1. Epidemiology and Global Health

Weakening of the cognition and height association from 1957 to 2018: Findings from four British birth cohort studies

  1. David Bann  Is a corresponding author
  2. Liam Wright
  3. Neil M Davies
  4. Vanessa Moulton
  1. Centre for Longitudinal Studies, Social Research Institute, University College London, United Kingdom
  2. MRC IEU, University of Bristol, United Kingdom
  3. K.G. Jebsen Center for Genetic Epidemiology, Department of Public Health and Nursing, NTNU, Norwegian University of Science and Technology, Norway
https://doi.org/10.7554/eLife.81099
Share this article
Cite this article
  1. David Bann
  2. Liam Wright
  3. Neil M Davies
  4. Vanessa Moulton
(2023)
Weakening of the cognition and height association from 1957 to 2018: Findings from four British birth cohort studies
eLife 12:e81099.
https://doi.org/10.7554/eLife.81099
  2. Download BibTeX
  3. Download .RIS

Abstract

Background:

Taller individuals have been repeatedly found to have higher scores on cognitive assessments. Recent studies have suggested that this association can be explained by genetic factors, yet this does not preclude the influence of environmental or social factors that may change over time. We thus tested whether the association changed across time using data from four British birth cohorts (born in 1946, 1958, 1970, and 2001).

Methods:

In each cohort height was measured and cognition via verbal reasoning, vocabulary/comprehension, and mathematical tests; at ages 10/11 and 14/17 years (N=41,418). We examined associations between height and cognition at each age, separately in each cohort, and for each cognitive test administered. Linear and quantile regression models were used.

Results:

Taller participants had higher mean cognitive assessment scores in childhood and adolescence, yet the associations were weaker in later (1970 and 2001) cohorts. For example, the mean difference in height comparing the highest with lowest verbal cognition scores at 10/11 years was 0.57 SD (95% CI = 0.44–0.70) in the 1946 cohort, yet 0.30 SD (0.23–0.37) in the 2001 cohort. Expressed alternatively, there was a reduction in correlation from 0.17 (0.15–0.20) to 0.08 (0.06–0.10). This pattern of change in the association was observed across all ages and cognition measures used, was robust to adjustment for social class and parental height, and modeling of plausible missing-not-at-random scenarios. Quantile regression analyses suggested that these differences were driven by differences in the lower centiles of height, where environmental influence may be greatest.

Conclusions:

Associations between height and cognitive assessment scores in childhood-adolescence substantially weakened from 1957–2018. These results support the notion that environmental and social change can markedly weaken associations between cognition and other traits.

Funding:

DB is supported by the Economic and Social Research Council (grant number ES/M001660/1); DB and LW by the Medical Research Council (MR/V002147/1). The Medical Research Council (MRC) and the University of Bristol support the MRC Integrative Epidemiology Unit [MC_UU_00011/1]. NMD is supported by an Norwegian Research Council Grant number 295989. VM is supported by the CLOSER Innovation Fund WP19 which is funded by the Economic and Social Research Council (award reference: ES/K000357/1) and Economic and Social Research Council (ES/M001660/1). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Editor's evaluation

This paper provides valuable evidence for a weakening of the association between cognitive ability and height from 1957 to 2018 in the UK. The authors find the strength of the association declined over this time frame. These associations were further attenuated after accounting for proxy measures of social class. This paper is a solid contribution to debates about how genetic, environmental, and social factors have affected the joint distribution of height and cognitive ability over the last 60 years.

https://doi.org/10.7554/eLife.81099.sa0

Introduction

Cognitive ability is a potentially important determinant of health, with associations between higher cognition and favorable subsequent health outcomes repeatedly documented in childhood, adolescence, and across adulthood (Deary and Batty, 2007). This body of evidence—termed by some as ‘cognitive epidemiology’—includes associations of lower cognition and anthropometric measures such as shorter height (Marioni et al., 2014; Keller et al., 2013). Shorter height is in turn associated with a range of disease outcomes and, like cognitive development, is influenced by early life factors (Monteiro and Victora, 2005; Walker et al., 2007; Perkins et al., 2016; Batty et al., 2009; Sudfeld et al., 2015).

Recent studies have suggested that links between higher cognition and taller height can be large or to a sizable extent explained by shared genetic influences (Marioni et al., 2014; Keller et al., 2013; Silventoinen et al., 2012; Vuoksimaa et al., 2018), with others suggesting a role of assortative mating (Keller et al., 2013; Beauchamp et al., 2011). Keller et al., 2013 for example estimated that approximately half of the correlation was accounted for by assortative mating, and half by shared genetic factors. However, this evidence is largely derived from the analysis of single cohorts. Historical data can yield insight into the importance of the environment, even for highly heritable traits. Indeed, while height is highly heritable (Jelenkovic et al., 2016), substantial increases in average height occurred in the 20th century (Meredith, 1976; Cole, 2000), likely due to improved nutrition and a declining impact of early life infectious disease (Walker et al., 2007; Perkins et al., 2016) which in particular affected early life leg growth (Cole, 2000). Cognitive test scores are also highly heritable (Plomin and Deary, 2015) yet have also increased in the 20th century (the ‘Flynn’ effect) (Bratsberg and Rogeberg, 2018) suggesting that both factors are responsive to environmental change. Shared developmental processes may affect both outcomes (Sudfeld et al., 2015; Mansukoski et al., 2020), such as growth influencing height and brain size (Vuoksimaa et al., 2018). Improvements in nutrition and other shared developmental determinants may have weakened associations between cognition and height (Gale, 2005), as suggested by a weakening of links between low birth weight and cognition from 1958–2001 (Goisis et al., 2017a). A declining association between cognition and height has been suggested in the analysis of a subsample of eastern Danish male conscripts (1939–1967) (Teasdale et al., 1989), yet data from more heterogeneous and contemporary samples is required to strengthen conclusions and aid generalizability.

We conducted a cross-cohort study to investigate if the associations of cognition with child-adolescent height have systematically changed over time (1957–2018). We used four British birth cohort studies, each containing prospectively ascertained assessments of cognition and height. We hypothesized that improvements in the shared determinants of both cognition and height would lead to the weakening of associations across time; see Figure 1 for a causal diagram. We also hypothesized that cognition-height associations would be strongest amongst those in the lowest height centiles, where environmental influences on impaired height development may be particularly pronounced.

Figure 1
Download asset Open asset
Illustrative causal diagrams of two alternative processes which may either (or both in some combination) generate associations between child cognition and height: (right) shared genetic and environmental factors and (left) assortative mating of parental cognition and height (within and across each trait).

Methods

Study samples

Four birth cohort studies conducted in Britain with participants followed-up in childhood and adolescence were used. These were born in 1946 (MRC National Survey of Health and Development, 1946 c), 1958 (National Child Development Study, 1958 c), 1970 (British Cohort Study, 1970 c), and 2000/02 (Millennium Cohort Study, 2001 c); further details are provided in the cohort profiles (Wadsworth et al., 2006; Power and Elliott, 2006; Elliott and Shepherd, 2006; Connelly and Platt, 2014). To aid comparability, analyses were restricted to singleton births born in England, Scotland, and Wales. Sample sizes for analyses, in each cohort and age, are shown in Table 1. Each study and sample sweep used in this manuscript has received ethical approval and obtained parental/participant consent according to guidance that was in place at the time of data collection from 1957–2018. These historic datasets are made available for anonymized observational research which can be published on a range of topics by bona fide researchers; further details are provided elsewhere (Shepherd and Gilbert, 2019a; Centre for Longitudinal Studies, 2014; Shepherd and Gilbert, 2019b) ethical review boards are as follows: 1946 c (Queen Square Research Ethics Committee: 14/LO/1073; and the Scotland A Research Ethics Committee: 14/SS/1009), 1958 c (London – Central: 12/LO/2010), 1970 c (London-Central Research Ethics Committee: 14/LO/0371), 2001 c (London – Central MREC: 13/LO/1786).

Table 1
Participant characteristics: Data from four British cohort studies.
Cohort study, birth year
1946 c1958 c1970 c2001 c
Cohort characteristics and outcomes
Sample size5,34817,49017,61717,057
Males (%)52.6%51.7%51.9%51.4%
Height at 11 years, mean (SD) z-score–0.41 (1.03)–0.25 (1.03)–0.18 (1.01)0.29 (1.00)
Height at 16 years, mean (SD) z-score–0.54 (1.03)–0.35 (1.02)–0.14 (1.16)0.27 (0.98)
Confounding variables
Social class
 I Professional3.0%5.4%6.1%5.0%
 II Managerial and technical14%18.2%23.6%43.1%
 III Skilled Manual46.7%43.4%44.5%21.7%
 III Skilled Non-Manual9.1%9.3%9.2%13.2%
 IV Partly Skilled19.0%17.6%12.5%14.0%
 V Unskilled8.2%6.1%4.1%2.9%
Maternal education, % post compulsory20.86%25.1%34.5%52.8%
Maternal height, mean (SD) cm160.81 (6.38)161.99 (6.45)161.28 (6.63)164.14 (6.93)
Paternal height, mean (SD) cm172.68 (8.07)174.47 (7.42)175.21 (7.51)178.28 (7.23)

  1. Note: Data are imputed (32 imputations); see methods for details and exact age of measurement in each cohort.

Measures

Height

As described elsewhere (Johnson et al., 2015), height was obtained at the following ages in each study: ages 11 and 16 in the 1946 c and 1958 c; ages 10 and 16 in 1970 c; ages 7, 11, 14, and 17 in 2001 c. Height was directly measured in all instances, except age 16 years in 1970 c in which either self-report or objectively measured height was obtained (see sensitivity analyses below). To facilitate cross-cohort comparisons in our main analysis height variables were converted to age and sex-adjusted z-scores using growth reference charts (the UK90 reference panel; the childsds package in R Vogel, 2018). As a robustness check, we also converted height to cohort-specific z-scores at each age (i.e. without reference to an external panel such as the UK90 as used in the main analysis), percentile ranks, and ridit scores (see below).

