Brent et al.--
There are two issues you face now--the difference between the
hypothesis tests Michael Frone and I have discussed, and multiple
hypothesis testing. On the latter, see -help _mtest- (note leading
underscore). On the former, tests of equality of proportions
implemented using linear regression cannot be right (proportions are
not means, except in a very limited sense), unless they accord exactly
with the correct test by chance, so you should place no faith in
results from -svy: reg- if they differ from -svy: tab- IMHO. As for
-svy: logit- it is possible to get quite odd results, or no result at
all (consider the case where the observed proportion is zero). In
cases where -svy: tab- and -svy: logit- differ, I think further
investigation is warranted.
The main issue is much more general than Brent's specific case--he
asks how he can test if some proportion in a population is equal to
the proportion in a subpopulation. If you knew with certainty the
proportion of the population in the subpopulation, the two tests we
have been discussing are identical. Call the proportion satisfying
the relevant condition (diagnosed with ADHD, married, what have you)
in the population p and the proportion of the population in the
subpopulation (living in Michigan, living in sin, what have you) a,
and the proportion satisfying the relevant condition in the
subpopulation q, and in the balance of the population (the complement
of the subpop) r, so
p = aq + (1-a)r
and the two nulls we have discussed are Ho: p=q and Ho: q=r but Ho:
p=q is equivalent to Ho: aq+(1-a)r=q which is equivalent to Ho:
(1-a)r=q-aq which is equivalent to Ho: q=r if a is fixed. Which is
why folks test the equivalence of proportions across a subpop and the
balance of the pop, and call it testing the subpop versus the
population, I suppose. But implementing the test of Ho: q=r is much
more straightforward.
i.e. Use -svy: tab- as far as possible.
On 12/5/06, Brent Fulton <[email protected]> wrote:
Using my sample, I wanted to use a statistical test to determine whether or
not to reject the null hypothesis, where there would be 50 null hypotheses
(one for each state). Michael Frone's Tuesday, December 05, 2006 8:27 AM
(PST) email provides a method for this. (Need to further investigate Austin
Nichols's reply.)
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
