Model Checking in R

[danpi...@googlemail.com]

Hi everybody
,
I was wondering if anyone has come across a decent description of the
model checking process in R.

I am trying to fit a Poisson Generalized Linear Model for Count Data
and am still rather inexperienced with these things (and I guess GLMs
in general). When asked to plot a given model, R outputs a series of
four graphs (residuals x fitted, QQ, ...). Now, I find these plots
rather hard to interpret. The most thorough discussion I know of how
these things should look like is Crawley 2007 (The R book), but the
comments there are a little "impressionistic" at times ("This plot
should look like a sky at night” Crawley 2007:357; "[nonconstancy of
variance] shows up as a triangular scatter (like a wedge of cheese)”
Crawley 2007:359...).

To put it bluntly: I am having trouble deciding if what I see is
sufficiently similar to a sky at night or a wedge of cheese...;) Any
hint as to where I can find a model answer for (Poisson Generalized
Linear) Model checking and the interpretation of those plots would be
tremendously appreciated.

best,
daniel

[danpi...@googlemail.com]

oh, I just noticed that Stefan has uploaded these plots onto this site
some time ago. I guess this means that the new book will provide some
discussion...now, amazon tells me the book is coming out on Dec 15 -
is that realistic?

On 23 Okt., 12:37, "danpire...@googlemail.com"

[Stefan Th. Gries]

> oh, I just noticed that Stefan has uploaded these plots onto this site some time ago.

I am not sure which plots you mean especially since ...

> I guess this means that the new book will provide some discussion ...

... the book is not gonna discuss Poisson regression.

> now, amazon tells me the book is coming out on Dec 15

I am in the middle of completing the index and Mouton told me that, if
I can get the crc to them in Nov, and I will do that, then the
publication will happen as planned.

Cheers,
STG
--
Stefan Th. Gries
-----------------------------------------------
University of California, Santa Barbara
http://www.linguistics.ucsb.edu/faculty/stgries
-----------------------------------------------

[danpi...@googlemail.com]

I was referring to the plots labelled as "Abbildung 65:
Modelldiagnostische Grafiken".
While the graphs are not from a Poisson model, I am guessing they are
from some type of regression model and so their discussion might still
be useful to me (as R provides similar diagnostic plots for these
models (the first three plots are of the same type ... the last one is
different)).

I guess -conceptually speaking- I get what these graphs are supposed
to show, but I am having trouble interpreting my own models: basically
"model quality" appears to me to refer to a rather vague quantity and
so it would be great to be able to see experienced interpreters
interpret many models.

> ... the book is not gonna discuss Poisson regression.

well, I'll buy it anyway ;)

[Stefan Th. Gries]

> basically "model quality" appears to me to refer to a rather vague quantity and so it would be great to be able to see experienced interpreters interpret many models.

Well, some of the literature does suggest that model selection is to
some degree more of an art than anything. I guess, at some level of
complexity, playing around with different approaches, leaving out
different points identified as outliers etc. becomes necessary, which
cannot always be done as rigorously as one may want. A similarly
sometimes controversial case in point is the way in models are
selected, AIC/BIC, p-values, ...? Or when to eliminate random effects.
Just my $0.02 tho ...

>> ... the book is not gonna discuss Poisson regression.

> well, I'll buy it anyway ;)

That's the spirit ;-)

[danpi...@googlemail.com]

> Well, some of the literature does suggest that model selection is to
> some degree more of an art than anything.

right...I guess I'm gonna try and bring out the artist in me then ;)

thanks again,
best
daniel