The challenges and opportunities of our daily lives require most of us to harness specific cognitive function abilities for successful navigation. These cognitive capacities include verbal working memory (WM; fluency to process mentally and change auditory data in real-time) [Reference Cowan, Rouder, Blume and Saults1], spatial cognition (ability to cognitively generate, sustain, retrieve, and alter visual objects) [Reference Klencklen, Despres and Dufour2], and processing speed (capacity to fulfill rudimentary or intricate tasks that require persistent attention efficiently) [Reference Eckert3]. These cognitive indices have been reliably correlated with various markers of biopsychosocial health. Diverse health indicators include mood regulation [Reference Nigg4,Reference Zainal and Newman5], physical health [Reference Quinones, Gallegos, Lin and Heffner6,Reference Zainal and Newman7], school performance [Reference Cook, Braaten and Surman8], work productivity [Reference van Dijk, van Rhenen, Murre and Verwijk9], and quality of personal and professional relationships [Reference Knight and Baune10]. This is likely because spatial cognition, verbal WM, and processing speed are related to myriad basic attentional and complex higher-order executive functioning skills [Reference Salthouse11] and emotion modulation tactics [Reference Groves, Kofler, Wells, Day and Chan12]. Thus, enhancing our understanding of the predictors and outcomes of these aspects of cognitive function is essential.

It has been theorized that factors which facilitate the growth of psychological well-being, such as life satisfaction (LS), may prevent future long-term decline in cognitive abilities [Reference Fiocco and Yaffe13,Reference Ratigan, Kritz-Silverstein and Barrett-Connor14]. LS is the subjective global appraisal of the degree to which cumulative experiences thus far have been, on balance, fulfilling, and rewarding in the realms of health, relationships, career, and lived environments [Reference Veenhoven, Glatzer, Camfield, Møller and Rojas15]. Relatedly, empirical evidence has shown that LS is a construct distinct from dispositional negative affect or worry (tendency to assume the worst) and depression (prolonged low mood) at a single time point and across long developmental periods [Reference Hounkpatin, Boyce, Dunn and Wood16,Reference Johnson, Neyer and Finn17].

Reduction in LS may lead to subsequent cognitive function deterioration across long durations via suboptimal lifestyle choices such as infrequent exercise and unhealthy diet and nutrition [Reference Grant, Wardle and Steptoe18,Reference Kim, Park, Sun, Smith and Peterson19]. Failure to engage in health-promoting activities could lead to increased inflammation in executive functioning and related brain regions and weakened cardiovascular strength, thus causing long-term adverse effects of aging on cognitive function [Reference Baumann, Tchicaya, Lorentz and Le Bihan20–Reference Zainal and Newman23]. Such ideas are consistent with scar theories which argue that behavioral inactivity and repetitive negative thinking that may accompany a reduction in LS could over time negatively impact cognitive function abilities [Reference Zainal and Newman24,Reference Zainal and Newman25]. This process may occur through increased allostatic load (i.e., stress-linked wear and tear of the hypothalamic-pituitary-axis and cognitive function-related brain regions) [Reference D’Amico, Amestoy and Fiocco26] across long periods. Thus, a decrease in LS may dovetail with worsening of spatial cognition, verbal WM, and processing speed in the long term.

As yet, six longitudinal studies have tested the aforesaid idea that reduced LS and related concepts would be associated with subsequent cognitive function decline. For example, among older Swedish community adults, diminished LS coincided with a decrease in self-rated cognitive capacity from working life to post-retirement [Reference Hansson, Buratti, Thorvaldsson, Johansson and Berg27]; however, whether LS would forecast performance-based cognitive function scores could not be deduced from that study. Other studies suggest such a possibility. Using latent growth curve modeling, Wilson et al. [Reference Wilson, Boyle, Segawa, Yu, Begeny and Anagnos28] and Salthouse [Reference Salthouse29] observed that between persons, lower scores on LS and related indices (e.g., satisfaction with relationships) predicted a more rapid decline in processing speed, global cognition, and memory over 6–13 years in early, middle, and late adulthood. Likewise, using hierarchical linear modeling, within persons, an 8-year decrease in psychological well-being was correlated with reductions in processing speed, memory, and global cognition among English midlife and older community adults during that period [Reference Allerhand, Gale and Deary30]; despite that, lead–lag relations were not examined in the study. A Switzerland-based study [Reference Ihle, Ghisletta, Gouveia, Gouveia, Oris and Maurer31] that tested lead–lag associations showed that reduced LS predicted a 6-year decline in set-shifting (i.e., the adeptness to change from one mental mode to another [Reference Miyake and Friedman32]); however, due to its two-wave design, change-to-future change relations could not be inferred. Relatedly, within persons, a lower prior level of LS resulted in a greater 1.5–2-year decline in perceptual speed among older German adults across six waves of assessment [Reference Gerstorf, Lövdén, Röcke, Smith and Lindenberger33].

