Ian Marschner

@f2harrell @BerryConsultant In the same way that posterior distributions do in Bayesian inference

@IanMarschner @BerryConsultant How do they bring actionable, simply interpretable evidence for more than trivial effects?

Dr. Andrew Thomson explains why borderline clinical trial results represent the hardest decisions in the field. When the point estimate falls below the minimally clinically important difference, findings may be statistically significant, but possibly not clinically relevant. This triggers intense scrutiny—questions emerge around estimands, estimation methods, missing data, and the impact of a single patient on results. These cases require rigorous scrutiny and thorough analysis. Discover more on the full episode: berryconsultants.com/resource/33-a-…

@f2harrell @BerryConsultant Confidence distributions allow you to do the same sort of thing in a frequentist context

@BerryConsultant Because of imprecision of estimates we can't know from a point estimate just below the MCID that the true effect < MCID. That's why we need Bayesian P(effect > trivial effect threshold), e.g., P(effect > MCID/3) discourse.datamethods.org/t/effect-sizes…

@FOXFOOTY Amazing that the number 1 issue is not on the list. 18 teams. 22games.

Incoming AFL football boss Greg Swann has several key ticket items for 2026 ⬇ READ MORE: bit.ly/40kUlcZ

@KertViele @DrToddLee @ABsteward @Inox94 @BradSpellberg @YukiKotani5 @ADAlthousePhD @JasonConnorPhD This is another example of why we shouldn’t use RAR for a 2-arm comparison. It was perhaps less clear when this study was designed but now is very clear

@DrToddLee @ABsteward @Inox94 @BradSpellberg @YukiKotani5 @ADAlthousePhD @JasonConnorPhD and finally 3) many sites refused to randomized to control, which generated smaller sample sizes on the control arm. In addition, the skewed RAR probabilities actually were in line with prior expectations, which may have generated less suspicion (I don't know this).

🔥REMAP-CAP RCT🔥 @yourICM Among patients admitted to the ICU with severe non-pandemic CAP and either cardiovascular or respiratory failure there was a low probability that a 7-day fixed-duration course of hydrocortisone improved 90d mortality @remap_cap link.springer.com/article/10.100…

Non-concurrent controls occur in adaptive platform trials. Our paper gives a new more rigorous definition based on randomization cohort rather than time. Cohort adjustment and implications for databases & software are discussed. @austrim_cre @TrialsCentre onlinelibrary.wiley.com/doi/10.1002/bi…

@JAMAOnc Treatments that improve survival provide greater opportunity for adverse events to occur. See the Limitations Section: "improved survival with ARSIs could manifest as a higher captured incidence of CV events. It is impossible to account for this given lack of time to event data"

Meta-analysis of 24 randomized controlled trials demonstrates an increased risk of CV events amongst patients with advanced and metastatic prostate cancer commencing androgen receptor signaling inhibitors. ja.ma/3LBSGYJ

Recent @TheLancet trial uses confidence distribution to conclude “the confidence that the risk ratio is lower than 1 is 97.2%”. Full confidence distribution plot in supplementary material. Excellent way to present trial results as described in Reference 24 thelancet.com/journals/lance…

@KertViele @syctong @GuyattGH I agree that small concurrently randomised cohorts are a challenge for reporting. I would interpret this as an argument for avoiding design features that produce small cohorts e.g. I would argue we should avoid frequent interim analyses that lead to frequent design adaptations

@syctong @GuyattGH I think this is valuable. Very useful and insightful to recognize the independent “between interim” cohorts and the meta analysis connections. I think people will need to be cautioned not to overreact to noise in the smaller samples within cohort.

