Cognitive reserve (CR) is often operationally defined as a complex structure of latent variables. Here, we present a structural model that analyzes the effect of CR on three cognitive domains: episodic memory, working memory, and general cognitive performance. We developed and analyzed a structural equation model (SEM) to study CR and cognitive performance in 326 participants over 50 years old with subjective memory complaints. The CR construct was found to consist of two factors: (a) educational level and (b) lifestyle. The model revealed that CR had significant direct effects on episodic memory, working memory, and general cognitive performance, and indirect effects on episodic memory via working memory. As a latent construct, CR is related to cognitive performance in participants over 50 years with subjective memory complaints, and it should therefore be considered in the evaluation and diagnosis of such people.

## Introduction

### Cognitive Reserve and Cognitive Performance

Cognitive reserve (CR) is defined as an active process whereby the brain adapts to a situation of deterioration by using cognitive processing resources to compensate for the deficits (Stern, 2009). As CR is a non-observable construct, it is usually studied by latent variable modeling, which allows analysis of the relationship between the different indicators associated with CR and the presence or the absence of cognitive deterioration (Jones et al., 2011; Satz, Cole, Hardy, & Rassovsky, 2011). The most commonly used indicators of CR are observable variables related to the educational level (e.g. professional status and intelligence) and others related to lifestyle (e.g. participation in leisure, social, and cognitive activities; Stern, 2009). These indicators have been shown to be associated with a delay in manifestation of the symptoms of mild cognitive impairment (MCI; Lojo-Seoane, Facal, & Juncos Rabadán, 2012). Thus, people with a high level of education (Kawano et al., 2010; Mejía, Gutiérrez, Villa, & Ostrosky-Solís, 2004) and an active lifestyle (Verghese et al., 2006; Wilson, Scherr, Schneider, Tang, & Bennett, 2007) have better cognitive performance and a lower risk of cognitive decline.

Interest in studying the validity of the CR construct and its relation to cognitive functioning has increased in recent years. The validity of the construct can be analyzed by taking into account two aspects: convergent validity, which establishes that the variables grouped in this construct must be strongly or moderately related, and discriminant validity, which establishes that the variables that constitute the construct should be weakly related to other variables grouped in different constructs (Salthouse & Davis, 2006). Regarding convergent validity, research has been carried out to determine the variables or indicators that constitute the CR construct. Regarding discriminant validity, different neuropsychological tests and tasks have been used in many studies to analyze the relationship between CR and other constructs that represent cognitive performance. In previous studies involving the convergent validity of CR, samples with normal cognitive ageing, MCI, and Alzheimer's disease (AD) were compared (Bennett, Schneider, Arnold, & Wilson, 2006; Boyle, Wilson, Schneider, Bienias, & Bennett, 2008; Mitchell, Shaughnessy, Shirk, Yang, & Atri1, 2012; Siedlecki et al., 2009). Although only a few of the indicators proposed in the literature were used in these studies, moderate to strong relationships between CR and various cognitive domains were demonstrated. Bennett and colleagues (Bennett et al., 2006; Boyle et al., 2008) observed that years of education and social network size exert a strong mediation effect on cognitive performance in people with AD in domains such as speed of perception, working memory, episodic memory, and visuospatial skills. Siedlecki and colleagues (2009) observed moderate to high correlations between CR and all three domains in a study in which they used structural equation modeling (SEM) to analyze the relationship between CR (considering years of education, Peabody vocabulary test scores, and performance of the National Adult Reading Test as indicators) and cognitive performance (in the domains memory, processing speed, and executive function) in participants with normal cognitive ageing. Mitchell and colleagues (2012) applied confirmatory factorial analysis to a group with normal cognitive ageing and a group with cognitive impairment. The four-factor model, which included CR (with years of education and performance of the National Adult Reading Test as indicators) and three cognitive performance factors (memory/language, attention, and processing speed/executive function), provided an excellent fit for the control group and an adequate fit for the cognitive impairment group.

The results of the above studies indicate that CR would be a valid construct represented by several indicators in samples of older adults with different patterns of cognitive performance/cognitive impairment and that there is a moderate to strong relationship between CR and performance in several cognitive domains in these samples. Such relationships may make it difficult to differentiate between CR and other cognitive constructs. Bennett and colleagues (2006) considered that the processing resources construct (speed of perception and working memory) can be considered as CR. Siedlecki and colleagues (2009) observed that, in their models, CR indicators were also significant in the executive function construct, which led them to question the discriminant validity of the CR construct and to suggest that some aspects of executive function, such as cognitive flexibility, may form part of CR because they are influenced by life experiences. Nevertheless, we assume, like Mitchell and colleagues (2012), that CR and cognitive performance are different constructs. Stern (2009, p. 2017) states that “… active models such as CR suggest that the brain actively attempts to cope with brain damage by using pre-existing cognitive processes or by enlisting compensatory processes” and that “… this compensation may help maintain or improve performance.” Therefore, the different use of these processes would explain why some individuals cope better than others with brain pathology and display better cognitive performance. We also assume that measures of CR should be related to variables or indicators that take into account the use of cognitive processes throughout the lifespan.