Cognitive function

Given differences in the cognitive tests administered to each cohort (Moulton et al., 2020), we screened and identified measures that captured the same broad cognitive domains in childhood and adolescence. Assuming the relative ranking was similar, this would enable valid cross-cohort comparisons of associations with outcomes. Differences in the cognitive tests completed could potentially bias cross-cohort comparisons of cognition-height associations. Thus, we used multiple tests assessing different cognitive domains, measured at two different ages in order to examine whether our main findings were robust to the cognitive domain tested or the age chosen. A summary is provided below—information on the tests used, procedures, and scoring is provided in Supplementary file 1; further information on the origins and technical details of these tests is available from elsewhere (Moulton et al., 2020).

Age 10/11 years: verbal reasoning test scores. These capture general verbal ability, verbal knowledge, and reasoning. In the 1946 c and 1958 c, this was measured using the verbal items from the General Ability Test (GAT) (Pigeon, 1964). In the younger cohorts tests from the British Ability Scales (BAS), the Word Similarities test—the precursor to the Verbal Similarities (BAS II)—were administered in the 1970 c and 2001 c, respectively (Elliot et al., 1978; Elliott et al., 1996).

Age 14/16 years: reading/vocabulary test scores. The Watts-Vernon Reading Test (WVRT) was administered in the 1946 c and the Reading Comprehension Test in the 1958 c was devised by the National Foundation for Education Research (NFER) to parallel the aforementioned WVRT; both tests measure reading and word comprehension (Shepherd, 2012). The Applied Psychology Unit (APU) Vocabulary Test was administered in both the 1970 c and 2001 c and measures word knowledge and meaning (Closs, 1976a).

Furthermore, mathematic tests were administered in each cohort and used as an alternative indicator of cognitive performance. These were measured at 15/17 years in each cohort, at 7 years in 2001 c and 10/11 years in the 1946 c, 1958 c, and 1970 c. Age 10/11 years and age 7 years in 2001 c: in the earlier cohorts the mathematics tests were specifically designed for each of the cohorts and measured contemporary number skills and mathematical calculations (1946 c Pigeon, 1964, 1958 c Shepherd, 2012, 1970 c Parsons, 2014, 2001 c Chaplin Gray et al., 2010). For the 2001 c an adaptation of the NFER Progress in Maths Test was administered. The Age 15/17 years: mathematical tests devised by the NFER were administered in the 1946 c (Pigeon, 1964) and 1958 c (Shepherd, 2012) and the APU Arithmetic Test in the 1970 c (Closs and Hutchings, 1976b); all tested numerical skills and problem-solving, along with geometry and trigonometry in the earlier cohorts. Number Analogies, a short version of the Quantitative Reasoning Battery (GL Assessments) was administered in the 2001 c.

Since test performance differs by age (Moulton et al., 2020), we constructed age-adjusted cognition measures by exporting the residuals from a linear regression model with test score as the outcome and age as the exposure variable (linear term). At age 16 years in 1970 c, vocabulary tests were administered either in the home or in school—since test scores differed by mode, we also adjusted for a binary indicator of mode of administration in these models. Finally, to aid the interpretation of effect sizes—and comparisons across age and cohort—in main analyses, we standardized all cognition measures to ridit scores (Mackenbach and Kunst, 1997) individuals in each category were assigned a value based on the proportion of the population with lower cognitive scores. This results in values between 0 and 1 with higher scores equating to better cognitive performance. In regression models coefficients thus show the mean difference in height comparing those with the lowest and highest cognition scores (termed the slope index of inequality in the health inequality literature Mackenbach and Kunst, 1997). As a robustness check, we also repeated analyses converting cognition scores to (test and cohort-specific) z-scores and percentile ranks.

Potential confounders

We hypothesized two sets of confounding variables—family socioeconomic characteristics, and parental height. Childhood socioeconomic position (SEP) was indicated by the father’s occupational social class, measured at 10/11 years (highest parental occupational was used for the 2001 c). To aid cross-cohort comparability, the Registrar General’s Social Class was used to classify social class—from I (professional), II (managerial and technical), IIIN (skilled nonmanual), IIIM (skilled manual), IV (partly-skilled), and V (unskilled) occupations. Maternal education was also used (ascertained at age six in 1946 c NSHD, at birth in the 1958 c NCDS and 1970c BCS, and at nine months – 14 years in 2001 c MCS) with a binary indicator of whether the mother had left education at the mandatory leaving age (14 years old from 1918, 15 from 1944, and 16 from 1972).

Parental height was reported in early life— six years (1946 c), 11 years (1958 c), 0 & 10 years (1970 c), nine months – seven years (2001 c)—with those with particularly low height removed (height <1.4 m; n=14, 18, 12, 10, respectively in each cohort). Parental cognition was not measured in each cohort, precluding its investigation.

Statistical analysis

First, we conducted descriptive analyses and estimated bivariate correlations between cognition, social class, and height. Second, we regressed height on individual cognition variables (measured at the same age) adjusting for sex using linear regression since findings were similar in either sex. We did not anticipate childhood cognition and height to be directly causally related, but rather associated due to the influence of common environmental and genetic causes (see Figure 1). To test our hypotheses, we sequentially adjusted for parental SEP (class and education) and parental height (mothers and fathers).

We next used quantile regression models (Koenker and Bassett, 1978) to investigate whether mean differences observed in linear regression models arose due to differences in the lower quantiles (reflecting impaired height growth, which may be more pronounced in older cohorts). We estimated the cognition-height association at each decile, again adjusting for sex and sequentially for parental SEP and parental height.

To address missing data, multiple imputations by chained equations was completed (separately in each cohort/sex), with 32 imputations performed (burn-in = 10 iterations). To improve the plausibility of the missing at random assumption, additional auxiliary variables were included in imputation models (e.g. childhood BMI and parental education). OLS results were pooled using Rubin’s rules (Rubin, 1976), while quantile regressions were pooled using the MI-then-bootstrap procedure (Bartlett and Hughes, 2020) with confidence intervals derived using the percentile method (200 bootstraps per imputation). The main findings did not differ when complete case analyses were conducted (results available upon request). Since the 1946 c and 2001 c are stratified study designs, analyses using these cohorts were conducted using study-specific design weights.

Additional and sensitivity analyses

As differences in missing data may bias cross-cohort patterns of association (Bann et al., 2021), in addition to using multiple imputations we carried out sensitivity analyses using pattern mixture modeling Leurent et al., 2018; our main OLS models were re-run altering imputed height values by a range of constant values (–1 to + 1 SD). This enabled investigation of the robustness of the main results to non-random missingness in the height data.

We also carried out a series of other robustness checks. First, we repeated all models using different procedures for harmonizing height (raw, z-score, ridit, and rank score) and cognition variables (z-score and rank score). Second, we repeated analyses for males and females separately. Third, as the ethnic mix of the 2001 c greatly exceeds that of older cohorts, we investigated if ethnic composition differences may have contributed to the main inferences drawn by conducting additional analyses restricted to the majority (European ancestry) population. Fourth, as the height at age 16 was collected by self-report for some participants in the 1970 c, we repeated main analyses using only the subset of those with measured height.

The associations of cognition and height could be biased by differences in the measurement of cognition in each cohort. Consequently, we carried out further sensitivity analyses to examine the role of measurement in explaining our results. First, we constrained each set of comparisons to have the same number of cognition response items; those from the cohort with the smallest observed range of scores (e.g. 22 items for verbal scores at 10/11 years; 1970 c had 22 test items, all other cohorts had 40–41 items and were thus rounded down; non-integer values were rounded up or down to the nearest integer; see Supplementary file 1). Second, we took random subsets of individual items to produce tests of the same size. This was only possible for the 1970 c and 2001 c, for which item-level data were available.

Results

Descriptive statistics

Childhood and adolescent height were greater in each successive cohort (Table 1, Supplementary file 1). While differences in the cognitive tests administered preclude a comparison of absolute levels of cognition by cohort, the different tests used provided variation within each cohort and at each age (Figure 2). Cognitive test scores were strongly moderately positively correlated with each other, with the size of the correlation weakening across time (Figure 2—figure supplement 1). Higher social class was associated with taller height and higher cognition scores; these correlations also appeared to weaken over time, as did positive correlations between maternal and paternal height (Figure 2—figure supplement 1). The correlation between parental height and parental education was weakest in the 2001 c, but generally similar in the other cohorts (Figure 2—figure supplement 2).

