Cognitive reserve (CR) refers to how flexibly and efficiently one can make use of the available brain reserve estimated by brain size or neuronal count (Stern, 2002). CR has been defined as a hypothetical construct that can be studied by latent variables related to life experience (Jones et al., 2011). Standard proxies for CR include years of schooling, job complexity or occupational attainment, crystallized intelligence (vocabulary level), literacy, engagement in leisure activities, social and cultural participation and integrity of social networks (Sánchez et al., 2011; Tucker and Stern, 2011; Lojo-Seoane et al., 2012; Giogkaraki et al., 2013). Many of these proxies have been included in a comprehensive questionnaire that quantifies CR by considering three main areas: education, working activity and leisure activities (Nucci et al., 2012). Increased frequency of such activities at any time of life may provide the brain with more resources to enable it to compensate for the damage or stressful situations (Freret et al., 2015; Lenehan et al., 2015b; Gelfo et al., 2017). The CR construct is therefore used to explain why two people can have different clinical manifestations of the same disease (Stern, 2012). The protective effect of CR has been studied in elderly populations such as healthy people, people with subjective memory complaints, and patients with mild cognitive impairment (MCI) (Constantinidou et al., 2014; Mondini et al., 2016; Franzmeier et al., 2017a,b), and some studies have found that CR influences progress to dementia (Ghisletta et al., 2006; Sumowski et al., 2014; Vaughan et al., 2014).

Cross-sectional studies have shown that CR is associated with a high level of cognitive performance, suggesting that individuals with greater CR may have more resources available to confront cognitive decline. Mitchell et al. (2012) found that in healthy and memory-impaired older adults, CR was a valid construct comprising years of education and reading ability and was positively correlated with performance in several cognitive domains such as memory/language, attention and processing speed/executive function. Giogkaraki et al. (2013) used structural equation modelling (SEM) to examine how CR was related to age and cognitive functions in healthy older adults and found that CR exerted a moderating role in decreasing the direct negative effect of age on executive functions and episodic memory.

Subjective cognitive complaints (SCCs) are considered important for establishing subjective markers of decline and for determining the influence of SCCs on changes in cognitive performance. The report of SCCs in adults has been suggested to be an early sign of clinical relevance to determine early symptoms of cognitive impairment (Ávila-Villanueva et al., 2016). SCCs constitute an important criterion for diagnosis of MCI (Petersen et al., 1999; Petersen, 2004; Albert et al., 2011) and individuals with SCCs are more likely to develop dementia than those without (Reisberg and Gauthier, 2008; Mitchell et al., 2014). Taking into account the protective effect of CR, study of the influence over time of CR on cognitive performance of people with SCCs is very important (Van Oijen et al., 2007).

In a cross-sectional study, Lojo-Seoane et al. (2014a) used SEM to examine the relationship between CR as a construct comprising educational and lifestyle variables and cognitive performance in older adults with SCCs. Their findings revealed a model of the relationships between CR and cognitive performance with significant direct effects of CR on working memory (WM), general cognitive performance (GCP) and episodic memory (EM), and a significant indirect effect of CR on episodic memory (EM) through WM, suggesting the importance of this particular construct.

The level of performance of WM tasks has been demonstrated to be a sensitive measure for differentiating between normal cognitive aging and MCI, and between MCI and Alzheimer’s disease (AD) (Economou et al., 2007; Belleville et al., 2008; Gagnon and Belleville, 2011). Evidence for the mediating effect of WM on the relationship between CR and cognitive performance has been obtained in a study comparing a sample of 70 participants with amnestic MCI and a control group of 139 participants (Constantinidou et al., 2014). These researchers examined the relationship between age and decline in EM by taking into account the mediator effects of WM. The findings confirmed that the effect of age on episodic memory was mediated by working memory. Additionally, the mediating role of working memory was more important in the amnestic MCI group.

