4
$\begingroup$

Say you do this in R:

g <- rgamma(5000, 4)
t <- rt(5000, 5)

So, now you've got data from the gamma and $t$-distributions.

Then you model these. I use 'gam' from 'mgcv'.

m1 <- gam(g ~1, family=Gamma)
m2 <- gam(g ~1, family=scat)

m3 <- gam(20 + t ~1, family=scat)   # add 20 to get positive values
m4 <- gam(20 + t ~1, family=Gamma)  # ^^^

Then you plot the 4 corresponding normal qq-plots of residuals. (I used resid() to get these.) Above models have nothing to do with the normal distribution, so why is it that just by looking at the normal qq-plots of the residuals, I can tell easily which model is the correct one?

enter image description here

As you can see, the qq-plots of model 1 and 3 (the correct models), are normal. The others are not.

$\endgroup$
  • 1
    $\begingroup$ You should try to make your example reproducible (e.g. include library(mgcv) ...) $\endgroup$ – Glen_b Jun 6 '17 at 1:11
5
$\begingroup$

The function you called returns deviance residuals by default. (resid is an alias of residuals which when called on a gam object invokes residuals.gam; see its help)

These are typically considerably more normal looking than raw residuals ($y_i-\hat{\mu_i}$).

For a gamma random variable, the deviance residuals would be

$r_D(i)=\operatorname{sign}(y_i-\hat{\mu}_i)\sqrt{-2\nu [\log(\frac{y_i}{\hat{\mu}_i})-\frac{y_i-\hat{\mu}_i}{\hat{\mu}_i}]}$

(though presumably it would be estimating $\nu$ from the deviance)

In particular, for your model, $\hat{\mu}_i$ will be $\bar y$, and since the sample size is very large we might reasonably approximate it by $\mu$.

If you look at the function $t(x)=\operatorname{sign}(x-1)\sqrt{x-\log(x)-1}$, in the vicinity of $1$ (NB $r_D(i) \propto t(y_i/\hat{\mu}_i)$), it's rather similar to (a linear transformation of) a cube root:

plot of sign(x-1) sqrt(x - log x - 1) vs x and (13/6) (x^(1/3) - 1) vs x, which are almost coincident on (0,4)

The cube root is an approximate symmetrizing transformation for the gamma, sometimes called the Wilson-Hilferty transformation); note that Anscombe residuals for the gamma are $3(\sqrt[3]{x} - 1)$ applied to $y/\hat\mu$. Both transformations ($t$ and the cube root) would be expected to produce close-to-normal results for gamma variates.

(in implementation $r_D(i)$ may also be adjusted for the observation's influence on its own fitted value by dividing by $\sqrt{1-h_{ii}}$ -- however those are constant for your examples)


In the case of the scaled-t (which is not exponential family), it's not immediately clear from the residuals.gam function what residuals are being used in that case, but it would not be surprising if they were similarly a kind that would be more normal-looking than raw residuals.

$\endgroup$
2
$\begingroup$

There is nothing about a Q-Q plot that is intrinsically tied to the normal distribution. It is a plot of the quantile points of one distribution against the quantile points of another. If those two distributions are the same, the points will line up. If they are different, they will not.

If you make one of those distributions the normal distribution, then it is a useful tool to tell whether the other is also normal. If you make one of those distributions a gamma distribution, then it is a useful tool to tell whether the other is also gamma. And so on, for any other distribution.

$\endgroup$
  • 1
    $\begingroup$ The point of the question is, why are the residuals of the two correctly fitted models normally distributed and not, as one might suppose, gamma resp. t-distributed? The model does not involve the normal distribution in any ways, so why are residuals normally distributed, as evidenced by the normal qqplot? $\endgroup$ – Sarah Jun 5 '17 at 23:18
  • 3
    $\begingroup$ "There is nothing about a Q-Q plot that is intrinsically tied to the normal distribution." — Yes, but these plots do use normal distributions for the "Theoretical Quantiles", if OP is describing them correctly. $\endgroup$ – Kodiologist Jun 6 '17 at 0:54
  • $\begingroup$ @Sarah Are you sure the "Theoretical Quantiles" are for normal distributions? $\endgroup$ – Kodiologist Jun 6 '17 at 0:55

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.