Study of the impact of CR on cognitive performance in people at risk of developing AD, in people who meet the criteria for MCI (Petersen, 2004), and in those with subjective memory complaints (Geerlings, Jonker, Bouter, Adèr, & Schmand, 1999) can be used to develop successful and accurate means of diagnosis as a high CR may mask the symptoms of cognitive decline. MCI is basically diagnosed by the evaluation of episodic memory (Dubois & Albert, 2004; Gauthier et al. 2006), although the evaluation of general cognitive performance is useful because other domains or cognitive processes may also be affected (Albert et al., 2011; Petersen, 2004). It is also important to evaluate the role of working memory in complex tasks that involve simultaneous processing and storage (Belleville, Sylvain-Roy, de Boysson, & Ménard, 2008; Conway, Kane, & Engle, 2003; Daneman & Carpenter, 1980; Waters & Caplan, 1996).

The main objectives of the present study were to develop a structural model including most of the CR indictors proposed in the literature, such as level of education, reading habits, occupation, social activity, and vocabulary knowledge (Jones et al., 2011; Kawano et al., 2010; Lojo-Seoane et al., 2012; Satz et al., 2011; Stern, 2009; Verghese et al., 2006; Wilson et al., 2007), and to use this model to analyze the impact of CR on cognitive performance. We propose using this model to evaluate whether all of the indicators included converge to form the CR construct. We also studied how CR affects other constructs measured by different indicators of cognitive performance. We applied the model to data from a sample of people over 50 years old attending primary care centers with subjective memory complaints, which is one of the proposed criteria for progression to cognitive MCI and which possibly indicates a risk of developing AD (Albert et al., 2011).

## Methodology

### Participants

The participants in the study were volunteers over 50 years old who were attending primary care centers with subjective complaints of memory loss and who were participating in a larger longitudinal study of cognitive decline. The study met with the approval of the Research Ethics Committee of the Xunta de Galicia, Spain. In total, 699 volunteers were recruited for the study. The volunteers were required to sign an approved informed consent form and were subjected to extensive clinical and neuropsychological evaluation. Volunteers who met any of the following exclusion criteria did not take further part in the study: prior diagnosis of depression or other psychiatric disturbances (according to DSM-IV criteria), prior diagnosis of neurological disease, including probable AD or other types of dementia (according to NINCDS-ADRDA and DMS-IV criteria), previous brain damage or brain surgery, undergoing chemotherapy, prior diagnosis of diabetes type II, sensorial or motor disturbances, and consumption of substances that might affect normal performance of the tasks. The participants included in the final sample had Mini-Mental State Examination (MMSE) scores higher than 20 (the cutoff measures were based on age- and education-related norms).

After the exclusion/inclusion criteria were applied, the final sample of participants comprised 326 subjects, of which 218 were women (66.87%) and 108 men (33.13%). Information about demographics, CR proxies, and neuropsychological performance of the sample is included in Table 1. In the total sample, the 13.5% presented mda-MCI, 7.7% na-MCI, 10.7% a-MCI, and 68.1% presented normal cognitive scores according to the current MCI criteria (Petersen, 2004, updated by Albert et al., 2011).

Table 1.

Means and standard deviations on demographic, CR, and neuropsychological measures

Mean Standard deviation
Age 66.62 9.03
Years of education 9.45 4.44
Occupational attainment 3.12 1.06
Reading habits 3.19 1.12
Social activities 2.22 1.13
Cultural activities 1.67 1.12
WAIS vocabulary test 47.05 13.44
Peabody picture-vocabulary test 59.96 17.18
CVLT short-term free recall 8.51 3.72
CVLT short-term cued recall 9.68 3.53
CVLT long-term fee recall 9.30 3.91
CVLT long-term cued recall 10.06 3.60
Counting span correct items 25.69 11.64
Counting span correct series 2.48 1.12
Listening span correct items 13.64 7.12
Listening span correct series 1.23 1.12
MMSE 26.49 3.03
CAMCOG orientation 9.65 1.70
CAMCOG attention 7.40 1.76
CAMCOG total 85.85 10.31
Mean Standard deviation
Age 66.62 9.03
Years of education 9.45 4.44
Occupational attainment 3.12 1.06
Reading habits 3.19 1.12
Social activities 2.22 1.13
Cultural activities 1.67 1.12
WAIS vocabulary test 47.05 13.44
Peabody picture-vocabulary test 59.96 17.18
CVLT short-term free recall 8.51 3.72
CVLT short-term cued recall 9.68 3.53
CVLT long-term fee recall 9.30 3.91
CVLT long-term cued recall 10.06 3.60
Counting span correct items 25.69 11.64
Counting span correct series 2.48 1.12
Listening span correct items 13.64 7.12
Listening span correct series 1.23 1.12
MMSE 26.49 3.03
CAMCOG orientation 9.65 1.70
CAMCOG attention 7.40 1.76
CAMCOG total 85.85 10.31

Notes: WAIS = Wechsler Adult Intelligence scale; CVLT = California Verbal Learning Test; MMSE = Mini-Mental State Examination; CAMCOG = Cambridge Cognitive Examination.

### Instruments

To collect the data on CR, an ad hoc questionnaire was constructed and administered to the subjects in an interview. The following measures were considered as dependent variables: (a) total number of years of formal schooling; (b) occupational attainment, which evaluates the complexity of the profession to which the participants have dedicated most of their working life, according to the protocol outlined in the Network for efficiency and standardization of dementia diagnosis (NEST-DD) project (Garibotto et al., 2008), on a scale of 1–6 (where 1 = no occupation, 2 = unqualified worker, 3 = housewife, 4 = qualified worker, shop-keeper, low-ranking civil servant, employee, small business employee, office worker, or sales person, 5 = middle-ranking civil servant or manager, small business owner, teacher, or specialist in subordinate position, and 6 = high-ranking civil servant or director, university lecturer, self-employed with high level of responsibility); (c) reading habits, a measure that evaluates the frequency of reading during the last 3 years via one question with responses on a scale of 0–4 (where 0 = never, 1 = occasionally, 2 = once a week, 3 = twice a week, 4 = every day); (d) frequency of social and cultural activities, which evaluates participation in these types of activities during the last 3 years via two questions with responses on a scale of 0–4 (where 0 = never, 1 = rarely, 2 = occasionally, 3 = often, and 4 = always).

