2
$\begingroup$

There is one medical data set which contains 3 treatments: A, B, and C. The response variable is Yes/No. However, due to a specific medical reason, the dataset needs to be split into several sub-datasets. The major research question is to compare the difference of proportions between $P_A$, $P_B$ and $P_C$.

If I did not need to split data, I could use a "simultaneous confidence interval" {Agresti et al., (2008)} to perform a multiple comparison between these three treatments. However, I can't find an appropriate statistical method to compare multiple proportions when a dataset is split into several sub-datasets. Can someone please give me a hint on this?

Many thanks in advance.

$\endgroup$
3
  • $\begingroup$ Your activity page still shows you have accepted none. You can accept answers by clicking the green tick next to an answer to a question you have asked (if you just clicked the 'normal distribution' like picture, you upvoted it, which is also nice, but not the same as accepting it). $\endgroup$
    – Nick Sabbe
    Nov 18, 2011 at 14:43
  • $\begingroup$ Hi Nick, after I click the activity page, It does not show anything. What should I from there? I did not see any green tick. Thanks. Tu $\endgroup$
    – Tu.2
    Nov 18, 2011 at 17:52
  • 1
    $\begingroup$ If you look at e.g. my answer below, you have the number of votes probably top left of that (for now that's a 2 - it has an arrow up and down around it that look like normal distributions). If I'm not mistaken, just below that, there should be an empty tick right below that. As you are the one who asked the question, you are the only one who can say which of the answers was the one that helped you most (or more technically, which answer was most to the point). You do that by clicking in the empty tick, which 'accepts' that answer (and fills the tick). $\endgroup$
    – Nick Sabbe
    Nov 20, 2011 at 21:46

1 Answer 1

2
$\begingroup$

Not exactly a ready-made solution, and there are as always a lot of things not accounted for here (regarding the way the treatments were randomised over you subsets), but:

Build a logistic regression model with dummies for treatments B and C (or any set of two that will make interpretation the easiest) and dummies for all but 1 of your subset indicators. If you have reason to think the treatments may work differently in the different subsets (and it is on the safe side to think so), also include the interactions.

In R, this is done through mdl1<-glm(out~treatment*group, data=yourdataset)

Then build a model not holding the treatments: mdl2<-glm(out~group, data=yourdataset)

Now you can do a likelihood ratio test between the two models (check ?anova) to see if there is a reason to include any of the treatment effects/interactions (perhaps you could do this in two steps to first try to eliminate the interactions).

If you want to answer other research questions (e.g. differences per group), you need to compare other sets of models, but the idea is always the same: compare models with and without the variable that represents what you are interested in.

When you do this repeatedly, you are indeed reusing the same data to some extent to check different questions, so you need to do some form of multiple testing correction. Depending on your specific research questions (and of your design), (orthogonal) contrasts may be of aid to you. If not, you can resort to Family Wide Error Rate (if your sample sizes are large enough, the power may still be OK) or, if this is an option, to False Discovery Rate.

I repeat my warning that there are many issues with the methods presented here, but it should get you started.

$\endgroup$
0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.