What you do depends on what exactly is the research question. Nonetheless, to make this a typical logistic regression problem you need a yes/no outcome. You could form that by having the clinicians select a meaningful threshold, like TV reduced by 10%, yes/no. That makes for a logistic regression, but dichotomizing data often reduces power, and sort of muddies the interpretation. In my example, a reduction of 9.9% is a 'no' while 10.00 is a 'yes', but is there a meaningful difference in these two volumes? Relative % change = (TV1-TV3)/TV1 = 1-TV3/TV1. Even if TV is normally distributed the ratio TV3/TV1 almost surely isn't normal. From your data example it doesn't look like TV2 is used. So you could use the Kruskal-Wallis test, available with the Wilcoxon option in proc npar1way, to compare groups and not have to be concerned with parametric assumptions. If the overall test is significant, the DSCF for the pairwise comparisons. There is a literature, mostly based on simulations, suggesting that ANCOVA with TV1 as the covariate is the preferred analysis. Here is an easy read on the subject: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC34605/
... View more