The level of vocabulary knowledge was evaluated by the vocabulary test of the Wechsler Adult Intelligence scale (WAIS; Wechsler, 1988), which has a test–retest reliability of between 0.60 and 0.80 and a concurrent validity score of 0.82, and the Peabody picture vocabulary test (Dunn & Dunn, 1981), which has a test–retest reliability of 0.77 and a concurrent validity of 0.86. The total scores obtained in each of the tests were considered as dependent variables.

Episodic memory, acquisition, and recall of verbal material were evaluated with the Spanish version of the California Verbal Learning Test (Delis, Kramer, Kaplen, & Ober, 1987; Spanish version by Benedet & Alejandre, 1998). This test has proven to have adequate reliability (0.94) and validity (it explains 67% of the variance) in the validation study of the Spanish version of the test (Benedet & Alejandre, 1998). In this test, the scores for free recall in the short term and long term and cued recall in the short term and the long term were used as dependent variables.

Working memory was evaluated by two span tasks: (a) the counting span task (Case, Kurland, & Goldberg, 1982), and (b) the listening span task (Pickering, Baqués, & Gathercole, 1999), which is an adaptation of the reading span task developed by Daneman and Carpenter (1980). For both working memory span tasks, the measures number of total correct items and the total number of completed series were considered as dependent variables.

The general cognitive performance was evaluated with the following tests: (a) the MMSE (Folstein, Folstein, & McHogh, 1975; Spanish version by Lobo et al., 1999), which has proven to have good sensitivity (89.8%) and specificity (75.1%) in the evaluation of general cognitive decline; and (b) the Cambridge Cognitive Examination (CAMCOG- R; Roth et al., 1986), which evaluates general cognitive performance and performance in specific areas such as orientation, language, memory, attention, praxis, abstract thought, perception, and executive function and has shown good test–retest reliability (0.86), sensitivity (93%), and specificity (87%). We used total CAMCOG scores and the subscores in specific areas of orientation and attention, which are especially sensitive to aging and cognitive impairment (Cullum et al., 2000).

### Theoretical Structural Model for CR

The structural model proposed includes the most commonly used indicators of CR: years of education, occupational attainment, vocabulary knowledge, reading habits, frequency of social activity, and frequency of cultural activities. We considered two latent variables associated with the observable indicators: level of education, formed by the variables years of education, occupational attainment, vocabulary knowledge, and reading habits; and lifestyle, formed by the variables frequency of social activities and frequency of cultural activities (Jones et al., 2011), which are related to each other (no orthogonality). We also established a second order factor defined as the CR macroconstruct.

The structural model takes into account the direct effect of the CR construct in three cognitive domains: episodic memory, working memory, and general cognitive performance. In order to define these, we propose a formative system of indicators for each non-observable factor, so that for the general cognitive performance factor, the endogenous structure of the complete model will include the scores obtained in tests that measure diverse cognitive domains as well as the scores for attention and orientation, which are specific domains involved in the performance of most tasks; the factor verbal episodic memory includes the measures of short- and long-term free recall and cued recall, and finally, the working memory factor groups the scores obtained in counting and listening span tasks, which measure simultaneous storage and processing (Benedet & Alejandre, 1998; Case et al., 1982; Lobo et al., 1999; Roth et al., 1986).

The indirect effects of the impact of the CR variable on general cognitive performance and episodic memory are also defined in relation to working memory. Fig. 1 shows a complete diagram of the proposed structural model and the definition of the parameters that must be estimated to evaluate the viability of the model.

Fig. 1.

Diagram of the proposed structural model. MMSE = Mini-mental State Examination; CAMCOG = Cambridge Cognitive Examination.

Fig. 1.

Diagram of the proposed structural model. MMSE = Mini-mental State Examination; CAMCOG = Cambridge Cognitive Examination.

### Specification and Identification of the Proposed Structural Model

The different equations and assumptions specified in the structural model underlying the proposed theoretical model can be established from the diagram shown in Fig. 1. Thus, we can establish the following simultaneous equations for the structural part of the model:

\eqalign{{\eta _1} & = {\gamma _{31}}{\xi _3} + {\zeta _1} \cr {\eta _2} & = {\beta _{12}}{\eta _1} + {\gamma _{32}}{\xi _3} + {\zeta _2} \cr {\eta _3} & = {\beta _{13}}{\eta _1}{\gamma _{33}}{\xi _3} + {\zeta _3}.}

These expressions strictly follow the conditions of order and range defined in this type of the structural model. Both measurement models (exogenous and endogenous) can be described by the following expressions:

