On Wednesday 8 September 2021, the International Development Section and Medical Sections of the RSS held a session titled 'Addressing statistical and epidemiological issues in global health'.
Emily Webb, deputy director of the International Statistics and Epidemiology Group (ISEG) of the London School of Hygiene and Tropical Medicine (LSHTM), opened the session by briefly describing the work of the ISEG which was set up nearly 50 years ago. Their research focuses on epidemiology and control of major public health problems of low and middle income countries (LMICs). Group members are based in the UK and in several countries outside the UK, and work closely with researchers in many others.
A key component of the work of ISEG is capacity strengthening – it is one of their stated objectives to build a community of skilled statisticians and epidemiologists in LMICs. As part of this they offer fellowships to statisticians and epidemiologists from sub-Saharan Africa, something which is very popular and is being scaled up over the next five years. They also collaborate with a number of research institutions in LMICs to build capacity.
Emily concluded by flagging up the ISEG 50th Anniversary symposium which will be held virtually in June 2022.
The second speaker was Schadrac Agbla from the University of Liverpool, a former recipient of an ISEG fellowship. His topic was the important issue of non-adherence in randomised control trials, in other words how the outcome of those trials might be affected by individuals receiving the wrong treatment, or no treatment at all. Some indeed may have been expecting one treatment and in fact got the other, which can lead to confounding in certain circumstances. He added a practical dimension to the debate by bringing in cluster randomised trials (CRT) – for example a hospital of patients where the hospital is the unit of randomisation.
Statistics offers a number of methods for dealing with this, and Schadrac wanted to compare them. He conducted 2500 CRT simulations of 1000 individuals in 50 clusters of 20, and tested out the Wald estimator, a two-stage least square estimation with both the Hubert-White-Rogers and Moulton’s corrections, and a Bayesian multi-level mixed (BMM) model.
The results showed that the first three models behaved in a very similar way, but the BMM model had the best performance particularly when non-adherence was at the individual level and the intracluster correlation coefficient for the outcome was large, providing estimates with small to negligible bias and coverage close to nominal level.
The third speaker was Paul Mee from LSHTM. Paul had been working on Dengue with the Centre for Arbovirus Discovery, Diagnostics, Genomics and Epidemiology (CADDE) in Brazil when Covid-19 struck, and the research was quickly reconfigured to look into the more pressing question. The aim was to create additional resources to support those responding to Covid-19. The available dashboards, while good at national and broad sub-national level, did not take into account the dynamics of the epidemic sweeping Brazil, which affected municipalities in quite different ways, linked to their age structures, levels of poverty and over-crowding, and other risk factors.
Using data from Brazil.io, a site which makes publically-available data more accessible, the team developed a data processing and analysis pipeline and a data visualisation app for the more than 4,000 municipalities in Brazil. This provides, on a daily basis, visual summaries, and associations with key covariates, makes estimates of R, and allows the visualisations to be run backwards and forwards over time.
The app clearly has applications beyond Covid and work continues.
Neal Alexander from LSHTM spoke fourth, on spatial analysis of cluster-randomised trails. The issue here is the effect that proximity can have on behaviour. Many models assume that there is an independence between clusters, and this has been shown not always to be the case. For example, the vaccine status of your neighbours can spill over into behaviour and immunity in your own cluster, and spacing the clusters to avoid this may not be feasible.
So the work looked at two things: how to measure the degree to which a cluster is surrounded; and how to achieve a simple marginal interpretation of the parameters while avoiding “spatial confounding”, in other words the effect proximity has on estimates of the fixed effects, which includes inflating their variance.
After explaining some of the theory, Neal described an application involving the vectors of dengue when controlled with insecticide-treated curtains and water tank covers. There was a clear proximity effect on the results of the trials – they were there, and they were significant.
This was a thought-provoking session and demonstrated the innovative work that ISEG is doing on statistics and epidemiology in global health issues.
Author
Phil Crook is meetings secretary of the RSS International Development Section and a consultant development statistician. He previously worked for the Department for International Development as a statistics adviser.