First, you've got a set-up problem on the left.
With PHREG, the first variable is the days until something happened. The second variable is whether what happened was a censoring event or a real event. So your ANC500 needs to be restructured to combine the dates to engraftment with days to death and days to last follow-up (or declaration of failure to engraft). Then your Cens_ANC500 can have multiple codes representing the different conditions at the end.
Because you have (at least) three endpoints, you have a competing risks situation and need to either do a multiple event model or a sensitivity analysis (combining different event types).
It may be conceptually easier to reframe your model to have the outcome of interest being "failure to engraft" and then you just censor on engraftment. Censoring on death often leads to a "waiting time bias" that needs to be carefully evaluated (See the article by Mitch Gail about the Stanford Heart Transplant data (Annals of Internal medicine, 1972) for the classic example.).
On the right of the = go the candidate explanatory variables, age, sex, etc.
Doc Muhlbaier
Duke