\eqalign{{X_i} & = {\Lambda _x}{\xi _k} + {\delta _i} \cr {Y_i} & = {\Lambda _y}{\eta _k} + {\varepsilon _i},}
assuming that the two-factor loading matrices (Λx and Λy) will take the following forms depending on the specifications of the two measurement models, for the exogenous model:
$${\Lambda _x} = \left[\matrix{ {{\lambda _{11}}} & 0 \cr {{\lambda _{21}}} & 0 \cr {{\lambda _{31}}} & 0 \cr {{\lambda _{41}}} & 0 \cr {{\lambda _{51}}} & 0 \cr 0 & {{\lambda _{62}}} \cr 0 & {{\lambda _{73}}} \cr }\right] ,$$
and for the endogenous model:
$${\Lambda _y} = \left[\matrix{ {{\lambda _{11}}} & 0 & 0 \cr {{\lambda _{21}}} & 0 & 0 \cr {{\lambda _{31}}} & 0 & 0 \cr {{\lambda _{41}}} & 0 & 0 \cr 0 & {{\lambda _{52}}} & 0 \cr 0 & {{\lambda _{62}}} & 0 \cr 0 & {{\lambda _{72}}} & 0 \cr 0 & {{\lambda _{82}}} & 0 \cr 0 & 0 & {{\lambda _{93}}} \cr 0 & 0 & {{\lambda _{10.3}}} \cr 0 & 0 & {{\lambda _{11.3}}} \cr 0 & 0 & {{\lambda _{12.3}}} \cr }\right] .$$

Likewise, the second-order factorial structure (CR) will be defined by the following vector:

$${\Lambda _\xi } = \matrix{ {{\lambda _{11}}} \cr {{\lambda _{21}}} \cr } ,$$
which represents the second-order factor loadings for each first-order factor. Finally, the specified model takes into account the usual assumptions in relation to the error variances (in both measurement and structural errors) as well as error independence. Thus, E(Xi) = E(Yi) = E(ξi) = 0 and Var(Xi) = Var(Yi) = Var(ξi) = 1 and therefore all of the quantitative variables were transformed by reduction and normalization. Likewise, E(εiεj) = E(δiδj) = E(ξδ) = E(ηε) = E(ζiζj) = 0, under the initial assumption of correlation between the measurement errors in relation to both exogenous and endogenous observable and latent variables.

As already mentioned, the specified model meets the conditions of order and range, and the process revealed over-identification with a total of 141 degrees of freedom. A total of 49 parameters can be estimated, determining a standardized system in which the variances of the latent variables are fixed in the unit.

### Statistical Analysis

We used EQS 6.1 for Windows (Bentler, 1985) to study the structural model equations. To evaluate the model fits to the equations, we used the most commonly accepted parameters (Schreiber, Nora, Stage, Barlow, & King, 2006) in addition to the specific estimates for each parameter: the GFI (Goodness of Fit Index), which indicates a good fit when GFI ≥ 0.95; the AGFI (Adjusted GFI), a corrected index, which indicates a good fit when AGFI ≥ 0.95; the BBNFI (Bentler–Bonett Normed Fit Index), a standardized measure of comparative fit, and the BBNNFI (Bentler–Bonett Non-Normed Fit Index), a non-standardized measure of comparative fit; the CFI (Comparative Fit Index), an index of comparative fit, and the RMSE (Root Mean-Square Error). For the indicators of global fits (GFI, AGFI, BBNFI, BBNNFI, and CFI), values equal to or higher than 0.95 are recommended, whereas in the case of the RMSE, a good fit is assumed when the value is below 0.06, and the confidence interval is estimated for better interpretation. We also used the χ2 test for goodness of fit to analyze the structural fit between the matrix of initial correlations (R) between the observables and the reproduced matrix Σ of the same coefficients of correlation (from the decomposition rules derived from the structures of the equations that are defined on specifying the model). Although generally known, it must be taken into account that the estimation of the structural parameters is based on the minimization of the differences (R − Σ).

## Results

In view of the distributions of the 19 observable variables, the condition of a multinormal distribution was clearly compromised. In fact, only 6 of the 19 variables fitted a normal population distribution. Ory and Mokhtarian (2010) demonstrated the need to control the distribution of observable variables as well as the symmetry of the distributions and the kurtosis associated with the symmetry. Other studies have followed those carried out Lee in the 1990s (Palomo, Dunson, & Bollen, 2007; Poon & Lee, 1994) on the effect of the use of ordinal scales in estimating structural parameters. In all of the above-cited studies, the estimation methods are based on robust algorithms that guarantee little bias in the inference made about the population from the parameters studied. In the present study, we aimed to provide a solution to a rather complex question because most of the observable variables were ordinal and, moreover, the condition of multinormality was compromised. Considering the statistical properties of the observable variables, we therefore used the best estimation option available for analyzing the proposed model. Ory and Mokhtarian (2010) demonstrated the efficacy of classical solutions such as generalized least squares, the potential usefulness of maximum likelihood (ML; for application to distributions that are not particularly asymmetric) and estimations corrected by Bootstrap techniques (Poon & Lee, 1994). The development of Bayesian solutions, with or without prior information functions, provided some success in this respect. Thus, the generation of simulations with different sample sizes demonstrated a certain degree of independence of the solutions and an important degree of resistance to the presence of asymmetry and, even more importantly, a clear decrease in the standard errors of the estimates. Palomo and colleagues (2007) provide a good example of this focus as they use a large part of the strategy included in routines of the TETRAD type (http://www.phil.cmu.edu/projects/tetrad/). Finally, the use of structural models in unfavorable conditions has recently been proposed. We refer here to Asymptotically Distribution-Free methods based on algorithms with few assumptions and close to the simplicity of m-estimators. According to Ory and Mokhtarian (2010), the simplicity and applicability of these methods to cases that are difficult to resolve should make them particularly suited to problems such as that addressed in the present study.

