In survival analysis, is it OK to standardize length of stay by dividing by length of followup? I am reviewing an article, so I cannot give a lot of detail, but the study looked at length of stay for patients who had very different amounts of follow up, in a situation where multiple admissions were possible. The authors divided LOS by length of follow up. They said they did this because they wanted to look at other variables.  They used Cox regression.
This somehow seems wrong to me, but I can't figure out exactly why. It seems that they could use a method designed for recurrent events, but I am not expert enough in those to say "You HAVE to do it this way" vs. "You COULD do it this way" vs. "it MIGHT BE better to do it this way" or to just say that what they did is OK.
 A: Unless there's some subtlety I'm missing, I don't see how such "standardization" could be valid.
Cox models might not take absolute event times into account, but they do depend on the correct ordering of event times. As you describe this, two individuals with the same dates of hospital admission and discharge (even over multiple admissions) would differ in terms of modeled event times if one simply had longer follow-up than the other. That makes no sense in a survival model. Standard Cox models certainly can take "other variables" into account, even if the values of those variables change over time. But the ordering in time among individuals needs to be respected.
I'd recommend asking the authors to either (a) document from the literature that their approach is nevertheless valid, or (b) use methods known to be appropriate for this type of data. As I suspect that transitions both into and out of the hospital are of interest, you could point them to the multi-state vignette of the R survival package, or the tools provided by the mstate package.
This seems to fit the structure of an "Alternating Two-State Process," discussed by Cook and Lawless in Section 6.5 of The Statistical Analysis of Recurrent Events, Springer, 2007.
