Bidirectional Associations Between Adiposity and Cognitive Function: A Prospective Analysis of the Canadian Longitudinal Study on Aging (CLSA)

Abstract

Background

Theoretical perspectives suggest that adiposity and cognitive function may be bidirectionally associated, but this has not been examined in a large-scale data set. The current investigation aims to fill this gap using a large, representative sample of middle-aged and older adults.

Methods

Using data from the Canadian Longitudinal Study on Aging (N = 25 854), the bidirectional hypothesis was examined with 3 indicators of cognitive function (ie, executive function, processing speed, and verbal fluency) and adiposity (ie, waist circumference [WC], body mass index [BMI], and total fat mass). We used multivariate multivariable regression and structural equation modeling to assess the prospective associations between adiposity and cognitive indicators.

Results

Analyses revealed that higher baseline WC was associated with higher Stroop interference at follow-up for both middle-aged (standardized estimate, β = 0.08, 95% confidence interval [CI] 0.06, 0.10) and older adults (β = 0.07, 95% CI 0.04, 0.09). Similarly, higher baseline Stroop interference was also associated with higher follow-up WC in middle-aged (β = 0.08, 95% CI 0.06, 0.10) and older adults (β = 0.03, 95% CI 0.01, 0.06). Effects involving semantic fluency and processing speed were less consistent. The earlier effects were similar to those observed using other adiposity indicators (eg, BMI and total fat mass) and were robust to adjustment for demographics and other cofounders, and when using latent variable modeling of the adiposity variable.

Conclusion

Evidence for a bidirectional relationship between adiposity and cognitive function exists, though the associations are most reliable for executive function and primarily evident at midlife.

The prevalence of obesity is rising sharply around the world because of increasingly sedentary lifestyles and excess intake of calorie-dense foods (1). The World Health Organization reported that more than 1.9 billion people (≥18 years) worldwide are overweight, and over 650 million of them are obese (2). Based on a report by Statistics Canada, approximately 63.1% of the adult Canadians were either overweight (36.3%) or obese (26.8%) in 2018 (3); in the United States, corresponding figures were 31.1% for overweight and 42.5% for obesity (4). Excess adiposity is recognized as a risk factor for numerous chronic diseases, including type 2 diabetes mellitus (T2DM), heart disease, and cancer (5). Beyond effects on the endocrine, cardiovascular, and immune systems, obesity may also affect the central nervous system adversely (6). Promoting optimal brain health is a public health priority considering the growing number of older adults and an associated increase in neurodegenerative disease incidence globally (7).

A significant body of research suggests a negative association between obesity and cognitive performance (8,9). The detrimental effects of obesity are especially evident in the domains of short-term memory and executive function (9–11). For example, individuals with higher body mass index (BMI) perform poorly on verbal learning tasks (eg, delayed recall and recognition of words) compared with those with lower BMI (10,12). Performance decrements are also observed among obese individuals for tasks that require significant executive control requirements to complete, such as concept formation and set shifting in Wisconsin card sorting test (13). On a structural brain level, obesity is associated with reduced white matter integrity in prefrontal cortex (PFC) tracts involved in executive control (14). Beyond executive control, obesity adversely affects other domains, including psychomotor function, selective attention, decision making, planning, and problem-solving (9). Finally, clinical epidemiology studies document reliable associations between obesity and cognitive impairment (eg, dementia) among older adults (15).

Studies examining the link between obesity and cognitive function are typically undertaken with the assumption of a unidirectional relationship between the two, such that obesity predates—and contributes to—future cognitive dysfunction. However, there remains the possibility of a bidirectional influence (16). To illustrate, executive function, a cognitive process strongly linked with the lateral PFC (17), enables individuals to control thoughts, actions, and emotions (18). Impaired lateral prefrontal function can lead to increased impulsivity and reduced inhibitory control; it has been shown to be predictive of eating tendencies (19), emotional eating (20), and binge eating (21) among members of the general population. A causal role of lateral PFC on eating indulgence has also been demonstrated in experimental studies. For example, experimentally attenuating the dorsolateral PFC using suppressive brain stimulation results in disinhibited eating in the laboratory context (22), which can be enhanced further in the presence of facilitative cues to consume (23). Likewise, large-scale population studies have demonstrated a correlation between lateral prefrontal morphology and body composition in early life, long before the brain could register the cumulative impact of decades of adiposity (24). The earlier experimental and epidemiological findings underscore the possibility that relationships between executive function and obesity may be bidirectional, when considered over extended periods of time (6,16). However, this proposition has not been tested in an explicit manner in a large, population-representative data set with sufficient statistical power to detect subtle, cumulative effects.

Aside from eating behavior, several other factors could influence or mediate the hypothesized bidirectional association, including physical activity, hypertension, and T2DM, as each of these are also associated with cognitive performance (25,26). Similarly, an argument can also be made that people with lower executive function may find it disproportionately challenging to consistently enact an active lifestyle (27), given the inherent requirements for self-organization, planning, and inhibiting indulgence in distracting sedentary activities (eg, screen time). Accordingly, it is plausible that both eating and physical activity may mediate brain health associations with adiposity, and further that this could occur with or without the development of clinical mediating conditions (ie, hypertension and T2DM).

The current investigation aims to test bidirectional associations between adiposity and cognitive function, using a population-representative sample of middle-aged and older adults in Canada. Consistent with prior research, we anticipate that higher adiposity at baseline will be associated with lower follow-up cognitive function; conversely, however, we also anticipate that lower baseline cognitive function will be associated with significantly higher follow-up adiposity. Given that older adulthood involves longer exposure to potential adverse effects of obesity on the brain, we expect that any bidirectional associations between adiposity and cognitive function would be more prominent in the middle-aged subsample than the older adult sample. We also hypothesize that the earlier-mentioned bidirectional associations would be mediated through lifestyle behavior (eg, physical activity and diet) and adverse health conditions (eg, high blood pressure [BP] and diabetes).

Method

Data Source and Study Settings

This study involved a 3-year prospective analysis of the baseline and first follow-up data from the Canadian Longitudinal Study on Aging (CLSA). The recruitment process of the CLSA commenced in 2010 and completed baseline data collection in 2015. The current analysis used the comprehensive cohort from within the CLSA, which includes 30 097 Canadians aged between 45 and 85 years. These participants were interviewed both at home and at hospital-based data collection sites for comprehensive assessments. The comprehensive cohort participants were recruited from provincial health registries and by telephone sampling (random digit dialing). The CLSA exclusion criteria were as follows (a) inability to communicate in 1 of the 2 national languages, English or French; (b) cognitive impairment at the time of recruitment; (c) resident of the 3 territories; (d) full-time member of the Canadian Armed Forces; (e) resident in a long-term care institution; and (f) living on Federal First Nations reserves or other First Nations settlements. As initially planned, the follow-up data for the CLSA will be collected every 3 years for a period of 20 years. Currently, baseline and first follow-up data collection have been completed and released; second follow-up data were not yet available to the researchers. The details of the study design and recruitment process have been published elsewhere (28). The present study received ethics approval from the Office of Research Ethics, University of Waterloo (ORE# 41434). The data access application was approved by the CLSA Data and Sample Access Committee (Application# 1906024).