In view of this, we opted to generate a matrix of polychoric correlations between the variables included in the model and to use this matrix to establish the estimations using the supposedly robust ML technique. The standardized results of the previously described process are shown in Fig. 2.

Fig. 2.

Standardized estimates of each free parameter in the proposed structural equation model of CR.

Fig. 2.

Standardized estimates of each free parameter in the proposed structural equation model of CR.

In the measurement model, we can observe the suitability of the indicators selected to measure CR. The CR indicators were successfully grouped into two factors or latent variables, which we denominated educational level and lifestyle (Fig. 2). For the factor educational level, the indicators with the highest factor loadings were the scores obtained in the WAIS (λ = 0.82; p < .001) and Peabody vocabulary tests (λ = 0.81; p < .001) and years of schooling (λ = 0.78; p < .001). Occupational attainment (λ = 0.58; p < .001) and reading habits (λ = 0.50; p < .001) had lower factor loadings in the educational level. For lifestyle, cultural activities had a higher factor loading than social activities. Moreover, the relationship between educational level and lifestyle was significant (ф = 0.172; p = .024).

The CR construct comprised the factors educational level and lifestyle. The factor loadings for each latent variable in the construct are shown in Fig. 2. Although the educational level had a greater weighting in CR (λ = 0.514; p < .001) than lifestyle (λ = 0.211; p = .041), both factor loadings were significant.

The structural part of the proposed model enables a reflective analysis of the relation between the CR, as a construct, and the three cognitive performance domains. CR had a significant impact on general cognitive performance (γ = 0.966; p < .001), and it was also significantly related to working memory performance (γ = 0.685; p < .001) and episodic memory performance (γ = 0.382; p < .001; Fig. 2). Moreover, CR also had an indirect effect on general cognitive performance and episodic memory, via working memory. This indirect effect was equally significant for episodic memory (β = 0.220; p = .03) but not for general cognitive performance (β = 0.049; p = .37). Finally, the goodness-of-fit measures obtained for this model were reasonably satisfactory (Table 2). The values of the fit indices GFI (0.945), AGFI (0.955), BBNFI (0.947), and BBNNFI (0.951) were appropriate (very close to >0.95). The CFI value (0.921) was lower, and the mean square error value (RMSE = 0.002) was suitable (well below 0.06). The chi-square value (546.12) has a proper significance (.09) very close to 0.10, indicating a good fit. Also, the ratio between chi-square and degrees of freedom was rather high (3.879) as a good fit is indicated by a value of less than 3.

Table 2.

Fitting indices derived from robust ML estimation applied to the structural model shown in Fig. 1

Indicator Estimate
Goodness of Fit Index 0.945
Adjusted Goodness of Fit Index 0.955
Bentler Bonnet Normed Fit Index 0.947
Bentler Bonnet Non-Normed Fit Index 0.951
Comparative Fit Index 0.921
Coefficient of Determination 0.523
Root Mean Standard Errors 0.006
Standardized Root Mean Standard Errors 0.002
χ2 with df = 141 546.12 (p = .09)
Ratio χ2/df 3.873
Indicator Estimate
Goodness of Fit Index 0.945
Adjusted Goodness of Fit Index 0.955
Bentler Bonnet Normed Fit Index 0.947
Bentler Bonnet Non-Normed Fit Index 0.951
Comparative Fit Index 0.921
Coefficient of Determination 0.523
Root Mean Standard Errors 0.006
Standardized Root Mean Standard Errors 0.002
χ2 with df = 141 546.12 (p = .09)
Ratio χ2/df 3.873

In order to control for the effects of age, partial correlations with factorial scores for CR, episodic memory, working memory, and cognitive status have been calculated controlling for age. As expected, age significantly correlates with all the factors, but when the effect of age is controlled correlations between CR and cognitive performance factors remain (Table 3).

Table 3.

Partial correlations between factorial scores for CR, episodic memory, working memory, and cognitive status controlling for the effects of age

Cognitive reserve Episodic memory Working memory Cognitive status Age
Zero-order correlations
Cognitive reserve
Episodic memory .05
Working memory .29** .01
Cognitive status .27** −.028 .01
Age −.13* −.44** −.30** −.15**
Controlled for age
Cognitive reserve
Episodic memory −.01
Working memory .26** −.15**
Cognitive status .26** −.10 −.04
Cognitive reserve Episodic memory Working memory Cognitive status Age
Zero-order correlations
Cognitive reserve
Episodic memory .05
Working memory .29** .01
Cognitive status .27** −.028 .01
Age −.13* −.44** −.30** −.15**
Controlled for age
Cognitive reserve
Episodic memory −.01
Working memory .26** −.15**
Cognitive status .26** −.10 −.04

*Significant correlation at <.05.

**Significant correlation at <.01.

## Discussion

The results appear to support the structural model proposed for CR in relation to episodic memory, working memory, and general cognitive performance in adults over 50 years old with subjective memory complaints. In the model, the CR construct is formed by grouping the indicators into two inter-related latent variables denominated educational level and lifestyle, the first of which has the higher factor loading in the construct. All of the indicators had significant loadings in the latent variables educational level and lifestyle and provide evidence for the convergent validity of the CR construct. The present study makes two contributions regarding the composition of the CR construct. First, we have included a large number of indicators in the CR construct, as most variables reported in the literature have been shown to influence CR. This provides us with a more complete view of CR construct and minimizes the risk of overlooking any activity that contributes significantly improve CR. Second, we have shown that the structure of the CR construct can be successfully described by using at least two first-order latent variables—educational level and lifestyle. Although these variables are interrelated, each has its own structure.

