Pooling data for a McNemar test

[amorphia]

Hi all,

I have a very straightforward design for an experiment and I want to
do a straightforward McNemar test, but I'm not 100% certain it is
acceptable to pool data in the way I propose to.

First I measure subjects' baseline preference for either A or B (as a
binary variable), then I perform one of either of two manipulations,
both of which are essentially the same kind of manipulation except one
is predicted to shift preference towards A and one towards B, then I
measure the post-manipulation preference.

In a sense therefore there are two groups, one predicted to shift
towards preferring A and one towards B. But I'm not really interested
in this difference: I just want to know if my manipulation generally
works to shift preference.

It is obvious to me that each group can be individually McNemar
tested, i.e. compare within each group the proportion who shift A to B
against the proportion who shift B to A. My question is, can I do one
single test with the groups pooled, comparing the proportion shifting
in the predicted direction against the proportion shifting in the non-
predicted direction?

Cheers,

Ben

[Bruce Weaver]

McNemar's test is equivalent to a two-category chi-square goodness of
fit test, where the two categories are change in one direction versus
change in the other direction, and H0 states that equal proportions
fall in the two categories. From this point of view, I see no problem
with a test where the two categories are "change in the predicted
direction" and "change in the other direction".

--
Bruce Weaver
bwe...@lakeheadu.ca
http://sites.google.com/a/lakeheadu.ca/bweaver/
"When all else fails, RTFM."

[RichUlrich]

On Wed, 24 Sep 2008 08:31:11 -0700 (PDT), amorphia
<spam.o...@gmail.com> wrote:

>Hi all,
>
>I have a very straightforward design for an experiment and I want to
>do a straightforward McNemar test, but I'm not 100% certain it is
>acceptable to pool data in the way I propose to.
>
>First I measure subjects' baseline preference for either A or B (as a
>binary variable), then I perform one of either of two manipulations,
>both of which are essentially the same kind of manipulation except one
>is predicted to shift preference towards A and one towards B, then I
>measure the post-manipulation preference.
>
>In a sense therefore there are two groups, one predicted to shift
>towards preferring A and one towards B. But I'm not really interested
>in this difference: I just want to know if my manipulation generally
>works to shift preference.
>
>It is obvious to me that each group can be individually McNemar
>tested, i.e. compare within each group the proportion who shift A to B
>against the proportion who shift B to A. My question is, can I do one

I don't understand. If the groups are created according to
a binary preference for A or B at the start -- How do you construct
a McNemar test on a group that is all A? How can they shift
in "the non-predicted direction" if they already are A?

>single test with the groups pooled, comparing the proportion shifting
>in the predicted direction against the proportion shifting in the non-
>predicted direction?
>

--
Rich Ulrich