Lung function and cognitive ability in children: a UK birth cohort study

Background Decreased adult lung function is associated with subsequent impairment in cognition. A similar relationship in early life could be of great policy importance, since childhood cognitive ability determines key adult outcomes, including socioeconomic status and mortality. We aimed to expand the very limited data available on this relationship in children, and hypothesised that reduced lung function would be longitudinally associated with decreased cognitive ability. Methods Lung function was measured at age 8 (forced expiratory volume in one second (FEV1), forced vital capacity (FVC); % predicted), and cognitive ability was measured at ages 8 (Wechsler Intelligence Scale for Children, third edition) and 15 (Wechsler Abbreviated Scale of Intelligence), in the Avon Longitudinal Study of Parents and Children. Potential confounders were identified as preterm birth, birth weight, breastfeeding duration, prenatal maternal smoking, childhood environmental tobacco smoke exposure, socioeconomic status and prenatal/childhood air pollution exposure. Univariable and multivariable linear models (n range=2332–6672) were fitted to assess the cross-sectional and longitudinal associations of lung function with cognitive ability, and change in cognitive ability between ages 8 and 15. Results In univariate analyses, both FEV1 and FVC at age 8 were associated with cognitive ability at both ages, but after adjustment, only FVC was associated with full-scale IQ (FSIQ) at ages 8 (β=0.09 (95% CI 0.05 to 0.12; p<0.001)) and 15 (β=0.06 (0.03 to 0.10; p=0.001)). We did not find evidence of an association between either lung function parameter and interval change in standardised FSIQ. Discussion Reduced FVC, but not FEV1, is independently associated with decreased cognitive ability in children. This low-magnitude association attenuates between ages 8 and 15, while no association is evident with longitudinal change in cognitive ability. Our results support a link between FVC and cognition across the life course, possibly due to shared genetic or environmental risk, rather than causation.

To facilitate estimation of the likelihood of our results being subject to selection bias due to missing data, following a method outlined by Cornish et al, 2 we fitted logistic regression models with the outcome of being a complete case at age 8 and 15. These models included as covariates all variables in our main multivariable linear models.

Supplementary results
Supplementary table 1 shows the results of multivariable models including sex, age and height. Results changed very little with their addition, with identical point estimates for the effect of FEV1 and FVC at age 8 on FSIQ at ages 8 and 15, with some minor changes to confidence intervals and p values. Point estimates for the effect of lung function parameters on interval change in standardised FSIQ differed slightly, but remained very close to zero, with very large p values. interaction with the effect of FEV1 on subscale scores at age 8, we also fitted separate models for males and females with these outcomes, in order to determine whether any particular facet of cognitive ability could explain this observation. For simplicity, we have not displayed these here, but the pattern was similar, with strong evidence for an association between FEV1 at age 8 and all subscale scores in females, but not in males.
No specific subscale score appeared to explain the association between FEV1 and FSIQ in females at age 8 more than others. Given this isolated gender-specific cross-sectional association of FEV1 with FSIQ, we ran models to include asthma diagnosis and childhood wheezing phenotype as covariates in females only, to see if these variables might explain it. Results changed very little, with β=0.09 (0.03, 0.16; p=0.006), indicating that these variables did not explain the association. Supplementary table 4 shows the results of the univariable linear regression models for all participants with data for exposures and FSIQ outcomes, but not necessarily covariates. In this model, estimates for the associations of FEV1 and FVC with FSIQ at ages 8 and 15 were slightly higher than in the main univariate analysis, which was restricted to participants with complete data (e.g. for FVC and FSIQ at 8 years β = 0.13 [0.10, 0.16; p<0.001], compared to 0.11 [0.07, 0.15; p<0.001] in the main analysis). As in the main analysis, there was little to no evidence of an association of either FEV1 or FVC with change in standardised FSIQ between 8 and 15 years of age.