Both latent variables comprise measures related to the richness of experiences that individuals have had throughout their life and not to cognitive process or domains. The CR construct proposed by Mitchell and colleagues (2012) was represented by only two indicators, years of schooling, and reading habits. The construct proposed by Siedlecki and colleagues (2009) added a third indicator measured by performance of the Peabody vocabulary test. Variables related to lifestyle were not considered in either of these studies; we included indicators associated with participation in social and cultural activities, which have proven to have a positive effect in delaying the appearance of symptoms of cognitive decline (Verghese et al., 2006; Wilson et al., 2007).

In addition to years of schooling, indicators of educational level also include scores on the Peabody picture vocabulary test and the WAIS vocabulary test, which is a classical measure of crystallized intelligence represented by verbal and general knowledge (Horn & Cattell, 1967). Occupational attainment and years of schooling are CR indicators that have been shown to have a protective effect on probable AD and MCI (Garibotto et al., 2008). Frequency of reading, which we grouped as an indicator of educational level, has been shown to have a protective effect on cognitive decline and is associated with lifestyle (Verghese et al., 2006; Wilson et al., 2007).

Our results point to a close relationship between CR and general cognitive performance. CR has a direct positive effect (0.96) on general cognitive performance measured with the MMSE and CAMCOG screening tests. This relation indicates that higher educational levels and an active lifestyle are related to a general improvement in the execution of cognitive performance tasks. This signifies that the CR construct influences general cognitive performance, but it does not signify that both constructs constitute the CR or that some aspects of general cognitive performance may form part of the CR construct, as apparently suggested by Siedlecki and colleagues (2009).

CR also had a strong impact on working memory (0.68), indicating that it contributes to a better performance of complex tasks that involve simultaneous processing and storage of information (Daneman & Carpenter, 1980). A similar close relationship has also been observed by Siedlecki and colleagues (2009) and Mitchell and colleagues (2012) for the executive function construct, which is consistent with the large overlap between neuropsychological measures of working memory and executive function (Salthouse & Davis, 2006). CR had less effect on episodic memory (0.38) indicating that the resources activated by educational level and lifestyle have less effect on the capacity to remember words in the short and long term than on working memory or general cognitive functioning. This may explain the widely observed finding that deterioration of episodic memory is one of the first symptoms of MCI that progresses to AD (Albert et al., 2011; Dubois et al., 2007; Petersen, 2004). These results are consistent with those reported by Mitchell et al. (2012), who demonstrated a close correlation between the CR construct and cognitive performance and concluded that CR is a valid construct, which is distinct from (although related to) other cognitive domains.

The proposed structural model also enables the indirect effect of CR on cognitive performance to be tested via working memory. The working memory span tasks used in this study are characterized by being involved in simultaneous processing and storage of information (Daneman & Carpenter, 1980), reflecting the capacity to retain information in the memory, despite distractions (Conway et al., 2003; Waters & Caplan, 1996). Performance of this type of task therefore predicts performance in other relevant cognitive domains on a continuum between normal cognitive aging and cognitive decline (Belleville et al., 2008). CR has a positive effect on the performance of working memory span tasks and therefore an indirect effect on cognitive performance, especially episodic memory, which would benefit from the greater availability of processing resources (Salthouse, 1991).

Proxy variables and latent variable models have been used to attempt to operationalize the hypothetical construct of CR (Jones et al., 2011). Taking into account the complexity of the relations between the CR proxies in the CR theoretical framework, and also taking into account the effect of other CR and non-CR variables in the ability to cope with cognitive demanding tasks in the daily life, our model cannot be considered a final one, but an empirical approximation to the CR theoretical framework by measuring the influence of a wide set of CR proxies in episodic memory, working memory, and cognitive status instruments in middle-aged and older adults with memory complaints. Other proxies have been included in different studies in the field and other approaches have been taken, driving to different but complementary models. In example, Bennett and colleagues (2006), Boyle and colleagues (2008), and Mitchell and colleagues (2012) study relations between CR and cognitive performance in a clinical population with AD. In a sample of individuals with normal cognitive ageing, Giogkaraki, Michaelides, and Constantinidou (2013) observe an indirect effect of CR in reducing the negative effect of age on cognitive performance. Other proxies of CR, such as socioeconomic status and participating in cognitive activities (Levi, Rassovsky, Agranov, Sela-Kaufman, & Vakil, 2013; Reed et al., 2011), have been also successfully related to cognitive performance. Being more circumscribed, our approach allows us to sum evidence to a needed research field. Apart from other possible CR proxies and relations between proxies, the most important limitation of the present study is that it is cross-sectional, and therefore, it did not allow study of the influence of CR over time. We believe that the application of the model to longitudinal data currently being collected as part of a wider project by our research team will provide new information about the impact of CR on the process of decline in people with subjective memory complaints toward MCI and thereafter to AD. Other limitation of our study is the lack of genetic information and other biomarkers that could be useful in the characterization of risk of dementia in our sample. Finally, on the completion of the analysis, we have found that a more accurate collection of data about CR proxies is possible, especially those included in the lifestyle factor. In future studies, we are improving the measure of these proxies by adapting to Spanish and applying the CR Index questionnaire (Nucci, Mapelli, & Mondini, 2012).

