Date: Mon, 13 Aug 2007 11:21:14 -0700
Reply-To: Andrew Hill <hill.andrewd@GMAIL.COM>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: Andrew Hill <hill.andrewd@GMAIL.COM>
Subject: Re: modelling presence/absence of fish using LOGISTIC/GLIMMIX
Content-Type: text/plain; charset=ISO-8859-1
To quote a friend of mine "Fish are just like trees, except you can't
see them and they move."
I am not quite sure what the underlying distribution is. I would guess
that you are assuming a Poisson with regard to time. If that is the
case then the longer the time period the more fish you get as a matter
of course because Hour < photoperios < day < moon phase < year. So
probablity of seeing a fish is: year > moon phase > day > photo period
So you are kind of confounding yourself.
One way to test the moon phase would be to assign periods of the same
length that are out of sync with the moon phase and see if these are
significant. This would show if the moon phase is just happenstance or
if its impacts are real.
Also, you may not want to use step-wise. You can hunt though the
archives to see why not. David Cassell has written a lot about the
problems caused the assumptions it violates.
Anyway, hope this helps.
On 8/10/07, mike <firstname.lastname@example.org> wrote:
> Hi all,
> I'm trying to model some presence/absence data using logistic and
> glimmix. What I've got is two separate years of data, each fish is
> coded as '0' or '1' for each hour from the beginning to the end of
> each study. I used a stepwise LOGISTIC regression to find significant
> variables from among the ones I expected might be affecting
> probability of presence(ID--of the fish, day of the study, hour of
> day, year, moonphase, photoperiod, and several environmental
> variables). I then moved to GLIMMIX, so that I could run ID (which
> identifies each fish uniquely) as a random variable.
> My first problem is that moonphase accounts for a large amount of
> variability (second only to ID), but I suspect this is not for
> ecological reasons, but rather that it just 'changes' more slowly that
> DAY, i.e., DAY increases in steps of 1, but moonphase can be the same
> for several days at a time. I suspect that 'more has happened' in
> that time than happens from day-to-day. Is there a way for me to
> 'block' days that would more realistically model what is actually
> Also, I am uncertain how to treat YEAR--the day*year interaction comes
> up significant, i.e., day 1 in year 1 is not equivalent to day 1 in
> year 2. Can I run YEAR as a random variable? I must mention that
> 'year information' is contained in the ID, i.e., I know which fish
> were tagged in which year.