How to test this?

Hi everybody,

We are four students from the University of Nijmegen. We are applied psychologists. We are making a thesis about smoking behavior in front of entrances. There is a building were we do our intervention. Ofcourse we are doing a pre-post test here. Because we want to control for weather circumstances we use a control building on the same campus also pre-post test. We can't measure on an individual level. So we are observing if people smoke inside the smoking area or outside the smoking area. Categorical variables.

So we have

Building 1
pre-intervention postintervention
Inside area A C
Outside area B D

The same for Building 2

If we want to see if our intervention worked Chi-Square is OK. But how do we control for building 2 (like an ANCOVA for non-equivalent group design with controlgroup)? Is this possible with Chi-square or a different kind of analysis.

I hope someone can help us.

Thank you very much!


TS Contributor
Mantel Haenszel-Test: crosstabulate Inside/Outside x Pre-intervention/Post-intervention and use "Building" as layer. It will be tested (amongst other things) whether the association between inside/outside and time of measurement is different between buildings.

Or, use binary logistic regression with inside vs. outside as dependent variable, and as predictors: time of measurement, building, and the building*time interaction.


Thanks for you response Karabiner.

We are working on it, but the homogeniteity of variance is violated for the Mazel-Haenszel Test. Is there a way to correct for this?


TS Contributor
What do you mean by homogeneity of variance?
If you mean homogeneity of the odds ratios - then
that would be what you are looking for: a test of
whether pre-post changes are the same for both

With Kind regards

Again we are very thankful for your help!

OK, haha we cant find any info on this test. Andy Field doenst describe it in his book.

Which output do we need for effect size?

So we can conclude that there is an effect for intervention when controlling for pretest differences?



TS Contributor
Maybe you'll be better off if you use
binary logistic regression
with inside vs. outside as dependent variable, and as predictors
"time of measurement", "building", and the building*time interaction.

With Kind regards