Clinical trials, models and simulation

summary

Data fusion is a problem with many facets, and being able to combine studies of different types or designs promises to revolutionize the estimation of causal effects. There are many methodological avenues of interest here, including generalized meta-analysis that uses data-driven approaches to determine how much information each dataset is providing. Possible applications - all aimed at enhancing power through data fusion - include combining data from a randomized trial with historical trials of the same treatment and/or relevant observational data, meta-analysing trials conducted in different geographical regions, and even synthesizing trial results with genetic associations from resources like the UK Biobank, where a variant can be used to proxy the pharmaceutical effect.

Another core aim of this theme will be to develop techniques for realistic simulation of clinical trials where we can specify the causal effect, which is a surprisingly difficult problem. The issue is that the causal effect is almost always marginal over covariates, but simulation is typically conducted in a conditional manner. Being able to do this in a way that mimics real data, at least in the covariate structure and missingness patterns, adds another layer of complexity. Applications would be to things like sample size calculation or to deciding which of a series of possible methods to apply to a given real dataset.

References

• Lin et al. Combining Experimental and Observational Data through a Power Likelihood. arXiv:2304.02339, 2023 to appear, Biometrics.

• Colnet et al. Causal inference methods for combining randomized trials and observational studies: a review. Statistical Science, 39(1), pp.165-191, 2024.

• Parikh et al. Validating causal inference methods. In ICML (pp. 17346-17358). PMLR, 2022.

• Evans and Didelez. Parameterizing and simulating from causal models. JRSS-B, 2024.