Correct statistical test for 2 group within a group for two continuous variables

[Annabelle Coleman]

Dear STATA List users,

I have a statistical test question that I have been trying to tackle.

I have two main comparable groups: those with and without a PEG insertion. Within each of these groups, I have matched sub-disease groups including Advanced, Moderate, Early and Unknown.

I am trying to compare whether weight has changed over time in these participants, between PEG and no-PEG, within each disease subgroup. e.g., is weight different between time point 1 and time point 2 between PEG versus no-PEG, in "advanced" subgroup.

The variables I am currently working with includes Weight1, Weight2, PEG group, and sub-disease group where weight is a continuous variable and the groups are categorical.

I have been wondering how to do this, and have realised that the data is longitudinal. To use the weight data in a multiple linear regression (taking into account age and gender as covariates), would I use the difference between the two weight data variables?

Therefore the code would look like:

regress weightdifference i.PEG_group Age gender

My main question is about whether looking at the difference in weight as a comparable measure is the correct approach?

I look forward to your responses.

Kind regards,
Annabelle

[Carlo Lazzaro]

Annabelle:
if you're dealing with a longitudinal study, my guess is that you'll be better off with a panel data regression with fixed effect (see -xtreg,fe-).

[Joseph Coveney]

The wisdom of using change scores comes up frequently on the list. See this post for a recent discussion of the topic. You'd probably be better off fitting a longitudinal model as suggested there. Something along the lines of reshape long Weight (you don't show anything and so I don't know what your patient identifier is called in your dataset), and then fitting a linear mixed model with a three-way interaction term of percutaneous endoscopic gastrostomy treatment group, disease severity category and observation time point predictors.

[Rich Goldstein]

I agree very strongly with Carlo Lazzaro and Joseph Coveney ; in addition, I note that your model will be subject to regression to the mean issues making it, at best, very hard to interpret