Figure 2 with 2 supplements see all
Download asset Open asset
Distribution of cognition scores, by cohort and test.

Observed data.

Main analyses

Higher cognition scores were associated with taller height—this association was found across all cohorts at both 11 and 16 years (Figure 3). These associations were stable or slightly larger between the 1946 c and 1958 c cohorts, and weakened thereafter, being substantially weakest in the youngest (2001 c) cohort. For example, after adjustment for sex, the mean difference in height associated with verbal reasoning score at 10/11 years was 0.57 SD (95% CI=0.45, 0.7) in the 1946 c, 0.59 SD (95% CI=0.53, 0.65) in the 1958 c, 0.47 SD (95% CI=0.40, 0.54) in the 1970 c, and 0.29 SD (95% CI=0.21, 0.36) in the 2001 c. Note that cognition scores were rescaled to ridit scores (0–1) to aid comparability: these effect estimates, therefore, show the difference in height comparing the highest with the lowest cognition score. Expressed alternatively, there was a reduction in the correlation coefficient between verbal reasoning scores and height at 10/11 years from 0.17 in 1958 c (0.15–0.20) to 0.08 in 2001 c (0.06–0.10). Testing via the inclusion of cohort*cognition interaction terms also supported this finding (Figure 3—figure supplement 1). In particular, estimates from the 1946 c were least precise owing to its smaller sample size, and confidence intervals overlapped with the larger 1958 c. The pattern of weakening association across time was found at both 11 and 16 years and across all measures of cognition (verbal reasoning, maths, and reading/vocabulary).

Figure 3 with 4 supplements see all
Download asset Open asset
Association between height and cognition in four birth cohort studies.

Cognition scores are the independent variables and height (age-adjusted z-scores) are the dependent variables; estimates are from separate linear regression models (95% CIs): each estimate shows the mean difference in height (z-score) comparing the highest with lowest cognition test score. Models were sequentially adjusted for sex (left panel); sex, mother’s education, and father’s social class (middle panel), and additionally for maternal and paternal height (right panel).

Associations between cognition and height were partly attenuated after adjustment for parental SEP and partly attenuated further by parental height (Figure 3) middle and right panels, respectively. However, the associations did not attenuate completely to zero after this adjustment, and the weakening of the cognition-height association observed across each cohort was similar before and after the adjustment. Attenuation (in absolute terms) when adjusting for parental SEP and parental height was broadly similar in the 1946 c, 1958 c, and 1970 c, but was smallest in the 2001 c (Figure 3—figure supplement 2).

Additional analyses: Quantile regression results

In quantile regression models adjusting for sex, there was evidence that associations between height and cognition were stronger at lower centiles of height, particularly in older cohorts (Figure 4). That is, the average (mean) differences reported previously using linear regression models may have been driven by differences in the lower part of the height distribution, as hypothesized. This was especially pronounced in the 1958 c: for verbal scores at age 11, cognitive ability was associated with 0.73 SD (95% CI=0.63, 0.83) greater height at the 10th percentile and 0.47 SD (95% CI=0.37, 0.58) greater height at the 90th percentile. Differences across centiles were less marked in the 2001 c; corresponding figures in 2001c were 0.34 SD (95% CI=0.22, 0.45) and 0.27 SD (95% CI=0.14, 0.41), respectively. Qualitatively similar results were obtained when adjusted sequentially for childhood SEP and parental height.

Figure 4
Download asset Open asset
Association between height and cognition in four birth cohort studies.

Cognition scores are the independent variables and height (age-adjusted z-scores) are the dependent variables; estimates are from separate quantile regression models (95% CIs): each estimate shows the difference in height (z-score) comparing the highest with lowest cognition test score, repeated for each decile of height (x-axis), cohort (panel columns) and measure of cognition (panel rows). Quantile regression results are interpreted similarly to those from linear regression. For example, the 50th centile shows the estimated difference in height at the median (linear regression would show the mean difference). Results were drawn from pooled quantile regression models (32 imputed datasets) with adjustment for sex. Shaded areas are 95% CIs (200 bootstraps per imputed dataset; centile method) for the specific cohort and cognition measure examined. Dashed lines show estimates for other cohorts to aid cross-cohort comparisons.

Sensitivity analyses

Among participants whose height was measured on at least one occasion, there were small differences in height according to whether an individual was lost to follow-up or not (<0.3 SD). However, repeating main analyses altering imputed values by a constant factor to attempt to account for non-random missingness (i.e. pattern mixture modeling) generally yielded qualitatively similar results, even where the factor was large (i.e.>0.8 SD; Figure 5): smallest associations were estimated in the 2001 c, and largest associations were estimated in the 1946 c or 1958 c.

Figure 5 with 3 supplements see all
Download asset Open asset
Result of pattern mixture models.

Difference in the association between height and maths @ age 11 across the cohort. Derived from pooled OLS models (32 imputed datasets) including adjustment for sex. OLS models repeated with imputed height values adjusted by a constant factor given in columns and rows. The color of each square indicates the difference in the association between height and cognition between two cohorts with the coefficient for the cohort on the horizontal subtracted from that on the horizontal (with given adjustment factors used). Height harmonized across cohorts using growth charts from a 1990 UK growth study. Cognition scores harmonized using ridit scoring (range 0–1).

Repeating analyses using different procedures to score cognition (z-score, ridit scores, and percentile ranks) and height (raw, percentile rank, z-scores, ridit scores) variables in order to compare estimates across cohorts also yielded qualitatively similar results to the main OLS analysis, as did limiting analyses with the 2001 c to white participants only, or the 1970 c to those with measured rather than self-reported height at 16 years (results available upon request). Finally, results were similar when constraining similar tests to have the same range (Figure 3—figure supplement 3) or sampling an equal number of items from similar tests (Figure 3—figure supplement 4).

Discussion

Summary of findings

The association between higher cognition and height weakened substantially from 1957–2018. Associations were at least half as strong in the youngest (2001 c) compared with the oldest (1946 c and 1958 c) cohorts. Quantile regression analyses suggested that height-cognition associations (and potentially their weakening across time) were driven by differences in lower quantiles of height. Associations were only partly attenuated when controlling for parental SEP or parental height.

Our findings build on earlier work, such as a study of Danish male conscripts (Teasdale et al., 1989), which also reported a weakening of the cognition-height association from 1939–1967. We extend these findings by use of more recent data in more generalizable samples, utilizing multiple cognitive tests, undertaking both linear and quantile regression analyses, and accounting for both parental social class and height.

Explanation of findings

Just as multiple processes could explain why cognition itself has increased across time (the Flynn effect) (Pietschnig and Voracek, 2015), multiple processes could explain why associations between cognition and height have weakened across time. First, changes could have been driven by changes across time in the shared determinants of height and cognition. In particular, improvements in nutrition and reductions in the incidence (and/or impact of) infectious disease may have weakened this association, particularly given the stronger magnitude of the association at lower height centiles. This finding is consistent with declines in the socioeconomic patterning of height observed across this time period, as found in this study and reported elsewhere (Bann et al., 2018).

We adjusted for multiple indicators of parental SEP available in each cohort (social class and education), and this may only partly capture the socioeconomically-graded environmental factors responsible for this association; further, we cannot fully rule out the potential for residual confounding by SEP itself (i.e. that height and cognition are correlated because of common effects of the familial environment). In particular, an apparent weakening of the correlation between parental social class and offspring cognition across time (Paterson, 2021) could have led to more residual confounding and thus stronger associations between height and cognition in older cohorts. In support of this, we found some evidence that attenuation upon adjustment for SEP was more pronounced in older cohorts. However, differences in the measurement of SEP across time may have operated in the other direction. While the SEP measures were constructed to be comparable in each cohort, the social class schema adopted in our study originated in the study of 20th century job roles (Bland, 1979); analogously, parental education may have been a more distinguishing trait in earlier cohorts (before the broader expansion of education). Thus, measurement error of SEP and resulting residual confounding may have instead been greater in the younger cohorts. The robustness of the cross-cohort weakening in association after adjustment for parental height (an indicator of genetic liability to taller height) is similarly consistent with the hypothesis that environmental factors which influence early life height and cognition growth may be at least partly responsible for the weakening in association observed rather than genetic factors related to height.

Changes in mating patterns across time could also theoretically contribute to our finding of weakening in the association between cognition and height. A weakening of cross-trait assortative mating by cognition and height across cohorts could drive a weakening in the associations between these factors that we observed. However, we are not aware of empirical evidence for this and are unable to directly test this owing to a lack of data for parental cognition. While we were unable to investigate changes in cognition-height assortative mating, parental height correlations were broadly similar in the 1946 c, 1958 c, and 1970 c and slightly weaker in the youngest cohort; correlations between paternal height and maternal education were also weaker in the youngest cohort. We also lack data to test whether other changes to mating patterns—such as reductions in genetic markers of inbreeding across time (Nalls et al., 2009) may explain our findings. In the social science literature, evidence for changes in assortative mating by socioeconomic factors paints a complex picture: studies have suggested either stability (Henz and Mills, 2018), increases in assortative mating, or mixed findings (Schwartz, 2013). Further research utilizing genetic and parental data is warranted to investigate these possibilities.