The relationship between CR and cognitive performance has also been investigated in longitudinal research conducted to examine whether individuals with greater CR may have more resources available with which to confront cognitive decline over time, as well as to test the potential masking of early symptoms of cognitive impairment (Ghisletta et al., 2006; Sumowski et al., 2014; Vaughan et al., 2014; Lenehan et al., 2015a). Lenehan et al. (2015a) conducted a review of studies aimed at analyzing the relationship between education (as a proxy for CR) and age-related cognitive decline. The reviewed studies provide evidence indicating that people with a higher level of education will continue to perform at a higher level of cognitive functioning, which may delay the onset of impairment in the future. Sumowski et al. (2014) conducted a longitudinal study of participants with multiple sclerosis who were evaluated at baseline and after 4–5 years, observing that CR protected against cognitive decline in GCP and memory. Ghisletta et al. (2006) investigated the relationship between engagement with activities (as a proxy for CR) and cognitive performance and observed that participation in cognitively demanding activities was associated with a more gradual decline in cognitive performance in a sample of very old participants (80–85 years of age). Vaughan et al. (2014) examined associations between cognitive activities (such as reading books, playing games, and using computers) carried out during the 12 months prior to the study and cognitive performance in a sample of 393 elderly women, who were tested three times at intervals of 1 year. Although these authors observed an association between CR and cognitive performance at baseline, they did not observe any effect of cognitive activity on latent change in cognitive performance at follow-up. In short, according to the hypothesis of CR, people with greater CR would tolerate greater pathological burden, which would be compensated because they have more neuronal resources. However, due to this greater pathological burden, it is likely that when they initiate the symptoms of dementia, experience a more rapid progression of the disease than those with lower CR. That is, the CR benefits by delaying the onset of cognitive deficits, but when they finally appear, the progression is faster (Wilson et al., 2010; Lojo-Seoane et al., 2012).

As far as we are aware, no longitudinal studies of the effects of CR on cognitive performance in people with SCCs have been conducted to date. Taking into account the importance of this population in relation to cognitive impairment, and considering the previous cross-sectional findings by Lojo-Seoane et al. (2014a), we examined the statistical effects of CR (measured at baseline) on the natural course of cognitive performance in a sample of older adults with SCCs, over a period of about 18 months. In accordance with Lojo-Seoane et al. (2014a), the CR construct comprised educational and lifestyle variables. The effect of CR was tested in the three constructs of cognitive performance: EM, WM, and GCP. The specific aims of this research were: (A) To determine whether the relationships between CR and cognitive performance (positive direct effects of CR on WM, GCP, and EM and indirect effect of CR on EM through WM) tested in the previous model including baseline data (Lojo-Seoane et al., 2014a) were maintained in the analysis of the corresponding follow-up data. (B) To analyze the role of CR on cognitive performance at follow-up through the three cognitive domains measured at baseline. (C) To test the role of working memory at follow-up to confirm it mediating role. In keeping with previous research, we predicted that at follow-up (1) the CR model proposed in the previous study will be useful for measuring the relationships between CR and cognitive performance at follow-up, (2) CR will have significant and positive effects on cognitive performance at follow-up, and (3) CR will help to maintain a higher level of WM, which will have a positive influence on GCP and episodic memory over time.

