This study used data from wave 2 (2004–2005) to wave 7 (2014–2015) of the ELSA, a prospective and nationally representative cohort of men and women living in England aged 50 years and over . A detailed description of the goals, design and methods of the ELSA has been published elsewhere . A flow chart of participant selection for the present study population based on inclusion and exclusion criteria is shown in Fig. 1. A total of 9432 individuals attended the wave 2 survey of the ELSA. Of these, 1766 were excluded from the present study because they did not have a nurse visit (clinical assessment). A further 1883 individuals were excluded for the following reasons: they had missing HbA1c results (n = 1851), they did not complete all of the cognitive tests (n = 17) or they had a confirmed diagnosis of dementia and/or Alzheimer’s disease at baseline (n = 15). An additional 594 individuals were excluded from the main analysis (but were included in a sensitivity analysis) because they were lost to follow-up from waves 3 to 7. The remaining 5189 participants (2329 men and 2860 women) with complete baseline data and at least one reassessment of cognitive function (waves 3–7) were included in the analyses reported here.
The ELSA was approved by the London Multicentre Research Ethics Committee (MREC/01/2/91). Informed consent was obtained from all participants.
Participants underwent a memory assessment through immediate and delayed recall of ten unrelated words. Both immediate and delayed recall scores ranged from 0 to 10, with higher scores indicating better memory performance. Immediate and delayed recall tests have been shown to have good construct validity and consistency . A composite memory score was created by summing the scores of the two individual memory tests. Executive function was assessed by a verbal fluency task in which participants were required to orally name as many animals as they could in 60 s. The task is well documented to be both reliable and valid, and has previously been used as an indicator of executive function for the ELSA population . The score for this task was the total number of words produced, excluding repeat words and non-animal words. Orientation was assessed by asking four questions regarding the date, i.e. day of month, month, year and day of week, and scoring one point for each correct answer. Generally, higher scores indicate better cognitive function.
To enable comparison across cognitive tests, z scores standardised to wave 2 were generated for individual tests by subtracting the mean score at wave 2 from the participant’s test score at each wave and dividing by the SD of the wave 2 scores. A composite global cognitive z score was calculated for each participant by averaging the z scores of the three tests and re-standardising to wave 2 using the mean and SD of the global cognitive z score at wave 2. A z score of 1 would therefore describe cognitive performance that is 1 SD above the mean score at wave 2. For all cognitive tests, we used standardised values in the regression analysis to allow for comparisons of regression coefficients across cognitive tests.
Measurement of HbA1c
In wave 2, blood samples were collected and sent to the Biochemistry Department at the Royal Victoria Infirmary, Newcastle, UK for laboratory analysis . Total HbA1c was measured by the Haematology Department at the Royal Victoria Infirmary using a Tosoh G7 analyser (Tosoh, Tokyo, Japan) . The analytical methods used for HbA1c measurement in the UK are required to be traceable to the work carried out in the Diabetes Control and Complications Trial (DCCT), part of the National Glycohemoglobin Standardization Program in the USA. The Secondary Reference Laboratory at the University of Minnesota was the main analytical laboratory for the DCCT work.
Definition of diabetes and prediabetes
Diabetes was defined as an HbA1c level ≥47.5 mmol/mol (6.5%), a self-reported physician diagnosis of diabetes or current use of glucose-lowering therapy. Among participants without diabetes, we defined prediabetes as an HbA1c level in the range 38.8–46.4 mmol/mol (5.7–6.4%), according to the 2014 American Diabetes Association guidelines . In participants with diabetes, HbA1c levels were further categorised using a standard clinical cut-off value of 53.0 mmol/mol (7.0%) to test the effect of glucose management on subsequent cognitive decline .
Covariates shown by previous studies to be associated with both HbA1c levels and cognitive function were selected for our analyses. These covariates included age, sex, total cholesterol, HDL-cholesterol, triacylglycerol, circulating high-sensitivity C-reactive protein (CRP), BMI, education, marital status, depressive symptoms, current smoking, alcohol consumption, hypertension, CHD, stroke, chronic lung disease and cancer. Details of covariates are available in the ESM Methods.
The results are presented as percentages for categorical variables and means ± SDs for normally distributed continuous variables. The results for high-sensitivity CRP and triacylglycerol are presented as medians with interquartile ranges because their distribution was highly skewed. The cross-sectional associations between HbA1c levels and cognitive scores at baseline were tested using multiple linear regression models, and linear mixed models were used to evaluate longitudinal associations. We also conducted longitudinal analyses to calculate the mean difference in the rate of change in cognitive scores (SD/year) and compared categories of baseline diabetes status using non-diabetic participants with normal HbA1c levels (<38.8 mmol/mol [5.7%]) as the reference group. Linear mixed models can incorporate all available follow-up data, account for the fact that repeated measures in the same participant are correlated with each other, and handle missing data. In the two models that we ran, both the intercept and the slope were fitted as random effects to account for inter-individual differences at baseline and different rates of change in cognitive function over the follow-up period. The first model included HbA1c levels (or diabetes status), time (years since baseline), time × HbA1c interaction, age (years) and sex (male or female). The time × HbA1c interaction term indicated differential change by each one unit increment in HbA1c from baseline to the end of the study. The second model additionally adjusted for baseline total cholesterol (mmol/l), HDL-cholesterol (mmol/l), triacylglycerol (mmol/l), high-sensitivity CRP (nmol/l), BMI (kg/m2), education (below level 3 National Vocational Qualification [NVQ3]/General Certificate of Education [GCE] A level, or above or equal to NVQ3/GCE A level), marital status (currently living alone or not), depressive symptoms (yes or no), current smoking (yes or no), one or more alcoholic drinks once or more per week (yes or no), hypertension (yes or no), CHD (yes or no), stroke (yes or no), chronic lung disease (yes or no) and cancer (yes or no).
We used a multiple imputation, chained-equations method to replace missing data for cognitive assessments during follow-up (waves 3–7) and used all available data from 5783 participants in the sensitivity analyses. Variables used to impute the missing values of cognitive scores included participants’ baseline information (age, sex, education, marital status, BMI, current smoking, alcohol consumption, diabetes and stroke) and baseline cognitive scores. For each longitudinal analysis, we created 20 imputed data sets and combined the results using the MIANALYZE procedure of SAS version 9.4 (SAS Institute, Cary, NC, USA). To detect differences in the rate of change in cognitive z scores between individuals with diabetes and those completely free of diabetes, we conducted another sensitivity analysis that excluded 261 participants with incident diabetes during follow-up.
Statistical analyses were performed using SAS software, version 9.4 (SAS Institute). All analyses were two-sided; an alpha value of 0.05 was considered the threshold for statistical significance.