The multi-purpose and multidisciplinary cohorts used cognition tests which differed slightly in each cohort. It is, therefore, possible that differences in testing could have either: (1) generated the pattern of results we observed, such that if identical tests were used the association between cognition and height would otherwise have been similar in each cohort; in contrast to previous findings which reported using identical tests Teasdale et al., 1989; or (2) biased our results, such that if identical tests were used the decline in the association between cognition and height would have been less marked than we reported. While we cannot directly falsify this alternative hypothesis given our reliance on historical data sources, a number of lines of reasoning suggest that the first scenario is unlikely. First, our results were similar when using four different cognitive tests (spanning mathematical and verbal reasoning); any bias which generated the results we observed should be similarly present across all 4 tests. Other things being equal, one would expect that more discriminatory tests (i.e. those with a greater number of items) would have higher accuracy and thus better measure cognition. Our results were similar when the youngest cohort had similar numbers of unique scores in cognitive tests compared with the oldest cohort (Verbal @ 11 years: n=41 in 1946 c, n=40 in 2001 c) and fewer unique scores (Maths @ 7/11: n=51 in 1946 c, n=21 in 2001 c). Our results were also similar in sensitivity analyses in which the number of response items was set to be the same in each cohort. Higher random measurement error in the independent variable (cognition) would lead to weakened observed associations with the outcome (height) (Hutcheon et al., 2010), yet we do not a-priori anticipate that this such error was higher in younger cohorts across all tests in such a manner that would have led to the correlation we observed.

Ensuring comparability of exposure is a major challenge across such large timespans. Reassuringly, our results are consistent with those from a previous study that reported the same tests being used (from 1939 to 1967) (Teasdale et al., 1989). However, even seemingly identical tests require modification across time (e.g. for verbal reasoning/vocabulary there is typically a need to adapt question items due to societal and cultural changes over time in vocabulary and numerical use); further, changes to education such as increases in testing may have led to increasing preparedness and familiarity with testing than in the past even where identical tests are used.

Interestingly, we observed a marked reduction in the correlation between cognitive tests across time (e.g. between verbal and maths scores). This trend has been reported in previous studies (Sundet et al., 2004; Kane, 2000) and warrants future investigation; it is consistent with evidence that IQ gains across time seemingly differ by cognitive domain (Pietschnig and Voracek, 2015) potentially capturing differences across time in cognitive skill use and development in the population. Previous studies using three (1958–2001 c) of the included cohorts have also reported changing associations between cognition (verbal test scores at 10/11 years) and other traits: a declining negative association with birth weight (Goisis et al., 2017a) and a change in direction of association with maternal age (from negative to positive) Goisis et al., 2017b; each finding has plausible explanations based on changes across time in relevant societal phenomena (improved medical conditions Goisis et al., 2017a and changes in parental characteristics Goisis et al., 2017b respectfully), yet also cannot conclusively falsify the notion that differences in tests used influences the results obtained. In this paper, we used multiple tests and sensitivity analyses to attempt to address this.

Strengths and limitations

Strengths of this study include the use of multiple nationally representative longitudinal cohorts from the UK—each with data on multiple cognitive test scores and measured height. We used four cohorts—the latest measure of the youngest cohort was 17 years—since height gains may occur up to 19–20 years (Rodriguez-Martinez et al., 2020), further work may be warranted in the future to test if the patterns of association remain in later adulthood. Our analytical strategy is also a strength, enabling us to test the robustness of our findings across multiple ages and multiple domains of cognitive test performance and statistical methods.

Limitations include the potential for bias due to missing data, which is more pronounced in the youngest cohort. However, multiple imputation models were used, with multiple auxiliary variables included to increase the credibility of the ‘missing at random’ assumption such models entail; and findings from pattern mixture modeling suggested that our main findings were robust to differential missingness of outcome data.

Conclusions

Associations between higher cognition and taller height in childhood-adolescence weakened markedly from 1957–2018. These findings support the notion that changes to environmental and societal factors across time may substantially weaken associations between cognition and other traits.

Data availability

Data files are available through the UK Data Service (1958c, 1970c, and 2001c; https://www.data-archive.ac.uk/) and via application (1946c; https://skylark.ucl.ac.uk/). The code used to run the analysis is available at https://osf.io/54unw/.

The following data sets were generated
    1. Wright L
    (2022) Open Science Framework
    ID 54unw. Weakening of the cognition and height association from 1957 to 2018.
    https://osf.io/54unw/

References

  2. Book
    1. Bann D
    2. Wright L
    3. Goisis A
    (2021)
    Investigating Change across Time in Prevalence or Association Using Observational Data: Guidance on Utility, Methodology, and Interpretation
    OSF Preprints.
    1. Bland R
    (1979)
    Measuring" social class" A discussion of the registrar-general’s classification
    Sociology 13:283–291.
  8. Book
    1. Centre for Longitudinal Studies
    (2014)
    National Child Development Study Ethical Review and Consent
    University College London Press.
  9. Report
    1. Chaplin Gray J
    2. Gatenby R
    3. Simmonds N
    (2010)
    Millennium Cohort Study Sweep 4 Technical Report
    Centre for Longitudinal Studies.
  10. Book
    1. Closs SJ
    (1976a)
    APU vocaubulary test (multiple choice format, 1986)
    Hodder and Stoughton Educational Ltd.
  11. Book
    1. Closs S
    2. Hutchings M
    (1976b)
    APU Arithmetic Test
    Hodder & Stoughton.
    1. Cole TJ
    (2000) Secular trends in growth
    The Proceedings of the Nutrition Society 59:317–324.
    https://doi.org/10.1017/s0029665100000355
    1. Deary IJ
    2. Batty GD
    (2007) Cognitive epidemiology
    Journal of Epidemiology and Community Health 61:378–384.
    https://doi.org/10.1136/jech.2005.039206
  15. Book
    1. Elliot C
    2. Murray D
    3. Pearson L
    (1978)
    British Ability Scales
    National Foundation for Educational Research.
  16. Report
    1. Elliott CD
    2. Smith P
    3. McCulloch K
    (1996)
    British Ability Scales Second Edition (BAS II) Administration and Scoring Manual
    NFER-Nelson.
  34. Book
    1. Moulton V
    2. McElroy E
    3. Richards M
    4. Fitzsimons E
    5. Northstone K
    6. Conti G
    7. Ploubidis GB
    8. Sullivan A
    9. O’Neill D
    (2020)
    A Guide to the Cognitive Measures in the British Birth Cohort Studies
    London, UK: CLOSER.
  36. Book
    1. Parsons S
    (2014)
    Childhood cognition in the 1970 British Cohort Study, CLS Working Paper. London: Centre for Longitudinal Studies
    University College London Press.
  40. Book
    1. Pigeon D
    (1964)
    Tests used in the 1954 and 1957 surveys
    In: Douglas JWB, editors. The Home and the School: A Study of Ability and Attainment in the Primary School. London, England: Macgibbon and Kee. pp. 129–132.
  46. Book
    1. Shepherd P.
    (2012)
    Measures of ability at ages 7 to 16
    National Child Development Study User Guide.
  47. Book
    1. Shepherd P
    2. Gilbert E
    (2019a)
    Millenium Cohort Study Ethical Review and Consent
    University College London Press.
  48. Book
    1. Shepherd P
    2. Gilbert E
    (2019b)
    British Cohort Study Ethical Review and Consent
    University College London Press.
  53. Book
    1. Vogel M
    (2018)
    A Combined Approach to Generate Age & Sex Dependent Reference Intervals in Pediatrics
    QUCOSA: Diss.

Decision letter

  1. Eduardo L Franco
    Senior and Reviewing Editor; McGill University, Canada
  2. Richard Border
    Reviewer; University of California, Los Angeles, United States

Our editorial process produces two outputs: i) public reviews designed to be posted alongside the preprint for the benefit of readers; ii) feedback on the manuscript for the authors, including requests for revisions, shown below. We also include an acceptance summary that explains what the editors found interesting or important about the work.

Decision letter after peer review:

Thank you for submitting your article "Weakening of the cognition and height association from 1957 to 2018: findings from four British birth cohort studies" for consideration by eLife. Your article has been reviewed by 3 peer reviewers, and I oversaw the evaluation in my dual role of Reviewing Editor and Senior Editor. The following individual involved in review of your submission has agreed to reveal their identity: Richard Border (Reviewer #1).

As is customary in eLife, the reviewers have discussed their critiques with one another and with the Editors. The decision was reached by consensus. What follows below is a lightly edited compilation of the essential and ancillary points provided by reviewers in their critiques and in their interaction post-review.

