The London School of Hygiene and Tropical Medicine (LSHTM), with the support of the RSS Medical Statistics Section, hosted the 28th Bradford Hill Memorial Lecture: Statistics as Alchemy: Turning Data into Gold, on the 11 April 2019. The event was part of the LSHTM MSc Medical Statistics 50th Anniversary.
Professor Michael Hughes’ lecture was entitled “Countering the HIV epidemic: statistical issues in treatment evaluation and policy.” Professor Hughes is Professor of Biostatistics at the Harvard TH Chan School of Public Health where he is also Director of the Center for Biostatistics in AIDS Research.
Professor Stuart Pocock, who is Professor of Medical Statistics at LSHTM, started by praising Austin (Tony) Bradford Hill for his impact on medical statistics, including in RCTs, his seminal text book, and lung cancer and smoking studies, before introducing Professor Hughes.
Professor Hughes began by referring to his time at LSHTM some 36 years ago in 1983, where he did the MSc followed by a PhD, before showing a plot of the number of HIV infections over time from 1987 to 2016 and asking why there was a marked improvement. To explore this, his talk composed of three main threads.
The first was to talk about the introduction of surrogates for accelerated endpoints. The US Food and Drug Administration (FDA) approval process in the late ‘80s to early ‘90s was concerned with a single class of drugs but from the mid-90s onward this changed to three drug treatment. Up to the early ‘90s drug approval was slow but then in 1992 the FDA allowed the use of surrogate endpoints. In response there was a major push in how to validate these. This was done, Professor Hughes explained, by investigating prognostic markers (viral load and CD4 counts) which focused on how a surrogate can predict the clinical endpoint and, more difficult to do, how the difference between randomised treatments in a surrogate endpoint is associated with the corresponding difference in clinical outcome, when evaluated across randomised trials in a meta-analysis.
Secondly, Professor Hughes considered the question of when to start treatment and highlighted work in this area which applied methods from causal inference in observational studies. The particular question was what CD4 threshold to use to start treatment. Earlier treatment could lead to better outcomes but increase adverse effects due to treatment. This was a controversial topic and split expert opinion as whether to use a <500 CD4 count or treat all patients irrespective of CD4 count. An NA-ACCORD study in the New England Journal of Medicine applied causal inference methods to this problem and was possibly the first time many in this field were introduced to these sorts of approaches in the context of defining treatment guidelines.
Ultimately, Professor Hughes explained, in January 2016 the guidance changed dramatically to treatment of all patients, due to large RCTs where people with CD4 counts greater than 500 were randomised between starting treatment when their CD4 count declined below 350 versus starting treatment immediately. Professor Hughes described recent important work by a collaboration of investigators from one of the RCTs and one of the observational study analyses that aimed to identify sources of differences in results between the two study approaches.
Finally, Professor Hughes provided a critique of some of the practical issues in evaluating new treatments in an era of non-inferiority trials. The FDA uses what are called 'snapshot' outcome measures which are composite outcomes but include a sizeable proportion of non-evaluable study participants (due, for example, to missing data and loss to follow-up) as non-responders. The push by the International Conference for Harmonisation to better identify estimands in clinical trials suggests that a different outcome measure may need defining. There is also a 'drift' in the population enrolled in trials of initial HIV treatment over time such that enrolled individuals are getting healthier and potentially easier to treat.
The proportion with high viral loads is going down, from approximately 50% to 20%, and so this will need to be considered in subgroup analyses that assess whether differences between randomised treatments appear non-inferior over a broad range of viral loads. Then Professor Hughes talked about applying methods from precision medicine to determine individually tailored decisions but also asked if a broad public health recommendation is satisfactory instead, which would obviously be preferable from a guidelines perspective particularly in resource limited settings.
After questions, Professor Stephen Evans, professor of pharmacoepidemiology at LSHTM, thanked the speaker and, in the spirit of a good study acronym, praised Professor Hughes’ MERIT (methods, educational work, RCTs, broad interests including with observational data and translating research), adding that his lecture went beyond this and showed real distinction.