iiinova — where your digital self keeps evolving

Stephen Burgess @stevesphd.bsky.social · 14/09/2026

Video replay officials in football (soccer) have to perform counterfactual reasoning when giving an offside decision. If the offside player [Fernandez] had not attempted to play the ball, would the defending player [Martinez] have made a different decision? from www.bbc.co.uk/sport/footba...

010

Stephen Burgess @stevesphd.bsky.social · 13/08/2026

In doing this, we may get a more valid estimate of a less relevant quantity, rather than a less valid estimate of a more relevant quantity. In many cases, the estimand is a good servant, but a poor master.

000

Stephen Burgess @stevesphd.bsky.social · 13/08/2026

We encourage a broader framing of target trial emulation: alongside “estimand-first” approaches, we should consider “design-first” or “assumption-first” approaches - where we start by asking "What quantity can we estimate reliably?".

100

Stephen Burgess @stevesphd.bsky.social · 13/08/2026

A natural experiment gives an estimate - but the estimate is often a means to an end, not the end in itself. It provides supportive evidence, rather than answering the question of interest. But we would often rather more reliable estimate, even if the estimand is less relevant.

100

Stephen Burgess @stevesphd.bsky.social · 13/08/2026

We argue that epidemiological studies can be characterized into early-phase and late-phase in the same way. A natural experiment may give an estimate that is not policy-relevant or translatable to applied practice. But it is still a relevant source of causal information.

100

Stephen Burgess @stevesphd.bsky.social · 13/08/2026

In drug development, the goal of a Phase 3 study is to estimate the impact of a well-defined intervention. Whereas the goal of a Phase 2 study is to provide proof-of-concept evidence supporting (or refuting) a given mechanism as a plausible target for intervention.

100

Stephen Burgess @stevesphd.bsky.social · 13/08/2026

Estimands are great! But often the quantity we want to estimate differs from what we are able to estimate reliably. The target trial framework is often presented with Step 1: "What quantity do you want to estimate?" - we characterize this as an "estimand-first approach".

100

Stephen Burgess @stevesphd.bsky.social · 28/07/2026

Many of the examples relate to dentistry, but the principles are universal: the credibility of a causal claim is determined not primarily by the statistical method, but by whether the underlying assumptions are considered plausible within the context of the scientific question.

011

Stephen Burgess @stevesphd.bsky.social · 28/07/2026

I think he agrees with me now that this was a touch ambitious, but the result is 19 pages that spans from asking questions to parameter identification (covering adjustment-based and design-based approaches) to analysis and beyond.

100

Stephen Burgess @stevesphd.bsky.social · 28/07/2026

I discouraged Nasir from writing this paper - I said there is a good reason why no-one in their right mind would write a practical overview on causal inference in a single paper. The topic is too large. He disagreed and wrote the paper anyway.

120

Stephen Burgess @stevesphd.bsky.social · 13/07/2026

Several limitations as ever, but great to see supportive evidence from human genetics for a biologically-plausible mechanism. Learnings for me about heterogeneity in oral cancers: different parts of the oral cavity get cancer via distinct mechanisms. Comments welcome!

020

Stephen Burgess @stevesphd.bsky.social · 13/07/2026

We tested whether these variants associated with different head and neck cancers. We saw positive associations for all cancer types, with significant estimates for oral cavity squamous cell carcinoma (OCSCC) and HPV-positive oropharynx squamous cell carcinoma (OPSCC).

100

Stephen Burgess @stevesphd.bsky.social · 13/07/2026

We consider two variants in gene regions with biological relevance to manganese levels, that associate with circulating manganese levels in blood, and associate with serum aspartate aminotransferase (AST), a biomarker known to increase under hepatocellular manganese overload.

100

Stephen Burgess @stevesphd.bsky.social · 13/07/2026

We wanted to know if there is evidence in a Mendelian randomization paradigm for an effect of manganese on risk of head and neck cancers. There is prior motivation for this hypothesis, as these cancers are disproportionately common in welders, who have exposure to manganese.

100

Stephen Burgess @stevesphd.bsky.social · 13/07/2026

This paper represents a paradigm I'd like to strive for: we spent 90% of this project working out what was the question of interest, the most appropriate genetic variants, relevant datasets, sensible validation analyses - and 10% of the time doing the analyses, which were relatively simple.

100

Stephen Burgess @stevesphd.bsky.social · 30/06/2026

Thanks to Linxuan Zhang for leading the work, all those at Lilly for their valuable input (particularly Zae, Yushi, and Pallav), and Dipender Gill at Sequoia for bringing everything together. Comments and questions are always welcome!

