How youth cognitive and sociodemographic factors relate to the development of overweight and obesity in the UK and the USA: a prospective cross-cohort study of the National Child Development Study and National Longitudinal Study of Youth 1979

Objectives We investigated how youth cognitive and sociodemographic factors are associated with the aetiology of overweight and obesity. We examined both onset (who is at early risk for overweight and obesity) and development (who gains weight and when). Design Prospective cohort study. Setting We used data from the US National Longitudinal Study of Youth 1979 (NLSY) and the UK National Child Development Study (NCDS); most of both studies completed a cognitive function test in youth. Participants 12 686 and 18 558 members of the NLSY and NCDS, respectively, with data on validated measures of youth cognitive function, youth socioeconomic disadvantage (eg, parental occupational class and time spent in school) and educational attainment. Height, weight and income data were available from across adulthood, from individuals’ 20s into their 50s. Primary and secondary outcome measures Body mass index (BMI) for four time points in adulthood. We modelled gain in BMI using latent growth curve models to capture linear and quadratic components of change in BMI over time. Results Across cohorts, higher cognitive function was associated with lower overall BMI. In the UK, 1 SD higher score in cognitive function was associated with lower BMI (β=−0.20, 95% CI −0.33 to −0.06 kg/m²). In America, this was true only for women (β=−0.53, 95% CI −0.90 to −0.15 kg/m²), for whom higher cognitive function was associated with lower BMI. In British participants only, we found limited evidence for negative and positive associations, respectively, between education (β=−0.15, 95% CI −0.26 to −0.04 kg/m²) and socioeconomic disadvantage (β=0.33, 95% CI 0.23 to 0.43 kg/m²) and higher BMI. Overall, no cognitive or socioeconomic factors in youth were associated with longitudinal changes in BMI. Conclusions While sociodemographic and particularly cognitive factors can explain some patterns in individuals’ overall weight levels, differences in who gains weight in adulthood could not be explained by any of these factors.

Across cultures, higher cognitive function was associated with lower overall BMI. In the UK, 44 1 standard deviation higher score in cognitive function was associated with a ~0. 2 Table 1.  intervals. The BMI latent variables included variables for overall BMI level, linear slope, and 164 quadratic slope. All analyses were carried out using the R package 'lavaan' 42 . 165 Youth SED, education, sex, cognitive function, and the interaction between sex and 166 cognitive function were regressed on each of these BMI latent variables. Income for each 167 time point was regressed directly onto the BMI measurements. Because there was 168 substantially more age variation among NLSY participants, additional control variables were 169 included in those models. These were age at cognitive testing in 1979, which was regressed 170 upon the BMI latent variables like cognitive function, and age during each follow-up wave,    (Table S1). In our main models, we included effects of  Overview of the results 208 We first describe the general appearance of the results, before reporting the results of 209 the models in each country separately. In both the US and the UK, and in men and women, 210 mean BMI increases from the 20s to the 50s, by about 4 to 5 BMI points (Table 1, Figure 2).

211
Men appear to have higher mean BMI than women ( Figure 2, upper panel). Americans have 212 higher BMI than the British from early adulthood, and the differences grow as adulthood 213 advances; Americans were more than half a BMI point higher than the British when  In the US sample, men had higher BMI level than women, but there were no main 223 effects of cognitive function, early life sociodemographic characteristics, or education (Table   224 2). There was evidence for an interaction between sex and cognitive function on BMI; the 225 model's results suggest that cognitive function is not related to BMI level in American men, 226 but, in American women, 1 standard deviation higher cognitive function is associated with a 227 more than half point lower BMI level. 228 We found no associations between early-life cognitive function or sociodemographic In the British sample, women had lower BMI levels than men, more than 0.8 points 236 all else being equal ( Table 2). Higher cognitive function in youth was associated with lower 237 BMI level; 1 standard deviation higher score in cognitive function was associated with ~0.2 238 points lower BMI. There was no sex by cognitive function interaction on BMI level. 239 Additionally, and unlike in the American sample, high youth SED and more educational attainment were associated with higher and lower BMI levels, respectively. We note that 241 these coefficients in the US samples are not dissimilar to those in the UK sample, and that the 242 UK sample has greater power. 243 As found in the American sample, there were no relationships between BMI growth 244 and cognitive function, youth SED, or education. However, there was an association between 245 BMI growth and sex, so that women had slower growth in BMI over time. Every decade, 246 men gained about 0.3 BMI points more than women, all else being equal. Sex was also 247 associated with the quadratic component of BMI growth, such that men's BMI growth tended 248 to level off in later ages, whereas women's linear BMI growth continued more linearly in 249 later ages. These British effects are visible in Figure 2.
To be consistent with earlier work 4 5 , including The Bell Curve 6 , we used the z-scored AFQT percentile score in our analyses.
In the NCDS, cognitive function was assessed using a general ability test, given to participants at age 11 7 . The test consisted of 40 verbal and 40 nonverbal items, and the total score was z-scored for analysis. The reliability is high (test-retest Cronbach's α = Youth SED Due to historic data collection limitation, youth SED was composed differently in each sample, but both measures have been previously validated. In the US, youth SES was the sum of z-scored variables for parental income, education, and occupation status. Like AFQT scores, this variable has been previously used and validated in previous work 4 6 .
In the UK, youth SED was the sum of six z-scored variables: father's social class at birth, father's social class at age 7, age at which the father left education, age at which the mother left education, parental housing tenure in childhood, and the number of people sharing a room in the household at age 7. This variable was composed to be consistent with earlier work 9 .

