Frank Harrell

19.4K posts

Frank Harrell

@f2harrell

Biostatistician/Professor/Founding Chair of Biostatistics, Vanderbilt U. Blog: Statistical Thinking:https://t.co/2BTEONzsfX @f2harrell on https://t.co/bsPN9JQNOS

Nashville, TN Katılım Ocak 2017

174 Takip Edilen30K Takipçiler

Frank Harrell@f2harrell·1d

@Stats_Veritas @CMichaelGibson There are many approaches but I prefer using Bayes to bring evidence for consistency of effects on a biomarker and on clinical events as exemplified here. Embed the continuous marker + events in one ordinal scale. fharrell.com/post/yborrow

English

Temariy@Stats_Veritas·2d

@f2harrell @CMichaelGibson Very nice discussion, Dr. Harrell. I am interested in learning more about how the Bayesian approach can circumvent the need for complicated analyses to establish the surrogacy of biomarker. Do you have a resource on how this is done in the Bayesian context?

English

Frank Harrell@f2harrell·2d

Thrilled to be interviewed by imminent cardiologist and clinical trialist Mike Gibson yesterday about the new draft Bayesian guidance at FDA: clinicaltrialresults.org/dr-frank-harre… #Statistics #bayes #pharma #clinicaltrials #rct @CMichaelGibson

English

3.5K

Frank Harrell@f2harrell·2d

@ParkLover_JA @CMichaelGibson Plus Bayesian methods are useful for earlier detection of bad treatments.

English

Jordan FatCat@ParkLover_JA·2d

@f2harrell @CMichaelGibson Smarter statistics, faster treatments. About time regulatory science caught up with methodology that's been proven for decades.

English

135

Frank Harrell retweetledi

Ed Krassenstein@EdKrassen·3d

Wow! After Donald Trump mocked Gavin Newsom and other people with dyslexia and learning disabilities, Newsom’s wife, Jennifer Siebel Newsom, just completely torched him. The end is the best part.

English

2.6K

10.3K

36.3K

804.9K

Frank Harrell@f2harrell·3d

@5_utr @free_radical28 @schneider_jw @BradSpellberg @DrToddLee @dnunan79 @BoussageonR @BJegorovic @NEJM @JAMA_current I.e., choose N so that two people with reasonably different priors interpret results the same.

English

Frank Harrell@f2harrell·3d

@5_utr @free_radical28 @schneider_jw @BradSpellberg @DrToddLee @dnunan79 @BoussageonR @BJegorovic @NEJM @JAMA_current And on that note, as exemplified in the link you provided later, it often makes sense to compute N such that the prior makes less than epsilon difference.

English

Boris Jegorović, MD, PhD, DTM&H@BJegorovic·5d

🥸 Expert opinion of the day: Corticosteroids for brain abscess. Low certainty of evidence - Strong recommendation 💪🏻 Paging guidelines police 🫡 @BradSpellberg @DrToddLee clinicalmicrobiologyandinfection.org/article/S1198-…

English

4.3K

Frank Harrell@f2harrell·3d

@samweisenthal @BoussageonR @dnunan79 @LGHemkens @GuyattGH @pash22 @AnilMakam @VickersBiostats Yes; good to start with the minimum absolute reduction in risk the patient would tolerate to put up with the side effects, as a way to specify the utility function indirectly.

English

Samuel Weisenthal@samweisenthal·3d

@BoussageonR @dnunan79 @LGHemkens @GuyattGH @pash22 @AnilMakam @VickersBiostats @f2harrell For the individual patient, though, these risks are only part of the picture, because every person has a unique utility function

English

Boussageon Rémy@BoussageonR·3d

Okay, let's start a discussion: what kind of evidence E is necessary to justify an intervention? I am a GP, and I claim to base my practice on E.. @dnunan79 @LGHemkens @lakens @GuyattGH @pash22 @AnilMakam @VickersBiostats @f2harrell

English

2.6K

Frank Harrell@f2harrell·3d

@matloff @BradSpellberg @Wiki_Guidelines It's best to personalize the calculations but you are right: judicious choices of representative patients can lead to semi-accurate absolute risk reduction estimates for the patient at hand.

English

Norm Matloff 你有冇諗清楚呀?@matloff·3d

@f2harrell @BradSpellberg @Wiki_Guidelines But you can form confidence intervals for several representative types of patience. You seem to be setting up a straw man here.

English

Brad Spellberg@BradSpellberg·3d

Completely agree @f2harrell (as usual). 2 Venn diagrammed discussions happening simultaneous: 1) how to interpret the data cited; 2) how to construct a guideline related to the data. My comments were for #2, not #1. A brief string related to guidelines vs. data...@Wiki_Guidelines

Frank Harrell@f2harrell

@BradSpellberg @DrToddLee @dnunan79 @BoussageonR @BJegorovic Best interpretation of a 0.95 CI for an odds ratio being [0.5, 1.3]: We know no more after looking at the data than we did before we collected it.

English

7.7K

Frank Harrell@f2harrell·3d

@LGHemkens @BoussageonR @dnunan79 @GuyattGH @pash22 @AnilMakam @VickersBiostats No, I see doctors delay decisions, even the watchful waiting employed regarding surgery, all the time. Or start on a low dose of a drug, or do a short-term trial of a drug, wait for one more diagnostic test, ... Often no decision (for now) is the best decision.

English

Lars G. Hemkens@LGHemkens·3d

@f2harrell @BoussageonR @dnunan79 @lakens @GuyattGH @pash22 @AnilMakam @VickersBiostats You have to act or not to act. "We don’t have information" is irrelevant, you still need to decide.