Simultaneously, theorists have proposed that factors impeding goal attainment, such as a decline in cognitive function, may predict a reduction in LS across long durations [Reference Baumeister, Schmeichel and Vohs34,Reference Diener, Suh, Lucas and Smith35]. This is because such executive functioning-related capacities are essential to engender and sustain progress and a sense of fulfillment in school, career, relationships, and other realms reflective of biopsychosocial health across time [Reference Diamond36,Reference Diener, Oishi and Tay37]. It is also plausible that increased cognitive function, such as processing speed, spatial cognition, and verbal WM, are protective against depletion of LS across development as these resources aid optimal problem definition, planning, and skillful decision-making [Reference Biederman, Petty, Fried, Doyle, Spencer and Seidman38,Reference Swanson and Fung39]. Such cognitive resources-related strategic approaches include sacrificing short-term relief or pleasure to fulfill long-term aspirations [Reference Hofmann, Schmeichel and Baddeley40]. Overall, it is tenable that a decline in verbal WM, processing speed and spatial cognition would predict a subsequent decrease in LS.

Thus far, seven studies have tested the proposition that a decline in LS would be related to future cognitive function deficits or vice versa. One earlier study evidenced no relation between LS and global cognition 6 years later in middle- to older-aged adults [Reference Comijs, Dik, Aartsen, Deeg and Jonker41]. Conversely, middle-aged and older men who displayed stronger verbal WM and processing speed reported higher LS or positive affect following 7–12 years [Reference Rouch, Achour-Crawford, Roche, Castro-Lionard, Laurent and Ntougou Assoumou42,Reference Siedlecki, Yazdani, Minahan and Falzarano43]. Similarly, another study that used ordinary least squares (OLS) regression demonstrated that higher baseline spatial cognition or processing speed scores predicted greater LS after 3 years in younger, middle-aged, and older adults [Reference Enkvist, Ekström and Elmståhl44]. Moreover, stronger WM ability predicted higher trait and daily levels of LS among college students [Reference Pe, Koval and Kuppens45]. In addition to performance-based cognitive function measures, two recent studies [Reference Hansson, Buratti, Thorvaldsson, Johansson and Berg27,Reference Hansson, Buratti, Thorvaldsson, Johansson, Berg and Henkens46] showed that a decline in subjective cognitive capacity was associated with a later reduction in LS among adults transitioning to retirement.

Despite the progress made thus far to understand the relations between LS and cognitive function, literature still presents with several limitations the current study aimed to remedy. First, most studies testing LS–cognitive function relations were cross-sectional (e.g., [47–49]), thus precluding causal inferences due to the absence of temporal precedence. Second, most prior studies on this topic used OLS regression, latent growth modeling, and cross-lagged panel models, which permit conclusions on between-person, but not within-person, associations [Reference Ghisletta and McArdle50,Reference Hamaker, Kuiper and Grasman51]. This is problematic as theories above on LS–cognitive function connections imply how within-person changes in each unique variable predict one another across the lifespan. Furthermore, relations between- and within-persons can differ in magnitude and direction [Reference Zainal and Newman5,Reference Zainal and Newman52], underlining the need to tether within-person analytic methods to test preceding theories. Therefore, the present study used bivariate dual latent change score (BLCS) modeling. BLCS modeling is an advanced longitudinal structural equation modeling (SEM) technique that could examine within-person change-to-future change relations among unique variables [Reference McArdle, Grimm, van Montfort, Oud and Satorra53,Reference Usami, Hayes and McArdle54]. BLCS reduces biases arising from between-person effects, measurement error, autoregressive self-feedback loops, baseline and prior scores, and regression to the mean [Reference Klopack and Wickrama55]. BLCS thus allows a better understanding of how a change in cognitive function at a time lag is linked to a change in LS at a future time lag (and conversely) [Reference Hounkpatin, Boyce, Dunn and Wood16].