Any CLSA participants who could not walk without assistance or for whom English or French was a second language were excluded from the present statistical analyses, to remove these confounds from the measurement of the language-mediated cognitive tests (ie, Stroop task and animal naming). Likewise, individuals diagnosed with clinical conditions that might affect cognitive function trajectories (ie, multiple sclerosis, dementia, stroke, and Parkinson’s disease) were excluded from the analysis. Exclusion criteria were applied on the baseline measures. Following the earlier exclusions, the effective sample for the present analyses comprised 25 854 participants from the CLSA comprehensive cohort (Supplementary Figure 2, Supplementary Table 1 and 2).

Measures of Adiposity

Waist circumference

Waist circumference (WC) was measured around the position of the natural indent in the waist area (halfway between the last floating rib and the iliac crest) (29,30). This measurement was taken at the data collection sites. A higher WC indicates excess fat deposition around the midsection. According to the Centers for Disease Control and Prevention, an ideal WC for men is less than 40 inches, and for women is less than 35 inches (31). In general, the measures of centralized obesity (ie, WC) are considered preferred measures of adiposity because of their superior predictive validity, as compared with BMI (32).

Body mass index

Weight and standing (shoeless) height of the CLSA comprehensive cohort participants were measured at each data collection wave. Two measurements were taken for each variable and then averaged together (30). BMI was calculated as weight in kg/height in m². A BMI ranging between 18.5 and 24.9 is considered “normal,” whereas a BMI between 25 and 29.9 is deemed “overweight,” and over ≥30 is classified as “obese.”

Total fat mass

This measure of adiposity was assessed only at the CLSA baseline data collection using dual-energy x-ray absorptiometry (DXA) (30). Body fat percentage was calculated and provided in the data set as a continuous variable.

Measures of Cognitive Function

Stroop Neurological Screen Test

The Stroop Neurological Screen Test (SNST) was identified as the primary indicator for executive function within the CLSA. Stroop paradigms assess the ability to inhibit cognitive interference that arises when the processing of one stimulus is impeded by concurrent processing of another stimulus (33,34). This task is typically presented in 3 consecutive blocks. The first “neutral” block requires the participants to read neutral words (eg, chair, table, boat, and window). The second block is a “congruent” block where the participants are provided with a list of color words written in a manner that the font color and the name of the color are identical, such as green is written in green font. In the final “incongruent” block, a list of color words is provided, but the color words are printed in a manner that mismatches with the font color, such as green is written in red font. When the color word and font color mismatch, the time required to identify the font color increases considerably because people tend to read the color names automatically (33,34). This is denoted as the “Stroop interference effect” and is calculated by taking the difference of completion times between incongruent and congruent blocks. Higher Stroop interference is taken as an indicator of weak executive control. The CLSA implemented the Victoria version of the Stroop task (28,30,33,34). In the first block, participants were provided with a card that contained a list of neutral words written in different fonts. Participants were instructed to read the words. In the next block, participants were provided with a card containing a list of “X”s printed in different font colors. Participants were instructed to name the color of the font in which “X”s were printed. Finally, participants were provided with a card containing the name of color words printed in incompatible font color in the third block. Participants were instructed to name the color of the font while ignoring the meaning of the color words. The completion time was calculated in seconds for each block and provided in the data set. Among the 3 cognitive tasks included in this investigation, Stroop task is considered as the most robust measure of executive function (35).

Choice reaction time

Choice reaction time (CRT) primarily assesses speed of information processing (36), but performance on this task largely depends on several executive control components, such as working memory and attention (37,38). CLSA participants performed the CRT task at the data collection sites on a computer with a touch screen. Participants were presented with a horizontal row of 4-plus signs on the computer screen (30,39). One of the plus signs changed to a box after 1 000 milliseconds, and the participants were required to press the box on the touch screen as soon as possible (30,39). There were 10 practice trials in the task followed by 52 test trials (30,39). The reaction time (in milliseconds) of the participants was recorded automatically by the computer software. The mean reaction time was calculated as the average of the correct response of the test trials, excluding incorrect answers and timeouts. Higher reaction time indicates poor speed visual information processing.

Animal fluency test

Semantic fluency tasks are commonly used to measure memory store integrity, thereby facilitating diagnosis of disorders of aging, such as Alzheimer’s disease and other dementias (40). Such tasks involve verbally naming as many words as possible from a particular thematic category (eg, animals, fruits, or phonemic) in a specified period of time. The animal fluency test (AFT) assesses semantic fluency by asking individuals to name as many animals as they can in 60 seconds. One point is rewarded for mentioning each unique animal. As this task requires word retrieval (eg, animal names) while meeting certain constraints (eg, only animals, avoid repetition and proper nouns), people with sound cognitive function tend to produce more correct words (41). A score below 15 is generally taken to indicate impaired cognitive function (40). Beyond semantic fluency, performance on the AFT also requires some secondary demands involving executive control (42), but less so than the Stroop paradigm.

Within the CLSA, the comprehensive cohort participants performed the AFT during their in-home interview, and their responses were recorded (30,43). The recording was transcribed, and the animal names were coded based on scientific taxonomic classification. Animals with the same scientific taxonomic classification with variant names (eg, cougar and puma or salmon and salmon fish) were labeled with the same code. Animals with different scientific taxonomic classifications were labeled with unique codes. Scoring was conducted using a validated algorithm such that all unique codes received 1 point after excluding any matched lower taxonomic classifications. For instance, where participants mentioned “bird, parrot, pheasant,” only parrot and pheasant received a point but bird did not, because bird is the category that includes both parrot and pheasant (39).

Covariates and Mediators

Age

The CLSA participants were asked about their exact date of birth (43). Age was calculated using the date of birth and provided as integer values in the data set. For the purpose of this study, the analyses included examination of 2 age groups, corresponding with working age (45–65 years, n = 16 147) and retirement age (>65 years, n = 9 707). This approach was taken given the considerably different daily demands—in terms of exercise opportunities and eating constraints—between the 2 age groups, as well as the higher age-related cognitive deficits that would tend to selectively affect the latter age group.

Sex

Participants were asked to report their biological sex using the following item: “What was your sex at birth?” This variable was coded as 1 and 0, where 1 denotes male sex (43).

Ethnicity

The CLSA participants were asked about their cultural and racial backgrounds, such as White, Chinese, South Asian, Black, etc. This variable was coded as 1 and 0, where 1 denotes White, and 0 denotes all non-White ethnicities (43).