Education
Education was measured in age at which an individual left school in the NCDS, and highest grade achieved in the NLSY, which was only converted to age at which an individual left school for descriptive purposes in Table 1. Each variable was independently z-scored for use in our models.

Income
Net family income was used in both samples. Net family income was derived from a comprehensive set of income questions in both samples. The possible sources came from Income variables were taken from the same wave as BMI variables with one exception.
In the NCDS, income data were collected using a different instrument at age 55, the most recent available time point for BMI data. The age 55 measurement was missing more data than the age 50 measurement (652 more cases), and in addition to there being data collection inconsistencies at age 55, the variable had some problematic properties, e.g.
smaller mean and median income statistics compared to those at age 50. Considering that model fit was sub-par using the age 55 measure, but much better using the age 50 measure, we used the age 50 income measure in all of the analyses we present.

Methods -Missing Data
Study attrition and occasional non-response were present in both samples (see Table 1), though one of the benefits of our approach was that we could grapple with missing data with Full Information Maximum Likelihood estimation using observed information 11 .
Standard errors were calculated with first order derivatives and P-values were calculated with a χ 2 test, both of which are procedures that can handle missing data.

Instructions to authors
Complete this checklist by entering the page numbers from your manuscript where readers will find each of the items listed below.
Your article may not currently address all the items on the checklist. Please modify your text to include the missing information. If you are certain that an item does not apply, please write "n/a" and provide a short explanation.
Upload your completed checklist as an extra file when you submit to a journal.
In your methods section, say that you used the STROBE cohortreporting guidelines, and cite them as: follow-up, and analysed. Give information separately for for exposed and unexposed groups if applicable.  Figure 1 Descriptive data #14a Give characteristics of study participants (eg demographic, clinical, social) and information on exposures and potential confounders. Give information separately for exposed and unexposed groups if applicable. Table 1 Descriptive data #14b Indicate number of participants with missing data for each variable of interest Table 1 Descriptive data #14c Summarise follow-up time (eg, average and total amount) 9-10, Table 1 Outcome data #15 Report numbers of outcome events or summary measures over time. Give information separately for exposed and unexposed groups if applicable.     Across cohorts, higher cognitive function was associated with lower overall BMI. In the UK, 45 1 standard deviation higher score in cognitive function was associated with lower BMI (β=-

67
 This study used two independent cohorts from the US and UK, each with more than  diabetes and osteoarthritis, 5 as well as mental illnesses, such as depression, 6 7 anxiety, 8

238
Overview of the results 239 We first describe the ultimate structure of our latent growth curve models, second we 240 describe the general appearance of the results, and then report the results of the models in 241 each country separately. All models were a variation on the one presented in Figure 1. BMI 242 change over time was best modeled with quadratic and linear slope components in both shown in Table 2, and explained below.

247
In both the US and the UK, and in men and women, mean BMI increases from the 20s 248 to the 50s, by about 4 to 5 kg/m² (Table 1, Figure 2). Men appear to have higher mean BMI In the US sample, men had higher BMI level than women, but there were no main 261 effects of cognitive function, early life sociodemographic characteristics, or education (Table   262 2). There was evidence for an interaction between sex and cognitive function on BMI; the 263 model's results suggest that cognitive function is not related to BMI level in American men, 264 but, in American women, 1 standard deviation higher cognitive function is associated with a 265 more than half unit lower BMI level. to level off in later ages, whereas women's linear BMI growth continued more linearly in 287 later ages. These British effects are visible in Figure 2.