Essential revisions:

The critiques by the three reviewers contain challenges or reflections related to your study's findings that the tests of cognition have behaved differently across generations and how this affects the validity of your findings. Please submit a revised version that addresses these concerns directly. Although we expect that you will address these comments in your response letter, we also need to see the corresponding revision clearly marked in the text of the manuscript. Some of the reviewers' comments may seem to be simple queries or challenges that do not prompt revisions to the text. Please keep in mind, however, that readers may have the same perspective as the reviewers. Therefore, it is essential that you amend or expand the text to clarify the narrative accordingly.

Reviewer #1 (Recommendations for the authors):

The authors conducted a thorough analysis of the correlation between height and measures of cognitive abilities (what are essentially IQ test components) across four cohorts of children and adolescents in the UK measured between 1957 and 2018. The authors find the strength of the association between height and cognitive measures declined over this time frame--for example, among 10- and 11-year-olds born in 1958, height explained roughly 3% of the variation in verbal reasoning scores; this dropped to approximately 0.6% among those born in 2001. These associations were further attenuated after accounting for proxy measures of social class.

The authors' analyses were performed carefully and their observations regarding declining height / cognitive measure associations are likely to be robust if we interpret their results with an important caveat: these results reflect measurements aimed at assessing cognition rather than cognition itself. The importance of this distinction is evidenced by the changing correlation structure of the cognitive measures over time. For example, age 11 verbal / math scores were correlated at >= 0.75 at the first two time points but dropped to 0.33 at the most recent time point. Similar patterns are present for the other cognitive measures and time points. The authors' conclude that such changes are unlikely to impact their primary findings, but I'm less certain. For example, one interpretation of this finding is that older cognitive measures were simply worse at indexing distinct cognitive domains and instead reflected a combination of cognitive ability together with non-specific factors relating to opportunity, health, class, etc. Further, height was historically a stronger proxy for class and economic status than it is today (e.g., by capturing adequate nutritional intake, risk for childhood disease, etc.). Together, then, previously high height / cognitive measure correlations might reflect the fact that both phenotypes previously indexed socio-economic factors to a greater extent than they might today (which is still non-negligible).

Additionally, their findings add an interesting data point to a collection of recent results suggesting that the relationship between cognitive and anthropometric measures is complex and difficult to interpret. For example, studies using genetic markers to examine shared genetic bases have virtually all relied on methods assuming mating is random, which is not the case empirically. Howe et al. (doi.org/10.1038/s41588-022-01062-7) recently reported that the ostensible genetic correlation of -.32 between years of education and BMI attenuates to -.05 when using direct-effect estimates, which should theoretically be immune to the effects of non-random mating and other confounding variables. Likewise, Keller et al. (doi.org/10.1371/journal.pgen.1003451) and Border et al. (doi.org/10.1101/2022.03.21.485215) used very different approaches to arrive at the same conclusion that ~50% of the nominal genetic correlation between IQ and height could be attributed to bivariate assortative mating rather than shared causal biological factors. Given that assortative mating on both IQ measures and height involves many other traits (not just two as assumed in such bivariate models), the true extent to which height / IQ correlations reflect causal factors is plausibly even lower than these estimates suggest. For these reasons, I do not entirely agree with the authors' review of previous findings in the introduction, where they write "recent studies have suggested that links between higher cognition and taller height can be largely explained by genetic factors", though it is certainly true that this claim has been made.

This manuscript is well-written and reflects carefully conducted research. Below are my suggestions/concerns, in no particular order.

Throughout I would suggest that the authors replace the term "cognition", with something along the lines of "cognitive measures" or "measures of cognitive ability". This might seem pedantic, but the next point highlights why I believe it is important not to conflate tools attempting to measure constructs with the constructs themselves.

Figure S4 shows a striking pattern where the inter-correlations among cognitive measures appears to sharply decline over time. E.g., age 11 verbal / math scores were correlated at >= 0.75 in the first two cohorts. This drops to 0.33 in the 2001 cohort. I would ask the authors to discuss these findings, which are currently mentioned briefly in the discussion, in greater detail. Further, I'm not sure I agree with the authors' implication that the main findings are necessarily robust to differences in tests employed over time.

For example, one interpretation of this finding is that older cognitive measures were simply worse at measuring distinct cognitive domains and instead reflected a combination of the targeted phenotype together with non-specific factors relating to opportunity, health, class, etc. To what extent to historically higher estimates simply reflect poorer measurement? Put another way, to what extent might the decrease in cognitive measure/height correlation over time reflect that IQ tests, while surely still proxies for non-cognitive measures to some extent, are better than they used to be?

Please discuss in greater detail. Further, in Figure S67, it appears that the attenuation in height/cog association after accounting for SEP is larger in the earlier cohorts than the later cohorts. If the authors have adequate power to examine this, I think it would further add to the above discussion; OTOH, if the authors think their data isn't well suited to testing this hypothesis (whether due to the fact that SEP is an imperfect measure of socio-economic status/class, or on statistical grounds), it might warrant discussion as a future direction.

As highlighted in my public comments I think it's inaccurate to state that recent studies have suggested that height/cog measure correlations can be "largely explained by genetic factors". Of the four papers cited:

(2) – certainly does make this suggestion but relies on LDSC genetic correlation, which has been found to yield extreme overestimates of rg when random mating assumptions are violated.

(3) – used an ETFD and found roughly half of phenotypic correlation attributable to bivariate assortative mating (AM). While AM is "genetic" in the fact that it reflects correlated genetic liabilities, it need not reflect any shared causal biological factors across traits, which is how I think readers usually interpret such language.

(9) – genetic correlation estimates between height and IQ were non-significant for girls, bottom of CIs were.03 and.00 for the two boys' cohorts.

(10) – These results are complex: the authors reports report that the best fitting trivariate model of height, general cognitive ability (GCA), and cortical surface area (CSA) permitted the genetic correlation between GCA and height to be constrained to zero but don't test whether the CSA / height rg can also be constrained to zero. They also report that the height / GCA association is fully mediated via CSA, but that the CSA pathway only accounts for 39% of the phenotypic correlation between height / GCA.

Together with the additional results mentioned in the public-facing review, these findings paint a complicated picture that isn't currently reflected in the introduction and discussion.

I would suggest less causal-sounding language regarding the unclear links between height and cognitive measures. E.g., in "Cognitive ability is a potentially important determinant of health" replace "determinant" with "correlate".

Reviewer #2 (Recommendations for the authors):

The authors use birth cohorts with extensive cognitive assessments and height measurements along with data on parental height and socioeconomic status. The authors estimate that the correlation between height and cognitive ability has approximately halved in the last 60 years.

Quantile regression results suggest that this is due to a stronger association between low cognitive ability and short stature in older cohorts, potentially due to environmental factors that cause both and that have been removed by improvements in the environment in the last 60 years.

While this is a plausible hypothesis, the evidence presented in the manuscript is unable to rule out alternative hypotheses, such as changes in assortative mating.

The results in the manuscript will be of interest to researchers investigating how genetics and environment lead to correlations between cognitive and physical/health traits, and to researchers interested in the relationship between social and health inequalities.

While my sense of the evidence presented is that there is fairly solid statistical evidence for a trend where the correlation between cognitive ability and height declines over time, there is no formal quantification of this trend nor measurement of the uncertainty in the trend.

Similarly, the quantile regression plots in Figure 2 appear to show a trend across the height deciles for the two oldest cohorts, but no quantification of how strong this is nor what uncertainty exists is calculated. Furthermore, if the apparent trend in the quantile regression plots is true, wouldn't this imply a non-linear association between height and cognitive ability for the older cohorts? Can this be seen in the scatterplots or in a non-linear regression?

I think the authors could have done more with their data to investigate the contribution of assortative mating to the observed trend. Looking at Figure S4, it looks like the correlation between mother's education and father's height in the 2001 cohort is substantially lower than for previous cohorts. While cognitive ability may not be available for parents, one could look at, for example, father's education and mother's height across the cohorts and see if there is a downward trend in correlation.

https://doi.org/10.7554/eLife.81099.sa1

Author response

Reviewer #1 (Recommendations for the authors):

The authors conducted a thorough analysis of the correlation between height and measures of cognitive abilities (what are essentially IQ test components) across four cohorts of children and adolescents in the UK measured between 1957 and 2018. The authors find the strength of the association between height and cognitive measures declined over this time frame--for example, among 10- and 11-year-olds born in 1958, height explained roughly 3% of the variation in verbal reasoning scores; this dropped to approximately 0.6% among those born in 2001. These associations were further attenuated after accounting for proxy measures of social class.

The authors' analyses were performed carefully and their observations regarding declining height / cognitive measure associations are likely to be robust if we interpret their results with an important caveat: these results reflect measurements aimed at assessing cognition rather than cognition itself. The importance of this distinction is evidenced by the changing correlation structure of the cognitive measures over time. For example, age 11 verbal / math scores were correlated at >= 0.75 at the first two time points but dropped to 0.33 at the most recent time point. Similar patterns are present for the other cognitive measures and time points. The authors' conclude that such changes are unlikely to impact their primary findings, but I'm less certain. For example, one interpretation of this finding is that older cognitive measures were simply worse at indexing distinct cognitive domains and instead reflected a combination of cognitive ability together with non-specific factors relating to opportunity, health, class, etc. Further, height was historically a stronger proxy for class and economic status than it is today (e.g., by capturing adequate nutritional intake, risk for childhood disease, etc.). Together, then, previously high height / cognitive measure correlations might reflect the fact that both phenotypes previously indexed socio-economic factors to a greater extent than they might today (which is still non-negligible).

