Like sure I'd love to be able to design a trial to detect the smallest clinically relevant effect but I have $100k and I wouldn't wan't to miss a larger effect, assuming I have a reason to believe it's possible?.

Bracing myself for the flames but I have been wondering if @stephensenn's effect size for powering trials as the "effect you wouldn't want to miss" offers a more charitable view of the process of balancing budgets and statistical considerations in trial planning. 1/2.

I assume even less likely to see a *meaningful* change in direction, or change in direction if originally significant? Makes me wonder whether summaries of decision sensitivity to uncertainty captured in the estimate itself using threshold analysis or pred intervals would suffice.

Extra interesting given from what I can tell on quick analysis of data from Table 2 using penalized GLM there's no difference between high vs not high level evidence in terms of probability of update leading to sign change in the estimate.

@dnunan79 Just thinking in terms here of whether uncertainty + heterogeneity (+/- bias adjustment considering effect seems weak in this sample) already contains most of the decision uncertainty value of the rest of the GRADE process.

@dnunan79 Only other thing that bothers me is whether absolute RoR is interpretable across CoE here. Since OR is nonlinear won't we get differences in absolute RoR just as a function of absolute precision? So if high certainty generally higher precision abs RoR will be lower by design?.

Congrats @dnunan79 and team! This is an important contribution to discourse around value/validity of GRADE. Still processing but my initial takeaway is directionally promising but not as convincing as hoped

Anyone happen to know how to f/u on pathology in Nova Scotia? Friend of mine waiting 4 wks for path results on sentinal node biopsy from melanoma after waiting 7 wks for initial dx + confirmation margins werent clean. Is there something we can do to get this moving? @CBCNS

Quick little qualitative analysis dashboard for #TTPD. Hoping to inoculate against lazy takes. Powered by GPT4Turbo but hoping to serve as a starting point rather than any final word 🫡.

My team is hiring full-time remote biostats focused on analyses to support HTA apps, indirect treatment comparisons, etc. Feel free to reach out directly if you're interested!.

Without any irony I think there a possibly huge implications for #neoTwitter research @EBNEO @souvik_neo @DrMCampbellYeo.

Sometimes I wonder if I’m doing enough good in the world. Today the answer is yes. Just in time for #Swiftmas Presenting #SwiftSearch a GPT-3 powered semantic search for themes in @taylorswift13 songs. Finally you can know “what TS song am I?” #Swifties.

This is an initial draft so expect to find some bugs over the next few weeks. Feel free to share your thoughts/ideas here and at

If you work in HTA or do a lot of survival modeling you might find yourself a fair bit of model selection (AIC, plots, PH tests etc). I made a quick app to practice/explore when this might be harder/easier:

Me looking at continuous/ordinal data: PO seems pretty clearly violated here.Me simulating data under PO with similar sample size: Hmmm, I guess I don't know what the hell I'm talking about and should stop pretending I know better than @f2harrell .#epitwitter #statstwitter.

I feel like @laure_wynants tweeted something about traps in updating prediction models once and it's just been eating away at my soul ever since.

You fit sparkly new prediction model that says actually variables xyz are the most relevant. Now you implement this tool but suddenly all those high risk patients aren't being shuttled as effectively to therapy so outcomes are actually worse. Does this happen?.

Probably a dumb question but wonder what @laure_wynants @MaartenvSmeden @kdpsinghlab think: Say practice at a center has been influenced by good + simple prediction models (not formally implemented) that get lions share of high risk patients to effective therapy.

Don't even get me started on the relevance of local balance of benefits/harms and tolerance for risk.

Baseline risk varies a lot in MAs even with homogenous populations and strict protocols from RCTs. Have to get beyond idea that guideline panel can make decisions based on RRs and find a way to combine these models with locally relevant estimates for locally relevant decisions.