Is education associated with improvements in general cognitive ability, or in specific skills?

The Present Study

Here, we report an analysis of a longitudinal cohort (the Lothian Birth Cohort 1936) of over 1,000 individuals across a follow-up period of almost 60 years with intelligence measurements from both early and late in life. We investigated whether education is associated with relative improvements in the g factor extracted from a battery of 10 diverse cognitive tests (domain-general effects of education on cognitive development), or with improvements on only some of those tests (domain-specific effects of education). An advantage of the dataset used here is that we were able to build models of very long-term, lasting effects of education on lifetime cognitive development.

The three possibilities we tested are illustrated by Models A, B, and C in Figure 1 . All models control for prior intelligence, measured at age 11 years, before there was any major variation in educational duration in our sample. Higher childhood intelligence is hypothesized to predict both longer educational duration and higher g -factor scores in later life; these relationships are shown in the upper part of each model. In Model A, education is hypothesized to be associated with the subtests via the latent general factor, g , extracted from them. Model B, which also includes a path from education to g , is similar to Model A, except that it adds some specific associations between education and individual cognitive test scores. This model suggests that education raises all cognitive capabilities via g , but also, beyond these benefits, confers additional improvements on some specific tests. Finally, in Model C, education is hypothesized to be associated with the subtests via only domain-specific paths. Model C suggests that it is this direct improvement in some—potentially all—subtests that is reflected in the IQ score improvements found in previous studies (e.g., Brinch & Galloway, 2012 ), but that these specific improvements do not transfer to increases in general intelligence. We tested which models had better fit and predicted that, if education improves intelligence by raising g , either or both of Models A and B would have significantly better fit to the data than Model C.

Educational duration

Participants were interviewed about their number of years of formal, full-time education during the follow-up wave at age ∼ 70 years.

Analyses

The OpenMx package ( Boker et al., 2011 ) for R and Mplus v7.3 ( Muthén & Muthén, 1998–2014 ) were used to estimate and compare structural equation models of the types shown in Figure 1 . Full-information maximum likelihood estimation was used to adjust for missing data. As can be seen from the rightmost column of Table 1 , there were few missing data, with most of the total sample of 1,091 participants contributing data for each of the tests. The variance of the general intelligence factor was fixed at 1 to identify the model. To assess the absolute fit of each model, we calculated a range of indexes: root mean square error of approximation (RMSEA; values indicating good fit < .06; Hu & Bentler, 1999 ), comparative fit index (CFI; values > .95), and Tucker–Lewis index (TLI; values > .95). For model comparison (relative fit; our main analysis), we calculated the difference (Δ) in Akaike information criterion (AIC; Akaike, 1974 ) between the models, and also assessed the significance of this difference with the chi-square test. Finally, for testing the significance of individual paths within the models, we dropped them from the model (set their path weight to zero) and tested the significance of the resulting change in model fit, also using the chi-square test.

Results

Descriptive statistics and a correlation matrix for all variables examined are provided in Table 1 . All 10 cognitive tests administered at age ∼ 70 years had significant positive intercorrelations (range: r = .16 to .62; p s < .001). All were positively and significantly correlated with years of education (range: r = .14 to .53; p s < .001). All were positively and significantly correlated with IQ at age 11 (range: r = .28 to .69; p s < .001). IQ at age 11 correlated r = .42 with years of education ( p < .001).

We first tested the factor structure of the data. We ran Horn’s parallel analysis on the scores from the 10 IQ subtests, using 1,000 iterations of random data and eigenvalues at the 95th percentile ( Glorfeld, 1995 ). This showed that there was one factor ( g ; eigenvalue = 4.29) in the data. Nevertheless, we tested multiple alternative models derived from exploratory factor analyses extracting two, three, and four factors from the tests using direct oblimin rotation. The two-factor solution resulted in a “timed” factor reflecting Symbol Search, Block Design, Digit-Symbol Substitution and to a lesser extent Spatial Span, and a “nontimed” factor reflecting the remaining six tests. The three-factor solution had a “speed” factor (Symbol Search and Digit-Symbol Substitution), a “verbal memory” factor (Logical Memory and Verbal Paired Associates), and a “fluid intelligence” factor (the remaining six tests). The four-factor solution had the same factors of speed and verbal memory, but also a “reasoning” factor (Matrix Reasoning, Block Design, and MHT) and a “working memory” factor (Letter-Number Sequencing, Digit Span Backwards, and Spatial Span).

