The 30th Bradford Hill Memorial Lecture

On 20 May at 5:30pm Professor Anne Mills, Deputy Director and Provost of LSHTM, introduced the 30th Bradford Hill Memorial Lecture, which is held annually in memory of Sir Austin Bradford Hill, described by Richard Doll as 'the greatest medical statistician of the 20th century'.

This year’s lecture, for a second time held online, was given by Professor Bianca De Stavola, Professor of Medical Statistics at UCL Great Ormond Street Institute of Child Health, on the topic of 'Life course epidemiology – from Bradford Hill’s viewpoints to counterfactual comparisons'. In a nice coincidence, Professor De Stavola recently won the RSS Bradford Hill Medal for Medical Statistics.

Professor De Stavola began her talk by mentioning the famous Bradford Hill 'criteria' for cause and effect relationships, drawing attention to his comment that none of the nine viewpoints is on its own sufficient to establish or refute a cause-and-effect hypothesis. Bradford Hill spoke from his extensive experience as an applied statistician, as did Professor De Stavola, throughout her lecture explaining current approaches to causal inference using examples from life course epidemiology.  

Life course epidemiology is important for understanding the effects of lifetime exposures on health outcomes, and for suggesting interventions to improve health – but it is complex and challenging.  For instance, there are many possible pathways via which an exposure in early life may affect health outcomes in adulthood.  The effects of time-varying, interconnected exposures on outcomes are difficult to understand, even more so when detailed data over time are not available, as is often the case. 

A conceptual model can help to make sense of the complexities, and counterfactual reasoning can help specify relevant estimands – specifically where there are multiple pathways of effects, estimands relating to overall or specific pathways, such as total effects and controlled direct effects, may help address questions of substantive interest.  This is in contrast to the traditional approach of interpreting the parameters of an outcome regression model – among other issues, the traditional approach cannot deal with time-varying confounding.  Instead, a counterfactual approach aims to answer the question “how would the world have been, had something been different?” The potential outcomes framework can be used to formalise questions that involve interventions on one or more exposures, and to define estimands as contrasts of expectations of potential outcomes in alternative hypothetical worlds.

Professor De Stavola used an example on the pathways between children’s size from birth and adolescent eating disorders in adolescence, using longitudinal data from ALSPAC. A question of concern may be when to take preventive measures. Using directed acyclic graphs, she told a story beginning with a very simple conceptual model, which she gradually developed to give a more realistic understanding of the role of size for eating disorders. At first she considered a total causal effect of BMI at age 12 on a standardised score measuring binge eating at age 13, and found a strong effect of about ¼ a standard deviation, and also some evidence of a direct effect of birthweight (ie not involving BMI at 12). However, this is unlikely to reveal the true picture – BMI at age 12 is the most proximal measurement, but is it too late to intervene?

She considered a hypothetical intervention on the distribution of BMI at age 12, shifting it to what it would have been if the birthweight distribution had been shifted by specific amount, while holding birthweight itself constant, thus defining a type of indirect effect that could be identified even in the presence of time-varying confounding.

The estimate of the indirect effect indicated that a about half the total effect of birthweight was via BMI at age 12. But most life-course investigations involve multiple linked and time-varying exposures. She went on to include further measurements of children’s BMI, from age 7 to 12, combined as a joint mediator. Now the indirect effect, shifting the joint distribution of all the BMIs, accounted for a larger proportion of the total birthweight effect. However, the exposure of real interest is children’s growth, and their BMI are measurements, with error, of their latent growth processes. On modelling the latent growth process from age 7-12 as a mediator, it is apparent that latent size accounts for almost all the indirect effect, with rate of growth having almost no role; this clarifies the previous findings about the role of BMI at specific ages. Indeed, not being able to account for measurement error may lead to not being able to identify what really matters.

As in any observational study, unmeasured confounding is a concern in life course epidemiology. Sibling designs and Mendelian Randomisation can sometimes help, but caution is necessary – for instance, a key assumption in Mendelian Randomisation is likely to be violated in a setting where there are pathways between multiple exposures over time.
To draw her lecture to a close, Professor De Stavola returned to the Bradford Hill viewpoints and compared each of them with its current equivalent. There is a logical continuum between then and now, with more focus now on specifying the causal effect we are targeting. She concluded that while statistical methods and data have changed, the spirit and pragmatic approach to research have not.
Professor Nicholas Jewell of LSHTM thanked Professor De Stavola for a wonderful and insightful lecture, drawing attention to her academic achievements, her enormous role in promoting women in STEM, and her daily efforts in mentoring junior statisticians. The vote of thanks was resoundingly supported by the online listeners. 

A recording of the lecture is available on the LSHTM website, at

The speaker's slides are available here.

Report written by Nicola Fitz-Simon on behalf of the RSS Medical Section

Load more