We agree, it is possible that our results could in principle be explained by changes to the measures. We have provided further analysis to attempt to inform the likelihood of this suggestion and have expanded our discussion of this issue (Discussion, explanation of findings section; copied below).

First, we conducted additional sensitivity analysis repeating our main analysis using cognition measures in which the number of response options was set to be the same for each test (the lowest common denominator across all cohorts). This was tested in two separate approaches: 1) by reducing the number of categories to the same number in each cohort; and 2) or by picking a random sample of question items for each category. Our main findings were unchanged: described in “Additional and sensitivity analyses” section, FiguresS20-S21.

Regarding the suggestion that “high height / cognitive measure correlations might reflect the fact that both phenotypes previously indexed socio-economic factors to a greater extent than they might today” – we sought to account for this by adjustment for measured indicators of socioeconomic position, and found the trend remained after adjustment. As in other observational studies we cannot fully rule out the possibility of residual confounding however (Discussion, Explanation of findings paragraph 2).

“The multi-purpose and multidisciplinary cohorts used cognition tests which differed slightly in each cohort. It is therefore possible that differences in testing could have either: 1) entirely generated the pattern of results we observed, such that if identical tests were used the association between cognition and height would otherwise have been identical in each cohort; in contrast to previous findings which reported using identical tests20; or 2) biased our results, such that if identical tests were used the decline in association between cognition and height would have been less marked than we reported. While we cannot directly falsify this alternative hypothesis given our reliance on historical data sources, a number of lines of reasoning suggest that the first scenario is unlikely. First, our results were similar when using 4 different cognitive tests (spanning mathematical and verbal reasoning); any bias which generated the results we observed should be similarly present across all 4 tests. Other things being equal, one would expect that more discriminatory tests (i.e., those with a greater number of responses) would have higher accuracy and thus better index cognition. Our results were similar when the youngest cohort had similar numbers of unique scores in cognitive tests compared with the oldest cohort (Verbal @ 11 years: n=41 in 1946c, n=40 in 2001c) and fewer unique scores (Maths @ 7/11: n=51 in 1946c, n=21 in 2001c). Our results were also similar in sensitivity analyses in which the number of response options were set to be the same in each cohort. Higher random measurement error in the independent variable (cognition) would lead to weakened observed associations with the outcome (height),52 yet we do not a-priori anticipate that this such error was higher in younger across all tests in such a manner that would have led to the correlation we observed.

Ensuring comparability of exposure is a major challenge across such large timespans. Reassuringly, our results are consistent with those from a previous study which reported consistent tests being used (from 1939-1967).20 However, even seemingly identical require modification across time (e.g., for verbal reasoning/vocabulary there is typically a need to adapt question items due to societal and cultural changes over time in vocabulary and numerical use); further, changes to education such as increases in testing may have led to increasing preparedness and familiarity with testing than in the past even where identical tests are used.

Interestingly, we observed a marked reduction in the correlation between cognitive tests across time (e.g., between verbal and maths scores). This trend has been reported in previous studies53 54 and warrants future investigation; it is consistent with evidence that IQ gains across time seemingly differ by cognitive domain,45 potentially capturing differences across time in cognitive skill use and development in the population. Previous studies using three (1958-2001c) of the included cohorts have also reported changing associations between cognition (verbal test scores at 10/11 years) and other traits: a declining negative association with birth weight19 and a change in direction of association with maternal age (from negative to positive);55 each finding has plausible explanations based on changes across time in relevant societal phenomena (improved medical conditions19 and changes in parental characteristics,55 respectfully), yet also cannot conclusively falsify the notion that differences in tests used influences the results obtained. In this paper, we used multiple tests and sensitivity analyses to attempt to address this.”

Additionally, their findings add an interesting data point to a collection of recent results suggesting that the relationship between cognitive and anthropometric measures is complex and difficult to interpret. For example, studies using genetic markers to examine shared genetic bases have virtually all relied on methods assuming mating is random, which is not the case empirically. Howe et al. (doi.org/10.1038/s41588-022-01062-7) recently reported that the ostensible genetic correlation of -.32 between years of education and BMI attenuates to -.05 when using direct-effect estimates, which should theoretically be immune to the effects of non-random mating and other confounding variables. Likewise, Keller et al. (doi.org/10.1371/journal.pgen.1003451) and Border et al. (doi.org/10.1101/2022.03.21.485215) used very different approaches to arrive at the same conclusion that ~50% of the nominal genetic correlation between IQ and height could be attributed to bivariate assortative mating rather than shared causal biological factors. Given that assortative mating on both IQ measures and height involves many other traits (not just two as assumed in such bivariate models), the true extent to which height / IQ correlations reflect causal factors is plausibly even lower than these estimates suggest. For these reasons, I do not entirely agree with the authors' review of previous findings in the introduction, where they write "recent studies have suggested that links between higher cognition and taller height can be largely explained by genetic factors", though it is certainly true that this claim has been made.

We have revised our introduction to better reflect the complexity of previous findings and to note that this claim. Please also see our response below.

This manuscript is well-written and reflects carefully conducted research. Below are my suggestions/concerns, in no particular order.

Throughout I would suggest that the authors replace the term "cognition", with something along the lines of "cognitive measures" or "measures of cognitive ability". This might seem pedantic, but the next point highlights why I believe it is important not to conflate tools attempting to measure constructs with the constructs themselves.

We have checked the text throughout and additionally used the term ‘cognition scores’ to try to address this point in several instances. In other instances, we felt that the word cognition was a succinct summary of the intended meaning (ie, a necessary tool used to capture the otherwise unmeasured construct); if the editorial team felt that we should make further amendments we would be happy to consider this.

Figure S4 shows a striking pattern where the inter-correlations among cognitive measures appears to sharply decline over time. E.g., age 11 verbal / math scores were correlated at >= 0.75 in the first two cohorts. This drops to 0.33 in the 2001 cohort. I would ask the authors to discuss these findings, which are currently mentioned briefly in the discussion, in greater detail. Further, I'm not sure I agree with the authors' implication that the main findings are necessarily robust to differences in tests employed over time.

We have added to our discussion of this point. We are working on a separate paper which tackles this correlation in greater detail – it is consistent with studies in the Flynn effect literature, which suggest that increases across time have been divergent across cognitive domain (thus implying that correlations would weaken across time). Please see above regarding the general issue of the robustness of our findings to differences in tests used.

For example, one interpretation of this finding is that older cognitive measures were simply worse at measuring distinct cognitive domains and instead reflected a combination of the targeted phenotype together with non-specific factors relating to opportunity, health, class, etc. To what extent to historically higher estimates simply reflect poorer measurement? Put another way, to what extent might the decrease in cognitive measure/height correlation over time reflect that IQ tests, while surely still proxies for non-cognitive measures to some extent, are better than they used to be?

Please see above and the revisions for a broader discussion of this possibility. On balance we think it is unlikely for this to be the sole explanation for the pattern of results that we find across 4 separate tests (and sensitivity analyses).

Please discuss in greater detail. Further, in Figure S67, it appears that the attenuation in height/cog association after accounting for SEP is larger in the earlier cohorts than the later cohorts. If the authors have adequate power to examine this, I think it would further add to the above discussion; OTOH, if the authors think their data isn't well suited to testing this hypothesis (whether due to the fact that SEP is an imperfect measure of socio-economic status/class, or on statistical grounds), it might warrant discussion as a future direction.

We have added an additional supplementary information (Figure 3) which provides numerical figures on the degree of attenuation and discuss this (Discussion, Explanation of Findings paragraph 2).

As highlighted in my public comments I think it's inaccurate to state that recent studies have suggested that height/cog measure correlations can be "largely explained by genetic factors".

We have updated the introduction to better reflect findings from previous work (see Introduction paragraph 2). Please see below/the revised introduction (paragraph 2).

“Recent studies have suggested that links between higher cognition and taller height can be largely or to a sizable extent explained by shared genetic influences,2 3 9 10 with others suggesting a role of assortative mating.3 11 Keller et al. (2013) for example estimated that approximately half of the genetic association found in a single cohort was accounted for by assortative mating, and half by shared genetic factors.3”

Below we have provided quotes from each paper in support of their citation to support the statements made. We have refrained from additional discussion of methodological detail to retain the focus of the paper and ensure the introduction remained on-topic (eg, we do not test measures of brain structure (ref 10)).

Of the four papers cited:

(2) – certainly does make this suggestion but relies on LDSC genetic correlation, which has been found to yield extreme overestimates of rg when random mating assumptions are violated.

Marioni RE, Batty GD, Hayward C, et al. Common genetic variants explain the majority of the correlation between height and intelligence: the generation Scotland study. Behav Genet 2014;44(2):91-96.

Statement suggesting explained by genetic factors:

“This study identified a modest genetic correlation between height and intelligence with the majority of the phenotypic correlation being explained by shared genetic influences.”