000

Stephen Burgess @stevesphd.bsky.social · 30/06/2026

...and 4) if none of these approaches are possible, perform a naive analysis and estimate the likely magnitude of bias that could occur under realistically severe index event bias.

100

Stephen Burgess @stevesphd.bsky.social · 30/06/2026

2) if there are distinct variants affecting disease risk, consider performing a multivariable analysis with the risk factor and disease risk as exposures; 3) if you have access to individual-level data, consider performing an inverse-probability weighted analysis;

100

Stephen Burgess @stevesphd.bsky.social · 30/06/2026

Based on this, our suggested approach is: 1) if the same mechanisms affect disease progression as affect disease risk, perform an analysis for disease risk rather than for disease progression;

100

Stephen Burgess @stevesphd.bsky.social · 30/06/2026

If similar mechanisms affect disease risk and progression, then we are unlikely to find valid instruments for disease incidence. But in this case, we can just perform analyses for disease risk, as disease risk is easier to investigate (larger sample size, no index event bias).

100

Stephen Burgess @stevesphd.bsky.social · 30/06/2026

The multivariable methods performed well in simulations, but require a valid instrument for disease incidence (that does not have a direct effect on disease progression) to work well. Otherwise, they are flexible to implement. IVW worked better in the scenarios we considered.

100

Stephen Burgess @stevesphd.bsky.social · 30/06/2026

Slope-Hunter performed badly in all simulations, regularly either over-estimating or under-estimating the amount of bias, and hence over- or under-correcting. The method gave unbiased estimates in some cases, but with highly inflated Type 1 error rates.

100

Stephen Burgess @stevesphd.bsky.social · 30/06/2026

Heckman's method performed well in simulations, but requires both individual-level data and a valid instrument for disease incidence (that does not have a direct effect on disease progression). Software for implementation is not particularly flexible.

100

Stephen Burgess @stevesphd.bsky.social · 30/06/2026

Inverse-probability weighting performed okay in simulations, but requires individual-level data, and bias reduction depends on the predictive ability of the incident disease model.

100

Stephen Burgess @stevesphd.bsky.social · 30/06/2026

We compared inverse-probability weighting, Heckman's method, Slope-Hunter, and multivariable methods (IVW and CWBLS). These methods use different strategies to deal with the bias: model incidence, instrument incidence, estimate bias magnitude, and adjust for incidence.

100

Stephen Burgess @stevesphd.bsky.social · 30/06/2026

Genetic associations with disease progression can be affected by index event bias if they are only measured in individuals with an incident disease event (which they often are out of necessity). Several statistical methods aim to reduce index event bias. Do they work?

100

Stephen Burgess @stevesphd.bsky.social · 30/06/2026

Most genome-wide association studies look at disease incidence in population-based datasets. But drugs are typically given to selected individuals at high-risk of disease, or people with incident disease to reduce disease progression - gap between what we have and what we want.

100

Stephen Burgess @stevesphd.bsky.social · 30/06/2026

New publication "Comparison of methods for assessing effects of risk factors on disease progression in Mendelian randomization under index event bias" at @ajhgnews.bsky.social: authors.elsevier.com/sd/article/S.... Collaboration between Eli Lilly, Sequoia Genetics, and Cambridge University. Thread:

authors.elsevier.com

173

Stephen Burgess @stevesphd.bsky.social · 28/06/2026

Thanks to @andy_grant1 for leading this work, and to Ash Patel for supporting. Any feedback is welcome. Hope this method is useful to everypony!

100

Stephen Burgess @stevesphd.bsky.social · 28/06/2026

Weak instruments are rare in univariable (i.e. one risk factor) MR, as we typically select variants that are associated with the exposure at a GWAS-significant level. But weak instruments are much more common in multivariable MR, as we need to jointly predict all the exposures.

100

Stephen Burgess @stevesphd.bsky.social · 28/06/2026

Weak instruments aren't ideal in IV / MR, but data-driven attempts to deal with weak instruments often cause more problems than they solve. Using statistical methods that are robust to weak instruments (like this one!) is often a better solution.

110

Stephen Burgess @stevesphd.bsky.social · 28/06/2026

And by Celestia, the answer was yes! - we get much better bias and coverage properties than standard methods. The Bayesian procedure better accounts for uncertainty (error propagation, no asymptotic normality). As this is a cut-down version of MVMR-Horse, we called it MVMR-Pony.

110

Stephen Burgess @stevesphd.bsky.social · 28/06/2026

We observed the original method performs well with weak instruments. So we thought, if we remove the pleiotropic-robust aspect of the model, would we just have an MVMR method that has good statistical properties with weak instrument when the variants are valid?

