12
$\begingroup$

This paper uses generalised linear models (both binomial and negative binomial error distributions) to analyse data. But then in the statistical analysis section of the methods, there is this statement:

...and second by modelling the presence data using Logistic Regression Models, and the foraging time data using a Generalized Linear Model (GLM). A negative binomial distribution with a log link function was used to model the foraging time data (Welsh et al. 1996) and model adequacy was verified by examination of resi- duals (McCullagh & Nelder 1989). Shapiro–Wilk or Kolmogorov–Smirnov tests were used to test for normality depending on sample size; data were log-transformed before analyses to adhere to normality.

If they assume binomial and negative binomial error distributions, then surely they shouldn't be checking for normality of residuals?

$\endgroup$
  • 2
    $\begingroup$ Note that the errors aren't binomially distributed - each response is binomially distributed with a probability parameter given by the corresponding predictor values, as per the answers to one of your other questions. $\endgroup$ – Scortchi - Reinstate Monica Apr 3 '14 at 16:43
  • 3
    $\begingroup$ There's nothing in binomial or negative binomial regression than needs to be normal. If it's the response they transform, that may well be highly counterproductive; it will screw up the GLM. $\endgroup$ – Glen_b -Reinstate Monica Apr 4 '14 at 9:29
  • 1
    $\begingroup$ It's not clear from your quote what they're in fact testing for normality (are you sure it's the residuals?) or for what analysis they're transforming data (are you sure it's the GLMs?). $\endgroup$ – Scortchi - Reinstate Monica Apr 4 '14 at 9:55
  • $\begingroup$ I've expanded quote a little bit. Could someone confirm if what the authors of the paper did was wrong or right? $\endgroup$ – luciano Apr 4 '14 at 12:43
  • $\begingroup$ I'm afraid it's still not terribly clear - contact the authors for detail of how they carried out the analysis if it's not explained elsewhere in the paper or its references. $\endgroup$ – Scortchi - Reinstate Monica Apr 4 '14 at 12:53
16
$\begingroup$

NB the deviance (or Pearson) residuals are not expected to have a normal distribution except for a Gaussian model. For the logistic regression case, as @Stat says, deviance residuals for the $i$th observation $y_i$ are given by

$$r^{\mathrm{D}}_i=-\sqrt{2\left|\log{(1-\hat{\pi}_i)}\right|}$$

if $y_i=0$ &

$$r^{\mathrm{D}}_i=\sqrt{2\left|\log{(\hat{\pi}_i)}\right|}$$

if $y_i=1$, where $\hat{\pi_i}$ is the fitted Bernoulli probability. As each can take only one of two values, it's clear their distribution cannot be normal, even for a correctly specified model:

#generate Bernoulli probabilities from true model
x <-rnorm(100)
p<-exp(x)/(1+exp(x))

#one replication per predictor value
n <- rep(1,100)
#simulate response
y <- rbinom(100,n,p)
#fit model
glm(cbind(y,n-y)~x,family="binomial") -> mod
#make quantile-quantile plot of residuals
qqnorm(residuals(mod, type="deviance"))
abline(a=0,b=1)

Q-Q plot n=1

But if there are $n_i$ replicate observations for the $i$th predictor pattern, & the deviance residual is defined so as to gather these up

$$r^{\mathrm{D}}_i=\operatorname{sgn}({y_i-n_i\hat{\pi}_i})\sqrt{2\left[y_i\log{\frac{y_i}{n\hat{\pi}_i}} + (n_i-y_i)\log{\frac{n_i-y_i}{n_i(1-\hat{\pi}_i)}}\right]}$$

(where $y_i$ is now the count of successes from 0 to $n_i$) then as $n_i$ gets larger the distribution of the residuals approximates more to normality:

#many replications per predictor value
n <- rep(30,100)
#simulate response
y<-rbinom(100,n,p)
#fit model
glm(cbind(y,n-y)~x,family="binomial")->mod
#make quantile-quantile plot of residuals
qqnorm(residuals(mod, type="deviance"))
abline(a=0,b=1)

Q-Q plot n=30

Things are similar for Poisson or negative binomial GLMs: for low predicted counts the distribution of residuals is discrete & skewed, but tends to normality for larger counts under a correctly specified model.

It's not usual, at least not in my neck of the woods, to conduct a formal test of residual normality; if normality testing is essentially useless when your model assumes exact normality, then a fortiori it's useless when it doesn't. Nevertheless, for unsaturated models, graphical residual diagnostics are useful for assessing the presence & the nature of lack of fit, taking normality with a pinch or a fistful of salt depending on the number of replicates per predictor pattern.

