3

I am observing N devices over a period of time $[0,T]$ and counting the number of certain events $y_j$, for each device, during that period of time. I also have some specifications $x_j$ for each device (explanatory variables). I was thinking of conducting Poisson regression on these data. However, I realized that some devices fail before the observation period ends.

Is it still valid to use Poisson regression? Can I include a dummy variable that indicates whether the device failed or not?

kjetil b halvorsen
  • 77,844
Possum
  • 76
  • what do you mean events $Y_j$? is it a fail? If there is a fail then it is impossible that event $Y_j$ will happen I would not include those fails at all. For example you should not include dead people in your survival analysis. – Deep North May 04 '17 at 10:19
  • @DeepNorth No, $Y_j$ is not a fail. It is basically another non-fatal phenomenon. – Possum May 04 '17 at 10:27

1 Answers1

3

There are two situations here:

  1. You know when each device failed, so you know the total time it was functioning. Then you have, for each device a variable $OT_i$ (observation time $i$) that you use use (after logging) as an offset in the Poisson regression. See Offset in Poisson regression Or, if you know the OT only approximately, that could still be a useful approximation.

  2. You do not know, even approximately, when the device failed. Then the situation is more difficult, and we would need (much) more details and context for advice.

kjetil b halvorsen
  • 77,844