Income

The annual household income of the participants was used to assess income status. The participants were asked: “What is your best estimate of the total household income received by all household members, from all sources, before taxes and deductions, in the past 12 months?” (43). This was an ordinal variable with the following categories: <$20 000; $20 000–$50 000; $50 000–$100 000; $100 000–$150 000; and ≥$150 000 (43). The missing values were coded as “no response.”

Education

The level of education was measured at baseline, using 2 separate questions. At first, the participants were asked: “Have you received any other education that could be counted towards a degree, certificate, or diploma from an educational institution?” (43). Those individuals who answered “no” were considered to have an education level “high school or less.” Those participants who answered “yes” were asked a follow-up question: “What is the highest degree, certificate, or diploma you have obtained?” (43). The education variable was recoded with the following categories: “high school or less,” “certificate or degree below bachelor,” and “bachelor or above.”

Residence

The area of residence was classified as rural, urban core, urban fringe, urban population center outside a census metropolitan area and census agglomeration, secondary core, and postal code link to dissemination area (43). This variable was recoded as “rural” and “urban,” where urban included all the nonrural categories.

Comorbidity index

To adjust for the comorbidity load, we created a comorbidity index. Participants were asked whether a physician ever told them that they have chronic conditions (43,44). From the list of chronic conditions, we included 22 in the comorbidity index (Supplementary Materials). The index was created by summing across all chronic conditions included in the index. The comorbidity index was recomputed as required for the mediation analyses to exclude diabetes or hypertension when these chronic diseases were the target mediators.

Sleep duration

The participants were asked, “During the past month on average how many hours of actual sleep did you get at night?” The number of hours spent on sleeping provided in the data set as a numeric variable.

Physical activity

The CLSA participants were asked, “Over the past 7 days, how often did you take a walk outside your home or yard for any reason? For example, for pleasure or exercise, walking to work, walking the dog, etc.” (45). This was an ordinal variable with the following categories: “never,” “seldom (1–2 days),” “sometimes (3–4 days),” and “often (5–7 days).” The missing values were coded as “no response.”

Diet

The variables related to dietary behaviors were derived from the “Short Diet Questionnaire” (43,46). The intake of legumes, fruits, green salads, and carrot were selected as the surrogate of healthy food choice. The CLSA participants were asked: (i) “How often do you usually eat legumes: beans, peas, lentils?” (ii) “How often do you usually eat fruit (fresh, frozen, or canned)?” (iii) “How often do you usually eat green salad (lettuce, with or without other ingredients)?” and (iv) “How often do you usually eat carrots (fresh, frozen, canned, eaten on their own or with other food, and cooked or raw)?” (43). These variables were coded as per day, per week, per month, and per year. The variables were recoded with the following categories “daily,” “weekly,” and “rarely.” The average caloric density of these items is 54 kcal per 100 gm, according to standardized estimates (47). On the other hand, the intake of fries, snacks, pastries, and chocolate were selected to represent unhealthy food choices. Participants were asked, (i) “How often do you usually eat french fries or pan-fried potatoes, poutine?” (ii) “How often do you usually eat salty snacks (regular chips, crackers, …)?” (iii) “How often do you usually eat cakes, pies, doughnuts, pastries, cookies, muffins, …?” and (iv) How often do you usually eat chocolate bars? (43). These variables were also recoded as “daily,” “weekly,” and “rarely.” The average caloric density of these items is 434 kcal per 100 gm, according to standardized estimates (47).

Blood pressure

BP (systolic and diastolic) was measured 6 times for each participant using the BpTRU BPM200 BP Monitor (30). The average of systolic BP and the average of diastolic BP (excluding first reading) were provided separately in the data set as in units of millimeters of mercury (mmHg).

Diabetes

To know about diabetic status, participants were asked, “Has a physician ever told you that you have diabetes, borderline diabetes or that your blood sugar is high?” (29). Participants who responded “yes” were asked a follow-up question about their type of diabetes, that is, Type 1, Type 2, or neither. The diabetes variable was derived from these questions and recoded as “Type 2”, “other type,” and “none.”

Statistical Analyses

Statistical analyses were performed using R statistical software version 4.1.0 (48). To assess bidirectional associations, we used multivariate multivariable regression and structural equation modeling (SEM). The statistical analyses were adjusted for analytic weight as per the recommendation of the CLSA (49). The BMI variable was power transformed by −0.7 for variance stabilization and/or nonnormality of the residuals of the model.

The brain-as-outcome path (adiposity → cognition) was assessed using baseline WC or BMI^−0.7 or DXA as an independent variable and three follow-up cognitive tasks (ie, Stroop, AFT, and mean reaction time [MRT]) as dependent variables in the multivariate multivariable regression analyses. The brain-as-predictor path (cognition → adiposity) was assessed using baseline Stroop interference/AFT/MRT as an independent variable and 2 follow-up adiposity measures (ie, WC and BMI^−0.7) as dependent variables. A total of 3 models were assessed for each regression path. Model 1 was the unadjusted model; Model 2 was adjusted for major sociodemographic characteristics (eg, age, sex, income, and education); and Model 3 was the fully adjusted model controlled for all included covariates (eg, age, sex, income, education, ethnicity, residence, physical activity, comorbidity load, and sleep duration; Tables 1 and 2).

Table 1.