A better comprehension of the within-person factors that predict the decline in LS and cognitive function is imperative. Globally, societies are currently grappling with increasing economic, psychosocial, and public health burdens linked to diverse mental illnesses (e.g., depression and anxiety) and neurocognitive disorders (e.g., dementia) that compromise LS [56–58]. Studies that identify the risk factors for decline in cognitive function and LS may be helpful for developing effective prevention, early identification, and treatment strategies. Moreover, within-person change lies at the heart of clinical science [Reference Zainal and Newman52,Reference Wright and Woods59]. Thus, based on the preceding theories and data, we used BLCS modeling to test lead–lag associations between LS and cognitive function across five time points spanning 23 years in a community adult sample. First, we hypothesized that within persons, a decline in LS at a prior time lag would be significantly related to decreased verbal WM, processing speed, and spatial cognition at the next adjacent time lag (Hypothesis 1). Likewise, we predicted that within-person decline in processing speed, verbal WM, and spatial cognition at a previous time lag would be substantially associated with a reduction in LS at the adjacent time lag (Hypothesis 2).

Method

Participants

The current study examined participants (n = 560) who gave permission to undergo face-to-face neuropsychology testing and tocomplete in-person self-report measures such as the LS questionnaire. Forty participants were removed from our analyses as they were diagnosed with dementia, leaving a final sample of 520 participants.Footnote ¹ Of the 520 participants, 54.8% identified as probable or definite fraternal twins, 30.8% as identical twins, and the remaining 14.4% mentioned they did not know the nature of their twinship. In addition, 52% were reared together, and the remaining 48% were reared apart. The current sample were middle- to older-aged adults at baseline (M age = 59.61 years, SD = 8.98, range = 40.57–84.11; refer to Table S1 in the Supplementary Material [SM] for demographics at each time point). Females comprised 58% (n = 302) of the sample. Concerning formal education, 8.46% received elementary to lower secondary school, 21.20% had a high school education, and 70.40% attended college or higher academia, which included courses to acquire specialized skill sets (e.g., forestry and business).

Procedures

Using the Swedish Adoption/Twin Study of Aging publicly available dataset, the present study conducted a theoretically informed secondary analysis. The Karolinska Institutet provided ethical approval, and all participants gave voluntary informed consent. We were interested in the relations between LS and cognitive function (processing speed, verbal WM, and spatial cognition scores) instead of genetic heritability [Reference Petkus, Reynolds, Wetherell, Kremen and Gatz62]. The study comprised five waves of assessment during these years: 1987 (Time 1; T1), 1990 (Time 2; T2), 1993 (Time 3; T3), 2004 (Time 4; T4), and 2007 (Time 5; T5). Due to funding deficits, a sizeable 11-year gap occurred between 1993 and 2004, and in-person cognitive testing was unavailable during this period. This led to five available time points with four sequential time lags ranging from 3 to 11 years. Thus, the present study examined if within-person change at a time lag (ΔT) in a variable was associated with change at the next sequential time lag (ΔT + 1) in another variable. The following measures were administered at all five time points.

Measures

Life satisfaction

An 8-item self-report measure of LS adapted from [Reference Wood, Wylie and Sheafor63]) was used. Respondents endorsed items on a 5-point Likert scale, ranging from 1 = strongly disagree to 5 = strongly agree. It showed satisfactory-to-good between-person reliability (Cronbach’s α = .70–.79) and within-person reliability (α = .70) in the current study. Furthermore, the measure demonstrated good retest reliability (r = .79) [Reference Wood, Wylie and Sheafor63], as well as strong convergent and discriminant validity [Reference Harris, Pedersen, Stacey, McClearn and Nesselroade64,Reference Lehto, Karlsson, Lundholm and Pedersen65]. Table S2 in the SM lists the items in this measure and details how the eight items in the current scale were derived from the original 13-item scale [Reference Harris, Pedersen, Stacey, McClearn and Nesselroade64].