(3) – used an ETFD and found roughly half of phenotypic correlation attributable to bivariate assortative mating (AM). While AM is "genetic" in the fact that it reflects correlated genetic liabilities, it need not reflect any shared causal biological factors across traits, which is how I think readers usually interpret such language.

Keller MC, Garver-Apgar CE, Wright MJ, et al. The genetic correlation between height and IQ: Shared genes or assortative mating? PLoS genetics 2013;9(4):e1003451.

Statement suggesting explained by genetic factors + assortative mating:

“We used this model to demonstrate that the phenotypic correlation between two potentially sexually selected traits in humans, IQ and height, is largely genetic in nature, and that both shared genes and assortative mating contribute importantly to it”

(9) – genetic correlation estimates between height and IQ were non-significant for girls, bottom of CIs were.03 and.00 for the two boys' cohorts.

Silventoinen K, Iacono WG, Krueger R, et al. Genetic and environmental contributions to the association between anthropometric measures and IQ: a study of Minnesota twins at age 11 and 17. Behav Genet 2012;42(3):393-401.

Statement suggesting explained by genetic factors:

“head circumference and total height showed stronger associations with IQ than anthropometric measures focused on specific body parts. We found some evidence that these associations were mainly genetic origin, which may indicate the role of endocrinological factors or genetically based susceptibility to environmental stressors shared by family members.”

(10) – These results are complex: the authors reports report that the best fitting trivariate model of height, general cognitive ability (GCA), and cortical surface area (CSA) permitted the genetic correlation between GCA and height to be constrained to zero but don't test whether the CSA / height rg can also be constrained to zero. They also report that the height / GCA association is fully mediated via CSA, but that the CSA pathway only accounts for 39% of the phenotypic correlation between height / GCA.

Together with the additional results mentioned in the public-facing review, these findings paint a complicated picture that isn't currently reflected in the introduction and discussion.

Vuoksimaa E, Panizzon MS, Franz CE, et al. Brain structure mediates the association between height and cognitive ability. Brain Structure and Function 2018;223(7):3487-94.

Statement suggesting explained by genetic factors:

“Both height and GCA have substantial heritability and their association is largely due to shared genetic effects”

I would suggest less causal-sounding language regarding the unclear links between height and cognitive measures. E.g., in "Cognitive ability is a potentially important determinant of health" replace "determinant" with "correlate".

We agree it is absolutely essential to be very careful with our use of causal language. We believe that “potentially important determinant” reflects the implied causal link and the uncertainty. We refer to https://ajph.aphapublications.org/doi/10.2105/AJPH.2018.304337 for an illustration of the benefits of being explicit regarding causal links, even if this is challenging using observational data.

We have checked the manuscript throughout regarding the causal language used and edited where it was felt that our intended meaning was unclear. On reflection we have opted to retain the causal language where causal links are intended to be implied. We would be happy to revise further if we have overstated or misled in our claims re causation.

Reviewer #2 (Recommendations for the authors):

The authors use birth cohorts with extensive cognitive assessments and height measurements along with data on parental height and socioeconomic status. The authors estimate that the correlation between height and cognitive ability has approximately halved in the last 60 years.

Quantile regression results suggest that this is due to a stronger association between low cognitive ability and short stature in older cohorts, potentially due to environmental factors that cause both and that have been removed by improvements in the environment in the last 60 years.

While this is a plausible hypothesis, the evidence presented in the manuscript is unable to rule out alternative hypotheses, such as changes in assortative mating.

The results in the manuscript will be of interest to researchers investigating how genetics and environment lead to correlations between cognitive and physical/health traits, and to researchers interested in the relationship between social and health inequalities.

While my sense of the evidence presented is that there is fairly solid statistical evidence for a trend where the correlation between cognitive ability and height declines over time, there is no formal quantification of this trend nor measurement of the uncertainty in the trend.

We now include additional statistical tests to compare estimates in each cohort (Figure 3). We have opted to include this in supplemental material given the large number of tests included already.

Similarly, the quantile regression plots in Figure 2 appear to show a trend across the height deciles for the two oldest cohorts, but no quantification of how strong this is nor what uncertainty exists is calculated. Furthermore, if the apparent trend in the quantile regression plots is true, wouldn't this imply a non-linear association between height and cognitive ability for the older cohorts? Can this be seen in the scatterplots or in a non-linear regression?

We included 95% confidence intervals in our quantile regression analyses which provide an indication of uncertainty. We believe that given the substantial amount of analyses (across 4 historical cohorts and 4 cognition tests; 23 supplemental results) further work would be best placed to undertake additional statistical exploration of both quantile regression and non-linear associations. We would be happy to reconsider this if requested.

I think the authors could have done more with their data to investigate the contribution of assortative mating to the observed trend. Looking at Figure S4, it looks like the correlation between mother's education and father's height in the 2001 cohort is substantially lower than for previous cohorts. While cognitive ability may not be available for parents, one could look at, for example, father's education and mother's height across the cohorts and see if there is a downward trend in correlation.

We now include in Figure 2 cross-cohort investigation of the correlation between parental height and maternal education. We find that the correlation is similar across 1946c, 1958c, and 1970c, yet is weaker in 2001c. We comment on this in the paper (see revised discussion, explanation of findings section). Interpretation of these results is complicated by measurement error in parental education (typically reported for both parents by mothers). Further, interpretation may be further complicated by reductions in the socioeconomic patterning of height across time (see https://www.thelancet.com/journals/lanpub/article/PIIS2468-2667(18)30045-8/fulltext). Future would which focuses on assortative mating could investigate these issues.

https://doi.org/10.7554/eLife.81099.sa2

Article and author information

Author details

  1. David Bann

    Centre for Longitudinal Studies, Social Research Institute, University College London, London, United Kingdom
    Contribution
    Conceptualization, Data curation, Supervision, Funding acquisition, Validation, Investigation, Methodology, Writing – original draft, Writing – review and editing
    For correspondence
    david.bann@ucl.ac.uk
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-6454-626X

  2. Liam Wright

    Centre for Longitudinal Studies, Social Research Institute, University College London, London, United Kingdom
    Contribution
    Conceptualization, Resources, Data curation, Software, Formal analysis, Investigation, Visualization, Methodology, Writing – review and editing
    Competing interests
    No competing interests declared

  3. Neil M Davies

    1. MRC IEU, University of Bristol, Bristol, United Kingdom
    2. K.G. Jebsen Center for Genetic Epidemiology, Department of Public Health and Nursing, NTNU, Norwegian University of Science and Technology, Trondheim, Norway
    Contribution
    Conceptualization, Investigation, Methodology, Writing – review and editing
    Competing interests
    No competing interests declared
    Additional information
    Joint Senior Authors
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-2460-0508

  4. Vanessa Moulton

    Centre for Longitudinal Studies, Social Research Institute, University College London, London, United Kingdom
    Contribution
    Conceptualization, Data curation, Investigation, Methodology, Writing – review and editing
    Competing interests
    No competing interests declared
    Additional information
    Joint Senior Authors
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-7400-2522

Funding

Medical Research Council (MR/V002147/1)

  • David Bann

Economic and Social Research Council (ES/M001660/1)

  • David Bann
  • Vanessa Moulton

Economic and Social Research Council (ES/K000357/1)

  • Vanessa Moulton

Medical Research Council (MC_UU_12013/1)

  • Neil M Davies

Medical Research Council (MC_UU_12013/9MC_UU_00011/1)

  • Neil M Davies

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Ethics

Each study and sample sweep used in this manuscript has received ethical approval and obtained parental/participant consent according to guidance which was in place at the time of data collections from 1957 to 2018. These historic datasets are made available for anonymized observational research which can be published in a range of topics by bona fide researchers; further details are provided elsewhere; 25-27 ethical review boards are as follows: 1946c (Queen Square Research Ethics Committee: 14/LO/1073; and the Scotland A Research Ethics Committee: 14/SS/1009), 1958c (London - Central: 12/LO/2010), 1970c (London-Central Research Ethics Committee: 14/LO/0371), 2001c (London - Central MREC: 13/LO/1786).

Senior and Reviewing Editor

  1. Eduardo L Franco, McGill University, Canada

Reviewer

  1. Richard Border, University of California, Los Angeles, United States

Publication history

  1. Preprint posted: December 14, 2021 (view preprint)
  2. Received: June 15, 2022
  3. Accepted: March 19, 2023
  4. Version of Record published: April 6, 2023 (version 1)

Copyright

© 2023, Bann et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 281
    Page views
  • 10
    Downloads
  • 0
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Open citations (links to open the citations from this article in various online reference manager services)

Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)

  1. David Bann
  2. Liam Wright
  3. Neil M Davies
  4. Vanessa Moulton
(2023)
Weakening of the cognition and height association from 1957 to 2018: Findings from four British birth cohort studies
eLife 12:e81099.
https://doi.org/10.7554/eLife.81099

Categories and tags

Research organism

Further reading

    1. Epidemiology and Global Health

    Predictive performance of multi-model ensemble forecasts of COVID-19 across European nations

    Katharine Sherratt, Hugo Gruson ... Sebastian Funk
    Research Article Updated

    Background:

    Short-term forecasts of infectious disease burden can contribute to situational awareness and aid capacity planning. Based on best practice in other fields and recent insights in infectious disease epidemiology, one can maximise the predictive performance of such forecasts if multiple models are combined into an ensemble. Here, we report on the performance of ensembles in predicting COVID-19 cases and deaths across Europe between 08 March 2021 and 07 March 2022.

