3

Let us say I would like to adjust for centre effects in the analysis of data from a multicentre clinical trial.

In the fixed effects approach one simply chooses one centre as reference and includes a dummy variable for each of the remaining centres.

A problem with that method is that the number of parameters increases at the same rate as the number of centres.

In my example, I have 300 centres and 2 patients in each, and I think that the following R error is due to the above-mentioned problem:

In fitter(X, Y, strats, offset, init, control, weights = weights,  :
Ran out of iterations and did not converge

Question

Is there a statistical/mathematical reason why the estimation procedure does not converge, or is it because the computational requirement is beyond the capacity of the software/my computer?

PS I have fitted a Cox proportional hazards model but I guess this is not very relevant with respect to the question.

ocram
  • 19,898
  • 5
  • 76
  • 77
  • IMHO, if you have only two patients per centre, you do not have enough data to calculate centre effects *by far*, and if R gave you any results, they’d probably be useless or misleading. You could cluster your centres into a few categories, if that makes sense with your data. – mzuba Sep 13 '11 at 13:02
  • @mzuba: Indeed, I realize "I do not have enough data"... but I am looking for an argument that is more rigourous...if any... thx – ocram Sep 13 '11 at 13:42
  • Related Q : [How to conduct conditional Cox regression for matched case-control study?](http://stats.stackexchange.com/questions/2748/how-to-conduct-conditional-cox-regression-for-matched-case-control-study) – onestop Sep 13 '11 at 14:06
  • @onestop: Thank you for the related Q. Actually, frailty models offer a (better) alternative to the fixed effects approach. However, I'd like to deeply understand why the fixed effects approach fail to converge. – ocram Sep 13 '11 at 15:00

0 Answers0