Consider a survey on two schools, where the same students are asked, in the begining of the school year and at the end, whether they like sports, or not. In one of the schools (intervention) there are sport promotion activities during the school year. The question of interest is whether the proportion of change (doesn't like --> like, or likes --> doesn't like: it should be symmetrical, right?) in the school subject to the intervention is statistically different from the change observed in the control school. That is, if the intervention is able to generate 'more change', with statistical significance, in the intervention school.
Within a school, I believe one can apply the McNemar test for a paired 2x2 table. But both schools can have (within school) a significant change. So, how can we test if the intervention is associated with a statistically significant change, when compared to the change observed in the control group?
I wonder if the answer can be extended to a situation where the answers are classifications of 3 or more classes (e.g., doesn't like, indiferente, likes, ...).
I thank you in advance!, and hope this questions maybe helpful to many others, besides me.
In your design, "school" is confounded with "intervention". I wouldn't give too much on the data. Only when the difference is huge, it will be suggestive for an intervention effect (one still has to discuss critically the diffences beween the schools [structure, location, social embedding, etc]), and then you will see the difference with your naked eye.
I'd recommend a simple F-Test via ratio or OR as power chi-square distributed.
Hope this helps,
Matt
You can use the SAS proc freq (for the McNemar test, i.e. with the agree option) to give you the Summary Statistics (overall Kappa, Test for equal kappas, and Cochran's Q for the interaction) controlling for School like in this example:
options ps=400 ls=110 nodate nonumber pageno=1 formdlim="";
data a;
input school hus $ wif $ count;
cards;
1 yes yes 20
1 yes no 5
1 no yes 10
1 no no 10
2 yes yes 20
2 yes no 5
2 no yes 10
2 no no 10
;
proc freq order =data;
weight count;
tables school*hus*wif / agree;
run;
I would try an ordinal mixed model with intervention and time under interacction. Since you have 2 observations for the same student you have to add the individuals has a random intercept. The interacction result would tell you whether the students who had intervention like more sports in time 2, than the students who havent. Additionally you could include some confounding factor to the analysis.
In R you can use the ordinal package with the function clmm.
- Badges
- Science topic