Discrete survival analysis and time-varying covariates

[Chris Rooney]

Hello all,

Much of the focus in Stata is on continuous time survival data and there really isn't much information regarding incorporating TVCs in discrete time survival analysis. My 'problem' is the following (not really specifically to do with Stata - more of a layout issue):

My reading of the scant exercises on discrete time survival analysis is that one can use some commands to expand the dataset into a person-period dataset. However, I have a variable 'Residence' (1 = in university residence, 0 otherwise) which is recorded for each year that a student is at university. I am unsure how to lay it out - should Year 1, Year 2 etc. be separate column variables? Also, I have a series of cohorts (students who started in 2006, students who started university in 2007 etc) which effectively means the 'Year 1' would be different depending on when the person entered university. Of course, the more 'natural way' to lay out the data is to record the Residence status in one column, and use the row to separate the years. However, what will Stata do when I use the command to expand it into a person-period dataset?

I've attached an example. Obviously, I know that when I actually do the survival analysis, the format will be like in the second part in the image. My question is can I put it in that format before using the person-period command?



I suppose this question also applies to cumulative GPA after year 1, year 2 etc...

[wbuchanan]

Please read the FAQ prior to posting. Although everyone is willing to help where possible, many of us - myself included - appreciate the professional curtesy extended by using your full real name in your posts (or at least some reference to your name). One thing that I would think could be problematic is the use of cumulative GPA rather than period specific GPA and how it introduces a term that is highly correlated and dependent on itself over time. The same may also be true regarding residency status depending on the college; my undergrad university - Berklee College of Music - would not allow students to move into on campus housing if the student elected to live off campus at any point (e.g., on campus as freshman, off campus for sophomore year, no longer eligible for on campus housing). I'd also venture to say, without having done much research in this right now, that there are likely to be several folks within the Stata user community that have and do regularly deal with these issues on a regular basis. It might be a bit more helpful if you provide some context around the goal(s) of your study to get some better advice from the user community. There are also two books available from StataPress dealing with different approaches to survival analysis that could provide you with style guidance that you are looking for as well.

[Chris Rooney]

Thanks for the reply. I've had a look at the user options and it seems I can't change my username. Am I correct in assuming that I will have to contact a moderator for my username to be changed?

Yeah, good point about cumulative GPA. The aim of my research is to look at factors which explain the likelihood of someone graduating.

[Stephen Jenkins]

(1) to change your username is easy: contact the administrator through the "CONTACT US" link in the strip at the bottom of the webpage (the strip is dark blue in my browser). Please action this, for the reasons cited.

(2) Data organisation: I am puzzled about what you're asking. Once you have the data in "person-year" format, then you can 'slot in' any time-varying covariate you like. The data for each student (in your case) occupies as many rows as the number of years that s/he is at risk of experiencing of experiencing the event of interest (time from entry to college until graduation from college?). Explanatory variables (of all types) are in the 'columns' (though I would avoid this spreadsheet-like nomenclature). "Residency status" (binary indicator) is one of these variables. If Joe Student is in residence in year t of his "spell" (corresponding to calendar year 2010, say), then "Residency status" = 1 and, if he is not, "Residency status" = 0.
Have you studied the resources on such issues, e.g. http://www.iser.essex.ac.uk/survival-analysis ?