Supplementary discussion
Inclusion of age, height and sex in our multivariable models resulted in near identical results for the estimates of the association of FEV1 and FVC at age 8 with FSIQ at age 8 and 15, and very minimal change in the estimates of the association with interval change in standardised FSIQ. These results indicate that these variables had indeed been adequately controlled for in the original analysis.
When we investigated for interaction of the association between exposure and outcome with sex in the multivariable regression models, we found that while there was strong evidence of a cross-sectional association between FEV1 and FSIQ at age 8 in females, in males there was none. This sex-specific association was not explained by asthma diagnosis or childhood wheezing phenotype. The association was not explained by any specific facet of cognitive ability, as measured by the WISC-III subscale scores. It is difficult to explain why decreased FEV1 should be cross-sectionally associated with decreased FSIQ in girls, but not boys, at age 8. FEV1 is an effort independent measurement, meaning it is very unlikely that this association can be explained by differences between males and females in the ability to perform the procedure correctly. The association was not explained by asthma diagnosis or wheezing phenotype, despite the fact that these were associated with reduced FSIQ in separate univariate modelling. Both asthma and pathological wheezing phenotypes were more prevalent in males than females (eg 17.6% of boys at age 8 had a diagnosis of asthma, versus 12.7% of girls), so if this were the explanation for the sex-specific association it would be expected that the association would be stronger in boys. The sex-specific association for FEV1 with FSIQ did not remain evident at age 15. Overall, the significance of this finding is unclear.
Our univariable models including all participants with data for exposure and outcome, but not necessarily covariates, had slightly larger coefficients than the main univariable analysis, and narrower confidence intervals. Although a larger sample size confers greater precision, it is likely that these effect estimates are inflated by confounding, for which it is not possible to control in the additional individuals included, due to a lack of covariate data. On the other hand, the fact that the coefficients are larger in this more unselected group, than in the main univariable analysis, may also support the notion that any selection bias, from analysing an increasingly select and affluent subsample, is likely to be towards the null. This is discussed further below.
Estimated coefficients for the association of FVC with FSIQ were only very slightly attenuated by inclusion of asthma and wheezing phenotype as covariates. This indicates that these diagnoses do not account for much of the variance in IQ explained by FVC. The notion that asthma and wheezing are important in explaining the relationship between FVC and FSIQ is also contested by a loss of precision, as demonstrated by the widening of confidence intervals in the estimates from models including these variables (although this might also be due to a smaller sample size). Another explanation for the attenuation of the coefficients in these models might be the aforementioned effect of selection bias towards the null, due to a more select subsample. The evidence from our results against asthma and wheezing being important explanators of the relationship between FVC and FSIQ is unsurprising, because FEV1 is a better measure of the airflow obstruction which characterises asthma and bronchospasm.
It is useful to highlight the clinical differences between FEV1 and FVC to potentiate the formulation of hypotheses as to why we have detected an effect on childhood cognition from the latter but not the former.
