAt				ΔNAt – 1 ➔ Δ Processing Speedt
	β	[95% CI]	d		β	[95% CI]	d
	−4.41***	[−6.16, −2.65]	−0.20		−0.19***	[−0.24, −0.15]	−0.48
Within-person proportional change (β)	Processing Speedt – 1 ➔ ΔProcessing Speedt				NAt – 1 ➔ ΔNAt
	0.02	[−0.11, 0.15]	0.02		0.06	[−0.16, 0.28]	0.03
	Processing Speed				NA
Initial status	β	[95% CI]	d		β	[95% CI]	d
Mean	41.86***	[40.89, 42.83]	2.30		12.50***	[12.29, 12.72]	2.50
Variance	105.77***	[89.48, 122.05]	0.03		3.99***	[3.30, 4.67]	0.48
Constant change (α)
Mean	−3.59	[−8.72, 1.55]	−0.05		−0.97	[−3.77, 1.83]	−0.02
Variance	6.95**	[2.68, 11.23]	0.10		0.50**	[0.29, 0.71]	0.28
Residual error
${\sigma }_e^2$	25.75***	[21.98, 29.52]	0.14		2.93***	[2.60, 3.26]	0.81
Model fit
χ2	153.063
df	40
p	< .001
CFI	.943
TLI	.932
RMSEA	.083
SRMR	.092

Note.

^*p < .05;

^**p < .01;

^***p < .001;

NA = negative affectivity; CI = confidence interval; df = degrees of freedom; CFI = confirmatory fit index; RMSEA = root mean square error of approximation. All estimates have accounted for nesting within twins.

Discussion

This study moves clinical science forward by examining the relations between within-person 3- to 14-year change in NA and subsequent 3- to 14-year change in spatial cognition, verbal WM, or processing speed (and vice versa) across 23 years and 4 time lags. Analyses showed that within-person increase in NA at each preceding time lag was related to decline in spatial cognition, verbal WM, or processing speed (and conversely) at the sequential time lag with small-to-moderate effects. Notably, prior NA or cognitive functioning scores, regression to the mean, and between-person variation could not account for the results. Moreover, our measures showed strong psychometric properties at both the between- (α = .73–.96) and within-persons (α = .71–.91) levels, and acceptable 14- to 33-day retest reliability (r = .53–.90). In addition, LMI analyses showed that the measures were capturing the intended constructs via similar measurement properties at each wave. Taken together, this suggests that observed changes over time were reliable. Further, concordant with research showing small associations between within- and between-person results (Fisher et al., 2018), our within-person findings were largely inconsistent with the small, null, or even positive NA-cognition between-person links (refer to Table S2). Also, our LCS findings showing the relations between change in NA and subsequent change in cognitive functioning (and vice versa) were not consistent with findings from alternative cross-lagged models that tested whether levels predicted subsequent levels. Thus, our findings are specific to within-person change. Developmental psychopathology (Hur et al., 2019) and scar theories (Ottaviani et al., 2016) could explain these findings.

Concordant with scar models, rise in NA was related to future decline in spatial cognition, verbal WM, and processing speed. Plausibly, increased NA likely gave rise to long-term adverse effects on neurobiological regions (e.g., prefrontal cortex, medial temporal lobe, and occipital regions implicated for these cognitive domains) and cardiovascular systems, thereby correlating with cognitive decline across decades (e.g., across 10 to 35 years; Johansson et al., 2010; Kubzansky et al., 1997). Further, our results parallel three prior studies using bivariate dual LCS models. First, in a sample of German community adults aged 70 years or older, 2-year reduction in subjective positive well-being was associated with sharper future 2- to 4-year decline in perceptual speed over 13 years (Gerstorf, Lövdén, Röcke, Smith, & Lindenberger, 2007). Extending that finding, Australian residential care older adults’ 2-year increase in depressive symptoms was related to subsequent 2- to 4-year decreased perceptual speed across 15 years (Bielak, Gerstorf, Kiely, Anstey, & Luszcz, 2011). Similarly, within-person 9-year rise in pathological worry was related to future 9-year reduced global and specific executive functioning (Zainal & Newman, 2020). Our 23-year study extends these investigations by testing reciprocal relations between 3–14 year change in unique cognitive constructs (processing speed, verbal WM, spatial cognition) and future 3- to 14-year change in trait NA.

Notably, our bivariate dual LCS-based findings concur with prior within-person analyses using hierarchical linear modeling and evidencing inverse, within-person, prospective trait NA-cognition relations. For instance, in older adults, higher initial level of depressive symptoms dovetailed with subsequent 1 to 2-year reduction in processing speed, executive functioning, and global cognition across 7 years (Allerhand, Gale, & Deary, 2014). Likewise, in middle-aged community adults in the United States, higher initial level of trait NA was related to larger future 2- to 9-year reductions in visuospatial ability, verbal WM, and processing speed indexed by identical measures used herein (Caselli et al., 2016). Future prospective cognition-NA studies using similar within-person methods could determine if such pattern of results hold.