Our definition of the CR construct on the basis of indicators related to the richness of educational, social, and cultural experiences and their validation in the SEM suggests that a detailed assessment of these experiences could have benefits in the diagnosis of people with subjective memory complaints. In this regard, we believe that the proposed structural model may help to identify the direct and indirect effects of CR on cognitive performance during the continuum from normal aging to dementia. Such information may have important implications in the field of prevention, diagnosis, and early treatment of cognitive decline of adults with subjective memory complaints.

## Funding

This work was financially supported by the Spanish Directorate General for Science and Technology under Projects (SEJ2007-67964-CO2-01 and PSI2010-22224-C03-01) and by the Galician Government: Consellería de Industria e Innovación/Economía e Industria (PGIDIT07PXIB211018PR and 10PXIB2011070 PR).

None declared.

## References

Albert
M. S.
DeKosky
S. T.
Dickson
D.
Dubois
B.
Feldman
H. H.
Fox
N. C.
The diagnosis of mild cognitive impairment due to Alzheimer's disease: Recommendations from the National Institute on Aging-Alzheimer's Association workgroups on diagnostic guidelines for Alzheimer's disease
Alzheimer's & Dementia

2011
7
3
270
279
Belleville
S.
Sylvain-Roy
S.
de Boysson
C.
Ménard
M. C.
Sossin
W. S.
Lacaille
J.-C.
Castellucci
V. F.
Belleville
S.
Characterizing the memory changes in persons with mild cognitive impairment
Progress in brain research

2008
169
365
375
Amsterdam: Elsevier
Benedet
M. J.
Alejandre
M. A.
TAVEC: test de aprendizaje verbal España- complutense

1998
TEA Ediciones
Bennett
D. A.
Schneider
J. A.
Tang
Y.
Arnold
S. E.
Wilson
R. S.
The effect of social networks on the relation between Alzheimer's disease pathology and level of cognitive function in old people: A longitudinal cohort study
The Lancet Neurology

2006
5
5
406
412
Bentler
P. M.
Theory and implementation of EQS: A structural equations program

1985
Los Angeles
BMDP Statistical Software
Boyle
P. A.
Wilson
R. S.
Schneider
J. A.
Bienias
J. L.
Bennett
D. A.
Processing resources modify relation of AD pathology with other cognitive systems
Neurology

2008
70
17
1534
1542
Case
R.
Kurland
M. D.
Goldberg
J.
Operational efficiency and the growth of short-term memory span
Journal of Experimental Child Psychology

1982
33
3
386
404
Conway
A. R. A.
Kane
M. J.
Engle
R. W.
Working memory capacity and its relation to general intelligence
Trends in Cognitive Science

2003
7
12
547
552
Cullum
S.
Huppert
P. A.
McGee
M.
Denning
T.
Ahmed
A.
Paykel
E. S.
et al.
Decline across different domains of cognitive functions in normal ageing: Results of a longitudinal population-based study using CAMCOG
International Journal of Geriatric Psychiatry

2000
15
853
862
Daneman
M.
Carpenter
P. A.
Individual differences in working memory and reading
Journal of Verbal Learning and Verbal Behavior

1980
19
4
450
466
Delis
D. C.
Kramer
J. H.
Kaplen
E.
Ober
B. A.
Californian verbal learning test. Manual-research edition

1987
San Diego
The Psychological Corporation
Dubois
B.
Albert
M. L.
Amnestic MCI or prodomal Alzheimer disease
Lancet

2004
3
4
246
258
Dubois
B.
Feldman
H. H.
Jacova
C.
Dekosky
S. T.
Barberger- Gateau
P.
Cummings
J.
Research criteria for the diagnosis of Alzheimer's disease: Revising the NINCDS-ADRDA criteria
Lancet Neurology

2007
6
8
734
746
Dunn
L. M.
Dunn
L. M.
Peabody picture vocabulary test- revised

1981
Circle Pines
American Guidance Service
Folstein
M. F.
Folstein
S. E.
McHogh
P. R.
“Mini-mental state.” A practical method for grading the cognitive state of patients for the clinician
Journal of Psychiatric Research

1975
12
189
198
Garibotto
V.
Borroni
B.
Kalbe
E.
Herholz
K.
Salmon
E.
Holtoff
V.
Education and occupation as proxies for reserve in aMCI converters and AD FDG-PET evidence
Neurology

2008
71
17
1342
1349
Gauthier
S.
Reisberg
B.
Zaudig
M.
Petersen
R. C.
Ritchie
K.
Broich
K.
Mild cognitive impairment
Lancet

2006
367
9518
1262
1270
Geerlings
M. I.
Jonker
C.
Bouter
L. M.
H. J.
Schmand
B.
Association between memory complaints and incident Alzheimer's disease in elderly people with normal baseline cognition
American Journal of Psychiatry

1999
156
4
531
537
Giogkaraki
E.
Michaelides
M. P.
Constantinidou
F.
The role of cognitive reserve in cognitive aging: Results from the neurocognitive study on aging
Journal of Clinical and Experimental Neuropsychology

2013
35
10
1024
1035
Horn
J. L.
Cattell
R. B.
Age differences in fluid and crystallize intelligence
Acta Psychologica

1967
26
107
129
Jones
R. H.
Manly
J.
Glymour
M. M.
Rentz
D. M.
Jefferson
A. L.
Stern
Y.
Conceptual and measurement challenges in research on cognitive reserve
Journal of the International Neuropsychological Society