In a series of confirmatory factor analyses, we compared models including these factors to a baseline model with one general factor and five significant residual correlations between subtests. The models either included the two, three, or four factors (correlated together) instead of a general factor, or were hierarchically arranged with g as a second-order factor, or had a nested (bifactor) arrangement in which they were included in addition to g but were defined as orthogonal to it (see Schmiedek & Li, 2004 ). The fit of the alternative models ranged from poor to excellent (all RMSEA < .105, all CFI > .908, all TLI > .858). The version with the best absolute fit, the bifactor model with four subfactors, had significantly better fit than the baseline model, ΔAIC = 393.69, χ 2 (1) = 407.69, p < .001.

However, two of the factors had only two indicators, making them no more informative than a residual covariance between the subtests. We tested whether a more parsimonious model could be constructed using only one factor and residual covariances. Using modification indexes calculated in Mplus, we found five residual covariances that were significant in the baseline model. Four of these described clear content overlap in the tests (between Matrix Reasoning and Block Design, Logical Memory and Verbal Paired Associates, Digit-Symbol and Symbol Search, Digit Span Backwards and Letter-Number Sequencing) and one was unexpectedly negative (between the MHT and Spatial Span). This model had excellent absolute fit (RMSEA = .05, CFI = .981, TLI = .972) and fit significantly better than the best-fitting bifactor model, ΔAIC = 57.82, χ 2 (1) = 59.82, p < .001. The one-factor model with residual correlations was supported by exploratory factor analysis (parallel analysis) and also was the most parsimonious of the models tested. We thus used it in all of the models below.

Using the one-factor model as the base, we tested the three types of model shown in Figure 1 . In all three models, the path from age 11 IQ to years of education was significant (standardized path weights = .43, p values < .001 for all three models;), as were the paths from age 11 IQ to g (standardized path weights = .69, .69, and .74 for Models A, B, and C, respectively; age 11 IQ thus explained 48%, 48%, and 55% of the variance in later life g in the three models, respectively; p s < .001).

In Model A, shown in Figure 2 , the path from years of education to g was significant (path weight = .15, p < .001, explaining 2.25% of the variance). As shown in Table 2 , which provides fit indexes for each of the three models, Model A had good fit to the data. Model B, shown in Figure 3 , also contained a significant path from education to g (path weight = .14, p < .001, explaining 1.96% of the variance), and also two additional direct paths from education to Logical Memory (path weight = .08, p = .006) and to Digit-Symbol Substitution (path weight = .06, p = .01). Model B’s fit to the data was also good (see Table 2 ). It was significantly better than that of Model A, ΔAIC = 9.18, χ 2 (2) = 13.18, p = .001, indicating that the inclusion of the two direct paths from education to the subtests improved model fit. Note that the percentage variance explained in each of the subtests can be calculated by subtracting the residual variance of each from 1 (e.g., in Model B, 30% of the variance in Logical Memory was explained by g and by education together).

For Model C (shown in Figure 4 ), we began with a model with paths from education to all subtests, but not to g . We were able to drop three nonsignificant direct paths from education to Spatial Span, Digit Span Backwards, and Letter-Number Sequencing with no significant decrement in model fit. We retained the remaining seven paths, the strongest of which was the path from education to the Logical Memory (path weight = .15 p < .001). The path weights of the other direct relationships between education and the subtests ranged from .06 to .12 ( p s < .04). As shown in Table 2 , Model C also had good fit to the data.

We then compared Model C to the previous models. It had significantly better fit than both Model A, ΔAIC = 19.08, χ 2 (6) = 31.08, p < .001; and Model B, ΔAIC = 9.90, χ 2 (4) = 17.90, p = .001. Therefore, the model that had no path from educ

Discussion

We aimed to address the generality of education’s effect on cognitive development. Structural equation modeling of data from a large sample of individuals followed up across the life course from childhood to old age suggested that education is associated with specific IQ subtests, rather than with the general factor of intelligence. Our analysis had the advantage of controlling for intelligence prior to variation in the length of schooling, and used a wide variety of cognitive subtests to give a reliable indicator of g . The effect found was not dependent on one particular analysis strategy; it was robust to the inclusion or exclusion of additional paths in our three theoretical models.

