The effect of mid-life insulin resistance and type 2 diabetes on older-age cognitive state: the explanatory role of early-life advantage

The Medical Research Council (MRC) National Survey of Health and Development (NSHD, also known as the British 1946 birth cohort) recruited a representative sample of 5362 men and women born in England, Scotland and Wales in 1 week in March 1946 [21]. It is the oldest British birth cohort with repeated data collected since birth. The most recent data collection, the 24th, was conducted between 2014 and 2015, when participants were aged 68–69 years [22]. At age 69 years, following a postal questionnaire at age 68 years, participants still alive and with a known current address in mainland Britain (n = 2698) were invited to have a home visit; 2149 (80%) completed a visit. More detailed information and a flow diagram about the follow-up rates and attrition of participants at the latest stage of recruitment is given in Kuh et al (2016) [22] and electronic supplementary material (ESM) Fig. 1. For the most recent data collection, we obtained ethical approval from Queen Square Research Ethics Committee (REC) (14/LO/1073) and Scotland A REC (14/SS/1009). All participants gave written informed consent.

The ACE-III, a test of cognitive state [23], was used as a primary outcome measure at age 69 years. The ACE-III is divided into five domains: attention and orientation (scored 0–18); verbal fluency (0–14); memory (0–26); language (0–26); and visuospatial function (0–16). Thus, the maximum total score is 100. A customised version of the ACE-III was administered by iPad using ACEMobile (www.​acemobile.​org, version 1, accessed 1 January 2015); where this was not possible, a paper version was used. All offline scoring was undertaken by trained personnel. Of the 2149 participants with a home visit at age 69 years, 32 refused or were unable to undertake the ACE-III. Of the remaining 2117, 35 undertook it but did not complete it and data from 353 participants were corrupt through equipment failure. Thus, complete ACE-III data were available for 1729 participants, 80% of those who received home visit (ESM Fig. 1). For those interviewed at age 69 years, there were no statistical differences in variables included in this analysis for those with and without missing ACE-III data (data not shown).

Verbal memory was tested at age 69 years using a 15 item word-learning task; each word was shown for 2 s. When all words were shown, the study member was asked to write down as many words as possible. This was performed over three identical trials, and the total number of words correctly recalled was summed (maximum score = 45).

Known type 2 diabetes status was based on self-reports of doctor-diagnosed type 2 diabetes or use of oral glucose-lowering medication up to age 60–64 years. Previous work in the NSHD shows good validity of self-reported diabetes compared with general practitioner notes [24]. Medication use was recorded at ages 36, 43, 53 and 60–64 years and coded to the British National Formulary [25]. For this analysis, those with known type 1 diabetes were excluded (n = 6).

HbA_1c was measured in blood taken at age 60–64 years and 69 years by ion exchange HPLC on a Tosoh analyser (Tosoh Bioscience, Tessenderlo, Belgium). A fasting blood sample was collected at age 60–64 years. Samples were analysed for glucose (measured by enzymatic assay using hexokinase coupled to glucose 6-phosphate dehydrogenase, using a Siemens Dimension Xpand analyser, Siemens Medical Solutions, Erlangen, Germany) and insulin (measured by fluoroimmunoassay using a PerkinElmer AutoDELFIA analyser, PerkinElmer, Waltham, MA, USA). HOMA2-IR was calculated [26].

Blood samples from participants at age 53 years were genotyped using MetaboChip, a custom Illumina iSelectarray (San Diego, CA, USA) that includes ~200,000 SNPs and covers the loci identified by genome-wide association studies (GWASs) in cardiometabolic diseases, including rare variants identified by the 1000 Genomes Project [27]. Quality control analysis of genotyped samples has been previously reported [28].

Three PRSs were calculated. A type 2 diabetes PRS has previously been derived for NSHD study members [29]. In brief, a genetic risk score was computed using the published coefficients for 65 SNPs identified by a prior GWAS for type 2 diabetes [30, 31]. We additionally derived a PRS for insulin resistance using 17 previously demonstrated genome-wide significant SNPs [32] and for hyperglycaemia using ten previously demonstrated SNPs [32] using PRSice [33]. This calculates the sum of the number of risk alleles (unweighted score) carried by each person, and weights it based on previously published coefficients (weighted score). As is standard practice, SNPs with a minor allele frequency <0.01 were excluded.

