Situation: clinical trial of cancer, two treatment arms, trt1, trt2, outcome (or say endpoint), PFS (progression free survival), totally about 600 days.
Use SAS 9.2 proc lifetest, we get two KM survival curves. Those two curves crossed several times over time.
Sorry, I cannot paste the plot here.
At first, trt1 curve is above trt2 curve (so trt1 is better, right?), then till around day 75, trt1 down to trt2, I mean they touch together, this is first time they touch together. Then trt1 is above trt2 again, till about day 125, then trt1 down to trt2 again, the second time they touch together. After this point, trt2 is above trt1 basically with a few times they touch together.
Also, by plot option in proc lifetest, we get the hazard function plot with Epanechnikov kernel and bw = 15. Basically, at first, trt1 hazard curve is lower than trt2 one; then trt2 curve is above trt1 curve till day 220.
Log-rank test in proc lifetest for strata is not significant at all, p_value = 0.83, also not significant for the Wilcoxon test, p_value = 0.57.
We also used the proc phreg, using only trt var as covariate. It is still not significant at all, p_value, ~0.83 for all three tests in phreg.
proc phreg also: Parameter estimate for trt variable: p_value is 0.83 and Hazard ratio is 0.959.
So my questions:
1. Can we say the proportional assumption is violated? If yes, can we say report single hazard ratio does not make sense in this situation?
2. For survival, can we say, at first, trt 1 is better than trt2, a few point they are same; then later, trt2 is better than trt1?