Descriptive Statistics of the Sample

Variables	Overall Weighted Mean/ Percentage (95% CI)	Middle-aged Weighted Mean/Percentage (95% CI)	Older Adults Weighted Mean/ Percentage (95% CI)
Baseline assessment
Animal fluency	20.69 (20.60, 20.77)	21.67 (21.57, 21.78)	17.74 (17.61, 17.86)
Stroop interference	9.81 (9.72, 9.89)	8.78 (8.69, 8.87)	12.90 (12.72, 13.08)
Mean reaction time	787.60 (785.23, 789.97)	759.48 (756.79, 762.17)	871.96 (867.91, 876.00)
Body mass index	27.77 (27.69, 27.86)	27.80 (27.70, 27.91)	27.69 (27.57, 27.80)
Waist circumference	92.76 (92.53, 92.98)	92.32 (92.04, 92.60)	94.07 (93.74, 94.40)
Total fat mass	33.47 (33.34, 33.60)	32.76 (32.60, 32.92)	35.61 (35.42, 35.81)
Age	59.31 (59.16, 59.45)	54.61 (54.51, 54.71)	73.32 (73.19, 73.44)
Comorbidity	2.44 (2.41, 2.47)	2.22 (2.18, 2.25)	3.11 (3.06, 3.16)
Sleep duration	6.82 (6.80, 6.84)	6.78 (6.75, 6.80)	6.94 (6.91, 6.98)
Sex
Female	50.51 (49.74, 51.28)	49.14 (48.19, 50.09)	54.59 (53.41, 55.77)
Male	49.49 (48.72, 50.26)	50.86 (49.91, 51.81)	45.41 (44.23, 46.59)
Ethnicity
Non-White	2.09 (1.87, 2.31)	2.22 (1.94, 2.50)	1.68 (1.39, 1.98)
White	97.91 (97.69, 98.13)	97.78 (97.50, 98.06)	98.32 (98.02, 98.61)
Income
No response	5.27 (4.95, 5.60)	4.23 (3.86, 4.59)	8.40 (7.73, 9.07)
<$20 000	4.00 (3.73, 4.26)	3.36 (3.05, 3.67)	5.90 (5.36, 6.44)
$20 000 to <$50 000	16.79 (16.27, 17.31)	11.74 (11.18, 12.31)	31.86 (30.74, 32.98)
$50 000 to <$100 000	31.62 (30.91, 32.32)	30.26 (29.40, 31.12)	35.69 (34.54, 36.83)
$100 000 to <$150 000	21.52 (20.86, 22.18)	24.57 (23.74, 25.41)	12.39 (11.60, 13.17)
<$150 000 or more	20.80 (20.14, 21.47)	25.84 (24.99, 26.69)	5.76 (5.22, 6.30)
Education
High school or less	13.94 (13.42, 14.45)	11.19 (10.60, 11.78)	22.12 (21.11, 23.13)
Below bachelor	39.95 (39.19, 40.70)	40.60 (39.67, 41.53)	38.00 (36.84, 39.16)
Bachelor or above	46.12 (45.35, 46.88)	48.21 (47.26, 49.15)	39.88 (38.72, 41.04)
Residence
Urban	90.98 (90.53, 91.42)	90.55 (90.00, 91.10)	92.26 (91.57, 92.94)
Rural	9.02 (8.58, 9.47)	9.45 (8.90, 10.00)	7.74 (7.06, 8.43)
Physical activity
No response	3.42 (3.16, 3.68)	3.19 (2.89, 3.50)	4.10 (3.64, 4.56)
Never	12.31 (11.81, 12.80)	11.58 (10.98, 12.19)	14.46 (13.64, 15.29)
Seldom (1–2 days)	14.70 (14.15, 15.26)	15.03 (14.34, 15.71)	13.73 (12.89, 14.58)
Sometimes (3–4 days)	18.13 (17.52, 18.73)	18.13 (17.39, 18.87)	18.11 (17.18, 19.04)
Often (5–7 days)	51.45 (50.68, 52.21)	52.06 (51.12, 53.01)	49.60 (48.40, 50.79)
Three-year follow-up
Animal fluency	20.56 (20.47, 20.65)	21.53 (21.42, 21.63)	17.48 (17.35, 17.61)
Stroop interference	3.15 (3.11, 3.19)	2.66 (2.62, 2.70)	4.73 (4.64, 4.81)
Mean reaction time	803.71 (800.92, 806.49)	772.55 (769.50, 775.59)	905.18 (899.93, 910.44)
Body mass index	27.88 (27.79, 27.97)	28.00 (27.89, 28.11)	27.49 (27.36, 27.62)
Waist circumference	92.98 (92.74, 93.22)	92.77 (92.47, 93.07)	93.64 (93.27, 94.00)
Change scores
Animal fluency	−0.34 (−0.42, −0.26)	−0.30 (−0.41, −0.20)	−0.46 (−0.58, −0.34)
Stroop interference	−6.48 (−6.57, −6.40)	−6.06 (−6.15, −5.96)	−7.85 (−8.04, −7.67)
Mean reaction time	21.98 (19.29, 24.67)	15.72 (12.60, 18.83)	42.42 (37.18, 47.65)
Body mass index	0.15 (0.12, 0.18)	0.24 (0.21, 0.27)	−0.14 (−0.19, −0.09)
Waist circumference	0.40 (0.30, 0.51)	0.61 (0.49, 0.73)	−0.25 (−0.41, −0.08)
Age	62.05 (61.90, 62.20)	57.56 (57.46, 57.67)	76.07 (75.94, 76.20)
Comorbidity	2.74 (2.71, 2.78)	2.51 (2.47, 2.55)	3.46 (3.41, 3.52)
Sleep duration	6.87 (6.85, 6.89)	6.84 (6.81, 6.86)	6.98 (6.94, 7.01)
Sex
Female	50.37 (49.57, 51.17)	49.02 (48.04, 49.99)	54.57 (53.33, 55.82)
Male	49.63 (48.83, 50.43)	50.98 (50.01, 51.96)	45.43 (44.18, 46.67)
Income
No response	4.93 (4.62, 5.25)	3.89 (3.53, 4.25)	8.19 (7.51, 8.86)
<$20 000	3.52 (3.26, 3.77)	2.89 (2.60, 3.18)	5.46 (4.92, 6.00)
$20 000 to <$50 000	16.45 (15.91, 16.99)	12.02 (11.44, 12.61)	30.29 (29.13, 31.46)
$50 000 to <$100 000	32.16 (31.43, 32.90)	30.60 (29.71, 31.48)	37.05 (35.84, 38.26)
$100 000 to <$150 000	21.08 (20.40, 21.76)	23.80 (22.95, 24.66)	12.58 (11.75, 13.41)
<$150 000 or more	21.85 (21.15, 22.56)	26.79 (25.90, 27.68)	6.42 (5.81, 7.03)
Residence
Rural	7.40 (6.96, 7.84)	7.95 (7.41, 8.49)	5.68 (5.05, 6.32)
Urban	92.60 (92.16, 93.04)	92.05 (91.51, 92.59)	94.32 (93.68, 94.95)
Physical activity
No response	0.06 (0.02, 0.10)	0.05 (0.01, 0.10)	0.08 (0.02, 0.14)
Never	16.46 (15.87, 17.05)	15.02 (14.32, 15.72)	20.95 (19.91, 22.00)
Seldom (1–2 days)	18.70 (18.06, 19.33)	18.89 (18.12, 19.67)	18.09 (17.11, 19.06)
Sometimes (3–4 days)	17.99 (17.38, 18.60)	17.75 (17.01, 18.49)	18.75 (17.77, 19.73)
Often (5–7 days)	46.79 (46.00, 47.58)	48.28 (47.31, 49.25)	42.13 (40.90, 43.37)

Note: CI = confidence interval.

Table 2.