100

Stephen Burgess @stevesphd.bsky.social · 28/06/2026

Two years ago, we developed a Bayesian approach called MVMR-Horse that fits a multivariable IV model using a horseshoe prior to allow for pleiotropic genetic effects - this is an overidentified model, and the prior gives model identification: x.com/andy_grant1/....

x.com

Andrew Grant (@andy_grant1) on X

New paper “A Bayesian approach to Mendelian randomization using summary statistics in the univariable and multivariable settings with correlated pleiotropy” now out in The American Journal of Human G...

110

Stephen Burgess @stevesphd.bsky.social · 28/06/2026

New preprint: "Multivariable Mendelian randomization with weak instruments: a comparison of Bayesian and frequentist methods" led by Andrew Grant: arxiv.org/abs/2606.26638. We present a method for multivariable Mendelian randomization (MVMR): MVMR-Pony. Thread follows:

arxiv.org

Multivariable Mendelian randomization with weak instruments: a comparison of Bayesian and frequentist methods

Weak instruments are a well known limitation for valid causal inference in Mendelian randomization studies. In the single exposure setting, weak instrument bias can be mitigated by selecting genetic i...

1185

Reposted by Stephen Burgess

American Society of Human Genetics (ASHG) @geneticssociety.bsky.social · 16/04/2026

In their recent @ajhgnews.bsky.social Review, @vkarhune.bsky.social, @stevesphd.bsky.social, & co discuss the key considerations and provide advice to produce a higher standard in planning, conducting, reviewing, and interpreting cis-Mendelian randomization studies: bit.ly/48DEz0H #ASHG

035

Stephen Burgess @stevesphd.bsky.social · 13/04/2026

Thanks Kaur! Your comments were really useful and helped ensure that we got the message across clearly. Much appreciated!

000

Stephen Burgess @stevesphd.bsky.social · 13/04/2026

Thanks to all co-authors, and also to reviewers who helped and supported this work (including @kauralasoo.bsky.social , who kindly self-identified!) - comments and feedback welcome!

020

Stephen Burgess @stevesphd.bsky.social · 13/04/2026

Cis-MR studies are not intrinsically superior to genome-wide MR studies, and algorithmically-performed cis-MR analyses will rarely be optimal. But when performed with care, cis-MR is a powerful tool to inform about putative causal effects.

123

Stephen Burgess @stevesphd.bsky.social · 13/04/2026

Validation can be statistical (e.g. colocalization), but also should be biologically-motivated (e.g. positive and negative controls). Robust MR methods are rarely conclusive, as if one variant is invalid, it is likely nearby variants will be invalid in a similar way.

100

Stephen Burgess @stevesphd.bsky.social · 13/04/2026

In particular, molecular exposure biomarkers measured in bulk tissues (usually blood) may reflect mechanisms in irrelevant tissues, whereas downstream biomarkers may be more specific.

110

Stephen Burgess @stevesphd.bsky.social · 13/04/2026

The exposure biomarker reflects the causal mechanism of interest. It is not necessarily the causal risk factor itself, and often choosing a downstream trait is better - if water has reached a proximal downstream station, then it must have passed though the causal mechanism.

100

Stephen Burgess @stevesphd.bsky.social · 13/04/2026

Important steps are: 1) defining the analysis question, 2) choosing an appropriate gene region, 3) choosing an appropriate biomarker of the causal mechanism, 4) choosing the optimal variant(s) in the selected gene region(s), 5) validating the variants.

100

Stephen Burgess @stevesphd.bsky.social · 13/04/2026

Our review provides guidance on how to perform reliable cis-MR analyses - the analysis is often not difficult to implement, and the major effort is ensuring that you are performing the optimal analysis, and reporting that analysis in a reasonable way.

100

Stephen Burgess @stevesphd.bsky.social · 13/04/2026

It's tempting to this of cis-Mendelian randomization analyses, particularly those using a single genetic variant, as simplistic, and analysis with multiple variants using large numbers of methods as more sophisticated. However, "design trumps analysis"...

100

Stephen Burgess @stevesphd.bsky.social · 13/04/2026

Our review "Integrating genetic data with biological insight: A practical guide to cis-Mendelian randomization" is now published at @ajhgnews.bsky.social - led by @vkarhune.bsky.social and Benji Woolf with critical insight from Dipender Gill and Pallav Bhatnagar. Thread follows:

1106

Stephen Burgess @stevesphd.bsky.social · 16/03/2026

Thanks to Ash for leading this work, and to Frank DiTraglia for asking difficult questions about the statistical methodology. All comments (and suggestions for applications) welcome!

000