When a government introduces a sugar tax or a city rolls out a needle-exchange programme, statisticians currently lack reliable tools to tell whether those interventions actually worked. This project builds new statistical methods to solve that problem. The core challenge is that public health evaluations rarely use clean, controlled experiments. Researchers must instead analyse messy time-series data from hospitals or regions, where some places get the intervention and others do not. Existing methods fail to handle four real-world complications: effects that vary wildly between different populations, hidden factors that shift over time, spillover effects where a policy in one area influences its neighbours, and situations where control groups disappear or policies get reversed. If successful, this research will give policymakers trustworthy, quantified estimates of whether an intervention caused a real change. The methods will be applied to three pressing UK cases: efforts to eliminate Hepatitis C among people who inject drugs by 2030, the impact of COVID-19 non-pharmaceutical interventions, and the effects of the Soft Drinks Industry Levy on sugary drink sales. The project will also produce free, easy-to-use software so that public health practitioners can apply these tools without needing a statistics degree.
View original technical description
Researchers and practitioners in the field of public health are routinely faced with the task of evaluating the effect of an intervention on an outcome of interest. Often, these evaluations rely on observational time-series data from a small number of units of intervention, such as hospitals or geographical regions, of which some receive the intervention (treated units) and some do not (control units). Increasing availability of such data has led to a pressing need for statistical methodology that can be used to draw causal conclusions in this context, whilst accounting for the problem of counfounding present in observational studies. Despite recent developments, current methods cannot accommodate important facets of an intervention that are typically of interest in a public health-related context. Firstly, the strong dependence of intervention effects on the characteristics of a unit, which leads to great effect heterogeneity. Secondly, the possibility that unobserved confounders change over time and that complex interactions exist between the observed confounders. Thirdly, the presence of correlations between neighbouring units, which, while could be accounted for in models to improve statistical power, also raise the question of how to deal with interference, i.e. the fact that the intervention will also affect the units surrounding those that have been treated. Fourthly, violation of the requirement that control units exist throughout the study, and that interventions cannot be withdrawn once delivered. I propose to develop statistical methodology that tackles these challenges, thus offering generic and much needed tools to assess public health interventions. These Bayesian methods will provide uncertainty quantification to aid policy making, and include easy-to-use software to facilitate future use. I will apply my methods to studies from the UK, arising from three key substantive areas: the evaluation of ongoing efforts to reduce Hepatitis C virus prevalence among people who inject drugs, which are part of UK's plan to eliminate the virus by 2030; the impact that various non-pharmaceutical interventions had in containing the spread of COVID-19 during the recent pandemic; and the effects of the Soft Drinks Industry Levy on drinks containing added sugar on improving health-related outcomes such as sales of products containing added sugar.
Plain English summaries and category classifications on this site are generated by AI and may not perfectly reflect the original research.
Is something wrong? Let us know