Adolescent self-organization predicts midlife memory in a prospective birth cohort study.

Adult Memory Test

At either the clinic or home visit, study members undertook the same tests of verbal memory and timed letter search given at ages 43, 53 ( Richards et al., 2004 ), and 60+ years. This consisted of a 15-item word-learning task devised by the NSHD. Each word was shown for 2 s. When all 15 words were shown, the study member was asked to write down as many of these as possible, in any order. Three repeated trials were conducted on each testing occasion. In this study, the sum of scores of correctly remembered items at each test trial from each assessment was used as an indicator for a latent variable construct representing memory performance. The distributions of these nine test items of verbal memory approximated a normal distribution in terms of minimal skewness and kurtosis. The mean skewness value was −0.11 (the skewness for a standard normal distribution is zero), with no item larger than 0.32 or lower than −0.50. The mean kurtosis value was 2.96 (the kurtosis for a standard normal distribution is 3), with no item larger than 3.10 or lower than 2.77.

Two different lists of words were applied to the memory assessment in an effort to avoid the impact of “practice” effects. At the first assessment (age 43 years), study members randomly received one of two different word lists, which were then alternated over the subsequent two data collections. To adjust for any potential bias introduced by word list, a binary variable representing test type was included in the analysis as a covariate. Because the most recent assessment took place between ages 60 and 64 years, age at assessment was incorporated into the model estimation to account for any differences in memory that might result from this. This was integrated into the latent growth curve model through the TSCORE command within Mplus 6.1 ( Muthén & Muthén, 2010 ) latent variable modeling software. The TSCORE function takes into account the individually varying times of observation resulting from the fieldwork data collection. Both the baseline measure and the slope representing memory decline were, in return, specified in an overarching structural equation model as outcomes in relation to childhood cognition, adolescent mental health, and educational attainment.

Adolescent Mental Health

Teachers were asked to rate study members on a three category response scale where they had to compare the study member’s behavior with that of “a normal child” at ages 13 and 15 years, using items that were forerunners of those used in the Rutter A scale ( Elander & Rutter, 1996 ; Rutter, Tizard, & Whitmore, 1970 ). Previous work has used factor analysis of these ratings to identify two dimensions, emotional and conduct problems, under traditional ( Rodgers, 1990 ) and categorical ( Colman et al., 2009 ) data factor analysis.

For our study, teacher rating data at age 13 and 15 years were resubjected to separate exploratory factor analysis at these ages. Item level data were modeled as ordinal using probit models for categorical outcomes in the statistical package Mplus 6.1 ( Muthén & Muthén, 2010 ). Some items clearly measured two different dimensions of mental health; for example, for the item “How does this child react to criticism or punishment?” teachers were asked to choose from unduly resentful , normal attitude to criticism and punishment , and tends to become unduly miserable or worried . To integrate such items into the factor analysis appropriately, we recoded these responses into two binary items, each representing a new variable compared with a normal child. These recoded binary items loaded substantially and uniquely on corresponding factors. Table 1 shows a list of the items recoded in this way, along with their factor loadings.

Examination of scree plots, eigenvalues, and model fit indices suggested a three-factor solution for this set of items, each representing emotional problems (e.g., gloomy and sad, extremely fearful), conduct problems (e.g., disobedience, evading truth to keep out of trouble), and self-organization. The self-organization factor was defined by items relating to attitude to work; concentration; neatness in work; and not daydreaming in class (see the Results section for details). Factor scores at ages 13 and 15 years were summed to create scales representing these dimensions, and to facilitate the interpretation, the new combined scales were then standardized to form z scores.

Childhood Cognition

Childhood cognitive ability at age 8 years was represented as the sum of four tests of verbal and nonverbal ability devised by the National Foundation for Educational Research ( Pigeon, 1964 ). These tests were (a) Reading comprehension (selecting appropriate words to complete 35 sentences); (b) Word Reading (ability to read and pronounce 50 words); (c) Vocabulary (ability to explain the meaning of 50 words); and (d) Picture Intelligence, consisting of a 60-item nonverbal reasoning test. We used confirmatory factor analysis to construct a scale summarizing these data. Model fit indices (see later description on fit indices) were: chi-square = 63.145 with 1 df , root-mean-square error of approximation (RMSEA) = .121, comparative fit index (CFI) = .994, Tucker-Lewis index (TLI) = .966. Factor scores were computed then standardized to a mean of zero with a standard deviation of one.

Educational Attainment

The highest level of educational qualification attained by age 26 years was coded by the U.K. Burnham scale ( Department of Education and Science, 1972 ) and recoded for the present analysis into no qualification ( n = 1,765, 39.8%); vocational only ( n = 353, 8.0%); “O level” (secondary level taken by public examination at age 15 years, n = 863, 19.5%); “A level” (advanced secondary level qualifications taken by public examination at 18 years, n = 1,040, 23.5%); and tertiary level (degree or equivalent, or higher degree, n = 411, 9.3%). Similar to adolescent mental health and childhood cognition, the educational attainment variable was standardized to a mean of zero with standard deviation of one.

