Cognitive function in early and later life is associated with blood glucose in older individuals: analysis of the Lothian Birth Cohort of 1936

LBC1936 is a community-dwelling sample of 1091 initially healthy individuals. All were born in 1936 and were followed up in four waves of one-to-one cognitive and health testing between 2004 and 2017, at mean ages 70, 73, 76 and 79 years. Details on the background, recruitment and data collection procedures are available [21, 22]. Participants provided written informed consent. Ethics permissions were obtained from the Multicentre Research Ethics Committee for Scotland (wave 1, MREC/01/0/56), the Lothian Research Ethics Committee (wave 1, LREC/2003/2/29) and the Scotland A Research Ethics Committee (waves 2–4, 07/MRE00/58).

When they were approximately 11 years old, 1028 members of the LBC1936 sat Moray House Test No. 12, a broad cognitive ability test that included word classification, proverbs, spatial items and arithmetic. The test correlated about 0.8 with the Terman–Merrill revision of the Stanford–Binet test, providing concurrent validity [23].

In older age, cognitive function is known to decline across multiple domains [24]. Age effects also act on overall cognitive function, a superordinate factor of the lower function domains [24]. We included tests to measure three important cognitive function domains that decline with age: visuospatial ability, processing speed and memory. We also tested crystallised ability, which remains relatively stable in later life [25]. All four domains also contribute to overall cognitive ability [26]. These relationships among cognitive tests and domains are described in the Statistical Analyses section below.

Cognitive functions in waves 1–4 were assessed using 14 individually administered cognitive tests at the same clinical research facility and using the same equipment and procedure for all four waves. The tests are fully described and referenced in an open-access protocol article [21]. The visuospatial domain consisted of matrix reasoning and block design from the Wechsler Adult Intelligence Scale (WAIS) [27], and spatial span forward and backward from the Wechsler Memory Scale (WMS) [28]. Processing speed was measured through symbol search and digit symbol substitution from the WAIS, plus four-choice reaction time [29] and inspection time [30, 31]. Memory was assessed using verbal paired associates and logical memory from the WMS [28], and the letter–number sequencing and digit span backward subtests of the WAIS [27]. Crystallised ability was measured through the National Adult Reading Test [32], Wechsler Test of Adult Reading [33] and a phonemic verbal fluency test [34].

For each individual during each wave, HbA_1c concentrations were measured using a Menarini HA-8160 analyser (Wokingham, UK). Diabetes diagnosis status was independently recorded. Twelve individuals with type 1 diabetes were excluded from further analyses.

Three variable sets were evaluated as potential confounders. The first set defined our baseline control covariates: sex, age (z scored), age 11 cognitive function (z scored), and years of education (z scored). The second set included the available cardiovascular risk factors, obtained from interview and genetic testing: high BP history (no/yes), cardiovascular disease (CVD) history (no/yes), smoking history (current/former/never), and the presence of an APOE*ε4 allele. The third set contained only BMI (z scored), recorded at every testing wave and introduced to our analyses separately from the other cardiovascular variables because BMI is strongly associated with HbA_1c and can have a confounding effect.

All four LBC1936 waves were analysed in latent growth curve models (LGCMs), a structural equation modelling technique that allows the user simultaneously to define and analyse multiple latent and measured variables [35]. For each variable modelled longitudinally, LGCMs model level (i.e. an intercept) and slope (i.e. a trajectory of change) variables. We modelled cognitive function using a hierarchical ‘factor of curves’ model, previously established with these data [26]. For each cognitive test we modelled a level (essentially the age 70 baseline) and a linear slope (the change between age 70 and 79, taking all four measurement occasions into account); for each cognitive domain (see Cognitive function, above) a latent level and linear slope variable were correspondingly composed of the individual tests’ latent level and latent slope variables. At the top of this cognitive hierarchy, latent levels and slopes for overall cognitive function were formed from the level and slope variables of each cognitive domain. For HbA_1c variables, the same approach was taken, but only a pair of level and slope variables were defined for this outcome variable, as no hierarchical structure exists. A structural diagram to illustrate these measured and latent variables is presented in Fig. 1. We also investigated quadratic slope variables for both cognitive function and HbA_1c, but models including quadratic components either would not converge successfully, or they fit very poorly and could not be trusted to produce reliable estimates.

