Long-term Intake of Gluten and Cognitive Function Among US Women

Key Points Question Is dietary gluten intake associated with cognitive function in the general population? Findings In this cohort study of 13 494 women at midlife without a celiac disease diagnosis, there was no statistical evidence of any association of long- or short-term gluten intake with cognitive function. Meaning These findings suggest that in the absence of celiac disease, restriction of dietary gluten to maintain cognitive function is not warranted.


Dietary assessment
Diet was assessed in 1991, 1995, 1999, 2003, 2007, 2011 and 2015 in Nurses' Health Study II (NHSII) using a semi-quantitative food frequency questionnaire (FFQ, eFigure 1). Participants were asked how frequently they consumed each food item in common portion size during the past year, with nine possible responses ranged from "Never or less than once per month" to "6+ per day".
To estimate servings of food groups consumed per day, we first converted the FFQ responses to servings/day and then summed up all food items in each food group. For example, whole grains included oats, brown rice, dark bread, whole-grain cracker, whole-grain cereal, bran, and wheat germ, while refined grains included white bread, white rice, muffin, pancake, pasta, and tortilla. We used the Harvard T. H. Chan School of Public Health nutrient database updated during each questionnaire cycle to calculate nutrient intakes. Previous validation studies in men and women have demonstrated comparable reproducibility (intraclass correlation coefficients: 0.49-0.71 in women) and modest correlation (Spearman correlation coefficient r: 0.36-0.75 in women) between FFQs and one-week food diaries. 1,2 Specifically, for gluten intake, we first identified food items with gluten-containing ingredients like wheat, wheat bran, wheat germ, rye, barley, cereal, and pasta, according to the ingredient lists on commercial product labels provided by manufacturers and recipes on cookbooks for home-cooked items. Then, we quantified the gluten content of the gluten-containing ingredients in all food item by multiplying the protein content of these ingredients with an approximate proportion of gluten (75%), in line with previous studies, 3,4 though the proportion of gluten in protein may be more variable in barley and rye than wheat. 5 We did not calculate the trace amount of gluten in oats and certain condiments like soy sauce since the quantity of gluten in these foods would be negligible compared to the total gluten intake. 6 Next, we added up gluten from all foods consumed to estimate total gluten intake.
In a recent validation study using 7-day dietary records, FFQ-derived gluten intake showed moderate to high validity (r: 0.58 for gluten, median r: 0.60 for gluten-rich foods). 7 Another validation studies also showed reasonable correlation between FFQ-derived and 7-day diet records-derived major sources of gluten (e.g., r: 0.79 for cold cereal). 8 Zong et al. has reported the trend of gluten intake over the past two decades in NHSII and major sources of gluten consumed in 2011, including whole-grain bread, pasta, cold cereal, pizza, and wheat products like pretzel, bagel, muffin, crackers, white bread, and tortilla. 4 To account for confounding from total energy intake and potential under-or over-reporting, we adjusted gluten intake for total energy intake using the residual method. As previously described, 9 absolute nutrient intake was regressed on total energy intake to get the nutrient residual, which was uncorrelated with the total energy intake, allowing direct evaluation of the variation in nutrient composition. To estimate long-term intakes, we averaged across FFQs preceding the cognitive assessment. For participants who completed cognitive assessment before the 2015 questionnaire cycle, we averaged daily gluten intake reported in 1991, 1995, 1999, 2003, 2007, and 2011. For the rest of participants, we averaged daily gluten intake reported in 1991, 1995, 1999, 2003, 2007, 2011, and 2015. We categorized participants into quintiles of energy adjusted gluten intake (0.1-<5.1, 5.1-<5.9, 5.9-<6.6, 6.6-<7.6,7.6-<18.3 g/day) to assess potential non-linear relationships and threshold of effect.

