Paired T-test adjusted for baseline Characteristics

#1
I am running a paired sample t-test (weight before and after treatment) however I want to see if the results are still valid when I adjust for sex, ethnicity and deprivation. Can anyone advise on how to do this?

Thank you
 

hlsmith

Less is more. Stay pure. Stay poor.
#3
Yup - @Dason is right on. What is your actual study question and was there a randomized intervention? There is currently a paper getting ready to come out soon in International journal of Epidemiology (dont know when) that will be relevant to you understanding the limitations of framing change point models.
 
#4
Run a regression instead
Thank you. Do you mean that I should take the difference in pre/post weight and make that the dependent variable and then run it with sex/ethnicity/deprivation as the independent variable? Would I need to run separate regression for all the independent variable or can I add them all into the same analysis?
 
#5
It w
Yup - @Dason is right on. What is your actual study question and was there a randomized intervention? There is currently a paper getting ready to come out soon in International journal of Epidemiology (dont know when) that will be relevant to you understanding the limitations of framing change point models.
Thank you. It was not randomised, it was just looking at intervention effect (natural experiment) but I want to see if that varied based on the baseline characteristics.
 

Dason

Ambassador to the humans
#7
Thank you. Do you mean that I should take the difference in pre/post weight and make that the dependent variable and then run it with sex/ethnicity/deprivation as the independent variable? Would I need to run separate regression for all the independent variable or can I add them all into the same analysis?
You would want to add them all into the same model I would think. You *could* manually compute the differences and use that as the dependent variable. That would be more analogous to what a paired t-test does.

But the other option is to use the post weights as the dependent variable and include the pre weights in the list of independent variables. Mathematically if the estimated coefficient for the pre-weight is 1 then you'd essentially be doing the same thing (for inference) as taking the differences directly.

If the coefficient is 0 it would suggest the pre weight is unrelated to the post weights so it would be more akin to how one might analyze a completely randomized design where we don't know the pre weights.

So by going this route you'd kind of let the data choose what is appropriate.
 
#8
So everyone got the intervention (natural experiment)?
Yes, the intervention was a weight loss programme. It was given to everyone and the outcome was how much weight they lost. I can see that the amount of weight loss was significant I just now want to see if it is still significant when I account for age, gender, ethnicity and deprivation.
 
#9
You would want to add them all into the same model I would think. You *could* manually compute the differences and use that as the dependent variable. That would be more analogous to what a paired t-test does.

But the other option is to use the post weights as the dependent variable and include the pre weights in the list of independent variables. Mathematically if the estimated coefficient for the pre-weight is 1 then you'd essentially be doing the same thing (for inference) as taking the differences directly.

If the coefficient is 0 it would suggest the pre weight is unrelated to the post weights so it would be more akin to how one might analyze a completely randomized design where we don't know the pre weights.

So by going this route you'd kind of let the data choose what is appropriate.
This is really helpful - thank you very much @Dason!
 

hlsmith

Less is more. Stay pure. Stay poor.
#10
You would want to keep the preterm in the model in case base weight influences how much weight they can and do lose. For example, if someone is 5 pounds overweight they only have the ability to lose 5 pounds. There may also be a nonlinear relationship between pre-rweight and outcome, so this may need to be explored via polynomials or a spline term.

Side note, without a control group you can't conclusive attribute all changes in weight to the intervention.