# Thread: Change from baseline in proportion

1. ## Change from baseline in proportion

Dear all

Can anyone give me some advice for the following problem:
I'm analyzing a binary outcome (patients being able to perform a specific task yes/no). It is a longitudinal study with repeated measurements of the outcome and 2 treatment groups.
We want to know if the change in proportions of patients (from baseline) is the same in both treatment groups.
I thought to analyze the question with a GEE (with logistic link). The only disadvantage I see is, that we need to report all measurements then as OR. Also it is not clear to me if we report the treatment effect, this would be a difference in OR (difference in OR from baseline between the two groups). To report a difference in OR seems for me quite uncommon. Is there another way we could analyse the data?
Also I am not sure: Is usually the change from baseline for a binary outcome reported as a OR (this would be the case with the GEE).
Thanks in advance for any help.

Best regards
Lisa

2. ## Re: Change from baseline in proportion

Treatment randomized?

You might be able to look at risk differences.

3. ## The Following User Says Thank You to hlsmith For This Useful Post:

sternchen (08-06-2015)

4. ## Re: Change from baseline in proportion

Odds ratios compare groups by computing ratios of the odds of "success", so it makes sense to also compare effects across groups as ratios of odds ratios. In fact this is exactly what you get when you compute an interactions effect. See e.g. here.

5. ## The Following User Says Thank You to maartenbuis For This Useful Post:

sternchen (08-06-2015)

6. ## Re: Change from baseline in proportion

The study is not randomized.
How can I calculate risk difference taking into account the repeated measures?
Thank you.

7. ## Re: Change from baseline in proportion

MB, makes a point.

They are both just two different scales to present outcomes. In one you find the relative difference via division. In the other you just subtract the two values.

Without randomization, you may need to control for other variables in this model as well if they predict the outcome.

Also, other variables may be associated both with treatment assignment and outcome (if not, you at least have to show this is the situation). This falls under the assumption of "ignorable treatment assignment" given as: yo, y1 are independent of probability of treatment assignment given controlling for confounding covariates in the model. y0, y1 are the two outcomes.

 Tweet

#### Posting Permissions

• You may not post new threads
• You may not post replies
• You may not post attachments
• You may not edit your posts