Confounding in logistic regression

#1
I have a question about confounding that I hope someone can answer!

My data comes from patients who were treated with Drug A or Drug B. They were not randomized, so there are inequalities between the groups (Drug A group has higher rates of diabetes and is older than Drug B group). I have conducted a polytomous logistic regression to identify significant factors (including diabetes, age and the use of Drug A or B) predicting 4 different adverse events.

My question is, since patients were not randomized to Drug A or Drug B, is confounding an issue here with the results of the regression? I thought logistic regression controls for confounding (at least by the independent variables that have been added to the model) so it shouldn't be an issue, except for unrecognized confounders.

Thanks for any help you can provide!
 

Dragan

Super Moderator
#2
I have a question about confounding that I hope someone can answer!

My data comes from patients who were treated with Drug A or Drug B. They were not randomized, so there are inequalities between the groups (Drug A group has higher rates of diabetes and is older than Drug B group). I have conducted a polytomous logistic regression to identify significant factors (including diabetes, age and the use of Drug A or B) predicting 4 different adverse events.

My question is, since patients were not randomized to Drug A or Drug B, is confounding an issue here with the results of the regression? I thought logistic regression controls for confounding (at least by the independent variables that have been added to the model) so it shouldn't be an issue, except for unrecognized confounders.

Thanks for any help you can provide!

What are the values of your sample sizes?
 
#3
This isn't confounding. It would be confounding if all of drug A patients had diabetes and all of your drug B patients did not.

As long as a reasonable amount of samples in each drug treatment that had diabetes and did not have diabetes you can still sort out the effect of the drugs and the effect of diabetes.

Now the lack of randomization for something as serious as a drug effect is a problem. The idea behind randomization is not to help you control for something you noted like "this guy has diabetes". If you could some how measure everything about your experimental unit you wouldn't need randomization. What randomization is for is to spread out equally the effect of all the stuff you didn't think about.

In terms of the diabetes you have an unbalanced experiment but what the hay right? The scary part is you are going to declare the effect of a drug and you have no idea if there are any systematic differences between the populations, and the point of randomization was that it was to spread out equally between all the treatments all the stuff you didn't think to note.



According to the head of my department when I took a course with him as far as detirmining the effect of the drugs you have nothing without randomization. He joked that he would say "sorry" as a consultant when someone told him they didn't randomize. I have no idea if it was classroom dramatics or if he was serious. Who knows right?

PS I am only a masters student so I could just be flat out wrong on each and every point.
 
#4
Thanks for the replies!

Dragan, my total sample size is 6,000.

Rounds, I agree, randomization is almost always the way to go. I'm doing an analysis using cohort data and comparing it to a similar analysis using RCT data, to determine if/how predictors on my four outcome groups differ in each. I was getting a bit worried b/c I keep getting comments about how I need to be worried about the confounders I've added to my model which I don't quite understand since logistic regression accounts for these...it's the variables that haven't been added to my model that are the confounding issue.
 

Dragan

Super Moderator
#5
Thanks for the replies!

Dragan, my total sample size is 6,000.

Rounds, I agree, randomization is almost always the way to go. I'm doing an analysis using cohort data and comparing it to a similar analysis using RCT data, to determine if/how predictors on my four outcome groups differ in each. I was getting a bit worried b/c I keep getting comments about how I need to be worried about the confounders I've added to my model which I don't quite understand since logistic regression accounts for these...it's the variables that haven't been added to my model that are the confounding issue.

Jen: You're fine. :)