Multivariate Multivariable Regression of the Analytic Sample for Path b (Adiposity → Cognition) and c (Cognition → Adiposity)

Age Group	Outcome	Model 1		Model 2		Model 3
		β (95% CI)	p	β (95% CI)	p	β (95% CI)	p
Path b (WC → Cognition)
Middle-aged	Stroop	0.08 (0.06, 0.10)	<.001	0.04 (0.03, 0.06)	<.001	0.03 (0.01, 0.05)	.001
	AFT	−0.05 (−0.07, −0.03)	<.001	−0.01 (−0.03, 0.00)	.136	−0.01 (−0.03, 0.01)	.230
	MRT	0.01 (0.00, 0.03)	.099	0.02 (0.00, 0.04)	.089	0.02 (0.00, 0.04)	.035
Older adults	Stroop	0.07 (0.04, 0.09)	<.001	0.05 (0.03, 0.08)	<.001	0.03 (0.01, 0.06)	.011
	AFT	0.02 (0.00, 0.05)	.053	0.03 (0.00, 0.05)	.027	0.03 (0.00, 0.06)	.023
	MRT	0.00 (−0.03, 0.02)	.792	0.02 (−0.01, 0.04)	.225	0.01 (−0.02, 0.03)	.661
Path b (BMI−0.7 → Cognition)
Middle-aged	Stroop	−0.08 (−0.10, −0.06)	<.001	−0.05 (−0.06, −0.03)	<.001	−0.04 (−0.05, −0.02)	<.001
	AFT	0.05 (0.04, 0.07)	<.001	0.02 (0.00, 0.03)	.072	0.01 (0.00, 0.03)	.122
	MRT	−0.01 (−0.03, 0.01)	.273	0.00 (−0.01, 0.02)	.729	0.00 (−0.02, 0.02)	.962
Older adults	Stroop	−0.04 (−0.07, −0.02)	<.001	−0.03 (−0.06, −0.01)	.004	−0.02 (−0.04, 0.01)	.126
	AFT	−0.02 (−0.05, 0.00)	.066	−0.03 (−0.05, 0.00)	.019	−0.03 (−0.05, −0.01)	.017
	MRT	0.02 (−0.01, 0.04)	.149	0.00 (−0.02, 0.03)	.750	0.01 (−0.01, 0.04)	.271
Path b (DXA → Cognition)
Middle-aged	Stroop	0.08 (0.06, 0.10)	<.001	0.06 (0.04, 0.08)	<.001	0.05 (0.03, 0.07)	<.001
	AFT	−0.08 (−0.10, −0.06)	<.001	−0.04 (−0.07, −0.02)	<.001	−0.04 (−0.07, −0.02)	.001
	MRT	0.07 (0.06, 0.09)	<.001	−0.02 (−0.04, 0.00)	.105	−0.02 (−0.04, 0.01)	.193
Older adults	Stroop	0.05 (0.02, 0.07)	<.001	0.07 (0.04, 0.10)	<.001	0.05 (0.02, 0.09)	.002
	AFT	−0.06 (−0.09, −0.04)	<.001	0.00 (−0.03, 0.04)	.823	0.00 (−0.03, 0.04)	.849
	MRT	0.03 (0.01, 0.06)	.008	0.00 (−0.03, 0.04)	.942	−0.01 (−0.04, 0.03)	.657
Path c (Stroop → Adiposity)
Middle-aged	WC	0.08 (0.06, 0.10)	<.001	0.04 (0.02, 0.05)	<.001	0.03 (0.02, 0.05)	<.001
	BMI−0.7	−0.06 (−0.08, −0.04)	<.001	−0.04 (−0.05, −0.02)	<.001	−0.03 (−0.05, −0.01)	<.001
Older adults	WC	0.03 (0.01, 0.06)	.003	0.02 (−0.01, 0.04)	.145	0.02 (−0.01, 0.04)	.140
	BMI−0.7	−0.02 (−0.04, 0.01)	.189	−0.02 (−0.05, 0.00)	.063	−0.02 (−0.04, 0.00)	.077
Path c (AFT → Adiposity)
Middle-aged	WC	−0.05 (−0.06, −0.03)	<.001	−0.02 (−0.03, 0.00)	.045	−0.01 (−0.03, 0.00)	.074
	BMI−0.7	0.06 (0.05, 0.08)	<.001	0.03 (0.01, 0.05)	.001	0.03 (0.01, 0.04)	.001
Older adults	WC	0.00 (−0.03, 0.02)	.662	−0.01 (−0.03, 0.01)	.254	−0.01 (−0.03, 0.01)	.493
	BMI−0.7	0.01 (−0.01, 0.03)	.414	0.01 (−0.01, 0.03)	.402	0.00 (−0.02, 0.03)	.701
Path c (MRT → Adiposity)
Middle-aged	WC	0.02 (0.01, 0.04)	.010	0.03 (0.01, 0.04)	.002	0.02 (0.01, 0.04)	.003
	BMI−0.7	−0.01 (−0.02, 0.01)	.555	0.00 (−0.01, 0.02)	.786	0.00 (−0.01, 0.02)	.567
Older adults	WC	0.00 (−0.03, 0.02)	.678	0.02 (0.00, 0.04)	.048	0.02 (0.00, 0.04)	.103
	BMI−0.7	0.03 (0.01, 0.05)	.015	0.01 (−0.01, 0.03)	.344	0.01 (−0.01, 0.04)	.236

Notes: Note 1: AFT = animal fluency task; MRT = mean reaction time, WC = waist circumference; BMI = body mass index; DXA = dual-energy x-ray absorptiometry (measure of total fat mass); CRT = choice reaction time; CI = confidence interval. Higher Stroop scores = worse executive function. Higher AFT scores = better semantic fluency. Higher MRT scores = worse processing speed. Cognition indicates cognitive function measured by Stroop task, AFT, or CRT. Path b indicates the association between baseline adiposity (ie, WC, BMI^−0.7, and DXA) and follow-up cognitive function (ie, Stroop, AFT, and MRT). Path c indicates the association between baseline cognitive function and follow-up adiposity.

Note 2: Model 1 is the unadjusted model. Model 2 is the partially adjusted model controlled for age, sex, income, and education. Model 3 is the fully adjusted model controlled for age, sex, ethnicity, income, education, residence, physical activity, comorbidity index, and sleep duration. All estimates are standardized coefficients. Bold fonts were used to identify the significant paths.

For SEM analysis, cross-lagged models were constructed using the “lavaan” and “lavaan.survey” R packages (50). The correlations among cognitive measures at baseline and correlations among cognitive measures at follow-up were small (r = −0.27 to 0.22) (Supplementary Table 3). Furthermore, cognitive variables included in this study measure different aspects of cognitive function despite the fact all tasks require some cognitive control requirements. On the other hand, correlations among adiposity measures at baseline and follow-up were relatively strong (r = ~0.8). For these reasons, cognitive variables were examined as a sole outcome or predictor in separate identical models, whereas WC and BMI were used as a latent construct in the SEM model. Figure 1 and Supplementary Figure 1 illustrates the cross-lagged paths constructed for the SEM. The diagonal lines in the figure indicate the primary paths of interest. Path b estimated the strength of the association between latent adiposity measure (ie, WC and BMI^−0.7) at baseline and cognition at follow-up, adjusting for the effects of baseline cognition and of the covariates. Path c estimated the strength of the association between cognition at baseline and latent adiposity measure at follow-up, adjusting for the effects of baseline adiposity and of the covariates. The primary set of models used full information maximum likelihood estimation, which means that all participants with at least baseline values on the variables of interest were included in the analysis regardless of whether those participants also have complete data at follow-up. This approach is suitable under the assumption that the data are missing at random (51). The standardized coefficients of the age subgroups (eg, middle-aged vs older adults) were compared by calculating z-score and corresponding p values (Supplementary Table 6).

