Research

Research interests

Introduction

As an educational data scientist with an interest in data science education, I am primarily interested in data science for causal inference, evidence generation, synthesis, and communication. In particular, I am interested in counterfactual prediction using causal machine learning techniques, the reproducibility of research findings, and if and how the results researchers reported are sensitive to analytical choices.

I collaborate with colleagues from multiple disciplines and institutions in order to solve some methodological challenges associated with social and educational interventions, for instance, the unintended consequences of randomisation in educational interventions as a result of pupils being randomly allocated to a treatment arm that might be sub-optimal, if not harmful, to them.

Tapping into recent development in machine learning, I have been working on predictive and personalised approaches to evidence, particularly, individualised treatment effect using experimental, observational, longitudinal, and simulated data. Understanding that what we observe right in front of our eyes is often ethnographically invisible, I employ mixed methods in my research and draw on inspirations from medical anthropology.

Policy Implications

A few months ago, when I presented our research on individualised evidence for education, some colleagues might understandably question its value or relevance to society. Today, the current COVID-19 global pandemic, at the time of writing, has led more than 150 countries to close their schools, affecting more than 1.5 billion children worldwide (source: The World Bank). Learning has never been so personalised, and individualised evidence has never been so important!

By showing where an intervention worked, for whom, and by how much, our research findings have implications for policy-makers in diverse areas of science and policy, such as education, public health, tax auditing, and medical trials. Beyond education, one area of particular relevance of our individualised approach is public health where policies that encourage certain behaviours for the public good are critical. Take, for example, the COVID-19 crisis: encouraging social distancing may take different forms and policy-makers would be well-advised to understand how different subgroups (e.g., the elderly, young people, and key workers) might respond to different messaging.

Our approach could go even further, in areas of rapid development and testing of drugs or vaccines, in order to better understand who benefits and by how much, even a vaccine that only works for a relevant subgroup (e.g., patients with underlying health conditions) would be a much-needed advancement to battle this deadly disease.

Nevertheless, we should not ignore that fact that a third of the world’s population relies heavily on traditional medicine and that folklores played significant roles in the advancement of science and medicine, e.g., cow fleas to eradicate smallpox, mosquitos that caused human blood to boil, and iodine against goitre. We survived and thrived by drawing on inspirations from nature and tales from the outposts. We can, again, today!

Research projects

Alan Turing Institute COVID-19 Rapid Response Data Science Taskforce (under review)

Team Members

Dr ZhiMin Xiao, Data Science Team Leader, 80% FTE

Dr Mohammad Golbabaee, Deputy Data Science Leader, 20% FTE

Dr Oliver Stoner, Deputy Data Science Leader, 50% FTE

Dr Christopher Boyle, Psychologist, 60% FTE

Dr Angela Cassidy, Expert in Ethics & History of Science, 40% FTE

Dr Zeliang Wang, Research Software Engineering Data Scientist Team Member, 40% FTE

Dr Lei Wang, Research Software Engineering, 20% FTE

Mr Ben Jones, Data Scientist Team Member, 20% FTE

Mr Charlie Kirkwood, Data Scientist Team Member, 20% FTE

Advisory Board Members

Dr Oliver Hauser

Professor Steve Higgins

Professor Mihaela van der Schaar

Brief Summary

We solve real-world problems using real-world data. Here, we focus on “what if” questions at individual patient level, namely, causal inference via causal machine learning to generate evidence that is actionable and individualised in Wave 1 (e.g. we can look at the effects of various drugs on individual patients). We then refine the methods in Wave 2 through performance validation using statistical and non-statistical theories, folklores in medical history and society, and simulated and empirical (ideally longitudinal and time series) data.

Before COVID-19, we had years of experience in detecting and estimating causal treatment effects. Some of us worked together, or are about to work together (e.g., ZhiMin, Oliver Stoner, van der Schaar just won an Exeter-Turing grant to estimate and validate individualised treatment effects). We treat causal inference as a missing data problem, and estimate causal effects, either aggregated or individualised, under the potential outcomes framework, meaning our factual and counterfactual predictions are beyond feature selection and input-to-output mapping.

In sum, we help decision-makers understand what will happen if and when they do something, but we also imagine what the outcomes would have been had they taken different courses of action (e.g., when to treat and/or discharge patients). Since we cannot validate counterfactual predictions using standard machine learning techniques, we will adopt the influence function approach without the need to access counterfactuals, and draw upon experiential knowledge of COVID-19, beliefs, and imaginaries of pandemics. These together can help bridge the gaps between what is real and what is unobservable.

Research grants

• 2020 Exeter University
  Machine learning to identify who benefits in policy trials, together with Professor Mihaela van der Schaar at the University of Cambridge and Dr Oliver Stoner of IDSAI, the University of Exeter. (approximately £10,000)
• 2019 Exeter University
  SSIS strategic discretionary research fund to estimate individualised treatment effect for randomised controlled trials using conventional regression models and machine learning algorithms. (£3,500)
• 2019 Exeter University
  With Katherin Barg and Lauren Stentiford: A scoping project to automatically extract data from recorded lectures at the University of Exeter. (£2,660)
• 2019 HEFCE
  Data Science Education. (£30,000)
• 2019 Exeter University
  Data and Science with R. (£2,000)
• 2019 Exeter University
  With Alexandra Allan, Karen Knapp, Susan McAnulla, and Matthew Newcombe: Education Incubator Project on flexible blended learning for mature students in the Graduate School of Education and Exeter Medical School. (£9,638)
• 2019 National Institute for Health Research
  With Obioha Ukoumunne and Michael Nunns from the Medical School: Cluster randomised trials and child psychiatric epidemiology studentship for three years at UK/EU level. (approximately £58,227)

Links

• OSF Preprints
• Google Scholar
• Twitter

Publications

Key publications | Publications by category | Publications by year

Refresh publications

External Engagement and Impact

Committee/panel activities

Emerging Applications Section of the Royal Statistical Society: Committee member.

Invited lectures

A tale of three gaps: Individualised treatment effect for better evidence in education. Jesus College & Faculty of Educaton, University of Cambridge, March 12th, 2020.

Uncovering Individualised Treatment Effect: Evidence from Educational Trials, Department of Computer Science, University of Bath, March 3rd, 2020.

Latest working paper presented at the Health Statistics Group, College of Medicine and Health, University of Exeter, January 7th, 2020.

The case against perfection in the mean: Why it is time for an individualised approach to evidence for education (Q-Step Centre, University of Exeter), November 8th, 2019.

The Case Against Perfection in the Mean: From Average to Individualised Treatment Effect in Randomised Controlled Trials for Education (Royal Statistical Society), September 23rd, 2019.

Pupil Advantage Index as an Alternative to Subgroup Analysis in RCTs for Education (Royal Statistical Society South West Local Group), February 13th, 2019.

Journal and book series Editorships and Editorial board membership

Associate Editor for the Digital Impacts Section of Frontiers in Human Dynamics

Teaching

I contribute to teaching and supervision of the following modules:

2019 - 2020:

(EFPM308) Preparing for Educational Research and Dissertation.

(EFPM913) Debating the Big Questions in Education.

(SSI2001) Learning from Work Experience (Undergraduate).

(EFPM005Z1) Preparing for Educational Enquiry.

2018 - 2019:

(EFPM308) Preparing for Educational Research and Dissertation.

(ERPM005) Designing and Communicating Research.

(SSI2001) Learning from Work Experience (Undergraduate).

(EFPM329a) Preparing for TESOL Enquiry and Dissertation.

(ERPM100) MSc Programme Dissertation Supervision.

Data and Science with R.

Supervision / Group

Postgraduate researchers