Put simply, HbA_1c level is analogous to blood glucose at the study’s beginning (age 70 years), and HbA_1c slope is analogous to the magnitude at which blood glucose increases over the following decade. Cognitive function level is analogous to cognitive function at age 70 and is known to be highly correlated with cognitive function at age 11 [26], demonstrating the stability of cognitive function across a lifetime. The slope of cognitive function represents the rate at which cognitive functions change (mostly decrease, in fact) over subsequent waves, which is otherwise known as cognitive decline.

Control covariates that applied to all waves of data (sex, age 11 cognitive function, years of education, APOE*ε4 status and smoking history) were regressed on to the intercept and slope variables for latent levels and slopes, including those for the domains of cognitive function, but not on to the levels or slopes of individual cognitive tests. Control covariates that were related to physical condition and could differ between waves (age, history of high BP, history of CVD, and BMI) were regressed on to the individual HbA_1c measurements.

Model coefficients were estimated using full information maximum likelihood, i.e. all data were used for all participants, even individuals who did not complete all waves. Standard errors were calculated using the robust Huber–White method, and p values were computed using the Yuan–Bentler scaled test statistic [36]. The false discovery rate correction for multiple testing was applied to the variables in each model that were not control covariates, which included the associations between the latent variables of cognitive function and its subdomains and HbA_1c. All analyses were conducted in the R programming language, version 3.3.2 (R Foundation for Statistical Computing, Vienna, Austria), using the latent variable analysis (lavaan) package for modelling [37].

Demographic and clinical characteristics of the study population at all four waves are presented in Table 1. Electronic supplementary material (ESM) Table 1 presents wave 1–3 descriptions of study completers only, i.e. those individuals who remained present in wave 4.

Individuals with type 2 diabetes were more likely than healthy individuals to have a history of CVD or high BP (Table 1). The type 2 diabetes group included more men, as well as more current and former smokers. Individuals with type 2 diabetes had higher BMI and HbA_1c levels; in all waves their mean HbA_1c was above the diagnostic threshold of 48 mmol/mol (6.5%). Individuals with type 2 diabetes had lower age 11 cognitive function and scored more poorly than healthy individuals on all cognitive tests across all four waves.

To illustrate cognitive decline across domains, we plotted the discrete overall cognitive function and domain scores across all four waves, with individuals categorised as either healthy or showing a clinical or biochemical (HbA_1c above diagnostic threshold) type 2 diabetes diagnosis at wave 1 (Fig. 2). Figure 2 represents study completers only, because including data from all study participants would bias the means of later waves in the positive direction, making plots unrepresentative of the sample. As expected, with the exception of crystallised ability, the curves show clear declines across all other cognitive domains in individuals with and without type 2 diabetes; however, cognitive function was worse in individuals with type 2 diabetes. Figure 2 is for illustration purposes only.

Figure 3 shows the relationships between the level and slope estimates of HbA_1c and overall cognitive function. Three models are presented, with control variables added cumulatively, as described in Methods. The diagrams present the results for overall cognitive function; domain-specific results can be found in ESM Table 2. The effects of control variables on the level and slope estimates are also presented in Fig. 3, whereas disease-relevant effects of control variables, measured at each wave, are presented separately in ESM Table 2. Model fit statistics of all LGCMs indicated a good fit (root mean square error of approximation [RMSEA] <0.33; standardised root mean residual [SRMR] <0.62; comparative fit index [CFI] >0.939; Tucker–Lewis index [TLI] >0.936; see ESM Table 3 for complete fit statistics).

Age 11 cognitive function made significant contributions to both cognitive function level (β > 0.4, coefficient of partial determination ρ² > 0.35) and HbA_1c level (β > −0.066, ρ² > 0.007) at age 70; that is, higher childhood cognitive function was associated with higher cognitive function and lower HbA_1c concentration at age 70. More education was also associated with better cognitive function in older age, though to a lesser degree (β > 0.12, ρ² > 0.026). These findings are consistent with those previously reported [13, 20, 38].