The general linear model comparing cognitive performance of participants who attended follow-up assessment and those who did not attend showed no significant differences in GCP variables {MMSE total = [F_(1,364) = 3.91, p = 0.05, ηp2 = 0.011, observed power = 0.505], CAMCOG-R Orientation = [F_(1,364) = 2.47, p = 0.11, ηp2 = 0.007, observed power = 0.348], CAMCOG-R Attention = [F_(1,364) = 0.87, p = 0.35, ηp2 = 0.002, observed power = 0.154], CAMCOG-R total = [F_(1,364) = 2.53, p = 0.11, ηp2 = 0.007, observed power = 0.355]}, or significant differences in WM {Counting span correct items = [F_(1,364) = 0.27, p = 0.6, ηp2 = 0.001, observed power = 0.081], Counting span correct series = [F_(1,364) = 0.8, p = 0.37, ηp2 = 0.002, observed power = 0.165], Listening span correct items = [F_(1,364) = 0.96, p = 0.32, ηp2 = 0.003, observed power = 0.165], Listening span correct series = [F_(1,364) = 0.20, p = 0.64, ηp2 = 0.001, observed power = 0.074]}. Significant differences were observed between those who attended the follow-up and those who did not in the EM variables {CVLT short term free recall = [F_1,364) = 7.69, p = 0.006, ηp2 = 0.021, observed power = 0.79], CVLT short term cued recall = [F_(1,364) = 12.08, p = 0.001, ηp2 = 0.032, observed power = 0.934], CVLT long term fee recall = [F_(1,364) = 8.27, p = 0.004, ηp2 = 0.022, observed power = 0.818], CVLT long term cued recall = [F_(1,364) = 7.91, p = 0.005, ηp2 = 0.021, observed power = 0.801]}. We also tested whether the two groups differed in two important CR proxies, i.e., years of formal schooling [F_(1,364) = 0.03, p = 0.854] and the WAIS Vocabulary Test [F_(1,364) = 1.69, p = 0.195]. In short, there were no differences in two important CR proxies between participants who attended and those who did not attend the follow-up, and only differences in the measures of episodic memory were observed, although with very a small effect size.

Mean values, standard deviations, skew and kurtosis, differences between baseline and follow-up scores and effect size (standardized measure of Cohen’s d) for neuropsychological measures are shown in Table 2. Although we observed statistically significant differences in performance on most neuropsychological measures between assessments, the size of the effect was so small that these differences were negligible. Moreover, although the skew and kurtosis parameters for some cognitive variables were outside the normal range, the residual values of all the proposed models were adjusted to normal and we therefore assumed that the data were normally distributed (Searly, 1987).

We examined the factorial structure of the three constructs of cognitive performance (EM, WM, and GCP) separately to search for any evidence for the invariance of the measurement model between T1 and T2 for each construct. The SEM fit indices for the three constructs were as follows: for EM, a value of χ² = 321.55 was obtained; p = 0.27 (RMSEA = 0.001 [90% CI = 00–003], CFI = 0.99, SMRS = 0.01) and all factor loadings were statistically significant (p < 0.001); for WM, χ² = 455.02; p = 0.41 (RMSEA = 0.001 [90% CI = 00–003], CFI = 0.99, SMRS = 0.008) and all factor loadings were statistically significant (p < 0.001); and for GCP, χ² = 328.12; p = 0.27 (RMSEA = 0.006 [90% CI = 00–003], CFI = 1, SMRS = 0.01) and again, all factor loadings were statistically significant (p < 0.001). These results show the invariance of the measurement model as defined for each construct and its structural stability in relation to the measures at T1 and T2.

Prior to studying the various models proposed, we examined the factorial structure of the two measurement models (exogenous for CR and endogenous for EM, WM, and GCP) including estimation of the correlation between two times to search for any evidence of stability of the measurement model between T1 and T2 for each construct. The fits for both measurement models were adequate: χ² = 252.87; df = 361; p = 0.9912; RMSEA = 0.042; CFI = 0.961; TLI = 0.949, for the exogenous model and χ² = 292.23; df = 276; p = 0.240, for the endogenous model. The correlations between T1 and T2 [r = 0.734 (p < 0.001) for EM, r = 0.856 (p < 0.001) for WM, and r = 0.614 (p < 0.001) for GCP] indicate the stability of the measures.

After testing the fit of the measurement models, we analyzed the hypothesized models that included structural effects between exogenous and endogenous factors. For the parameter estimations of all the models, we assumed the correlations between the variances of the measurement residuals of all the observable variables (exogenous and endogenous) as free parameters, in order to improve the fit of the data to the model and the total variance explained. Thus, in addition to the remaining usual assumptions of the SEMs (multivariate normal distribution), the statement E(δ_iδ_j) ≠ E(€_i€_j) ≠ 0 holds true for all models.