2011
17
4
1
9
Kawano
N.
Umegaki
H.
Suzuki
Y.
Yamamoto
S.
Mogi
N.
Iguchi
A.
Effects of educational background on verbal fluency task performance in older adults with Alzheimer's disease and mild cognitive impairment
International Psychogeriatrics, Special Issue: Focus on Training in Psychogeriatrics

2010
22
6
995
1002
Levi
Y.
Rassovsky
Y.
Agranov
E.
Sela-Kaufman
M.
Vakil
E.
Cognitive reserve components as expressed in traumatic brain injury
Journal of the International Neuropsychological Society

2013
19
6
664
671
Lobo
A.
Saz
P.
Marcos
G.
Día
J. L.
de la Cámara
C.
Ventura
T.
Revalidación y normalización del mini- examen cognoscitivo (primera versión en castellano del Mini-Mental Status Examination) en la población general geriátrica
Medicina Clínica

1999
112
20
767
774
Lojo-Seoane
C.
Facal
D.
O.
¿Previene la actividad intelectual el deterioro cognitivo? Relaciones entre reserva cognitiva y deterioro cognitivo ligero
Revista Española de Geriatría y Gerontología

2012
47
6
270
278
Mejía
S.
Gutiérrez
L. M.
Villa
A. R.
Ostrosky-Solís
F.
Cognition, functional status, education, and the diagnosis of dementia and mild cognitive impairment in Spanish-speaking elderly
Applied Neuropsychology

2004
11
4
194
201
Mitchell
M. B.
Shaughnessy
L. W.
Shirk
S. D.
Yang
F. M.
Atri1
A.
Neuropsychological test performance and cognitive reserve in healthy aging and the Alzheimer's disease spectrum: A theoretically driven factor analysis
Journal of the International Neuropsychological Society

2012
18
6
1071
1080
Nucci
M.
Mapelli
D.
Mondini
S.
Cognitive Reserve Index questionnaire (CRIq): A new instrument for measuring cognitive reserve
Aging Clinical and Experimental Research

2012
24
3
218
226
Ory
D. T.
Mokhtarian
P. L.
The impact of non-normality, sample size and estimation technique on goodness-of-fit measures in structural equation modeling: Evidence from ten empirical models of travel behavior
Quality and Quantity

2010
44
3
427
445
Palomo
J.
Dunson
D. B.
Bollen
K.
Lee
S.Y.
Bayesian structural equation modeling
Handbook of latent variable and related models

2007
New York
Elsevier
163
179
Petersen
R. C.
Mild cognitive impairment as a diagnostic entity
Journal of Internal Medicine

2004
256
3
183
194
Pickering
S. J.
Baqués
J.
Gathercole
S. E.
Batería de test de memoria de trabajo

1999
Barcelona
Laboratori de Memoria de la Universitat Autónoma de Barcelona
Poon
W. Y.
Lee
S. Y.
A distribution free approach for analysis of two-level structural equation model

1994
New York
Elsevier
Reed
B. R.
Dowling
M.
Tomaszewski Farias
S.
Sonnen
J.
Strauss
M.
Schneider
J. A.
et al.
Cognitive activities during adulthood are more important than education in building reserve
Journal of the International Neuropsychology Society

2011
17
4
615
624
Roth
M.
Tym
E.
Mountjoy
C. Q.
Huppert
F. A.
Hendrie
H.
Verma
S.
et al.
CAMDEX. A standardised instrument for the diagnosis of mental disorders in the elderly with special reference to the early detection of dementia
British Journal of Psychiatry

1986
149
6
698
709
Salthouse
T.
Mediation of adult age differences in cognition by reductions in working memory and speed of processing
Psychological Science

1991
2
3
179
183
Salthouse
T.
Davis
H. P.
Organization of cognitive abilities and neuropsychological variables across the lifespan
Developmental Review

2006
26
31
54
Satz
P.
Cole
M. A.
Hardy
D. J.
Rassovsky
Y.
Brain and cognitive reserve: Mediator(s) and construct validity, a critique
Journal of Clinical and Experimental Neuropsychology

2011
33
1
121
130
Schreiber
J. B.
Nora
A.
Stage
F. K.
Barlow
E. A.
King
J.
Reporting structural equation modeling and confirmatory factor analysis results: A review
The Journal of Educational Research

2006
99
6
323
337
Siedlecki
K. L.
Stern
Y.
Reuben
A.
Sacco
R. L.
Elkind
M. S. V.
Wright
C. V.
Construct validity of cognitive reserve in a multiethnic cohort: The Northern Manhattan Study
Journal of the International Neuropsychological Society

2009
15
4
558
569
Stern
Y.
Cognitive reserve
Neuropsychologia

2009
47
10
2015
2028
Verghese
J.
LeValley
A.
Derby
C.
Kuslansky
G.
Katz
M.
Hall
C.
et al.
Leisure activities and the risk of amnestic mild cognitive impairment in the elderly
Neurology

2006
66
6
821
827
Waters
G.
Caplan
D.
The measurement of verbal working memory capacity and its relation to reading comprehension
The Quarterly Journal of Experimental Psychology A: Human Experimental Psychology

1996
49A
1
51
79
Wechsler
D.
WAIS-R: Wechsler adult intelligence scale- revised

1988
Oxford
Psychological Corporation
Wilson
R. S.
Scherr
P. A.
Schneider
J. A.
Tang
Y.
Bennett
D. A.
Relation of cognitive activity to risk of developing Alzheimer disease
Neurology

2007
69
20
1911
1920