The Centre for Multilevel Modelling has a long history, first at the Institute of Education where it emerged from Professor Harvey Goldstein’s research team in the 1990s and for the past 20 years at the School of Education, University of Bristol. The Centre’s research looks into the use of statistical modelling to answer real world problems using data that have complicated dependency structures. Our research is wide-ranging from the development of new statistical methodology and supporting this with user friendly software to the application of this methodology to a whole range of disciplines. We have been fortunate to have had large amounts of funding over the years through the ESRC and particularly it’s NCRM programme. One important strand that has run through our research from the start has been the importance of a training element so that our methods gain maximum usage in the fields that require them. This has led to the research leading 2 related REF (results2021.ref.ac.uk/impact/7d9c92b8-8627-4c7d-9324-2625901b706f?page=1 and results2021.ref.ac.uk/impact/d710607c-2c84-4dfd-9d24-9eb3622cfedf?page=1 ) in the education and maths UoAs respectively.
In this talk I will describe some of the history of the centre and our software development work. I will talk about how this has changed in the 25+ years that I have been associated with the centre as fitting multilevel models has moved from being only available through specialist software to now being widely used in most disciplines. How this has impacted on our software development and training over the period. I will also talk about later work on automating statistical analysis and training material generation (funded by the ESRC and British Academy) and contrast this with AI and Chat-GPT that has probably largely superseded it.
Much use of multilevel models is with large secondary datasets and so if time allows I will talk briefly about the challenges of designing primary data studies that exhibit multilevel structures and how for example one can calculate required sample sizes for such studies via simulation.