Model 0 (Figure 4) showed significantly positive correlations between the different domains of cognitive performance at both baseline and follow-up. This model also showed significant direct effects of EM1 on EM2 (γ = 0.289; p < 0.001), WM1 on WM2 (γ = 0.321; p < 0.001) and GCP1 on GCP2 (γ = 0.402; p < 0.001). The goodness-of-fit indices for this model were not satisfactory, and the values of the fit indices CFI (0.89), TLI (0.913), and RMSEA (0.13) were also not adequate. Consequently, in addition to the direct effects estimated, we needed to improve model 0 by adding direct and indirect effects between measures, and the exogenous measurement model should also be included to evaluate the impact of CR on the aforementioned measures.

Analysis of the proposed nested models (Figure 3) yielded different values for each fit. The results obtained for maximum likelihood estimations with the aforementioned restrictions are summarized in Table 3. From the data presented in the table, we can infer that Model 3 provided the best fit. Indeed, both CFI (0.966) and TLI (0.976) values are adequate, as is the RSMEA estimation (0.059). The value of χ² is statistically significant (p < 0.001), but the value of χ²/degrees of freedom is very favorable (1.94). The AIC and BIC values are the lowest of all the models analyzed, confirming that Model 3 is the best of those evaluated.

Although some models yielded poor fits, we have included the values of the structural parameters for all the models (in Table 4) to simplify illustration of the structural effects specified in every model.

Model 1 reflects significant direct effects of CR on EM1 (γ = 0.442; p < 0.001), WM1 (γ = 0.312; p < 0.001) and GCP1 (γ = 0.297; p < 0.001). CR also had an indirect effect on GCP1 via WM1 (β = 0.219; p < 0.001) but not on EM1 (β = 0.033; p = 0.512). This model also revealed significant direct effects of cognitive performance constructs at baseline on cognitive performance constructs at follow-up (EM1–EM2, β = 0.512, p < 0.001; WM1–WM2, β = 0.672, p < 0.001; GCP1–GCP2, β = 0.601, p < 0.001).

In Model 2, the significant direct effect of CR on WM1 (γ = 0.431; p < 0.001) and GCP1 (γ = 0.328; p < 0.001) was maintained, as in Model 1, but the effect on EM1 was not (γ = 0.031, p = 0.421). Significant direct effects of WM1 on WM2 (β = 0.585; p < 0.001) and also on GCP2 (β = 0.301; p < 0.001), but not on EM2 (β = 0.102, p = 0.811), were observed. Significant direct effects of the cognitive constructs were maintained at baseline on the corresponding constructs at follow-up, as in Model 1, with similar intensity and equal direction.

Model 3 showed similar effects to those in Model 2 at baseline and for the baseline constructs on those at follow-up. This model also revealed significant indirect effects of WM2 on EM2 and GCP2 (β = 0.192; p < 0.05; and β = 0.166; p < 0.05, respectively).

Model 4 showed similar effects to those in Models 2 and 3 and also revealed significant indirect effects of WM2 on EM2 and GCP2 (β = 0.231; p < 0.001; and β = 0.399; p < 0.001, respectively); however, the direct effects of WM1 on EM2 and GCP2 were not significant (β = −0.04, p = 0.551; and β = −0.091, p = 0.577, respectively).

The data on fitting and parameter estimation in Table 3 therefore clearly establish Model 3 as the best model. This model is shown in Figure 5, in which the parameters are specified.

In summary, all models showed significant relationships between CR and cognitive performance, measured at baseline and at follow-up. Models 3 and 4 show that CR influences other cognitive constructs over time via working memory, with Model 3 providing the best fit to the data.

