Comparing proportions with Crosstabs McNemar when one proportion is zero (SPSS)

[u72947389283]

Hi,

Thanks in advance for your help!

I'm investigating 3 diseases in one population. The expectation is
that they will all have similar prevalence and so finding a significant
difference is meaningful.

Below are some sample prevalence values for diseases A, B, C in the
sample population. Let's say that the sample population size is 500.

A: None (0%)
B: 2%
C: 80%

I can successfully run a McNemar Chi-Square Test for diseases B and C
using Crosstabs. The prevalence of diseases B and C are significantly
different. However, when I try to do the same for diseases A and C, I
get an empty row for Crosstabs/McNemar and a footnote of "Computed only
for a PxP table, where P must be greater than 1". My table is [Has
disease A: yes, no]x[Has disease C: yes, no], and since there are no
subjects who do not have disease A, an empty row results.

Any advice? I think I'm setting up the table correctly for McNemar.
If so, is there another algorithm that can compare two proportions in
one population AND that accomodates zero (0%) as one of the
proportions?

Note that if I can find the correct algorithm, A versus C will likely
be significantly different (0% versus 80%), while A versus B probably
will not be significantly different (0% versus 2%).

(I'm running SPSS 14.0 for Windows)


Thanks for your help!

[d***@gmx.de]

If you use McNemar (actually not a chi2 Test but a binomial test of the
changes 0->1 vs 1->0) test under "nonparametric"->"compare two related
groups" it will work.
Some of spss little surprises to keep you busy...
Daniel

[Bruce Weaver]

Post by d***@gmx.de
If you use McNemar (actually not a chi2 Test but a binomial test of the
changes 0->1 vs 1->0) test under "nonparametric"->"compare two related
groups" it will work.
Some of spss little surprises to keep you busy...
Daniel

I suspect it gives either the normal or chi-square approximation for
large samples though, doesn't it Daniel?

[d***@gmx.de]

I am not sure, the normal approximation for the McNemar test is :
Z = (| n+ - n- | - 1)/sqrt( N )
n+ are the changes from 0 to 1 and n- the other way around. If one of
the two n's is 0, you still would get a result. The exact version of
McNemar in cross tabs does not work either.
Simiar stramge as that you can do a tie-corrected Spearman only with
"correlation" in crosstab but not with spearman under "correlations"...
Daniel

Post by Bruce Weaver

Post by d***@gmx.de
If you use McNemar (actually not a chi2 Test but a binomial test of the
changes 0->1 vs 1->0) test under "nonparametric"->"compare two related
groups" it will work.
Some of spss little surprises to keep you busy...
Daniel

I suspect it gives either the normal or chi-square approximation for
large samples though, doesn't it Daniel?
--
Bruce Weaver
www.angelfire.com/wv/bwhomedir

[u224466]

Post by d***@gmx.de
If you use McNemar (actually not a chi2 Test but a binomial test of the
changes 0->1 vs 1->0) test under "nonparametric"->"compare two related
groups" it will work.
Some of spss little surprises to keep you busy...
Daniel

This works! Thanks for your help.