Using pgmhaz with different times of becoming at-risk

[Gabriella Monahova]

-1down votefavorite

I am trying to estimate a discrete time proportional hazards model using the pgmhaz program in Stata (version 13.1). From this manual: https://www.iser.essex.ac.uk/files/t...-39-pgmhaz.pdf it seems like the program assumes that all individuals become at risk at the same time (t = 0). Is it possible to allow for different times of becoming at risk? Thanks!

[Steve Samuels]

Welcome to Statalist, Gabriella! To answer your question, we need a bit more information. Please describe the study design and question, including your population, your outcome, and what your your "times" are ( calendar months? years? something else?); what the time zero is; and why some individuals enter after time zero.

[Gabriella Monahova]

Hi Steve, thanks for your response!

I am studying decisions by firms to replace existing technology and I want to use a proportional hazards model since I am interested in the timing of replacement. I have a sample of firms which adopt and then possibly replace a technology. For each firm, I observe the year of initial adoption as well as the year of replacement for those firms that do replace. Thus, the "times" are years. Time zero is the first year in which I observe initial adoptions (a firm only becomes at-risk of replacement once it adopts) and some firms do not enter at time zero because their initial adoption happens later. I have both time-invariant and time-varying covariates that I want to include in the model.

Id be happy to provide additional information. Any help is greatly appreciated!

[Steve Samuels]

Okay, I understand. But you are mixing up two time-scales: calendar year and time from initial adoption. You can define the time zero to be the year of initial adoption, and then use the second scale (time from initial adoption). In discrete analysis, we don't use an actual time zero: pgmhaz8 requires that you have a sequence variable, specified with the seq(seqvar)that starts with 1. You can use calendar year of initial adoption as a fixed covariate.

Stephen Jenkins, who wrote pgmhaz8, has a wonderful web page of survival data information at:
https://www.iser.essex.ac.uk/resourc...sis-with-stata