Figure 1.

Paths of the cross-lagged models using Stroop interference and latent adiposity.Notes: Note 1: W1 and W2 indicate Wave 1 (baseline) and Wave 2 (3-year follow-up) measures, respectively. Adiposity indicates latent adiposity variable. Bolded arrows indicate statistically significant path coefficients; dotted arrows indicate nonsignificant path coefficients. BMI is power transformed by −0.7. All coefficients are standardized beta weights. *p < .05; **p < .001. Covariates include age, sex, ethnicity, household income, education, residence, physical activity, comorbidity, and sleep duration. Note 2: Description of the cross-lagged paths: (a) Path a: the association between baseline cognition and follow-up cognition, (b) Path b: the association between baseline adiposity and follow-up cognition, (c) Path c: the association between baseline cognition and follow-up adiposity, and (d) Path d: the association between baseline adiposity and follow-up adiposity.

Next, we conducted mediation analysis using lavaan R package (50). Mediation analysis was performed only for the cross-lagged paths of interest (path b and c). For path b (adiposity → cognition), BP and T2DM were examined as mediators, whereas for path c (cognition → adiposity), physical activity and diet were analyzed as mediators. Parallel mediation models were used where the mediator constitutes multiple variables (eg, diet and BP). Mediation analyses were adjusted for all covariates with appropriate survey weights applied to the analyses.

Results

Sample characteristics at baseline (Wave 1) and 3-year follow-up (Wave 2) are presented in Table 1. In general, middle-aged participants performed better on cognitive tasks at baseline, as compared with older adults. For example, older adults received statistically significantly lower scores on AFT (17.74, 95% confidence interval [CI] 17.61, 17.86 vs 21.67, 95% CI 21.57, 21.78) and higher scores on Stroop interference (12.90, 95% CI 12.72, 13.08 vs 8.78, 95% CI 8.69, 8.87) and MRT (871.96, 95% CI 867.91, 876 vs 759.48, 95% CI 756.79, 762.17) compared with middle-aged participants (Table 1). A similar trend was also evident at follow-up (Table 1). Both cohorts showed a statistically significant drop in Stroop interference scores (eg, 8.78 vs 2.66 in middle-aged and 12.90 vs 4.73 in older adults) from baseline to follow-up, likely reflecting a practice/familiarity effect or dropout bias (Table 1, Supplementaery Table 2). In terms of adiposity measures, older adults had slightly higher WC (94.07, 95% CI 93.74, 94.40 and 93.64, 95% CI 93.27, 94) at both waves of data collection compared with the middle-aged subsample (92.32, 95% CI 92.04, 92.60 and 92.77, 95% CI 92.47, 93.07). There was no statistically significant difference in BMI between middle-aged and older adults at baseline; however, a slightly higher BMI was observed for the middle-aged subsample at follow-up (28, 95% CI 27.89, 28.11) compared with older adults (27.49, 95% CI 27.36, 27.62; Table 1). When comparing baseline and follow-up adiposity measures, none of the age groups showed statistically significant difference in BMI or WC at follow-up (Table 1).

Multivariate multivariable regression models with WC as an independent variable and 3 cognitive tests as outcome variables (path b; adiposity → cognition) showed that higher baseline WC was statistically significantly associated with higher follow-up Stroop interference in unadjusted, partially adjusted and fully adjusted models in both middle-aged (standardized estimate, β _{Model 1} = 0.08, 95% CI 0.06, 0.10; β _{Model 2} = 0.04, 95% CI 0.03, 0.06; β _{Model 3} = 0.03, 95% CI 0.01, 0.05) and older adults (β _{Model 1} = 0.07, 95% CI 0.04, 0.09; β _{Model 2} = 0.05, 95% CI 0.03, 0.08; β _{Model 3} = 0.03, 95% CI 0.01, 0.06), whereas higher baseline WC was associated with better animal fluency scores in older adults only in partially adjusted and fully adjusted models (β _{Model 2} = 0.03, 95% CI 0.00, 0.05; β _{Model 3} = 0.03, 95% CI 0.00, 0.06; Table 2). In addition, higher baseline WC was significantly associated with higher follow-up reaction time in middle-aged adults in fully adjusted model (β _{Model 3} = 0.02, 95% CI 0.00, 0.04) but not in older adults. A similar pattern of findings was observed when other adiposity indicators were used in the model (ie, BMI^−0.7, DXA; Table 2).

Multivariate multivariable regression for path c (cognition → adiposity) indicated that higher baseline Stroop interference was associated with higher follow-up WC (β _{Model 1} = 0.08, 95% CI 0.06, 0.10; β _{Model 2} = 0.04, 95% CI 0.02, 0.05; β _{Model 3} = 0.03, 95% CI 0.02, 0.05) and lower follow-up BMI^−0.7 (β _{Model 1} = −0.06, 95% CI −0.08, −0.04; β _{Model 2} = −0.04, 95% CI −0.05, −0.02; β _{Model 3} = −0.03, 95% CI −0.05, −0.01) in middle-aged adults in all models. In the case of older adults, the only statistically significant association was between baseline Stroop interference and follow-up WC in the unadjusted model (β _{Model 1} = 0.03, 95% CI 0.01, 0.06; Table 2). Higher baseline AFT and lower baseline MRT were primarily associated with lower follow-up adiposity in middle-aged adults (Table 2).

To probe the presence of bidirectional relationships in a more parsimonious manner, we undertook SEM using a latent adiposity variable formed from both prospectively measured adiposity indicators (BMI and WC), predicting separately each cognitive construct (one indicator each). Supplementary Tables 4 and 5 summarizes the SEM models examining prospective associations between cognitive function and the latent adiposity variable. One of the primary paths of interest (path b; Figure 1) indicated that higher baseline adiposity was associated with higher follow-up Stroop interference, and this association was statistically significant for middle-aged adults (standardized estimate, β = 0.04, 95% CI 0.02, 0.06). However, in the case of the AFT in older adults, higher baseline adiposity was associated with significantly better AFT performance (0.04, 95% CI 0.02, 0.06) at follow-up, suggesting an advantage to verbal fluency conferred by adiposity (Supplementary Figure 3). Similarly, in the middle-aged subsample, higher baseline Stroop interference was associated with higher follow-up adiposity (path c; 0.01, 95% CI 0.00, 0.01). Additional statistically significant associations were observed between baseline and follow-up adiposity and cognitive function indicators, which are presented in Supplementary Tables 4 and 5, Figures 1 and 2, and Supplementary Figures 3 and 4. Overall, irrespective of the modeling approach, the bidirectional association between Stroop interference and adiposity was observed in the middle-aged subsample.

