Can I use an indicator variable to indicate a repeated measure? I have 2 groups of patients who did not do lose enough weight after their lap band procedure so they went on to either have a gastric sleeve resection (S) or a more aggressive roux-en-y surgery (R).
I have their initial weights (weight_i), the weight after the 1st surgery and before the 2nd surgery (weight_pre) and their weight after the 2nd surgery (weight_post).  Weight_post is my dependent variable for the model.  Covariates include sex, age, diabetes status, height, and type of surgery.
I'm using SAS and as expected, the 3 weight measurements are highly correlated and when I fit the full model their VIF is very high. I've looked at using PROC MIXED or PROC GLM with a REPEATED measures option but that doesn't help me with model selection.  
So I created an indicator variable for when the weight's were measured (0-initial, 1-pre, 2-post), created a long dataset with a new weight dependent variable and added the indicator to the model and did my model selection.  
Is this a cheap way of getting around it or am I doing something completely invalid??
I also reran the model with just the indicator = 1 or 2 and put the weight_i back into the model (by itself it has no problems with collinearity).  I thought this way would make the model more interpretable.
Would love any thoughts/feedback, thanks!
 A: As a first step one should plot the weights on histograms to see what kind of distributions they are. Next, I would break the patients that did not lose enough weight into two groups, so that we can compare the same patients to their own surgical results. Then we can ask if surgical procedure A decreased weight significantly or not using a paired test, which is more powerful than an unpaired test. Rather than regression, all we have to do is a Wilcoxon test for difference of signed medians, that is, if our histograms do not show normally distributed variables, and paired t-testing of difference of means if they are normal. Either way, that gives us two probabilities Surgery A and Surgery B. Then, we look at the probabilities and the more significant one is likely better. Then the average(normal) or median(not normal) difference in weight can be stated, and their confidence intervals given.
See comment below. Yes, but for paired data. It does not matter how many had one surgery or the other, but it does matter that each patient is used to compare his (her) pre-surgical and post-surgical results first for surgery A then repeat the process for a different number of patients for the surgery B group. Yes, indeed the Wilcoxon signed-rank test. 
There is also a Wilcoxon rank-sum test for independent variables. The Wilcoxon-Mann-Whitney test (for independent variables and different numbers of patients) you could use to compare differences in weight pre- and post-surgery in group A and to the differences in group B. That is, you sneak in the powerful paired information by taking the differences and then compare one group of differences to the other group of differences.