Spatial cognition

The Koh’s block design (BD) and card rotations (CR) tests assessed spatial cognition. Koh’s BD test parallels the Wechsler’s Adult Intelligence Scale [Reference Wechsler66] BD subtest; participants were instructed to replicate seven increasingly difficult square- and triangle-filled figures with two-colored blocks. In addition, the CR test required participants to correctly select, upon mental rotation, one out of four items that matched a target figure on display. Scores for both tests indicated the total completion time (in seconds) for each presented figure. These assessments have shown strong between-person reliability (α = .81–.85) and within-person reliability (α = .90) herein, high 33-day test–retest reliability (r = .90) [Reference Brand, Pieterse and Frost67], and good discriminant and convergent validity [Reference Finkel, Reynolds, McArdle and Pedersen68].

Processing speed

The widely used oral version of the symbol digit modalities (SDM) test [Reference Pahlen, Hamdi, Dahl Aslan, Horwitz, Panizzon and Petersen69,Reference Smith70] measured processing speed capacity.Footnote ² Participants orally indicated each integer (ranging from 1 to 9) linked to a specific unique symbol. The SDM test was administered in two blocks of 50 integers, each for 45 s. Possible scores ranged from 0 to 100. In the current study, the SDM test showed strong between-person (α = .90–.92) and within-person reliability (α = .90). Furthermore, it has shown strong 2-week retest reliability (r = .74), convergent and discriminant validity, with scores of measures of different constructs [71–73].

Verbal WM

The digit span (DS) test-backward was administered to assess verbal WM by instructing participants to register and recite three-to-nine number sequences of growing length in the reverse direction. The DS test-backward score was computed by summing the total number of digit sequences accurately recalled. Although DS test-forward measures verbal WM repetition, DS test-backward captures WM and attentional control processes reliably. It has excellent between-person (α = .93) and within-person reliability (α = .89), as well as retest reliability (α = .83) [Reference Wechsler74]. Moreover, it has strong discriminant and convergent validity [Reference Hilbert, Nakagawa, Puci, Zech and Bühner75,Reference Ryan, Townsend and Kreiner76].

Data analyses

All preprocessing and data analyses were conducted with R version 4.1.0 [77] and RStudio Version 1.4.1717. First, using the psych R package [Reference Revelle78], we inspected the data, determined that all variables were normally distributed, and observed no outliers (see Table S3 in the SM for descriptive statistics and correlation matrix of study variables). The psych R package [Reference Revelle78] was also used to conduct psychometric analyses (e.g., compute reliabilities or internal consistency of scores on measures). Next, a series of SEM analyses were performed. To evaluate SEM model fit, these indices were used: Chi-square with degrees of freedom (χ²(df)) and related p-value; confirmatory fit index (CFI) [Reference Bentler79]; and root-mean-square error of approximation (RMSEA) [Reference Steiger80]. Moreover, nesting within twins was accounted for by utilizing the lavaan.survey R package [Reference Oberski81]. Furthermore, maximum likelihood with robust sandwich estimators were used.

Figure S1 in the SM details the number of participants who completed the in-person cognitive function tests and the LS self-report at each time point. Missingness (a total of 24% of all observations) was managed by using multiple imputation with the raw item scores following recommended guidelines [Reference Graham82]. Data were aggregated across 100 multiply imputed datasets with the predictive mean matching algorithm, each with a maximum of 100 iterations; a gold standard approach in aging research [Reference MacDonald, Stawski, Schaie and Willis83]. Auxiliary variables that could have been related to dropout status (age, gender, education, baseline scores of cognitive functioning, and LS) were included in the multiple imputation models at all time points; a method that adjusts for any systematic trends in dropouts and yields parameter estimates aligned with the missing at random assumption. Moreover, multiple imputation (vs. complete case analysis) tends to produce unbiased, efficient, and precise parameter and standard error estimates that mirror those derived from complete datasets, even if total missingness exceeds 50% [Reference Lee and Shi84,Reference Madley-Dowd, Hughes, Tilling and Heron85].Footnote ³

