Co-authors: Here is a rough outline.  What is difficult about this is
that much of what we need to address, if we want to get an alternative
to Brophy & Joseph in the literature, is stuff that is off the main point
I have been pursuing.  Namely, we need to once again address absolute
vs. relative benefit, and we need to come up with some good text about
why GISSI and ISIS can't be pooled with GUSTO.  Here's a rough outline.
Suggestions/criticisms needed and welcome.

-Frank




           Placing GUSTO in Context using Bayesian Inference

                  F Harrell, K Lee, E Topol, R Califf

OUTLINE

Basic results and P-values

Limitations of P-values

Basics of Bayesian approach
    -   Brophy and Joseph laid this out really nicely
    -   Biggest single advantage is being able to estimate the
        probability of a worthwhile clinical benefit

Summary of Brophy & Joseph Paper

Where our approach differs
    -   Their analysis was based on absolute differences
    -   They assumed that since the study was powered to detect a
        0.01 mortality reduction, the GUSTO investigators assumed that
        a reduction had to be 0.01 to be clinically meaningful
    -   In using prior information (GISSI, ISIS), they did not recognize
        that different entry criteria, time frame, countries can result
        in a shift in base mortality patterns so that it is problematic
        to pool studies using absolute differences
    -   They point out the problems with meta-analysis, but in assuming
        that accelerated t-PA = usual t-PA, they in effect are doing what
        meta-analyzers do
    -   This leads them to assume 'poolability' when using the prior study
        results in prior distributions
    -   When the therapies are not given the same way as previous studies
        or the compounds are somewhat different, we favor emphasizing
        non-informative prior distributions (play ignorant!)

Bayesian estimates assuming non-informative (flat) prior distribution
    -   For each pair, estimate P{clinical benefit} and P{equivalence}
    -   For Accel t-PA compared with combined t-PA+SK
    -   For SK+IV Hep vs. SK+SQ Hep
    -   For Accel t-PA vs. combined SK
    -   For t-PA+SK vs. combined SK

Show sensitivity of results to varying prior distribution
    -   Use a normal prior distribution with mean at log odds ratio = 0
        (no effect) and varying degrees of sharpness (lower standard
        deviation)

Discussion