Date: Wed, 3 Nov 2004 13:59:15 -0800
Reply-To: huixin fei <hxfei@YAHOO.COM>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: huixin fei <hxfei@YAHOO.COM>
Subject: Re: chi square follow-up tests and type I correction
Content-Type: text/plain; charset=us-ascii
Dale suggested using the GLIMMIX procedure (SAS V9.1) a couple of month ago. I re-attached this message in this list.
> Dale McLerran <stringplayer_2@YAHOO.COM> wrote:Stephen,
> I don't believe that SAS has such a procedure. However, it could
> easily be implemented using data step code. Whether you should
> or not is another question. I am not sure that the Marascuillo
> multiple proportion comparison procedure is all that great. In
> the example given at the website that you referred to, the global
> test for differences among the proportions is significant at the
> alpha=0.05 level. However, none of the pairwise contrasts are
> significantly different. This seems quite peculiar, especially
> since the number of samples for each of the five lots being
> compared are equal. The global test has enough information to
> find significant differences in the sample defect rate, but there
> is not enough information to identify even one contrast where
> there is a difference?
> Now, how about an easily implemented alternative. An alternative
> to the chi-square test for equality of proportions would be to
> fit a logistic regression model with defects as the response and
> lot as the predictor variable. Now, the procedures LOGISTIC,
> CATMOD, and GENMOD can all fit a logistic regression, but none
> have an LSMEANS statement which allows multiple comparisons.
> (GENMOD has an LSMEANS statement, but it does not support
> multiple comparisons testing. The other two do not have any
> LSMEANS statement.) However, you can use the GLIMMIX macro - or
> the new GLIMMIX procedure! - to fit a logistic regression model.
> Both the GLIMMIX macro and procedure have LSMEANS statements which
> allow multiple comparison tests to be performed.
> Let me demonstrate with the data presented on the website that
> you referred to.
> /* Construct data of defect rates in 5 lots */
> data test;
> lot=1; defects=36; link binary;
> lot=2; defects=46; link binary;
> lot=3; defects=42; link binary;
> lot=4; defects=63; link binary;
> lot=5; defects=38; link binary;
> /* Output a record for each of the 1800 observations */
> do i=1 to defects;
> do i=1 to samples-defects;
> /* Construct the usual chi-square test that the */
> /* proportion of defects are the same across all lots */
> proc freq data=test;
> tables defect*lot / chisq;
> /* Use the GLIMMIX macro to */
> /* 1) construct a global test that can be compared */
> /* with the chi-square test formed above */
> /* 2) form and test pairwise comparisons, allowing */
> /* for multiple comparison adjustments */
> stmts=%str(class lot;
> model defect = lot;
> lsmeans lot / diff adjust=tukey;
> You will observe that the overall F-test for the effect of LOT
> has very nearly the same p-value as the chi-square test.
> Asymptotically, the two tests would be equivalent. We may
> actually prefer the overall F-test obtained from the logistic
> regression - especially if the sample sizes were small. Here,
> sample size is large enough that the two approaches yield very
> similar global test statistics.
> Having satisfied that the global test is providing an equivalent
> overall test, we can look at the pairwise comparisons obtained
> out of the LSMEANS statement. I employed a Tukey multiple
> comparison adjustment. You will observe that three comparisons
> are identified as significant
> 1) the contrast between lots 1 and 4
> 2) the contrast between lots 3 and 4
> 3) the contrast between lots 5 and 4
> In addition, the contrast between lots 2 and 4 is significant
> at alpha<0.075. None of the other contrasts even approach
> significance. Thus, our test identifies significant pairwise
> contrasts that the Marascuillo test does not.
> It is easy to generalize this to a model where you have multiple
> factors (lots within week and lots between weeks). I think that
> it would be best to construct your data with day of the week
> and week. Then you could test whether there were more failures
> on, say, Monday or on Friday. A priori hypotheses about day
> of the week may be dealt with employing GLIMMIX. You probably
> do not want to employ multiple comparison tests where an a priori
> hypothesis could be constructed. You would likely not have any
> a priori hypothesis that a certain weeks would have more failures
> than other weeks. Thus, you could use multiple comparison
> adjustments to test pairwise differences between weeks. You
> could test specific day of week contrasts without resorting to
> the penalty that a multiple comparisons procedure produces.
> Note that the GLIMMIX procedure is available only for SAS version
> 9.1 with Windows OS. The GLIMMIX macro has been distributed
> with SAS for quite a few years, so should be readily available.
> --- Stephen Arthur wrote:
> > Hello,
> > Is the Marascuillo multiple proportion comparison procedure
> > in
> > SAS?
> > http://www.itl.nist.gov/div898/handbook/prc/section4/prc474.htm
> > I did a search on the SAS website and the searches:
> > 1) Marascuillo
> > 2) +"multiple comparison" and proportion
> > turned up zero results.
> > http://sas.com/search/index.html
> > Also, does anyone know of a two-way Marascuillo multiple proportion
> > comparison procedure... maybe this doesn't make sense, I'll have to
> > think
> > about it some more.
> > Basically, I have an analysis situation where I want to test the
> > significance of proportions within a week, and between weeks. The
> > data is
> > very randomly collected (unbalance sample sizes and missing data),
> > these
> > are not controlled studies.
> > Thanks,
> > Stephen
> Dale McLerran
> Fred Hutchinson Cancer Research Center
> mailto: email@example.com
> Ph: (206) 667-2926
> Fax: (206) 667-5977
I really don't know the MULTTEST procedure at all. You might
be able to perform appropriate post hoc multiple comparisons
for proportions employing the MULTTEST procedure. I simply do
not know. Sorry.
--- huixin fei <firstname.lastname@example.org> wrote:
> Thank you very much for this information. I have to test the
> proportions among the
> treatments (and do “post hoc?multiple comparisons) in a biological
> environment, a
> similar problem. I tried to use multiple Z tests with Bonferroni’s
> Now I can try the GLIMMIX procedure. Could I use the Proc multtest?
> I just started to learn this procedure. Thanks.
> Huixin Fei
Serge Sévigny <serge.sevigny@PSY.ULAVAL.CA> wrote:Hi,
suppose we have a significant chi square on a 3 (a, b, c) by 4 ( 1, 2,
3, 4) crosstab. Now we know that the proportions differ somewhere. But
we want to know exactly where :
A) Are the proportions between a and b equal? Are the proportions
between a and c equal? Are the proportions between b and c equal?
1- what would be the most powerful test(s) to investigate this follow-up?
B) Then, suppose we find that the proportions between a and b are
different, how do we check if they are different at 1, or 2, or 3, or 4?
2- what would be the most powerful test(s) to investigate this follow-up?
3- For all follow-up tests, do I use a bonferonni correction?
4- Do you know of a good reference book on that matter?
Thanks for helping!
Do you Yahoo!?
Check out the new Yahoo! Front Page. www.yahoo.com/a