West conceptualises the lungs and thorax as an air pump, the function of which is dependent on stroke volume, resistance of the airways, and the force applied to the piston. 3 The latter is unimportant in forced expiration, due to dynamic compression of the airways, which means that flow is independent of effort. 3 FEV1 is the metric of airways resistance, which is affected by pathological processes which cause their narrowing or premature collapse. 3 The most common of these are asthma, which causes reversible bronchoconstriction due to hyperresponsiveness and inflammation of the airways, and chronic obstructive pulmonary disease (COPD), which causes airflow limitation due to chronically inflamed and narrowed airways (chronic bronchitis), and premature collapse due to the loss of radial traction caused by destruction of lung parenchyma distal to the terminal bronchiole (emphysema). 3 FVC is a measure of the 'stroke volume' of the air pump, which is determined by the capacity for the thoracic cage, respiratory muscles, pleura and lung parenchyma to fill the lungs with air, and then expel it, through alterations in intrathoracic pressure. 3 This can be negatively impacted by diseases of the thoracic cage, such as kyphoscoliosis, neuromuscular disorders such as muscular dystrophy, diseases of the lung parenchyma such as pulmonary and cystic fibrosis, and diseases of the pleura such as effusion or pneumothorax. A ratio of the FEV1 to FVC is often used to distinguish 'obstructive' from 'restrictive' patterns of spirometry, with the former showing a ratio of less than 0.8 due to airflow limitation (FVC may be normal or reduced, while total lung capacity measured by helium dilution may be increased), and the latter showing a normal or increased ratio, but in the presence of reduced FVC and total lung capacity. 3 We have noted an effect on childhood cognition from FVC, but not FEV1. However, as discussed elsewhere, our results do not appear to support causality. The most plausible explanation for the association may therefore be genetic and/or early developmental vulnerabilities shared between lung function and cognitive ability. Accordingly, it seems reasonable to suggest that the reason for FVC, but not FEV1, being associated with cognitive ability is that it may be a better correlate of respiratory and overall health in early life. As Another consideration is the increasing problem of childhood obesity, 4 because morbid obesity can cause a restrictive lung function defect with reduced FVC. Elevated body mass index (BMI) has been associated with reduced IQ, 5 although it is not clear if this is a causal relationship. BMI was not controlled for in the multivariable analysis, so it is possible that part of the observed association of FVC with FSIQ is explained by obesity. Estimates of the heritability of pulmonary function traits vary widely, but FVC has generally been assessed to be slightly more heritable than FEV1 or FEV1/FVC. 6 Another explanation for the observed association, for FVC but not FEV1, is that FVC might share more genetic architecture in common with cognitive ability than FEV1, with its slightly higher heritability potentially making this more likely.
Selection bias, due to study attrition and missing data, is an important source of bias in statistical estimates of relationships between variables derived from cohort studies. We conducted a complete-case analysis, which for an unbiased effect estimate, requires that data is either missing completely at random (MCAR i.e. no systematic missingness which could induce bias), or missing at random (MAR), i.e. missing dependent on the outcome or exposure only via measured covariates. Given the social patterning of missingness in the ALSPAC dataset demonstrated in table 1 and discussed in the main text, it is safe to assume that the MCAR assumption is not satisfied. We have controlled for SES, but this was imperfectly measured by two proxy variables (maternal education and housing tenure). It is probable that a failure to control for the unmeasured component of SES, as well as for other unmeasured factors related to both outcome and selection, mean that the MAR assumption is also unsatisfied. Cornish et al. showed that FSIQ at age 15 in ALSPAC is likely to be missing not at random (MNAR) i.e. missing dependent on its own (unknown) value, with those with higher IQs more likely to remain in the sample. 2 These findings are corroborated by the results of the logistic regression models we fitted using their method, to identify determinants of selection, which showed that not only is low maternal education negatively associated with inclusion, but that the odds of selection at each age are higher with increasing FSIQ scores, after adjustment for all other covariates. While it is not possible to demonstrate conclusively from the original dataset, these findings are strongly supportive of our outcomes being MNAR, a source of selection bias for which it is not possible to fully correct. According to the analysis of Cornish et al, 2 selection bias might have been reduced, and precision increased, by the use of educational data to impute missing values. Due to time constraints relating to thesis submission deadlines, we did not apply for educational records via data linkage for the purpose of multiple imputation (MI), meaning that there may be greater bias in our estimate than if we had performed MI. However, due to the likelihood of the outcome being MNAR, bias would not have been completely eliminated by MI, and it seems reasonable to assume that any selection bias affecting our estimates would be towards the null. This is because if proportionately more participants with comparatively lower IQ and lung function are lost to follow-up (which seems plausible due to the socially patterned nature of missingness), without completely adjusting for the cause of their deselection, then the observed association would be weakened. There is some evidence for the assertion that selection bias due to missing data is toward the null, from the fact that coefficients from univariable analysis limited to participants with complete data were of smaller magnitude than those from analyses including the less selected group of participants with exposure and outcome, but not necessarily covariate, data.