$\endgroup$
1
$\begingroup$

What they did is correct! I will give you a reference to double check. See Section 13.4.4 in Introduction to Linear Regression Analysis, 5th Edition by Douglas C. Montgomery, Elizabeth A. Peck, G. Geoffrey Vining. In particular, look at the examples on page 460, where they fit a binomial glm and double check the normality assumption of the "Deviance Residuals". As mentioned on page 458, this is because "the deviance residuals behave much like ordinary residuals do in a standard normal-theory linear regression model". So it make sense if you plot them on normal probability plot scale as well as vs fitted values. Again see page 456 of above reference. In the examples they have provided on page 460 and 461, not only for the binomial case, but also for the Poisson glm and the Gamma with (link=log), they have checked the normality of deviance residuals.
For the binomial case the deviance residual is defined as: $$r^{D}_i=-\sqrt{2|\ln{(1-\hat{\pi_i})}|}$$ if $y_i=0$ and $$r^{D}_i=\sqrt{2|\ln{(\hat{\pi_i})}|}$$ if $y_i=1$. Now some coding in R to show you how you can get it:

> attach(npk)

> #Fitting binomila glm
> fit.1=glm(P~yield,family=binomial(logit))
> 
> #Getting deviance residuals directly
> rd=residuals(fit.1,type = c("deviance"))
> rd
         1          2          3          4          5          6          7 
 1.1038306  1.2892945 -1.2912991 -1.1479881 -1.1097832  1.2282009 -1.1686771 
         8          9         10         11         12         13         14 
 1.1931365  1.2892945  1.1903473 -0.9821829 -1.1756061 -1.0801690  1.0943912 
        15         16         17         18         19         20         21 
-1.3099491  1.0333213  1.1378369 -1.2245380 -1.2485566  1.0943912 -1.1452410 
        22         23         24 
 1.2352561  1.1543163 -1.1617642 
> 
> 
> #Estimated success probabilities
> pi.hat=fitted(fit.1)
> 
> #Obtaining deviance residuals directly
> rd.check=-sqrt(2*abs(log(1-pi.hat)))
> rd.check[P==1]=sqrt(2*abs(log(pi.hat[P==1])))
> rd.check
         1          2          3          4          5          6          7 
 1.1038306  1.2892945 -1.2912991 -1.1479881 -1.1097832  1.2282009 -1.1686771 
         8          9         10         11         12         13         14 
 1.1931365  1.2892945  1.1903473 -0.9821829 -1.1756061 -1.0801690  1.0943912 
        15         16         17         18         19         20         21 
-1.3099491  1.0333213  1.1378369 -1.2245380 -1.2485566  1.0943912 -1.1452410 
        22         23         24 
 1.2352561  1.1543163 -1.1617642 
> 

Check here for the Poisson case as well.

$\endgroup$
  • 2
    $\begingroup$ Your example is an odd choice. Did you make a P-P or Q-Q plot of those deviance residuals; if so, what did you conclude? $\endgroup$ – Scortchi - Reinstate Monica Apr 3 '14 at 15:59
  • 5
    $\begingroup$ Point is in this case there'd be no sense in checking the normality of the residuals - they're clearly not normally distributed, nor should they be. It's only as the number of observations for each predictor pattern increases that the distribution of residuals (one residual being calculated per predictor pattern) tends to the normal. Similarly for a Poisson or negative binomial model - the counts need to be large-ish for the normal approximation to be good. $\endgroup$ – Scortchi - Reinstate Monica Apr 3 '14 at 16:22
  • 2
    $\begingroup$ The question is whether residuals from generalized linear models should be normally distributed. Your answer appears to be an unqualified "yes" (though your sources doubtless give the necessary qualifications, not every reader will check them). You then give an example in which there is no reason at all to expect the residuals to be normally distributed, even if the model were correctly specified: an unwary reader will assume that they should be & that, as they are clearly not, this is therefore an example of detecting model mis-specification by examining the residuals (though you ... $\endgroup$ – Scortchi - Reinstate Monica Apr 3 '14 at 20:16
  • 2
    $\begingroup$ ... haven't said it is). So I think the answer requires much clarification to be useful. $\endgroup$ – Scortchi - Reinstate Monica Apr 3 '14 at 20:17
  • 2
    $\begingroup$ IMO @Scortchi's comments are reasonable here. Looking at what I can see of the Montgomery book on google books preview they do make the QQ plot, but don't perform an actual normality test like mentioned by the original poster. Sure making the QQ plot is reasonable as a diagnostic test, but in pretty much all realistic circumstances even the deviance resids. won't be normal. $\endgroup$ – Andy W Apr 5 '14 at 0:09

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.