The findings indicate that education’s ability to raise intelligence test scores (as shown by, e.g., Brinch & Galloway, 2012 ) is driven by domain-specific effects that do not show “far transfer” to general cognitive ability. Such a result coheres with findings from Ritchie et al. (2013) , who, in the same participants who were assessed here, showed no association of education with elementary cognitive measures such as reaction and inspection time, despite an association with improved scores on more verbal IQ subtests. Our results are also broadly consistent with recent reviews concluding that training programs targeting the specific skill of working memory can improve performance on working memory (and closely related) tasks, but that this advantage does not seem to generalize to more distantly related skills such as reasoning and arithmetic (e.g., Melby-Lervåg & Hulme, 2013 ; though see Karbach & Verhaeghen, 2014 ). Finally, our results are in line with a study by Finn et al. (2014) , who showed in a longitudinal sample of schoolchildren that although the quality of the school they attended had effects on tests of directly taught subjects such as mathematics and English language, there was no relation of school quality to performance on tests of “fluid” ability such as processing speed, working memory, and reasoning. These findings, along with the results of present study, point to a conceptualization of education as a training program that develops particular intellectual abilities, but not more fundamental capacities such as the efficiency of cognitive operations.

A different result, demonstrating that education is associated with improvements in general ability, might be more encouraging to educators (see Adey, Csapó, Demetriou, Hautamäki, & Shayer, 2007 for a wide-ranging discussion of education and general ability). Our results were not, however, consistent with a g -related effect of education. We would nonetheless argue that, regardless of whether g is affected, domain-specific effects of education—for instance, on memory and reasoning ability—are still an important benefit for cognitive development. Improved ability on any of these cognitive measures may lead to important advantages in further education, occupational contexts, and everyday life. Our findings indicate that the two ostensibly opposing conceptualizations, of a largely general cognitive ability and a malleable IQ score, are not mutually exclusive.

A similar decoupling of IQ scores and g has been discussed in the context of the Flynn effect, the well-studied secular trend of increasing intelligence test scores across the 20th and 21st centuries (e.g., Flynn, 2009 ). A recent meta-analysis by te Nijenhuis and van der Flier (2013) concluded that the specific abilities shown to be improving across time tend to be those with lower g loadings. Our findings are consistent with the notion that increased compulsory education is one of the potential mechanisms of the Flynn effect (e.g., Rönnlund & Nilsson, 2008 ): Whereas education raises IQ scores, it—like the Flynn effect—does not appear to improve g . The independence from general ability of increases (and decreases) in IQ scores across time, and between groups, is included in the model proposed by Flynn (2009) .

The present study has a number of limitations. First, the measures of g were taken late in life: There was a substantial gap between completion of education and follow-up testing in the cohort. This allowed assessment of the developmental effects of education across almost 60 years, showing enduring associations with specific cognitive skills even after control for initial ability. However, the long time lag also means that a variety of processes may have accumulated to affect the cognitive abilities of the participants. These processes, which may differentially affect particular skills, include maturation, vocational opportunities, life experiences, and—because this particular sample was measured in later life—cognitive ageing ( Hedden & Gabrieli, 2004 ). On the other hand, as noted above, general intelligence is known to be highly stable across the life span after the age at which our childhood measure was administered ( Tucker-Drob & Briley, 2014 ), and the evidence for changes in the structure of g across the life span is inconclusive ( Batterham et al., 2011 ; de Frias et al., 2007 ; Li et al., 2004 ; Tucker-Drob, 2009 ; Tucker-Drob & Salthouse, 2008 ). Nevertheless, it remains possible that education has domain-general effects that are measurable earlier in life, which dissipate with age due to multiple, complex environmental, or biological effects. We would encourage researchers to test similar models to those examined here in samples of adults in midlife (prior to much of cognitive ageing), so long as they include a measure of prior intelligence and, at least at the later measurement, a sufficient range of cognitive tests so that a representative g factor can be extracted.

The range of cognitive tests indicating the g factor should be another focus for future research efforts. With the 10 tests that were administered to our sample, we were unable to produce multiple well-defined subdomains, and thus satisfactory hierarchical or bifactor models of intelligence. In test

Conclusions

The present study went beyond previous analyses of education and cognitive development, and tested whether general or specific aspects of later-life intelligence are associated with longer schooling. A model in which education had direct links to specific IQ subtests had significantly better fit to the data than models in which education was associated with the subtests via g . These findings, consistent with Spearman’s (1927) observation quoted at the outset of the present report, suggest that extended durations of education do not have domain-general effects on ability, but might still have the potential to raise some of an individual’s specific cognitive capabilities.