As a precursor to longitudinal SEM, we tested for longitudinal measurement invariance of all the measures to determine if their psychometric properties were reproducible across all time points [Reference Widaman, Ferrer and Conger86]. All levels of invariance were gradually tested in this order with increasingly restrictive models: configural (equal factor structure, varying factor loadings [λs], item intercepts [τs], and item residual variances [θs]); metric (equal factor structure, λs, varying τs, and θs); scalar (equal factor structure, λs, and τs, varying θs); and strict (equal factor structure, λs, τs, and θs). Measurement invariance was indicated by trivial changes in the practical fit indices; ΔCFI < −.010 or ΔRMSEA < +.030 when comparing the more (vs. less) constrained models [Reference Cheung and Rensvold87,Reference Zainal, Newman and Hong88].

Next, we conducted a series of LCS models to test Hypotheses 1 and 2. Latent change scores were computed by fixing the regression path between scores at a prior time point and the next adjacent time point equal to 1 [Reference Ghisletta and McArdle50]. This indicated that a certain amount of the subsequent score was equivalent to the prior score, and the residual score was construed as the latent change score. Equations (1) and (2) below denote univariate dual LCS models for LS and cognitive function variables. ∆C and ∆L signify the latent change in each cognitive function variable and LS at a time lag (∆T), respectively. The term α indicates the between-person constant change parameter of the slope of a variable, and β relays within-person proportional effects (autoregressive self-feedback loops of change in a variable forecasting future change in itself).

(1) $$ \varDelta {C}_{\left[\varDelta T\right]}\hskip0.35em =\hskip0.35em {\alpha}_C\hskip0.35em \ast \hskip0.35em {C}_S+{\beta}_c\hskip0.35em \ast \hskip0.35em {C}_{\left[\varDelta T-1\right]}. $$

(2) $$ \varDelta {L}_{\left[\varDelta T\right]}\hskip0.35em =\hskip0.35em {\alpha}_L\hskip0.35em \ast \hskip0.35em {L}_S+{\beta}_c\hskip0.35em \ast \hskip0.35em {L}_{\left[\varDelta T-1\right]}. $$

BLCS models were then conducted to test dynamic relations between LS and cognitive function within and between persons. By combining attributes of the cross-lagged panel and latent growth curve models, BLCS models examine change in each variable from one time point to an adjacent time point in a time-sequential manner across all time points [Reference Grimm and Ram89]. It computes within-person coupling effects (δ; cross-lagged effects of change in a variable at a time lag predicting change in another variable at a successive time lag). Simultaneously, BLCS models minimize biases stemming from measurement unreliability, regression to the mean, baseline scores and between-persons (or trait-level) intercepts, variances, constant change, and other model parameters [Reference Matusik, Hollenbeck and Mitchell90]. Equations (3) and (4) denote bivariate LCS models examined herein:

(3) $$ \varDelta {C}_{\left[\varDelta T\right]}\hskip0.35em =\hskip0.35em {\alpha}_C\hskip0.35em \ast \hskip0.35em {C}_S+{\beta}_c\hskip0.35em \ast \hskip0.35em {C}_{\left[\varDelta T-1\right]}+{\delta}_L\hskip0.35em \ast \hskip0.35em {L}_{\left[\varDelta T-1\right]}, $$

(4) $$ \varDelta {L}_{\left[\varDelta T\right]}\hskip0.35em =\hskip0.35em {\alpha}_L\hskip0.35em \ast \hskip0.35em {L}_S+{\beta}_c\hskip0.35em \ast \hskip0.35em {L}_{\left[\varDelta T-1\right]}+{\delta}_C\hskip0.35em \ast \hskip0.35em {C}_{\left[\varDelta T-1\right]}. $$