This follow-up study provides novel information about the impact of CR on cognitive performance over time in people with SCCs. First of all, our results confirm the robustness of the previous model of relationships between CR and cognitive performance (Lojo-Seoane et al., 2014a) and its structural stability at follow-up. This model enables reflective analysis of the temporal relationships between CR, as a construct composed of two latent variables (educational level and lifestyle), and three cognitive domains (GCP, episodic memory, and working memory) in people with SCCs. We suggest that this model may be useful for predicting the changes in cognitive performance of people with SCCs, taking into account that SCCs indicate a very early stage of MCI and dementia.

Considering the invariance of the measurement models presented, we defined the different models and analyzed the relationships between cognitive constructs. All of the models tested showed significant positive relationships between cognitive performance, measured at baseline and at follow-up, indicating that the performance at baseline is related to performance at follow-up. Models 3 and 4 clarified the effect of CR on cognitive performance via WM over time (Constantinidou et al., 2014; Lojo-Seoane et al., 2014a; Sumowski et al., 2014). Model 3 provided the best fit to the data, with favorable values for χ² and degrees of freedom, adequate values of CFI and TLI and estimated RMSEA, and low values of AIC and BIC. The effects of cognitive performance constructs at baseline on the same constructs at follow-up were significant in all models when CR was included. This shows an effect of CR on cognitive performance at follow-up that is measured indirectly through cognitive performance constructs at baseline and confirms hypothesis 2, which predicted significant and positive effects of CR on cognitive performance at follow-up.

The nested models specifically showed an effect of CR on GCP and WM at follow-up (GCP2 and WM2) via the same cognitive constructs as at baseline. These results are consistent with the findings reported in previous studies (Ghisletta et al., 2006; Van Oijen et al., 2007; Lojo-Seoane et al., 2012, 2014a,b; Mitchell et al., 2012; Giogkaraki et al., 2013; Facal et al., 2014), showing that CR can influence cognitive performance over time. However, our findings differ from those reported by Vaughan et al. (2014), who found that CR influenced cognitive performance at baseline but not after 2–3 years. Although the study by Vaughan and colleagues and our current study coincide regarding the invariance of structures of the corresponding measurement models, the discrepancy may be explained by taking into account that the measurement model used by the aforementioned authors to form the CR construct only included indicators related to lifestyle (such as reading books, playing games, and computer activities). The current model represents a more complete view of CR by integrating two latent variables: lifestyle, including indicators of cultural and social activities, and education, formed by indicators of years of education, occupational attainment, vocabulary knowledge and reading habits. Furthermore, Vaughan et al. (2014) used the digits backward test to estimate WM, while we used a more complex measure of WM, which may be a better estimate of the construct. Another difference between the two studies is the method used to analyze the influence of CR over time. We examined the influence of CR on cognitive performance by considering raw scores at baseline and at follow-up, whereas Vaughan et al. (2014) analyzed the influence of CR on the difference between scores at T2 and at T1.

Different methods of measuring CR are described in the relevant literature. While some authors use more measures related to the educational level such as vocabulary, years of schooling and reading ability (Giogkaraki et al., 2013; Facal et al., 2014), others use CR proxies related to lifestyle and leisure activities (Ghisletta et al., 2006; Vaughan et al., 2014). In this respect, the novelty of our contribution is the inclusion of numerous indicators related to two latent variables, i.e., educational level and lifestyle, which form the CR construct (Lojo-Seoane et al., 2014a).

