1
$\begingroup$

Normal QQ-plot

I have tried to fit a glm to some weather data and I got this weird qq-plot. What could this possibly mean? I am aware of how various skewed error distribution qq-plots should look like, but what I have here seems more bizarre.

$\endgroup$
5
  • 1
    $\begingroup$ What kind of GLM was it? GLMs shouldn't necessarily have normal residuals (see here). $\endgroup$ Apr 16, 2015 at 21:06
  • $\begingroup$ What is nF? Is it a limited range DV? $\endgroup$
    – iacobus
    Apr 16, 2015 at 21:28
  • $\begingroup$ Is this a possible duplicate? (However, gung's point is relevant; the interpretation may not matter much, depending on omitted details about the GLM) $\endgroup$
    – Glen_b
    Apr 17, 2015 at 4:53
  • 1
    $\begingroup$ Is this maybe a Poisson regression with small counts? $\endgroup$
    – Glen_b
    Aug 27, 2015 at 7:35
  • $\begingroup$ Simulate data from the estimated regression model, calculate residuals, repeat 1000 times, and plot those simulated residuals as a comparison distribution. Then we can see if this is typical residuals from your model, or not. There is in general no reason to expect normal residuals from an GLM. $\endgroup$ Aug 27, 2015 at 10:35

1 Answer 1

1
$\begingroup$

You've got some non-normal residual tails over there, especially a very heavy right one. Might want to consider taking a transform of your dependent variable. A log transform might help quite a bit, but no way to tell for sure without seeing the actual variable distribution, along with the rest of the residual plots (for both dependent and independent variables).

Some resources that might help you:

http://onlinestatbook.com/2/advanced_graphs/q-q_plots.html

https://www.youtube.com/watch?v=X9_ISJ0YpGw

$\endgroup$
5
  • 4
    $\begingroup$ Standard transformations won't work, because they will not be sufficiently non-differentiable at a particular value. We should instead suspect some form of a mixture or perhaps heteroscedasticity, and model it accordingly. $\endgroup$
    – whuber
    Apr 16, 2015 at 21:17
  • 2
    $\begingroup$ @whuber - fair point. I think I rushed into the suggestion because I was just working on a problem that had an initial qqplot similar to the one above and an exponentially-distributed dependent variable. To your point however, heteroscedasticity was present and it took the introduction of polynomials and interaction terms to deal with it. $\endgroup$
    – habu
    Apr 16, 2015 at 21:25
  • $\begingroup$ This is actually a Poisson glm with log link. I've even tried quasi poisson, but that doesn't change much. $\endgroup$
    – mackbox
    Apr 16, 2015 at 22:28
  • $\begingroup$ If you've used the log link, then the transformation is already factored in. As whuber said, this is one likely in need of dependent variable manipulation and/or introduction, so you'll have to dive deeper into the residual plots. $\endgroup$
    – habu
    Apr 16, 2015 at 22:32
  • 1
    $\begingroup$ quasi-poisson doesnt change the estimated model parameters, it only affects hypothesis tests/standard errors/confidence intervals. $\endgroup$ Aug 27, 2015 at 10:36

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.