Figure 2.

Bidirectional associations between latent adiposity and cognition by age groups (45–65 years and >65 years).Notes: W1 = Wave 1 or baseline measures; W2 = Wave 2 or 3-year follow-up measures; Adiposity = Latent adiposity variable.

Mediation analyses indicated that the relationship between baseline WC and follow-up Stroop interference (path b) was mediated through T2DM for both middle-aged (standardized estimate, β = 0.0090, 95% CI 0.0054, 0.0126) and older (0.0066, 95% 0.0021, 0.0110) adults (Table 3). Although systolic (0.0195, 95% CI 0.0094, 0.0297) and diastolic (−0.0148, 95% CI −0.0231, −0.0066) BP showed statistical significance as individual mediators in the middle-aged, the total indirect effects of BP were not found statistically significant in either group. For path c, the association between baseline Stroop interference and follow-up WC were mediated by total caloric consumption for middle-aged adults (−0.0006, 95% CI −0.0011, −0.0001) and pastries consumption for older adults (−0.0006, 95% CI −0.0011, −0.0001).

Table 3.

Mediation Analysis for the Path b (Adiposity → BP/T2DM → Cognition) and Path c (Cognition → Activity/Diet → Adiposity) Using Waist Circumference and Stroop Interference

Mediators	Middle-aged		Older Adults
	Indirect Effect (95% CI)	p	Indirect Effect (95% CI)	p
Path b (WCW1 → BP/T2DM → StroopW2)
Blood pressure
Total	0.0047 (−0.0018, 0.0111)	.154	−0.0006 (−0.0049, 0.0037)	.776
Systolic	0.0195 (0.0094, 0.0297)	<.001	0.0016 (−0.0038, 0.0070)	.568
Diastolic	−0.0148 (−0.0231, −0.0066)	<.001	−0.0022 (−0.0065, 0.0021)	.315
Type 2 diabetes	0.0090 (0.0054, 0.0126)	<.001	0.0066 (0.0021, 0.0110)	.004
Path c (StroopW1 → Activity/Diet → WCW2)
Physical activity	0.0000 (−0.0002, 0.0001)	.609	0.0000 (−0.0001, 0.0001)	.617
Diet
Total	−0.0006 (−0.0011, −0.0001)	.018	−0.0002 (−0.0013, 0.0009)	.679
Legume	−0.0002 (−0.0005, 0.0000)	.056	0.0002 (−0.0003, 0.0007)	.393
Fruit	0.0000 (0.0000, 0.0000)	.922	0.0001 (−0.0001, 0.0002)	.579
Salad	−0.0001 (−0.0003, 0.0001)	.161	−0.0002 (−0.0006, 0.0002)	.381
Carrot	0.0000 (−0.0001, 0.0001)	.634	−0.0002 (−0.0006, 0.0002)	.374
Fries	−0.0001 (−0.0002, 0.0001)	.310	0.0001 (−0.0002, 0.0004)	.407
Snack	0.0001 (−0.0001, 0.0002)	.272	0.0004 (0.0000, 0.0008)	.058
Pastries	−0.0003 (−0.0006, 0.0000)	.060	−0.0006 (−0.0011, −0.0001)	.013
Chocolate	0.0000 (0.0000, 0.0001)	.720	0.0000 (−0.0001, 0.0002)	.596

Notes: BP = blood pressure; T2DM = type 2 diabetes mellitus; CI = confidence interval; WC = waist circumference; W1 = Wave 1 or baseline measures; W2 = Wave 2 or 3-year follow-up measures. Path b: the association between baseline adiposity and follow-up cognition; path c: the association between baseline cognition and follow-up adiposity. Four decimal places were retained because of smaller values of the indirect effects and to clarify the direction of coefficient and CI. Total indicates the sum of individual indirect effects for the respective mediators. All estimates are standardized coefficients. Significant indirect effects indicate that the association between adiposity and cognition is, in part, mediated through that respective lifestyle behavior and chronic disease status.

Discussion

The purpose of this investigation was to probe for the possibility of a bidirectional relationship between adiposity and cognitive function using a large, representative sample of middle-aged and older adults. Using multivariate multivariable regression and cross-lagged latent variable modeling, we observed that higher baseline adiposity was associated with lower executive function at 3-year follow-up, with reliable associations observed for both age subgroups. In contrast, the association between baseline executive function and follow-up adiposity was statistically significant only in the middle-aged subsample. As such, our findings support a bidirectional association between cognition and adiposity, but primarily among midlife individuals, and with specific reference to executive function.

The earlier findings were robust following adjustment for a wide variety of confounders, including sociodemographic, medical, and lifestyle variables. The current findings significantly augment our knowledge about the relationship between adiposity and cognitive function and provide some information on the boundary conditions for any potential bidirectional relationships. Specifically, from a brain-as-outcome perspective, middle-aged adults showed evidence of an inverse association between obesity and executive function; from a brain-as-predictor perspective, worse executive function at baseline was shown to be associated with the accumulation of adiposity over a 3-year period for the same age group. The mediation analyses suggested that lifestyle behaviors and physical health conditions can be critical when considering the long-term bidirectional effects of adiposity and cognition. Diet and T2DM were emerged as statistically significant mediators for the brain-as-predictor path and brain-as-outcome path, respectively.

The current findings are consistent with previous studies documenting negative associations between obesity and brain health outcomes. Indeed, many prior cross-sectional and longitudinal studies have reported an association between obesity and cognitive dysfunction, as well as an association between obesity and brain pathologies that implicate the PFC (8,10,11,13–15). Evidence of bidirectional relationships between executive function and adiposity also supports prior theorizing (6). It is not clear why the bidirectional relationship was found only in the middle-aged subsample, but not in the older adults. However, it is possible that in older adults, much more of the variability in executive function is absorbed by medical comorbidities.

Despite our finding that higher adiposity at baseline was associated with lower performance on executive function at 3-year follow-up, the opposite was true for animal fluency in the older adult subsample. That is, obesity appeared to have a protective effect in relation to verbal fluency. Although this seems counterintuitive, it could be consistent with the literature on the so-called “obesity paradox” where it has been argued that weight gain at old age is protective from cognitive decline (52). The protective effects of mild adiposity can also be explained by prodromal weight loss in dementia. It has been documented that dementia and Alzheimer’s disease are usually preceded by years of unintentional weight loss (53,54). It is also possible that the animal naming effect is a selection bias or survivorship effect; this interpretation is supported by the fact that baseline associations between animal naming and adiposity were in the theoretically expected direction in a prior cross-sectional study conducted using CLSA baseline data (55).