Individuals were included in the analysis if they had: at least one measure of earlier-life factors (childhood SEP, childhood cognition, education); at least one measure of type 2 diabetes, insulin resistance or hyperglycaemia at age 60–64 years; and at least one measure of cognitive function (ACE-III or verbal memory) at age 69 years.

To explore whether the pattern of our results was affected by type 2 diabetes-related medication, we used two approaches. First, for those on diabetes medication, we replaced measured HbA_1c and HOMA2-IR scores with the maximum value within the cohort (used for all analyses where they are outcomes and predictors). Second, we re-ran the initial multivariable regression analyses using an additional covariate indicator for people on diabetes medication (n = 128) (ESM Fig. 4). Multivariable regression analyses were additionally re-run excluding participants with potentially clinically significant cognitive impairment (81 study members [5%] fell below the clinically validated ACE-III <82 threshold) and adjusting for duration of diabetes (ESM Figs 5, 6).

A higher PRS for type 2 diabetes was strongly and significantly associated with a higher likelihood of having known type 2 diabetes by age 60–64 years (OR 1.08, 95% CI 1.03, 1.11) and higher HOMA2-IR and HbA_1c levels at age 60–64 years (Table 2). There was little evidence that higher PRSs for type 2 diabetes, insulin resistance or hyperglycaemia were associated with lower ACE-III scores at age 69 years or lower verbal memory; the direction of regression coefficients suggested that higher risk scores were associated with lower cognitive state, though associations were weak and not significant at the 5% level (Table 2).

Figure 1 a shows simple path models between mid-life diabetes and its features and the ACE-III scores, adjusting for sex (similar to linear regression models adjusted for sex). Higher HOMA2-IR and known type 2 diabetes status at age 60–64 years were significantly associated with lower ACE-III scores (Fig. 1a). HOMA2-IR was the type 2 diabetes feature with the strongest coefficient with lower ACE-III scores (β = −0.09, 95% CI −0.15, −0.03) and lower verbal memory (β = −0.12, 95% CI −0.17, −0.06). HOMA2-IR was therefore subsequently selected for further modelling.

Figure 1 b shows life course path models representing associations between all covariables (childhood cognition, childhood SEP and education) and HOMA2-IR and later-life cognitive state. The numerical values refer to standardised regression weights whereby all paths are mutually adjusted and further adjusted for sex. Figure 2 and ESM Fig. 2, respectively, show path models replacing: (1) later-life cognitive state with later-life memory; and (2) HOMA2-IR with mid-life type 2 diabetes. Goodness-of-fit statistics indicated that all models were satisfactory (all models: RMSEA = 0.01, CFI = 1.0).

Path modelling revealed that when earlier covariables were mutually adjusted, the path from HOMA2-IR to the ACE-III was significantly attenuated (unadjusted β = −0.09, 95% CI −0.15, −0.03; adjusted β = −0.01, 95% CI −0.06, 0.03) (Fig. 1).

Paths from covariables to mid-life HOMA2-IR show mutually independent and significant paths from educational attainment (β = −0.16, 95% CI −0.23, −0.08) and childhood SEP (β = −0.07, 95% CI −0.14, −0.01) to HOMA2-IR at age 60–64 years, with the former path stronger than the latter (Fig. 1b). Paths from covariables to later-life cognitive state show direct mutually independent and significant paths from childhood SEP, childhood cognition and educational attainment to ACE-III scores at age 69 years (β = 0.07, 0.34 and 0.22, respectively), with the strongest path shown by childhood cognition.

Similarly, when verbal memory was the later-life cognitive outcome, the path from HOMA2-IR to memory was significantly attenuated when covariables were modelled (unadjusted β = −0.12, 95% CI −0.17, −0.06; adjusted β = −0.05, 95% CI −0.11, 0.01) (Fig. 2). The paths revealed a similar pattern, with education the strongest path coefficient to HOMA2-IR and childhood cognition the strongest path to later-life memory.

The pattern of findings was similar when type 2 diabetes was the intermediate mid-life diabetes feature (ESM Figs 2 and 3). The pattern of findings were similar but attenuated slightly when models were estimated adding in a covariate for those on diabetes medication up to age 69 years (ESM Fig. 4). Notably, when the analysis was additionally adjusted for diabetes medication use, the path from educational attainment to HOMA2-IR was largely attenuated (from β = −0.16 in the original model to β = 0.06 in the adjusted model) and became non-significant; the strongest path to HOMA2-IR in this model was childhood SEP (ESM Fig. 4). The pattern of findings for HOMA2-IR was similar when analyses were re-run excluding those with ACE-III scores <82 and adjusting for duration of diabetes (ESM Figs 5 and 6, respectively).