Furthermore, permitting within-person change-to-future change cross-construct coupling effects to be freely estimated across time-lags increases standard errors. Thus, following recommendations [Reference McArdle91], we fixed the cross-construct coupling effects to be equal. Figure 1 displays the graphical representation of a BLCS model and embedded parameters. Furthermore, to account for differences in time-lag length herein (i.e., 3 vs. 11 year), we conducted a chi-square (∆χ ²(df)) difference test [Reference Sardeshmukh and Vandenberg92] to examine if time-lag length significantly moderated critical parameter estimates (i.e., cross-construct coupling effects). Moreover, to determine if a change in LS predicting future change in cognitive functioning was stronger than vice versa, we conducted a ∆χ ²(df) test of BLCS models with and without equality constraints on the within-person coupling effects. In addition, given multiple comparisons of parameter estimates within and across BLCS models, we used the mixture distribution approach of alpha correction recommended by Ke and Wang [Reference Ke and Wang93] and Wang and Yang [Reference Wang, Yang, Emilio, Steven and Kevin94], which reduces the odds of Type I error in BLCS models. Based on simulations using this approach, these authors recommend that for our current sample size (N = 520) and our five measurement points, a value of p ≤ 0.029 would be the optimal cutoff to reduce the likelihood of Type I error [Reference Ke and Wang93,Reference Wang, Yang, Emilio, Steven and Kevin94]. Furthermore, Cohen’s d effect sizes were computed with the formula d = β/(SE(β)) × √(2/n) [Reference Dunlap, Cortina, Vaslow and Burke95], such that d values of 0.2, 0.5, and 0.8 indicated small, moderate, and large effects, respectively.

Figure 1. Bivariate dual latent change score model of LS and cognitive function.

Abbreviations: COG, global cognition; e, item-level residual error; LS, life satisfaction; T1, Time 1 (1987); T2, Time 2 (1990); T3, Time 3 (1993); T4, Time 4 (2004); T5, Time 5 (2007); α, between-person constant change; β, within-person proportional change (change in a variable predicting future change in itself); Δ, change in a variable; δ, within-person cross-domain coupling effect (change in a variable predicting future change in another variable); λ, factor loading.

Power analysis

A Monte Carlo power analysis mirroring study conditions (a gold standard recommended approach) was conducted using the RAMpath R package [Reference Zhang, Hamagami, Grimm and McArdle96]. To detect within-person change-to-future change cross-construct (coupling effects) and related parameter estimates (e.g., mean and variance) with conservative and small effect sizes ranging from |d| = 0.10 to 0.20, the current study had 97.4–100.0% power. Thus, the current study had an adequate sample size to conduct BLCS analyses.

Results

Longitudinal measurement invariance

As shown in Table 1, across all time points, the two-factor CFA model that included LS and each cognitive function variable displayed measurement invariance at the strict level (equal λs, τs, and εs). This meant that measures herein had comparable psychometric properties and could be construed along with the same metric at all time points. Thus, longitudinal SEM was appropriate for this dataset.

Table 1. Longitudinal measurement invariance of the four-factor model of LS and cognition.

Abbreviations: CFI, confirmatory fit index; LS, life satisfaction; RMSEA, root-mean-square error of approximation; WM, working memory; εs, residual variances; λs, factor loadings; τs, intercepts.

Univariate latent change score models

Table 2 demonstrates that the model fit indices for the ULCS models were good for LS (χ²(df = 13) = 50.479, p < 0.001, CFI = .963, RMSEA = .058), spatial cognition (χ²(df = 13) = 63.031, p < 0.001, CFI = .969, RMSEA = .086), verbal WM (χ²(df = 13) = 31.780, p < 0.001, CFI = .954, RMSEA = .053), and processing speed (χ²(df = 13) = 75.714, p < 0.001, CFI = .941, RMSEA = .097). Within persons, higher level of cognitive functioning at a prior time point was significantly associated with greater rise in cognitive functioning at the next time lag for all specific cognitive functioning indices (spatial cognition: β = 0.341, 95% CI [0.192, 0.491], p < 0.001, d = 0.278; verbal WM: β = 0.733, 95% CI [0.213, 1.253], p = 0.006, d = 0.172; processing speed: β = 0.234, 95% CI [0.114, 0.354], p < 0.001, d = 0.238). However, level of LS at a previous time point was not significantly related to future change in itself within persons (β = −0.043, 95% CI [−0.247, 0.161], p = 0.680, d = −0.036).

Table 2. Univariate latent change score models of each variable.

Abbreviations: CFI, confirmatory fit index; df, degrees of freedom; LS, life satisfaction; RMSEA, root-mean-square error of approximation; VSP, visual–spatial processing.

* p < 0.05.

^** p < 0.01.

^*** p < 0.001.