Although the brain-as-predictor view was inadequately explored in previous research, there is evidence of indirect paths through which impaired cognition can lead to adiposity. One such potential mechanism is the association between prefrontal function and food consumption. Executive function impairment is associated with poor decision-making related to food choice and consumption (19–21). Accordingly, people with cognitive impairments tend to gain weight over time, a potent indicator of the bidirectional relationship reported in previous studies (6). Experimental studies further confirm this association using noninvasive brain stimulation where suppression of the lateral PFC results in increased consumption of calorie-dense food (6,22,23). Our analysis suggests that such associations exist at the population level, at least in middle-aged people.

Likewise, T2DM was found as a statistically significant mediator for the association between baseline adiposity and follow-up cognition. As reported in previous studies, T2DM is associated with cognitive dysfunction, dementia, and Alzheimer’s disease in older adults (56,57). Therefore, it is highly plausible that obesity, together with T2DM hasten the progression of cognitive decline (58). Overall, mediation analysis suggests that the hypothesized bidirectional association could be influenced by modifying lifestyle behavior and physical health status.

Strengths of the current investigation include the use of a large-scale population data set, with substantial power to detect subtle effects. Likewise, given that the data set was representative of the general population to some extent, the findings may be generalized to the larger Canadian population, with the caveat that the sample was disproportionately urban. Furthermore, the analysis was adjusted for several important covariates, including comorbidity load. Finally, unlike most survey data, the adiposity measures were not self-reported, so there was less reporting bias associated with these measures.

There are several limitations of the present study. First, the CLSA lacks structural and functional brain imaging data in the waves available for this analysis. Imaging data is useful to examine for older adults as several reliable changes happen in brain structure within this age range (eg, gradual gray and white matter atrophy) (59); future CLSA waves will include such data and will enable questions about brain structure and function to be addressed directly. Second, some adiposity indicators within CLSA, such as total fat mass measured by DXA and waist–hip ratio, were not measured at the 3-year follow-up; therefore, we could not include these in our path c (cognition → adiposity) analysis. A prior study examining cross-sectional data from the CLSA baseline did include such measures (55), and found similar patterns of findings across all adiposity indicators. Third, as the CLSA comprehensive cohort participants were recruited from areas within 25–50 km radius of the data collection sites, and some participants were excluded based on exclusion criteria, our findings primarily represent those who live near the major urban centers, functionally mobile and do not have major neurological disorders (28,30). Fourth, we stratified our analysis using 2 broad age groups based on retirement age; however, it should be noted that the strategy of using only 2 age groups might affect the ability to identify other differences associated with age within each broad category (ie, individuals in their 80s might have significant differences in cognitive capabilities and adiposity status compared with those in their 60s). Fifth, a sizable number of the participants in the analytic sample were French speakers (n = 6 090); therefore, the possibility of differences in task performance because of different psychometric properties of English and French versions of the tasks cannot be excluded. Sixth, although the findings of cross-lagged analyses were consistent with regression-based analyses, it should be noted that the cross-lagged panel modeling depends on a number of assumptions (eg, synchronicity, stationarity, stability, and others) that are often violated or cannot be entirely met (60,61). Seventh, as expected in the longitudinal studies, one primary challenge for the CLSA is participant engagement and retention. Approximately 7% (n = 1 827) of the participants in our analytic sample were lost to follow-up. Although baseline characteristics of the retained and lost to follow-up individuals were quite similar and the attrition rate was low in absolute terms, it is possible that some findings are influenced by survivorship bias.

Finally, although the cross-lagged associations were reliable, they were very subtle. This implies the need for more extended follow-up intervals over which to properly assess the gradual and cumulative hypothesized reciprocal effects of brain and adiposity. A lag of 3 years between measurements within the current data set was minimally sufficient to detect the presence versus absence of statistically reliable cross-lagged effects. However, a minimum time lag of 10 years may be more ideal in order to properly assess the magnitude of such effects in absolute terms. This will be possible to examine with future waves of CLSA follow-up data, and with other population-level data sets.

Despite the earlier-mentioned limitations, the current study provides some evidence for bidirectional associations between adiposity and cognitive function, particularly for middle-aged adults. Further studies are needed to confirm these associations with other measures, populations, and longer follow-up intervals. Future studies should also examine adiposity–cognition relationships using structural and functional brain imaging data. Other cognitive variables could be considered in future studies to understand to what extent bidirectional associations exist in other cognitive domains. Finally, although the pattern of findings supports a bidirectional relationship between adiposity and executive function, the absolute magnitude of the associations is small over the 3-year window of the existing analysis. Given the high degree of stability in adiposity metrics over the 3-year follow-up window (r = ~0.9), relatively little variability in adiposity remained to be predicted by variables in the model, including executive function and other cognitive variables. It is expected that future waves of follow-up data over longer intervals will allow for a more statistically powerful test of the cumulative effects of both focal variables upon each other over time.

Conclusion

In summary, this investigation examined evidence for bidirectional relationships between indicators of cognitive function and adiposity from middle age to late life. Findings suggested modest but reliable evidence of a prospective association between baseline adiposity and follow-up executive function among middle-aged and older adults. There was evidence of bidirectional relationships, such that baseline executive function was associated with follow-up adiposity, although this was specific to middle-aged individuals. As such, the bidirectional relationship model was supported at middle age for executive function, but not other cognitive functions or in late life. Findings are largely consistent with prior theory proposing bidirectional relationships between prefrontal function and adiposity (6). Overall, the findings suggest that researchers, health care providers, and policymakers should consider the complexity of adiposity–cognition relationship beyond the dominant unidirectional assumption.

Funding

This research was made possible using the data/biospecimens collected by the Canadian Longitudinal Study on Aging (CLSA). Funding for the CLSA is provided by the Government of Canada through the Canadian Institutes of Health Research (CIHR) under grant reference: LSA 94473 and the Canada Foundation for Innovation, as well as the following provinces, Newfoundland, Nova Scotia, Quebec, Ontario, Manitoba, Alberta, and British Columbia. This research has been conducted using the CLSA Baseline Comprehensive Data set 4.1, under Application Number 1906024. The CLSA is led by Drs. Parminder Raina, Christina Wolfson, and Susan Kirkland. Data are available from the CLSA (www.clsa-elcv.ca) for researchers who meet the criteria for access to de-identified CLSA data. The opinions expressed in this manuscript are the author’s own and do not reflect the views of the CLSA.

The development, testing, and validation of the Short Diet Questionnaire were carried out among NuAge study participants as part of the CLSA Phase II validation studies, Canadian Institutes for Health Research (CIHR) 2006–2008. The NuAge study was supported by the CIHR, grant number MOP-62842, and the Quebec Network for Research on Aging, a network funded by the Fonds de Recherche du Québec-Santé.

Funding for the analysis and writing of the current manuscript was provided by the Natural Sciences and Engineering Research Council (NSERC) of Canada to the senior author (PH).

Conflict of Interest

None declared.