R help - Jun 2001 - binom.test appropriate?

Hi there, 

as part of a 2 x 2 contingency table analysis I would like to estimate
conditional probabilities (success rates) in a Bernoulli
experiment. In particular I want to test a null hypothesis p <= p0
versus the alternative hypothesis p > p0.

As far as I understand the subject, there are UMPU tests for these
types of hypotheses.

Now I know about R's "binom.test" but the help text is telling me
it
will test only simple null hypotheses of the form p = p0.

Could someone please give me a hint whether there is a suitable test
or about what I got wrong?

Thanks, Mirko. 

-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-
r-help mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html
Send "info", "help", or "[un]subscribe"
(in the "body", not the subject !)  To: r-help-request at
stat.math.ethz.ch
_._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._

Hi!

Prof Ripley wrote (marked by ''>'')

...
> But: the test for p = p0 vs p > p0 is the appropriate test for
> p <= p0 vs p > p0 within this family of tests, by the monotonicity
> properties.
Thanks, I meanwhile understand that monotonicity, a general property
of the power function w.r.t. exponential families of measures, makes
the binomial test work for our purposes.
> You mentioned a 2 x 2 table and UMPU, but did not say exactly what you are
> doing or how the data were sampled, nor how this hypothesis arises.  Under
> one set of assumptions, I believe the UMPU theory you mention tells you to
> use the binom.test for p = p0 vs p > p0, but it may be that other tests
> (Fisher''s exact test springs to mind) are more appropriate.  (And
the last
> U can be insidious, just as it can be for estimation.)
Let me be more elaborate on this point.  

We have a certain way of measuring a nonnegative real valued number
that we call "activity".  Having found out about the activity, we
experimentally observe some "response" that is a binary variable
"yes", "no". We have defined a threshold for the activity.
For an
activity smaller than the threshold, our "prediction" is
"yes",
otherwise it is "no".

Writing down the number of coincidences and non-coincidences of
"predictions" and "responses" we get the mentioned 2 x 2
matrix.

Now we would like to learn about the conditional probabilities

    p(response = x | prediction = y). 

In particular we would like to verify the alternative hypotheses

   p(response = "yes" | prediction = "yes") > p0

as well as

   p(response = "no" | prediction = "no") > p1

against the nulls "<= p0" and "<= p1" for certain
bounds p0 and p1
that came out of our CFO''s spreadsheet.

To this end we read the number of "response = ''yes''"
cases among the
"prediction = ''yes''" cases from the matrix (and do
so for "no" - "no")
and apply the binomial test.

Some questions that remain are 

     - Does this seem appropriate?

     - Would Fisher''s exact test be even more appropriate?

     - How could we optimize our threshold? 


Any comments are appreciated, thanks, Mirko.

--
Dr. M. Luedde <Mirko.Luedde at CellControl.De>
CellControl Biomedical Laboratories AG
Am Klopferspitz 19, 82152 Martinsried
+49-89-895275-0 +49-179-5252064 
-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-
r-help mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html
Send "info", "help", or "[un]subscribe"
(in the "body", not the subject !)  To: r-help-request at
stat.math.ethz.ch
_._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._