Our results show that the relationship between mid-life type 2 diabetes, insulin resistance and poorer later-life cognitive function or memory are likely to be confounded by the effects of earlier-life factors, in particular childhood advantage, associating with glycaemic status on one hand and late-life cognitive state on the other. Although it has previously been demonstrated that earlier-life influences, such as education, increase the risk of type 2 diabetes [8] and lower cognitive state and dementia [6], very few studies have investigated these shared influences in combination and how these relate to an association between type 2 diabetes and later-life cognitive state [14].

Using a PRS approach and, separately, a path model approach, we confirm and extend previous findings [15‐18] of a limited direct association between type 2 diabetes, hyperglycaemia and insulin resistance and later-life cognitive function. Within the same dataset, however, conventional regression analysis shows a strong association between mid-life type 2 diabetes and poor cognitive state. These discordant findings highlight the confounded nature of the observed association and the importance of considering life course influences. While the null association between metabolic PRS and later cognitive state could additionally be explained by the use of a weak genetic instrument and lack of power, and limitations of path analyses include assumptions about latent variables, directionality and unmeasured confounders, using these differential approaches to address the same underlying question—so called triangulation—helps to consolidate our findings [39]. Path analysis cannot prove causality, but it can be used to test models and disprove a model that postulates causal relationships among variables. Taking into account the strengths and limitations of both approaches, it is striking that both analyses suggest the same message: that there is no causal association between mid-life hyperglycaemia and later-life cognitive function at age 69 years.

In contrast to the null direct association between insulin resistance and subsequent cognitive function, we found that mid-life insulin resistance is associated with educational attainment and, independently, with childhood SEP. Mechanisms may include the adoption of healthier behaviours and, in particular, avoidance of obesity [40]. Further, we found that childhood cognition has a strong and independent effect on both older-age cognitive status and memory, with effect sizes equal to or stronger than those for education. A previous large-scale study associated genetic instruments for educational attainment and cognitive function separately with mid-life cognition: strong associations were observed for each of these with verbal-numeric reasoning, but not memory [16].

Other pathways related to type 2 diabetes or processes occurring as a complication of diabetes and its dysmetabolic precursors, such as small vessel disease and mixed vascular and neurodegeneration pathologies, may be aetiologically important in contributing to cognitive impairment/dementia risk [41].

The major strengths of this study include the unique resource of the NSHD cohort, with: direct measurement of life course factors, including the rarely available item childhood cognition, in a population-representative sample; and repeated collection of data on older-age cognitive state and glycaemic traits. In addition, we have sought to triangulate our results, taking into account the strengths and limitations of differential methods of PRS and path analysis.

Limitations include the relatively small sample size, precluding a formal Mendelian randomisation analysis, and differential loss to follow-up of those who were socioeconomically disadvantaged, which may result in an underestimation of the strength of associations between measures of childhood advantage and cognitive outcomes. In addition, there are relatively few people with diabetes, and HbA_1c, HOMA2-IR and cognitive state are mostly in a healthy range in this cohort at this age; subsequently, associations observed are expected to be weaker than they would be in more unhealthy ranges. We cannot wholly rule out reverse causality, i.e. that cognitive impairment may result in worse diabetes management. However, we also explored associations across the spectrum of insulin resistance and hyperglycaemia, where such a bias is less likely to be a problem. In addition, our path model analyses assumed that a normally distributed latent variable underlies responses of the ordinal scales used in the analyses (SEP and educational attainment) [20]. Comparisons between path estimates should be considered with this in mind.

Overall, our findings indicate that while mid-life type 2 diabetes and insulin resistance are associated with poorer later-life cognitive state, this association is, in part if not in full, confounded by earlier factors and is not consistent with a direct causal pathway of type 2 diabetes features per se to later-life lower cognitive state. Importantly, we also show that early-life measures of advantage, specifically education and cognition, appear to have independent effects on both older-age metabolic disease and cognitive state. These findings have important implications for developing interventions to reduce the risk of cognitive impairment. Despite efforts, and some promising findings [42], glucose-lowering agents are unlikely to be beneficial. Mechanisms accounting for the dual positive effects of prior advantage on mid-life insulin resistance and later-life cognitive state are not well understood and require further investigation.