12  Questions

1. Which definition of a p-value is more correct?
   1. the probability that the null hypothesis is true
   2. the probability that someone else’s data will be as extreme as mine were the null hypothesis true for them
   3. the probability that someone else’s data will be more extreme than mine were the null hypothesis true for them
2. The p-value is a conditional probability of data extremes. What does it assume (condition on)?
   1. the data
   2. the null hypothesis being true
   3. the data model
   4. the experimental design and sampling scheme are perfectly known so could be repeated indefinitely
   5. b, c, and d
3. Does the probability of extreme data directly inform you of the probability that a drug works?
4. Can traditional frequentist statistics provide a false positive probability?
5. Why is it useful to compute a degree of belief (probability) that the drug works?
   1. because it requires few assumptions
   2. because it is completely objective
   3. because it can incorporate extra-study information and feeds directly into an approval decision
6. What is needed to compute a degree of belief (probability) that the drug works?
   1. a prior probability distribution for the drug effect
   2. data
   3. a model for how the data are generated
   4. a, b, and c
7. What is an example of a useful compound probability, i.e., a probability of combinations of assertions?
8. What is the scale on which frequentist statistics places skepticism about evidence for efficacy?
   1. random chance
   2. chance that an effect is positive
9. What is the scale for Bayesian statistics?
   1. random chance
   2. chance that an effect is positive
10. What are simple reasons that Bayesian methods have no need or way to deal with multiplicity?
    1. Bayes deals with probabilities about parameters given current data
    2. Current data overrides the study’s data at an earlier point in time
    3. Asking a question about the effect of a drug on one endpoint should not influence the answer to the question about another endpoint
    4. The chance of getting more extreme data with multiple looks is different from the chance that a drug is effective
    5. all of the above
11. Is it possible to infer than a drug doesn’t work when p=0.7?
12. If the Bayesian posterior probability that a drug works is 0.93, what is the chance that it doesn’t work or actually causes harm?
    1. 0.07
    2. not computable from just this information