Two articles highlight big problems when platform trials include non-contemporaneous controls. In addressing A vs no A, analysis should only use randomizations that included possibility of A. Should never use controls of B vs non-B. nejm.org/doi/full/10.10… sciencedirect.com/science/articl…

@f2harrell @vandy_biostat Not always. If design choice reflects prior belief then Bayesian inference is affected by multiplicity. As Kass & Wasserstein said (JASA 1996 p.1359): “It could be argued that choice of design is informative and so the prior should depend on the design”

@f2harrell @vandy_biostat Note that absence of multiplicity issues in Bayesian inference requires the important assumption that you would use the same prior regardless of the design. Reference priors contravene this assumption, yielding different Bayesian inferences for sequential and fixed designs

Steve Ruberg has joined me as an author, and has made major improvements in the article about a simple football example that illustrates why there is no multiplicity with Bayesian sequential analyses: fharrell.com/post/nfl #Statistics #clinicaltrials @vandy_biostat

“Removing the advanced mathematics prerequisite does nothing to address the decline in mathematics enrolments at schools and sends the wrong signal to students.” science.org.au/news-and-event…

@KertViele Sounds like a great session. I agree that if anyone’s saying data should be thrown away that’s problematic. Also important to recognise that different types of data have different levels of quality. If we pool them then we need to understand the relative contributions of each

If you have a late flight Saturday at ISBS, come see our session on platform trials. I’ll be arguing data is a good thing and we shouldn’t throw it away. (Maybe a little more nuanced…but essentially)

@GuyattGH Here is closely related research conducted independently by PhD student Anne Lyngholm Soerensen published last year. Clearly an important topic journals.sagepub.com/doi/10.1177/09…

Recent important, useful innovation in conducting and presenting the results of subgroup analysis. Should become standard procedures. doi.org/10.1002/jrsm.1…

@MaartenvSmeden Not the probability of an hypothesis being true but perhaps the confidence? onlinelibrary.wiley.com/doi/full/10.10…

Once you realize p-values are probabilities relating to observing data rather than probabilities of an hypothesis being true, you are already doing significantly better than most people

@JasonConnorPhD @ADAlthousePhD You might look at these two old papers (and papers citing them), they sound relevant onlinelibrary.wiley.com/doi/abs/10.100… onlinelibrary.wiley.com/doi/abs/10.100…

Stats Q Group sequential trials look at X% of patients(info) for the final timepoint and calc the error spend. Is there an analogous group sequential method but you look at 100% of patients but at a 6m endpoint instead of 12m endpoint? Thanks!

@predict_addict @austrim_cre @TrialsCentre Thanks for the interesting references but you are talking about prediction whereas I'm talking about inference (a common difference between data scientists and statisticians). It's unclear what you mean by "lack validity" when the context is inference from a randomized experiment

Interesting, however statistical confidence distributions lack validity guarantees. Conformal Predictive Distributions solved that problem back in 2017 and can be used with any model, any data distribution and any dataset size. proceedings.mlr.press/v60/vovk17a.ht… arxiv.org/pdf/1911.00941… @valeman/how-to-predict-full-probability-distribution-using-machine-learning-conformal-predictive-f8f4d805e420?sk=c8eab1aeefba892777e6c3ee3afcf223" target="_blank" rel="nofollow noopener">medium.com/@valeman/how-t…

Confidence distributions are distributional summaries of evidence. They’re like Bayesian posteriors but without prior distribution assumptions. I wrote a tutorial paper in Statistics in Medicine for clinical trial statisticians @austrim_cre @TrialsCentre onlinelibrary.wiley.com/doi/10.1002/si…

@lakens Your opening argument is based on: “it is just as likely to observe a p-value of 0.001 as it is to observe a p-value of 0.999”. Both of these values have zero probability of being observed (for a continuous p-value distribution) so I’m not sure what the point is you are making

Common mistake, even among statisticians, to think p-values are measures of evidence. I explain why that is not a coherent way of thinking about p-values here osf.io/preprints/psya…

@syctong Need to be careful of non-proportional odds in such an analysis. The odds ratio for “considerable improvement” is about 12.4 compared to 4.7 for survival, yet they are assumed to be the same. Might be chance variation but important to check or else the odds ratio could be biased

@stephensenn @cjsnowdon In Australian 1970s classrooms we had to drink it warm on 35 degree days!

@cjsnowdon Spent four years of my childhood (aged 6-10) in the UK. The compulsory milk consumption still gives me nightmares. Thatcher milk snatcher was too late to save me.

Note the decline in sugar consumption.