Cognitive assessment
We used the CogState Brief Battery to assess cognitive function of participants. 10 The CogState Brief Battery was self-administered at home via a web-based interface and took about 15-20 minutes to complete. 10 As the battery were loaded from Internet and ran locally on participants' computers, the cognitive performance would not be affected by Internet connection and speed. Hardware differences would cause minimal time variation as well (on the order of milliseconds). 11 The validity of this unsupervised online battery has been well-demonstrated, [12][13][14] and its utility has been evaluated in large population-based epidemiological studies. 10,11 Moreover, the CogsState battery was sensitive in detecting mild cognitive impairment 12,15 and had comparable performance in supervised and unsupervised settings. 14 Consistent with previous studies, [11][12][13][14] we computed composite scores on these tasks as these scores may be more sensitive to detect cognitive variations than individual task scores. We averaged the z-scores on Detection and Identification to calculate psychomotor speed/attention score, averaged the z-scores on One Card Learning and One Back to calculate learning/working memory score, and averaged the z-scores on all four tasks to calculate the global cognition score. These composite scores were then re-standardized using z-normalization so that one unit increase in score indicated 1 SD higher than the sample mean. The composite score was not computed if any of the contributing task scores was missing. We excluded implausibly low score based on cutoffs of percent correctness for each task recommended by CogState (Detection: <0.8; Identification: <0.8; One Card Learning: <0.5; One Back: <0.7), because they were likely due to technical errors. These composite scores have been shown to have high test-retest reliability (r=0.95) in identifying adults with cognitive impairment. 12 Sumner et al. has shown that the standardized loadings for responses times on Detection (0.68) and Identification (0.95) for the psychomotor speed/attention factor and the standardized loadings for responses times on One Card Learning (0.69) and One Back (0.85) for the learning/working memory factor were high in a confirmatory factor analysis of NHSII cohort, 11 indicating that these composite scores were applicable in our cohort.

Covariates
Participants self-reported date of birth, race (White/non-White), husband's education (high school or below, college, graduate, unmarried or missing), family income (dollar), personal history of diseases (hypertension, diabetes, hypercholesterolemia, myocardial infarction, depression, cancer, and dementia), body mass index (BMI, kg/m 2 ), smoking status (never, past, current, and missing n=12), regular use of medications (antidepressant, aspirin, nonsteroidal anti-inflammatory drugs, postmenopausal hormone), use of multivitamin, and alcohol intake (g/day) in questionnaires. We averaged BMI, physical activity, total energy intake, and alcohol intake across questionnaire cycles from 1991 to the most recent questionnaire (2011, 2013, or 2015) prior to cognitive assessment, and acquired other covariates from the most recent questionnaire prior to cognitive assessment. Husband's education was measured in 1999 (missing n=391). We used husband's education as an indicator for socioeconomic status because participants in NHSII were nurses with similar education background. We included family income as a covariate to additional capture another domain of socioeconomic status of participants, since it has been found to be associated with diet quality and cognitive function. 16,17 We assessed current depression status based on self-reported depression diagnosis, treatment, and clinically relevant depressive symptoms (≥10) according to the validated 10-item Centre for Epidemiological Studies Depression Scale (CES-D-10), which was assessed in 2013. The CES-D-10 includes 3, 5, and 2 questions on depressed effect, somatic symptoms, and positive affect in the past week, respectively, with options ranging from "rarely or none of the time", which contributes 0 score, to "all of the time", which contributes 3 scores. 18 We included depression status as a covariate because it has been shown that depression was associated with cognitive deficits. 19 We computed the Alternative Health Eating Index 2010 (AHEI-10) as described previously, 20 without alcohol and whole grain components. The included AHEI-10 components each assigned a score of 0-10, with higher scores for higher intakes of vegetables, fruits, nuts and legumes, long-chain (n-3) fatty acids, and poly-unsaturated fatty acids, but lower scores for higher intakes of fruit juice and sugar-sweetened beverages, red and processed meats, trans fats, and sodium. 20 This score has been widely used in NHSII studies to assess diet quality 4,11 and was strongly associated with major chronic diseases like coronary heart disease, diabetes, and cancer. 20 Like other dietary variables, we averaged AHEI-10 scores across previous questionnaires.
For 22, 5, and 8 participants with missing information on family income, BMI, and AHEI-10, respectively, we imputed missing values with median values of these variables. For 2 participants with missing information on aspirin or non-steroidal anti-inflammatory drug use, we carried forward the last non-missing response.

Statistical analysis
We included the above covariates in our multivariable-adjusted linear regression model examining mean difference in each standardized cognitive score comparing across quintiles of gluten intake. We used a linear trend analysis to test the overall significance of the gluten variable and whether the cognitive score increased or decreased across gluten quintiles, by assigning the median quintile value to each gluten category (4.45 g/day, 5.48 g/day, 6.23 g/day, 7.03 g/day, 8.34 g/day) and testing this variable as a continuous variable. This trend test has been widely used in prior analyses of data from our cohort. 3,4,20,21 To test the robustness of this trend test, we modeled the continuous gluten intake variable in g/day to test its linear relationships with cognitive scores.