The relationship between cognitive function level at age 70 and cognitive function slope was robust and did not change when covariates were added (r = 0.002, ρ² = 4 × 10⁻⁶), but the association was small enough that we can only infer that cognitive function level and slope have no strong relationship. This is likely due to low covariance between cognitive function level and slope; or, in other words, regardless of an individual’s starting cognitive function level, cognitive function will consistently decline over the course of the eighth decade.

Significant associations of the same sign were found between HbA_1c slope and both level and slope of cognitive function in models. These results indicate that individuals with higher cognitive function at age 70 maintain lower blood glucose over the following decade, and these associations remain in the presence of all control covariates. The association between cognitive function level and HbA_1c slope was as large as r = −0.166 (ρ² = 0.03), which was found after controlling for the influence of age 11 cognitive function and education. The relationship between cognitive function slope and HbA_1c slope, while consistently positive and significant, was small: at most it was r = 0.008 (ρ² = 6.4 × 10⁻⁵) in the first model (Fig. 3a), which includes the fewest controls.

Associations between HbA_1c level and cognitive function slope were significant but were not consistent across models and were relatively small. Moreover, the association between cognitive function level and HbA_1c level was not always significant, nor was it in the same direction across models. We found no significant associations between HbA_1c level and slope in any of these models.

Associations between HbA_1c growth curve estimates and cognitive function at the domain level generally followed the associations found with overall cognitive function level and slope (Fig. 3; ESM Table 2), with the exception of crystallised ability, which, even in older individuals, tends not to decline over time. In the first model, with the fewest controls, all cognitive domain variables were significantly associated with HbA_1c level and slope, and this held true across subsequent models, except for the slope variables of memory and crystallised ability. Decline in memory and crystallised ability was not associated with either HbA_1c level or slope after controls were included.

We conducted sensitivity analyses on our three structural equation models; specifically, we fit the same models, excluding individuals from a wave if they had been given any type of diabetes diagnosis. The resultant models were very similar to the models presented in Fig. 3 (ESM Table 4); effect sizes were generally reduced, and no previously small and/or unstable associations became robust in these models. We also tested whether change in cognitive function that occurred between age 11 and age 70 was associated with HbA_1c levels, but while age 11 cognitive function significantly predicted HbA_1c, change in cognitive function did not (ESM Results, ESM Table 5).

Individuals might change their behaviour once they are aware that they have type 2 diabetes. To test whether or not higher functioning individuals may be better at caring for themselves after diagnosis, we ran a repeated-measures ANCOVA on the first two study waves, treating type 2 diabetes diagnosis, wave and overall cognitive function as independent variables and HbA_1c as the dependent variable. Type 2 diabetes diagnosis (ω² = 0.338, p < 0.0001), cognitive function (ω² = 0.004, p < 0.0001) and wave (ω² = 0.023, p < 0.0001) all had individual main effects, as did the type 2 diagnosis and wave interaction (ω² = 0.006, p < 0.0001). However, type 2 diagnosis and cognitive function did not significantly interact (ω² < 0.001, p = 0.742), suggesting that knowledge of one’s diagnosis is not related to higher functioning individuals taking better care of themselves. No additional interactions in the model were significant either. These results were robust to the inclusion of control covariates (ESM Table 6).

In much the same way that cognitive function might behaviourally impact post-diagnosis HbA_1c, individuals with type 2 diabetes who maintain low HbA_1c measurements may show cognitive differences from individuals with type 2 diabetes who do not maintain healthier HbA_1c. We also tested this is an ANCOVA framework, making cognitive function the outcome variable, and including HbA_1c, wave (1 or 2) and type 2 diabetes diagnosis as independent variables. We found main effects of type 2 diabetes diagnosis (ω² = 0.014, p < 0.0001) and HbA_1c (ω² = 0.006, p = 0.0001), but no effects of wave or any interactions. These effects were robust to the inclusion of covariates (ESM Table 7). Together, these findings suggest that type 2 diabetes diagnosis and HbA_1c are both associated with cognitive function, but neither changes in diagnosis nor whether an individual with type 2 diabetes has relatively low or high HbA_1c appear to be related to cognitive function.