Unmeasured third variables might also explain such scarring effects. Perhaps growth in NA dovetailed with lower sense of control over life circumstances, which has been shown to relate to problems with executing cognitive tasks two decades later (Caplan & Schooler, 2003). Simultaneously, increase in NA could over time adversely affect brain regions that control emotion regulation (e.g., ventromedial prefrontal cortex-amygdala connectivity; Makovac et al., 2016), thereby coinciding with cognitive decline. Future longitudinal behavioral and brain research could test these notions by including measures of perceived control and emotion regulation to explain within-person change-to-future change NA-cognition relations.

Notably, the fact that within-person growth in verbal WM, spatial cognition, and processing speed were associated with future decreases in NA suggests that these cognitive functioning domains relate to susceptibility to emotional distress. Findings are consistent with emotion regulation process models (Rudolph, Monti, & Flynn, 2018), which posit that cognitive dysfunction is linked to increased vulnerability to stress reactivity across adolescent and adulthood development. To date, at least three prospective studies using multiple linear regression analyses support the hypothesized cognitive dysfunction-emotion dysregulation connection across different age groups. For example, results match a prior report that poorer scores on spatial ability and processing speed tests were correlated with lower self-reported life satisfaction after 3 years in Swedish older adults (Enkvist, Ekström, & Elmståhl, 2013). Similarly, we extend data showing that lower verbal WM and processing speed was related to heightened post-traumatic stress symptoms four years later in Australian young adults (Parslow & Jorm, 2007). Moreover, our data expand on those of Vasterling et al. (2018) demonstrating that U.S. army veterans with weaker spatial reproduction and delayed recall capacities experienced greater post-traumatic stress severity following eight years. At the same time, decline in verbal WM, spatial cognition, and processing speed plausibly dovetails with reduced ability to attend to all pieces of evidence and reappraise life events in versatile ways, thereby predisposing one toward increased NA. Future multiple-wave observational study designs and within-person analyses could test these speculations.

Study limitations warrant attention. Basic assessments and single-item measures of verbal WM (backward DST) and processing speed (SDMT) were used herein. The few tests per cognitive domain could be remedied by future studies using multiple indicators per latent construct. Relatedly, although we could not rule out all measurement error and unexplained variance in NA and cognition, our latent variable analyses do help to substantially reduce measurement error. Also, no one unique latent variable model exists, thus future similar studies should examine other SEM models. However, given that our goal here was to determine if within-person change was associated with future change, LCS was the best choice. In addition, test complexity and its related level of cognitive load might either attenuate or augment the inverse relations among NA and cognitive functioning (Vytal, Cornwell, Arkin, Letkiewicz, & Grillon, 2013), and could be explored in upcoming investigations. Also, our findings may not extend to more culturally diverse or younger (vs. middle-aged or older) adults whose cognitive capacities on average display more temporal stability. Nonetheless, strengths include the large sample size, long study duration, and use of a potent latent variable method that tested dynamic, idiographic, NA-cognition change-to-future change relations. In sum, this 23-year study spanning from middle-to-older adulthood offers novel and essential data that increase in NA across one time lag is linked to a subsequent decrease in spatial cognition, verbal WM, and processing speed (and vice versa) across a different time lag. Clinical and cognitive sciences can profit from replicating, extending, and testing the ideas proposed herein.

Given the basic science nature of this study, upcoming studies could build on these findings by identifying clinical implications, such as taking steps toward establishing causality. A way forward could thus be to use randomized controlled trials (RCTs) to test if cognitive training could reduce trait NA, or if psychotherapies ameliorating NA could improve cognition. For example, an RCT (Wolinsky et al., 2009) evidenced that processing speed training or spatial cognition training (vs. no-contact, memory, or reasoning training) led to 30% reduced chance of clinically elevated depressive symptoms during the subsequent five years. On top of that, future meta-analytic work could holistically consider how benefits of cognitive training and other therapies apply to multiple unique cognitive functioning domains and trait NA in young (18–40 years), middle-aged (41–64), and older adults (65 years and older).

Figure 2. Bivariate dual LCS models of NA and verbal working memory.

Note. WM = verbal working memory; LCS = latent change score; NA = negative affect; T1 = time 1; T2 = time 2; T3 = time 3; T4 = time 4; T5 = time 5; λ = factor loadings; β = autoregressive feedback loop of within-person change in a construct predicting for future within-person change in itself; Δ denotes change in a construct. Of most interest to this study is the coupling parameter (δ_Y) that indicates the effect of prior within-person change in NA severity correlating with future within-person change in verbal working memory (and vice versa) while adjusting for other parameters in the model.

Highlights.

• Within-person 3-year increases in NA was associated with future 3- to 14-year depletions in spatial cognition, verbal WM, and processing speed.

• Within-person 3-year rise in spatial cognition, verbal WM and processing speed was related to future 3- to 14-year reduced NA.

• Such within-person change-to-future change coupling effects between cognitive facets and NA unfolded across 23 years.

• Significant small-to-moderate within-person NA-cognition change-to-future change relations were observed.

• Bivariate dual LCS, but not random-intercept cross-lagged panel models, yielded significant NA-cognition change-to-future change associations.