To assess whether gluten intake during different time periods relative to the cognitive assessment would produce different results, we first calculated gluten intake assessed during each 4-year FFQ cycle (≤4 years, 4-8 years, 8-12 years, 12-16 years, 16-20 years, and 20-24 years before cognitive assessment), average gluten intake in distant past (12-24 years), and average gluten intake in recent past (4-12 years), and then used each of these gluten variables in the linear model as the main exposure. To test whether change in gluten intake over time would affect cognitive function, we calculated the change in gluten intake from distant to recent past (mean: -0.87 g/day) and examined the mean difference in each cognitive score associated with per 1 SD (1.98 g/day) increase in gluten intake.
Other statistical analyses have been described in detail in the Methods section of main texts.  a Energy-adjusted gluten intake was cumulatively averaged from 1991 to the last questionnaire cycle preceding the cognitive assessment. b P-trend was calculated using the median gluten intake for each decile as a continuous variable. c Standardized scores for psychomotor speed/attention, learning/working memory, and global cognition were calculated by standardizing as the mean of the standardized scores of the following CogState battery tasks: Detection + Identification, One Card Learning + One-Back, and all four tasks, respectively. Higher scores indicate better performance, with one unit increase represents 1 SD higher than the mean. d Model was adjusted for age (years), race (white, non-white), body mass index (kg/m2), husband's educational attainment (high school or below, college, graduate school, unmarried/missing), family income (dollar), history of diabetes (yes, no), history of hypertension (yes, no), history of hypercholesterolemia (yes, no), history of myocardial infarction (yes, no), current depression status (yes, no), smoking (never, past, current, missing), aspirin or non-steroid anti-inflammatory drug use (yes, no), multivitamins use (yes, no), physical activity (metabolic equivalents/week), menopausal status and hormone use (pre-menopause, post-menopause and never/past/current user of hormone therapy), total energy intake (kcal/d), alcohol intake (g/d), and Alternative Healthy Eating Index score (excluding alcohol and whole grains). a Energy-adjusted gluten intake was cumulatively averaged from 1991 to the last questionnaire cycle preceding the cognitive assessment. Model was adjusted for age (years), race (white, non-white), body mass index (kg/m 2 ), husband's educational attainment (high school or below, college, graduate school, unmarried/missing), family income (dollar), history of diabetes (yes, no), history of hypertension (yes, no), history of hypercholesterolemia (yes, no), history of myocardial infarction (yes, no), current depression status (yes, no), smoking (never, past, current, missing), aspirin or nonsteroid anti-inflammatory drug use (yes, no), multivitamins use (yes, no), physical activity (metabolic equivalents/week), menopausal status and hormone use (pre-menopause, post-menopause and never/past/current user of hormone therapy), total energy intake (kcal/d), alcohol intake (g/d), Alternative Healthy Eating Index score (excluding alcohol and whole grains), and refined grain intake. Standardized score for psychomotor speed and attention was calculated by standardizing the mean standardized scores of two CogState battery tasks, Detection and Identification. A higher score indicated better performance, with one unit increase representing 1 SD higher than the mean. b P-trend was calculated using the median gluten intake for each quintile as a continuous variable. c P-interaction was estimated using an interaction term of gluten intake and the respective stratifying variable. b P-trend was calculated using the median gluten intake for each quintile as a continuous variable. c Standardized scores for psychomotor speed/attention, learning/working memory, and global cognition were calculated by standardizing as the mean of the standardized scores of the following CogState battery tasks: Detection + Identification, One Card Learning + One-Back, and all four tasks, respectively. Higher scores indicate better performance, with one unit increase represents 1 SD higher than the mean. d Model was adjusted for age (years), race (white, non-white), body mass index (kg/m 2 ), husband's educational attainment (high school or below, college, graduate school, unmarried/missing), family income (dollar), history of diabetes (yes, no), history of hypertension (yes, no), history of hypercholesterolemia (yes, no), history of myocardial infarction (yes, no), current depression status (yes, no), smoking (never, past, current, missing), aspirin or non-steroid anti-inflammatory drug use (yes, no), multivitamins use (yes, no), physical activity (metabolic equivalents/week), menopausal status and hormone use (pre-menopause, post-menopause and never/past/current user of hormone therapy), total energy intake (kcal/d), alcohol intake (g/d), and Alternative Healthy Eating Index score (excluding alcohol and whole grains). b P-trend was calculated using the median gluten intake for each quintile as a continuous variable. c Standardized scores for psychomotor speed/attention, learning/working memory, and global cognition were calculated by standardizing as the mean of the standardized scores of the following