288
In the UK sample, BMI level was associated with linear BMI slope, and linear slope 289 was associated with quadratic slope. This means that higher BMI level indicated greater  Modeling without income (Table S2)  participants showed less leveling off of BMI growth over time. 312 We also wished to see if our findings generalized to the entire populations represented 313 in these samples, so we refit our primary models (i.e. those described in Table 2) with all have only about 0.85 fewer BMI units than British men. A possible reason is visible in Figure   331 2: Americans weigh more in general and that allows Americans to have greater range of BMI 332 values, so for example, there is more opportunity for American men to have higher BMI than 333 the average American.

334
In their 20s, the lowest BMI American women approximately matched the lowest 335 BMI British women, but the American sample rapidly gained more weight, and by their 40s, 336 the average American is more than a full BMI unit higher. Diversity in BMI grows across 337 both samples as well: the SD of BMI among Americans goes from 4 units in the early 20s to  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59  60   F  o  r  p  e  e  r  r  e  v  i  e  w  o  n  l  y 16 338 more than 6 in the 50s, and among the British the SD rises from ~3 at age 23 to ~5.5 at age 339 55. BMI growth is faster in British men than British women, but that growth slowed in 340 participants' 40s and 50s, whereas in British women, growth was slower but more consistent.

341
No such effects were apparent in Americans. Both men and women tended to gain weight at 342 the same rate.

343
Associations between cognitive function and BMI levels were present in both

353
In addition to cognitive function, education and youth SED were also associated with 354 BMI levels in British people, but not in Americans. Having more education was associated 355 with lower BMI, to nearly the same extent as cognitive function. Coming from a less 356 deprived background was also associated with lower BMI, to a greater degree than either 357 cognitive function or education, though, again, only in the British. In the American sample, 358 education showed a similar effect size and a nearly overlapping confidence interval; 359 therefore, with more participants and thus power, we might have found that education was 360 similarly associated in the American sample. Some, but not all, of our sensitivity analyses 361 support this, so imprecision of variable measurement might also explain the width of this 362 confidence interval. Income was not associated with BMI at any point in our analyses, except  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59   cross-cultural provides further validation for the effects we found that spanned both of our 397 samples, thus demonstrating that these results are not unique to a particular nation or culture.

398
What is apparent is from our results is that cognitive function tested in youth has an 399 important association with BMI, from early adulthood into middle-age. BMI and cognitive 400 function are genetically correlated at r g = -0.13, 58 so some of the association we found might 401 be caused by genetic variation; however, some of the genetic risk of obesity can be reduced 402 through education. 59 It is therefore important to ask: do people learn things over the course of 403 their education that they use to live healthier lives? In much the same way, we might be able 404 to encourage what has been termed "phenocopying" 17 -that is, encouraging and enabling 405 people to follow the same strategies that individuals with higher cognitive function use to 406 look after their health with the hope of achieving the same results. In contrast to education, 407 which we only found to be associated with BMI in the UK, we found associations between 408 cognitive function and BMI in both samples, as did two prior studies. 58 59

409
The associations between cognitive function, sociodemographic factors, and BMI  probably best positioned to make a difference to individuals' lifelong health.

Methods -Income and measurement time
Income variables were taken from the same wave as BMI variables with one exception.
In the NCDS, income data were collected using a different instrument at age 55, the most recent available time point for BMI data. The age 55 measurement was missing more data than the age 50 measurement (652 more cases), and in addition to there being data collection inconsistencies at age 55, the variable had some problematic properties, e.g.

Methods -Missing Data
Study attrition and occasional non-response were present in both samples (see Table 1), though one of the benefits of our approach was that we could grapple with missing data with Full Information Maximum Likelihood estimation using observed information 11 .
Standard errors were calculated with first order derivatives and P-values were calculated with a χ 2 test, both of which are procedures that can handle missing data.

Instructions to authors
Complete this checklist by entering the page numbers from your manuscript where readers will find each of the items listed below.
Your article may not currently address all the items on the checklist. Please modify your text to include the missing information. If you are certain that an item does not apply, please write "n/a" and provide a short explanation.
Upload your completed checklist as an extra file when you submit to a journal.
In your methods section, say that you used the STROBE cohortreporting guidelines, and cite them as: Statistical methods #12b Describe any methods used to examine subgroups and interactions 8-9 Statistical methods #12c Explain how missing data were addressed 8-9 Statistical methods #12d If applicable, explain how loss to follow-up was addressed 8-9 Statistical methods #12e Describe any sensitivity analyses 12-13