Re: st: Fisher's exact test

[Marcello Pagano <pagano@hsph.harvard.edu>]

You can always use it, assuming the calculations are done correctly; as 
they are in Stata.

It gives the right answer when you are conditioning on both margins.  
Follow Fisher's dictum that says, If the margins do not contain 
information, then this test is fine.

Another way to view this is if you think of the two probabilities as p1 
and p2 (say the row probabilities that might represent success on two 
separate treatments) then you can rotate from these two parameters to 
two new ones, dif = p1 - p2 , and nui = p1 + p2. Knowing dif and nui is 
the same as knowing p1 and p2. Now the question becomes, if you are 
interested in dif, for example is dif=0, is knowing nui informative?  In 
the extremes, nui=0 or nui=2, then those are very informative.

If, on the other hand, the margins may be informative in general, then 
the unconditional tests (that do something with the nuisance, nui) can 
be more powerful than Fisher's test.

m.p.

p.s. No, there is no rule that says you cannot post in both, but we 
should try and drift to the new list. Which then raises the point, why 
did I respond here and not on the new list!


On 4/3/2014 1:35 PM, Gwinyai Masukume wrote:
> Thanks Israel. Would you have a reference for that?
>
> Apologies Prof Pagano - I was not aware of the doubling posting rule.
> Incidentally, I was reading Pagano in an attempt to find the rules
> governing appropriate use of Fisher's exact test. Any words from the
> man himself?
>
> With thanks,
> Gwinyai
>
> On 4/3/14, Marcello Pagano <pagano@hsph.harvard.edu> wrote:
> > Please do *not* post on both versions of Statalist.
> >
> > Thanks,
> >
> > m.p.
> >
> >
> >
> > On 4/3/2014 12:31 PM, Gwinyai Masukume wrote:
> > > Hi,
> > >
> > > According to some experts, "The traditional Fisher's exact test should
> > > practically never be used."
> > >    Lydersen S, Fagerland MW, Laake P. Recommended tests for association
> > > in 2 x 2 tables. Stat Med. 2009; 28(7):1159-75.
> > > http://www.ncbi.nlm.nih.gov/pubmed/19170020
> > >
> > > The convention seems to be if any of the boxes in a 2x2 table has less
> > > than 5 entities Fisher's exact test should be used.
> > >
> > > When is it correct/appropriate to use Fisher's exact test?
> > >
> > > With thanks,
> > > Gwinyai
> > > *
> > > *   For searches and help try:
> > > *   http://www.stata.com/help.cgi?search
> > > *   http://www.stata.com/support/faqs/resources/statalist-faq/
> > > *   http://www.ats.ucla.edu/stat/stata/
> > *
> > *   For searches and help try:
> > *   http://www.stata.com/help.cgi?search
> > *   http://www.stata.com/support/faqs/resources/statalist-faq/
> > *   http://www.ats.ucla.edu/stat/stata/
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/faqs/resources/statalist-faq/
> *   http://www.ats.ucla.edu/stat/stata/

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/