We analyzed the effects of WM on the other cognitive constructs at baseline and follow-up in Models 2, 3, and 4. The three models showed indirect effects of the CR cognitive constructs at baseline through the WM at baseline. Model 2 showed significant effects of WM at baseline on other cognitive constructs at follow-up and the best fit Model 3, and even Model 4, confirmed significant indirect effect of CR via WM at follow-up on episodic and GCP at follow-up. These findings confirm that CR also influences cognitive performance at follow-up through WM (hypothesis 3) and that episodic memory and GCPs are closely related to WM performance. These findings are consistent with previous findings emphasizing the important and discriminating involvement of WM in most cognitive assessment tasks (Engle, 2002; Conway et al., 2003; Gagnon and Belleville, 2011). Gagnon and Belleville (2011) studied the role of WM span across the retention interval, showing that the effect of retention interval on the amplitude of WM was a good measure for discriminating between healthy participants and people with MCI or AD. Verhaeghen (2011) also reported a series of meta-analyses on aging and executive control, concluding that WM is one of cognitive functions that best explain age-related differences in the performance of many cognitive tasks. Evidence for the mediating role of WM in the relationship of CR on other cognitive constructs or functions has been obtained in studies involving different types of neurological damage/disorders such as traumatic brain injury (Sandry et al., 2015), multiple sclerosis (Sandry and Sumowski, 2014) and MCI (Constantinidou et al., 2014). Sandry and Sumowski (2014) suggested WM capacity as a possible means whereby intellectual enrichment, as a proxy for CR, helps to preserve long-term memory in participants with multiple sclerosis. In a later study, Sandry et al. (2015) also found that WM capacity mediated the relationship between CR and long-term memory in participants with traumatic brain injury. In the same way, Constantinidou et al. (2014) reported that the effect of age on verbal episodic memory was mediated by WM capacity and observed that years of education influenced the rate of age-related decline in immediate verbal episodic memory. These findings help us to understand the relationship between CR and cognitive performance, considering WM as a mechanism by which CR may exert its protective effect on other cognitive domains. We speculate that CR will have a protective effect on WM, because education and lifestyle life contribute to the development of strategies throughout life that may improve the capacity to maintain and simultaneously manipulate information. We suggest that CR mainly protects functions involving greater availability of processing resources from WM. Regarding this suggestion, we hypothesize that the benefit of CR will be reduced when WM is impaired. However, further research on the effects of CR on cognitive performance is required in order to consider the deterioration in WM and to explain some findings indicating that CR is not associated with performance in some domains (Arcara et al., 2017).

Our findings confirm that CR model proposed in the previous study is useful for measuring the influence of CR on cognitive performance, specifically on Episodic Memory, Working Memory, and GCP, in a sample of adults with subjective cognitive complaints evaluated at baseline and at a follow-up after an interval of about 18 months. The findings also highlight the positive effect of CR on cognitive performance at baseline and at follow-up and confirm the mediating role of working memory on episodic memory and GCP at each of the evaluation times in people with subjective cognitive complaints, who may be at risk of suffering mild cognitive impairment. The mediating effect of working memory highlights that education, occupational attainment, reading habits and vocabulary, and participation in social and cultural activities, all of which are components of CR, may contribute to enhancing mechanisms that enable maintenance and simultaneous manipulation of information and favor performance of tasks involving other cognitive constructs at each of the evaluation times. The findings suggest the importance of participating in activities that help improve CR and taking into account CR in cognitive assessments. Likewise evaluating and training working memory in older adults with subjective cognitive complaints is important because the direct and indirect effects of CR on episodic memory and GCP may influence the possible progress from normal aging to MCI and dementia (Richter et al., 2015; Saunders et al., 2015).

This study has some limitations, because it includes only one follow-up evaluation and considers a whole sample without distinguishing between individuals with mild cognitive impairment and cognitively healthy individuals. In future studies, we will test this structural model over a longer period, taking into account a third evaluation which is being completed in our current longitudinal project. Further research is needed to explore the effect of CR on the progress of different subtypes of individuals with subjective cognitive complaints, such as those diagnosed with MCI and those confirmed as healthy controls, taking into account the possible transition between diagnostic states and possible progression to dementia (Facal et al., 2015). Although we have shown that there were no differences in two important CR proxies and only slight differences in measures of episodic memory between participants who attended and those who did not attend the follow-up, further longitudinal research is required to study attrition and to enable general conclusions about the effects of CR to be reached.

CL-S, DF, AP, and OJ-R designed the study, conducted the research, analyzed the data, contributed to writing the manuscript, and revised and approved the final version. JG-O conducted the research, analyzed the data, contributed to writing the manuscript and revised and approved the final version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.