Bivariate dual latent change score models

Life satisfaction and spatial cognition

Table 3 shows that the BLCS model examining within-person change-to-future change associations between LS and spatial cognition had good model fit (χ²(df = 40) = 118.780, p < 0.001, CFI = .968, RMSEA = .063). Within persons, prior greater decrease in LS at a 3–10-year time lag was significantly linked to higher future reduction in spatial cognition ability at the following time lag (β = −12.661, 95% CI [−22.892, −2.430], p = 0.015, d = −0.150). However, change in spatial cognition at a time lag was not significantly related to change in LS at the next adjacent time lag (β = −0.030, 95% CI [−0.072, 0.013], p = 0.177, d = −0.085). In addition, the cross-lagged change-to-future change association was significantly stronger for LS predicting spatial cognition ability than the reverse direction (Δχ²(df = 1) = 4.167, p = 0.041). Furthermore, time-lag length did not significantly moderate cross-construct coupling effects (Δχ²(df = 1) = 2.323, p = 0.127). Moreover, time-lag length differences (3 vs. 11 year) did not significantly moderate the change-to-future change cross-construct coupling effects between spatial cognition and LS (∆χ ²(df = 2) = 2.009, p = 0.366).

Table 3. Bivariate latent difference score models of spatial cognition and LS.

Abbreviations: CFI, confirmatory fit index; CI, confidence interval; df, degrees of freedom; LS, life satisfaction; RMSEA, root-mean-square error of approximation; Δ, change in a parameter estimate at one time lag.

* p < 0.05.

^*** p < 0.001.

Life satisfaction and verbal WM

Table 4 demonstrates that the BLCS model testing within-person change-to-future relations between LS and verbal WM had good model fit (χ²(df = 40) = 87.780, p < 0.001, CFI = .962, RMSEA = .052). Within-person decrease in LS at a 3–10-year time lag was significantly related to decrement in verbal WM during the adjacent subsequent time lag (β = 1.637, 95% CI [0.680, 2.594], p = 0.001, d = 0.208). Likewise, within-person decline in verbal WM at a 3–10-year time lag was significantly associated with reduction in LS at the next successive time lag (β = 0.357, 95% CI [0.051, 0.664], p = 0.022, d = 0.142). In addition, the change-to-future change relation for LS predicting verbal WM was significantly stronger than the reverse (Δχ²(df = 1) = 4.153, p = 0.042). Furthermore, time-lag length differences (3 vs. 11 year) did not significantly moderate the change-to-future change cross-construct coupling effects between verbal WM and LS (∆χ ²(df = 2) = 2.195, p = 0.334).

Table 4. Bivariate latent difference score models of verbal WM and LS.

Abbreviations: CFI, confirmatory fit index; CI, confidence interval; df, degrees of freedom; LS, life satisfaction; RMSEA, root-mean-square error of approximation; VWM, verbal working memory; Δ, change in a parameter estimate at one time lag.

* p < 0.05.

^** p < 0.01.

^*** p < 0.001.

Life satisfaction and processing speed

Table 5 conveys that the BLCS model determining within-person change-to-future change associations between LS and processing speed had good model fit (χ²(df = 40) = 147.622, p < 0.001, CFI = .948, RMSEA = .072). Within persons, decline in LS at a prior 3–10-year time lag was significantly related to reduction in processing speed at the next time lag (β = 11.437, 95% CI [6.261, 16.613], p < 0.001, d = 0.354), and vice versa (β = 0.040, 95% CI [0.026, 0.054], p < 0.001, d = 0.269). Furthermore, the magnitude of change-to-future change relations was significantly stronger for LS predicting processing speed than conversely (Δχ²(df = 1) = 4.359, p = 0.037). In addition, time-lag length differences (3 vs. 11 year) did not significantly moderate the change-to-future change cross-construct coupling effects between processing speed and LS (∆χ ²(df = 2) = 1.409, p = 0.494).

Table 5. Bivariate latent difference score models of processing speed and LS.

Abbreviations: CFI, confirmatory fit index; CI, confidence interval; df, degrees of freedom; LS, life satisfaction; RMSEA, root-mean-square error of approximation; Δ, change in a parameter estimate at one time lag.

* p < 0.05.

^*** p < 0.001.