CogState battery tasks: Detection + Identification, One Card Learning + One-Back, and all four tasks, respectively. Higher scores indicate better performance, with one unit increase represents 1 SD higher than the mean. d Model was adjusted for age (years), race (white, non-white), body mass index (kg/m 2 ), husband's educational attainment (high school or below, college, graduate school, unmarried/missing), family income (dollar), history of diabetes (yes, no), history of hypertension (yes, no), history of hypercholesterolemia (yes, no), history of myocardial infarction (yes, no), current depression status (yes, no), smoking (never, past, current, missing), aspirin or non-steroid anti-inflammatory drug use (yes, no), multivitamins use (yes, no), physical activity (metabolic equivalents/week), menopausal status and hormone use (pre-menopause, post-menopause and never/past/current user of hormone therapy), total energy intake (kcal/d), alcohol intake (g/d), and Alternative Healthy Eating Index score (excluding alcohol and whole grains). b P-trend was calculated using the median gluten intake for each quintile as a continuous variable. c Standardized scores for psychomotor speed/attention, learning/working memory, and global cognition were calculated by standardizing as the mean of the standardized scores of the following CogState battery tasks: Detection + Identification, One Card Learning + One-Back, and all four tasks, respectively. Higher scores indicate better performance, with one unit increase represents 1 SD higher than the mean. d Model was adjusted for age (years), race (white, non-white), body mass index (kg/m 2 ), husband's educational attainment (high school or below, college, graduate school, unmarried/missing), family income (dollar), history of diabetes (yes, no), history of hypertension (yes, no), history of hypercholesterolemia (yes, no), history of myocardial infarction (yes, no), current depression status (yes, no), smoking (never, past, current, missing), aspirin or non-steroid anti-inflammatory drug use (yes, no), multivitamins use (yes, no), physical activity (metabolic equivalents/week), menopausal status and hormone use (pre-menopause, post-menopause and never/past/current user of hormone therapy), total energy intake (kcal/d), alcohol intake (g/d), and Alternative Healthy Eating Index score (excluding alcohol and whole grains). a Energy-adjusted gluten intake was derived from individual assessment conducted in the respective year category, except for distant past and recent past intakes, which were cumulative averaged through 12-24 years and 4-12 years prior to cognitive assessment., respectively. Model was adjusted for age (years), race (white, non-white), body mass index (kg/m 2 ), husband's educational attainment (high school or below, college, graduate school, unmarried/missing), family income (dollar), history of diabetes (yes, no), history of hypertension (yes, no), history of hypercholesterolemia (yes, no), history of myocardial infarction (yes, no), current depression status (yes, no), smoking (never, past, current, missing), aspirin or non-steroid anti-inflammatory drug use (yes, no), multivitamins use (yes, no), physical activity (metabolic equivalents/week), menopausal status and hormone use (pre-menopause, post-menopause and never/past/current user of hormone therapy), total energy intake (kcal/d), alcohol intake (g/d), and Alternative Healthy Eating Index score (excluding alcohol and whole grains).
b P-trend was calculated using the median gluten intake for each quintile as a continuous variable c Standardized scores for psychomotor speed/attention, learning/working memory, and global cognition were calculated by standardizing as the mean of the standardized scores of the following CogState battery tasks: Detection + Identification, One Card Learning + One-Back, and all four tasks, respectively. Higher scores indicate better performance, with one unit increase represents 1 SD higher than the mean. d Distant and recent past intakes were mutually adjusted. Global cognition -0.01 (-0.02, 0.01) 0.62 a Energy-adjusted gluten intake was cumulatively averaged from 1991 to the last questionnaire cycle preceding the cognitive assessment. Model was adjusted for age (years), race (white, non-white), body mass index (kg/m 2 ), husband's educational attainment (high school or below, college, graduate school, unmarried/missing), family income (dollar), history of diabetes (yes, no), history of hypertension (yes, no), history of hypercholesterolemia (yes, no), history of myocardial infarction (yes, no), current depression status (yes, no), smoking (never, past, current, missing), aspirin or nonsteroid anti-inflammatory drug use (yes, no), multivitamins use (yes, no), physical activity (metabolic equivalents/week), menopausal status and hormone use (pre-menopause, post-menopause and never/past/current user of hormone therapy), total energy intake (kcal/d), alcohol intake (g/d), and Alternative Healthy Eating Index score (excluding alcohol and whole grains).
b Standardized scores for psychomotor speed/attention, learning/working memory, and global cognition were calculated by standardizing as the mean of the standardized scores of the following CogState battery tasks: Detection + Identification, One Card Learning + One-Back, and all four tasks, respectively. Higher scores indicate better performance, with one unit increase represents 1 SD higher than the mean.