    Methods:

    We used open-source tools to develop a public European COVID-19 Forecast Hub. We invited groups globally to contribute weekly forecasts for COVID-19 cases and deaths reported by a standardised source for 32 countries over the next 1–4 weeks. Teams submitted forecasts from March 2021 using standardised quantiles of the predictive distribution. Each week we created an ensemble forecast, where each predictive quantile was calculated as the equally-weighted average (initially the mean and then from 26th July the median) of all individual models’ predictive quantiles. We measured the performance of each model using the relative Weighted Interval Score (WIS), comparing models’ forecast accuracy relative to all other models. We retrospectively explored alternative methods for ensemble forecasts, including weighted averages based on models’ past predictive performance.

    Results:

    Over 52 weeks, we collected forecasts from 48 unique models. We evaluated 29 models’ forecast scores in comparison to the ensemble model. We found a weekly ensemble had a consistently strong performance across countries over time. Across all horizons and locations, the ensemble performed better on relative WIS than 83% of participating models’ forecasts of incident cases (with a total N=886 predictions from 23 unique models), and 91% of participating models’ forecasts of deaths (N=763 predictions from 20 models). Across a 1–4 week time horizon, ensemble performance declined with longer forecast periods when forecasting cases, but remained stable over 4 weeks for incident death forecasts. In every forecast across 32 countries, the ensemble outperformed most contributing models when forecasting either cases or deaths, frequently outperforming all of its individual component models. Among several choices of ensemble methods we found that the most influential and best choice was to use a median average of models instead of using the mean, regardless of methods of weighting component forecast models.

    Conclusions:

    Our results support the use of combining forecasts from individual models into an ensemble in order to improve predictive performance across epidemiological targets and populations during infectious disease epidemics. Our findings further suggest that median ensemble methods yield better predictive performance more than ones based on means. Our findings also highlight that forecast consumers should place more weight on incident death forecasts than incident case forecasts at forecast horizons greater than 2 weeks.

    Funding:

    AA, BH, BL, LWa, MMa, PP, SV funded by National Institutes of Health (NIH) Grant 1R01GM109718, NSF BIG DATA Grant IIS-1633028, NSF Grant No.: OAC-1916805, NSF Expeditions in Computing Grant CCF-1918656, CCF-1917819, NSF RAPID CNS-2028004, NSF RAPID OAC-2027541, US Centers for Disease Control and Prevention 75D30119C05935, a grant from Google, University of Virginia Strategic Investment Fund award number SIF160, Defense Threat Reduction Agency (DTRA) under Contract No. HDTRA1-19-D-0007, and respectively Virginia Dept of Health Grant VDH-21-501-0141, VDH-21-501-0143, VDH-21-501-0147, VDH-21-501-0145, VDH-21-501-0146, VDH-21-501-0142, VDH-21-501-0148. AF, AMa, GL funded by SMIGE - Modelli statistici inferenziali per governare l'epidemia, FISR 2020-Covid-19 I Fase, FISR2020IP-00156, Codice Progetto: PRJ-0695. AM, BK, FD, FR, JK, JN, JZ, KN, MG, MR, MS, RB funded by Ministry of Science and Higher Education of Poland with grant 28/WFSN/2021 to the University of Warsaw. BRe, CPe, JLAz funded by Ministerio de Sanidad/ISCIII. BT, PG funded by PERISCOPE European H2020 project, contract number 101016233. CP, DL, EA, MC, SA funded by European Commission - Directorate-General for Communications Networks, Content and Technology through the contract LC-01485746, and Ministerio de Ciencia, Innovacion y Universidades and FEDER, with the project PGC2018-095456-B-I00. DE., MGu funded by Spanish Ministry of Health / REACT-UE (FEDER). DO, GF, IMi, LC funded by Laboratory Directed Research and Development program of Los Alamos National Laboratory (LANL) under project number 20200700ER. DS, ELR, GG, NGR, NW, YW funded by National Institutes of General Medical Sciences (R35GM119582; the content is solely the responsibility of the authors and does not necessarily represent the official views of NIGMS or the National Institutes of Health). FB, FP funded by InPresa, Lombardy Region, Italy. HG, KS funded by European Centre for Disease Prevention and Control. IV funded by Agencia de Qualitat i Avaluacio Sanitaries de Catalunya (AQuAS) through contract 2021-021OE. JDe, SMo, VP funded by Netzwerk Universitatsmedizin (NUM) project egePan (01KX2021). JPB, SH, TH funded by Federal Ministry of Education and Research (BMBF; grant 05M18SIA). KH, MSc, YKh funded by Project SaxoCOV, funded by the German Free State of Saxony. Presentation of data, model results and simulations also funded by the NFDI4Health Task Force COVID-19 (https://www.nfdi4health.de/task-force-covid-19-2) within the framework of a DFG-project (LO-342/17-1). LP, VE funded by Mathematical and Statistical modelling project (MUNI/A/1615/2020), Online platform for real-time monitoring, analysis and management of epidemic situations (MUNI/11/02202001/2020); VE also supported by RECETOX research infrastructure (Ministry of Education, Youth and Sports of the Czech Republic: LM2018121), the CETOCOEN EXCELLENCE (CZ.02.1.01/0.0/0.0/17-043/0009632), RECETOX RI project (CZ.02.1.01/0.0/0.0/16-013/0001761). NIB funded by Health Protection Research Unit (grant code NIHR200908). SAb, SF funded by Wellcome Trust (210758/Z/18/Z).

    1. Epidemiology and Global Health

    Affectionate touch and diurnal oxytocin levels: An ecological momentary assessment study

    Ekaterina Schneider, Dora Hopf ... Beate Ditzen
    Research Article

    Background:

    Affectionate touch, which is vital for mental and physical health, was restricted during the Covid-19 pandemic. This study investigated the association between momentary affectionate touch and subjective well-being, as well as salivary oxytocin and cortisol in everyday life during the pandemic.

    Methods:

    In the first step, we measured anxiety and depression symptoms, loneliness and attitudes toward social touch in a large cross-sectional online survey (N = 1050). From this sample, N = 247 participants completed ecological momentary assessments over 2 days with six daily assessments by answering smartphone-based questions on affectionate touch and momentary mental state, and providing concomitant saliva samples for cortisol and oxytocin assessment.

    Results:

    Multilevel models showed that on a within-person level, affectionate touch was associated with decreased self-reported anxiety, general burden, stress, and increased oxytocin levels. On a between-person level, affectionate touch was associated with decreased cortisol levels and higher happiness. Moreover, individuals with a positive attitude toward social touch experiencing loneliness reported more mental health problems.

    Conclusions:

    Our results suggest that affectionate touch is linked to higher endogenous oxytocin in times of pandemic and lockdown and might buffer stress on a subjective and hormonal level. These findings might have implications for preventing mental burden during social contact restrictions.

    Funding:

    The study was funded by the German Research Foundation, the German Psychological Society, and German Academic Exchange Service.

    1. Epidemiology and Global Health

    Perceived barriers to cervical cancer screening and motivators for at-home human papillomavirus self-sampling during the COVID-19 pandemic: Results from a telephone survey

    Susan L Parker, Ashish A Deshmukh ... Jane R Montealegre
    Research Article

    Background: Home-based self-sampling for human papillomavirus (HPV) testing may be an alternative for women not attending clinic-based cervical cancer screening.

    Methods: We assessed barriers to care and motivators to use at-home HPV self-sampling kits during the COVID-19 pandemic as part of a randomized controlled trial evaluating kit effectiveness. Participants were women aged 30-65 and under-screened for cervical cancer in a safety-net healthcare system. We conducted telephone surveys in English/Spanish among a subgroup of trial participants, assessed differences between groups, and determined statistical significance at p<0.05.

    Results: Over half of 233 survey participants reported that clinic-based screening (Pap) is uncomfortable (67.8%), embarrassing (52.4%), and discomfort seeing male providers (63.1%). The last two factors were significantly more prevalent among Spanish versus English speakers (66.4% vs. 30% (p=0.000) and 69.9 vs. 52.2% (p=0.006), respectively). Most women who completed the kit found Pap more embarrassing (69.3%), stressful (55.6%), and less convenient (55.6%) than the kit. The first factor was more prevalent among Spanish versus English speakers (79.6% vs. 53.38%, p=0.001) and among patients with elementary education or below.

    Conclusions: The COVID-19 pandemic influenced most (59.5%) to participate in the trial due to fear of COVID, difficulty making appointments, and ease of using kits. HPV self-sampling kits may reduce barriers among under-screened women in a safety-net system.

    Funding: This study is supported by a grant from the National Institute for Minority Health and Health Disparities (NIMHD, R01MD013715, PI: JR Montealegre).

    Clinical trial number: NCT03898167.