0
$\begingroup$

I understand the deviance residuals $r_D$ for a Poisson GLM with log link function are given by $r_D = \mu_{ij} \log(\mu_{ij}/\hat{\mu}_{ij}) + (\hat{\mu}_{ij} - \mu_{ij})$

I was wondering though what the formula would be for a Poisson GLM with identity link function? Is it the same or not?

$\endgroup$
  • $\begingroup$ I don't think your formula is correct. (For example, note that the $i$th deviance residual is a function of the $i$th observed, $y_i$, which your fomula lacks). Where did your formula come from? $\endgroup$ – Glen_b Mar 18 at 11:12
  • $\begingroup$ I saw this formula mentioned in an answer posted here stats.stackexchange.com/questions/99065/… - what would be the correct formula then? $\endgroup$ – Tom Wenseleers Mar 18 at 12:04
  • $\begingroup$ I think that answer conflates two different things. I've put what I believe is the correct formula into my answer. $\endgroup$ – Glen_b Mar 18 at 12:07
1
$\begingroup$

The deviance of the $i$th observation - and hence the corresponding deviance residual - is determined by the distribution family; it is not affected by the link function (except in that the link function affects the estimate of $\mu_i$).

I think the $i$th deviance residual for the Poisson model is

$\text{sign}(y_i-\hat{\mu}_i)\sqrt{2\{y_i\log(y_i/\hat{\mu}_i)-(y_i-\hat{\mu}_i)\}}$

$\endgroup$
  • $\begingroup$ Many thanks for this - much appreciated - makes sense! $\endgroup$ – Tom Wenseleers Mar 18 at 12:04
  • 1
    $\begingroup$ It might be worthwhile to add that the deviance residual is defined to be zero when $y_i$ and $\hat\mu_i$ are both zero. $\endgroup$ – Gordon Smyth Mar 19 at 23:59

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.