English

Frank Harrell@f2harrell·3d

@5_utr @schneider_jw @BradSpellberg @DrToddLee @dnunan79 @BoussageonR @BJegorovic That's exactly how many trials should be interpreted (although you can still learn how to recruit patients and measure outcomes). The honesty is just not there for @NEJM @JAMA_current and virtually all other medical journals. One cause is unrealistic effect sizes in power calc

English

1.2K

NonsparseOncologist@5_utr·3d

@f2harrell @schneider_jw @BradSpellberg @DrToddLee @dnunan79 @BoussageonR @BJegorovic I wish there was more honesty in reporting, I would love to see “we learned nothing from this trial”, or “we were only able to rule out an absolute miracle”

English

708

Frank Harrell@f2harrell·3d

@dnunan79 @BoussageonR @LGHemkens @GuyattGH @pash22 @AnilMakam @VickersBiostats Observing an effect > m has little to do with anything, due to noise in the estimated effect. Bayesian posterior probabilities don't use point estimates.

English

108

David Nunan@dnunan79·3d

@f2harrell @BoussageonR @LGHemkens @GuyattGH @pash22 @AnilMakam @VickersBiostats And is it the case that: 1. the RCT design = trustworthy because of its bias-reducing properties 2. it doesn't matter if the RCT uses frequentist or Bayesian statistics? Ie. Both equally allow us to interpret a truth ("effective") if we observe an effect >m with p<0.05?

English

495

Frank Harrell@f2harrell·3d

@BradSpellberg @Wiki_Guidelines Mainly, with the addition of a bit of nuance that I haven't well stated.

English

Brad Spellberg@BradSpellberg·3d

@f2harrell @Wiki_Guidelines I think you are agreeing with my point? Or am i misunderstanding?

English

184

Frank Harrell@f2harrell·3d

@LGHemkens @BoussageonR @dnunan79 @lakens @GuyattGH @pash22 @AnilMakam @VickersBiostats No, decisions are >= 3 sided. One side is "we don't have enough information to make a decision" which leads to waiting for more data.

English

106

Lars G. Hemkens@LGHemkens·3d

@BoussageonR @dnunan79 @lakens @GuyattGH @pash22 @AnilMakam @VickersBiostats @f2harrell Decisions are two-sided. That’s what changes the underlying logic. What level of evidence E is necessary to justify not intervening?

English

184

Frank Harrell@f2harrell·3d

@BoussageonR @dnunan79 @LGHemkens @lakens @GuyattGH @pash22 @AnilMakam @VickersBiostats Corollary: A single double-blind well-executed randomized clinical trial with sufficient current sample size, where P(trustworthy) is virtually 1.0, is often the best basis for quantifying evidence for better than trivial effectiveness of a treatment.

English

856

Frank Harrell@f2harrell·3d

@BoussageonR @dnunan79 @LGHemkens @lakens @GuyattGH @pash22 @AnilMakam @VickersBiostats Right you are. That leads to a fuller Bayesian model that includes P(data are trustworthy), better modeled as a better of degree, by putting a prior on a bias parameter for each data source as in fharrell.com/post/hxcontrol

English

517

Frank Harrell@f2harrell·3d

@dnunan79 @doc_BLocke @NEJM @JAMA_current "Works" is ALWAYS a matter of degree. Any attempt to act otherwise will result in poor decision making.

English

David Nunan@dnunan79·3d

@f2harrell @doc_BLocke @NEJM @JAMA_current But if we don't have this degree of 'proof', then this discussion is purely theoretical (probability-based). Health professionals/decision makers will still want "Works/doesn't work", but we'll only ever get this 'in theory'. Then it comes down which 'in theory' we prefer

English

David Nunan@dnunan79·13 Mar

Franks dismay gets to the heart of how the majority of medial/health researchers (& users of their outputs) treat observed effects based on frequentist methods. Purists rightly pick the flaw - but it’s rare that real world consequences are shown. x.com/dnunan79/statu…

Frank Harrell@f2harrell

Oh my goodness how did @JClinEpi allow this? The authors are pretending that point estimates are true values! For this study the subjects should not have been given point estimates & you really can't answer the survey questions without Bayesian posterior probs. #Statistics

English

5.3K

Frank Harrell@f2harrell·3d

@dnunan79 @doc_BLocke @NEJM @JAMA_current Any study with a "large" p-value should compute P(practical equivalence). For most studies this will be around 0.5, showing that you don't know ANYTHING.

English

David Nunan@dnunan79·3d

@f2harrell @doc_BLocke @NEJM @JAMA_current If misinterpretations of frequentist stats means we still likely ineffective treatments in practice, would this not play out in the real world. Ie. we wouldn't see/much any benefit?

English

Frank Harrell@f2harrell·3d

@arthur_alb1 @JAMANeuro @Tufts_PACE Something like that but I can't wrap my head around the random effects being identifiable with only one observation per person.

English

Arthur Albuquerque@arthur_alb1·4d

@f2harrell @JAMANeuro @Tufts_PACE Random-intercepts by study (given that Kent's data comes from a IPD meta-analysis? Something along these lines, a bivariate 1-stage IPD meta-analysis: h_1ij = stroke hazard in patient i in study j h_2ij = AF hazard in patient i in study j

English

JAMA Neurology@JAMANeuro·5d

Among patients with #Stroke and #PatentForamenOvale, the PASCAL classification system identified those likely to benefit from closure and those at risk of increased atrial fibrillation. ja.ma/4bo2Rg4

English

123

12.4K

Keşfet

@Stats_Veritas @CMichaelGibson @ParkLover_JA @5_utr @free_radical28 @schneider_jw @BradSpellberg @DrToddLee