Interview 2019-07-15 for vinayakkprasad.com/plenarysession/

Transposed conditionals
 What is Type I error?
  Is NOT P(making a mistake in concluding the tx has + efficacy)
  Is P(making an assertion of evidence if efficacy=0 (and no harm))

Frequentistism: Replace something that is hard to do (quantify evidence for what you don't know, i.e., efficacy) with something that is easy to do (P(observing data more extreme than that observed IF no efficacy))
 Satisfies a common-sense requirement for decision making: condition on (make use of) what you already know to compute the probability of something you don't know

Frequentist - sampling - envision identical replications of experiment to make judgements on atypicality of data (degree of surprise) if the null is true; no protocol changes allowed
Bayesian: 
 The data were generated by a model, e.g., large patient-specific tendencies we don't understand + treatment received
 Try to uncover the hidden truth: what do we now know about what was
 in play that generated our data, other than the unexplainable
 patient-to-patient variation?  Quantify evidence for all possible
 levels of efficacy.

Formal way to incorporate outside information and skepticism
Unlimited data looks due to forward-information-flow (current) probs vs. P(would could have happen)


Rational interpretation of evidential information
 P(RFS benefit) > 0.95 or P(OS benefit) > 0.8
 P(RFS benefit & no severe toxicity) > 0.9
 P(RFS benefit > 20%)

Make bias/skepticism explicit
Freq multiplicity correction: Discount evidence for comparing A vs B because you compare C vs D 
Bayes: discount A vs B because of prior knowledge about A and B (C vs D irrelevant)

Eizabeth Ryan BMJOpen 2019   OPTIMISE
 Why is post prob so large when p=0.07?
 Multiplicity correction for protecting against a false claim of harm
 Immediate gain for Bayes being directional
 Imagine a study with 100 patients in each of tx A, B, 30 deaths for A 40 deaths for B
 How would you play the odds if you mom was being treated?  Which treatment would you choose if P(B better than A) = 0.51?
 Interpretations are definitely different between Bayes and freq.


Allen Pannell stated that there is no face validity of getting such a different result when using a "non-informative" prior
 Analysis did not consider that hypothesis is already true; it was a distribution
 He discounted the use of skeptical priors ("negative data")
 Just adding previous data to the dataset assumes same experimental design and no discounting.
 Credible intervals not necessarily tighter than confidence limits; Bayesians know how to take more uncertainties into account (e.g. non-normality; non-equal variance)

Against Inference and in favor of playing the odds