Statistical Analysis

We applied structural equation models ( Skrondal & Rabe-Hesketh, 2004 ; Tabachnick & Fidell, 2006 ) incorporating latent growth analyses ( Ferrer, Balluerka, & Widaman, 2008 ; Hancock, Kuo, & Lawrence, 2001 ) to describe the way that childhood cognition, the three adolescent mental health factors, and educational attainment were associated with verbal memory at 43 years (intercept) and rate of decline in verbal memory from 43 to 60+ years (slope). This model contained two main components: (a) paths from childhood cognition to the three adolescent mental health factors and to education; (b) simultaneous paths from these variables to memory baseline and rate of decline. We first examined the measurement properties of the memory construct over time. The goodness of fit of these models was evaluated by a range of recommended indices including the TLI ( Tucker & Lewis, 1973 ), the RMSEA ( Steiger, 1990 ), and the CFI ( Bentler, 1990 ). Comparative fit indexes and TLIs greater than .95 are often taken to indicate an acceptable model fit, whereas the RMSEAs less than .06 indicate good fit. We also present the chi-square statistic; however, this is highly sensitive to large sample sizes ( Browne & Cudeck, 1993 ), making it a less suitable index for this study.

Because memory at three different ages (43, 53, and 60+ years) provided the basis of age-related change, it is important to assess the extent to which the memory tests were comparable across these three sets of assessments. Thus, we conducted measurement invariance tests prior to fitting a latent growth model of memory. This was done through assessing model fit of a set of increasingly restrictive models. The model fit indices were compared to evaluate the degree of invariance of the measurement parameters in the models. The baseline model tested configural invariance, whereas the latent memory variable had the same number of factor indicators, that is, three trials at each occasion. In this model, the factor loadings were set to be freely estimated across the assessments. This model was a prerequisite for testing the next step, the metric invariance, where the factor loadings were constrained to be equal across assessments. This step ensured that the memory construct had the same substantive meaning across assessments. In the last step, scalar invariance was assessed by additionally constraining the intercepts of the indicators to be equal across assessments. This step validated the comparison of the latent means of latent memory variables across assessments, which was essential for fitting a latent growth curve model. We used the same range of fit indices to investigate models of measurement invariance. A restrictive model is preferred if the change in model fit indices are not significantly inferior to those of the less restrictive model. In terms of the RMSEA, the change should be less than .015 ( Chen, 2007 ). For CFI, the change should be less than .01 ( Chen, 2007 ; Cheung & Rensvold, 2001 ). We also controlled for the effect of test type with a multiple-indicator-multiple-cause (MIMIC) modeling approach ( Kaplan, 2000 ). Regression paths were specified between test type variable and latent memory variables at each assessment. Paths from test type to each of the memory test trial indicators were first constrained to be zero, assuming no effect from test type to memory indicators. Then modification indices were examined to evaluate whether this assumption was true. Paths associated with high modification indices were subsequently freed to control for any measurement bias in the latent memory variable items introduced by test type. Covariances of residual variances of test trials across assessments were estimated a priori, as it is expected that tests repeatedly assessed at different time points are auto-correlated.

Model estimation and mediation effects of adolescent mental health

All factor analyses and latent variable structural equation models were estimated using Mplus 6.1. The estimator used for the measurement invariance tests was maximum likelihood. The model estimator for the overall SEM was maximum likelihood with robust standard errors. In this study, educational attainment was a hypothesized mediator between adolescent mental health and memory decline. It is not possible in MPlus to request model output on the mediation effect when the robust maximum likelihood estimator is used together with TYPE = RANDOM and TSCORE. Hence, we calculated the mediation effect of adolescent mental health using the MODEL CONSTRAINT function in Mplus 6.1, following the mediation formula for continuous variables ( Hayes, 2009 ; von Soest & Hagtvet, 2011 ).

Adolescent Mental Health Symptoms

Exploratory factor analysis suggested three dimensions within the teacher ratings (see Method section), representing emotional problems (e.g., gloomy and sad, extremely fearful), conduct problems (e.g., disobedience, evading truth to keep out of trouble), and the new factor characterized as self-organization. The details of this analysis are presented in Table 1 . The putative self-organization factor was defined by four items relating to attitude to work; level of concentration; degree of neatness in work; and extent of daydreaming in class (see Table 1 ). For ease of interpretation, this factor was coded so that higher scores indicated better self-organization, whereas following convention higher scores for the emotional and conduct factors indicated more severe problems.

Longitudinal Measurement Invariance of Verbal Memory

We fitted a series of CFA models assessing the measurement invariance of latent memory variables as described above (see Table 2 ). Model m0 represented the configural invariance model, where all measurement parameters were freely estimated. Model m1 was the metric invariance model where factor loadings of memory latent variable were constrained to be equal across the three occasions of memory measurement. Model m2 was the scalar invariance model, for which, in addition to factor loadings, item intercepts were also constrained to be equal across assessment occasions. This was the most strict model and represented full measurement invariance necessary for the present investigation. Models m0, m1, and m2 all showed excellent fit indices, and little change was observed in the fit indices despite increasing model restriction, thus, indicating good measurement invariance of latent memory variables across time (see Table 2 ). Thus, it was valid to compare the latent memory variables across time, which provided a basis for fitting the latent growth curve models.

To investigate the effect of test type (the two different word lists in the memory test), in Model m3, we first constrained regression paths from test type to be 0 to each of the test-trial indicators (see Table 2 ). Modification indices were then examined to identify test trial indicators with high modification indices. In Model m4, paths associated with these indicators were allowed to be freely estimated. Model m4 represented the final measurement model specification for memory in all subsequent analysis. We fitted a CFA model incorporating all variables included in the present investigation. This model had good fit indices, demonstrating a close fit to the data (RMSEA = .017, CFI = .994, TLI = .99). Zero-order correlations of these variables are shown in Table 3 .