analysing data when the two groups are very different

[Kyra]

Hi,

I would like to study association between few outcomes (binary) and types of feed- bolus versus continuous in a group of patients.

I used multivariate logistic regression to do the analysis and we found significant results at 0.05 level.

However, some are making comments that our continuous fed group is sicker and very different from bolus group.

However , i have all the confounders adjusted for  in my logistic regression.

I am wondering how to addresss that concern.

Thanks

[PaigeMiller]

Typically in a study of this type, you randomly assign patients to a treatment (either bolus or continuous). This reduces (but does not eliminate) the possible differences between the groups affecting the result.

But if you have covariates (I think this is what you mean by a confounders) and you include them in the analysis, the analysis should (if you do it properly) come to conclusions about the treatment after adjusting for the covariates. 

So its not clear, on a statistical basis, why there is a concern over one group being sicker than the other group, if the health/sickness of the group is a covariate. You didn't specifically say what the covariates are, and if they measure the sickness of a group.

[Kyra]

Hi,

For example:
Birth weight was significantly different between the bolus and continuous
groups. I added that as a covariate in my multivariate model.
However, Bell's criteria another variable was not significantly different
between the two groups on univariate analysis. I did not adjust for Bell's
criteria.
Commenters are saying that even variable is not significantly different ,
it makes the two groups different. i had decided not to add those extra
variables to multivariable analysis since our sample size is very small.

Thanks,

[PaigeMiller]

---

@Kyra wrote:


Birth weight was significantly different between the bolus and continuous groups. I added that as a covariate in my multivariate model.

---

Does birthweight measure sickness? Why do the commenters bring up sickness anyway? Do you have a variable that measures sickness directly?

Commenters are saying that even variable is not significantly different , it makes the two groups different. i had decided not to add those extra variables to multivariable analysis since our sample size is very small.

Yes, of course, the two groups are different, they will never be EXACTLY the same, that's impossible. But that's why we do statistics and things like randomization and include the effect of covariates. But with a small sample size, you are limited in what you can do. There's only so much you can learn from the data and you can't after the fact make the study perfect and you can't account for ALL possible sources of difference.

So there's really no definitive answer here, if the opinion of some people are that this is a meaningful difference even if the statistics don't show it, well that's just the way these things go.

[StatDave]

This sounds like a problem that can be addressed by causal analysis. See the Overview section in the documentation of PROC CAUSALTRT.