fixed or random effects?

I'm modeling fish activity data with a gaussian distribution for scores
obtained from Principal Component Analysis, and have a little problem,
hopefully simple to resolve. My explanatory variables are group size, fish
length and temperature and I sampled in two consecutive years, in spring. My
problem is that I'm not sure whether I should consider year as a random or a
fixed effect. I wonder if you could help me.

For you the year factor will have only two levels and that is too few
to model the effect of year as a random effect.  When you incorporate
a random-effects term in a model you end up estimating a variance
component instead of trying to estimate coefficients in a linear model
expression directly.  Having only two levels of year will not allow
for a precise estimate of a variance component.  In fact, it will be a
horribly imprecise estimate.

There are no hard and fast rules of how many levels are required to be
able to estimate a variance component but fewer than 5 is too few and
more than 10 is adequate.  I have used as few as 6 levels but that was
on nicely balanced data from a designed experiment.  Observational
data that is highly unbalanced requires more care.

joana martelo <jmmartelo at ...> writes:

Hello R list

I'm modeling fish activity data with a gaussian distribution for scores
obtained from Principal Component Analysis, and have a little problem,
hopefully simple to resolve. My explanatory variables are group size, fish
length and temperature and I sampled in two consecutive years, in spring. My
problem is that I'm not sure whether I should consider year as a random or a
fixed effect. I wonder if you could help me.

[snip]

  As hinted in private e-mail, http://glmm.wikidot.com/faq tells you that
while you may *philosophically* want to treat two years as a sample from
a larger population of years, it is not computationally practical 
(nor will it get you much in terms of inferential power) to treat a factor
with only two levels as random: you should make it fixed.  This then has
the added advantage that you have only fixed effects, and you can use
lm() instead of getting into any of the complexities of mixed models.

Douglas Bates <bates at ...> writes:

On Mon, Oct 1, 2012 at 1:10 PM, joana martelo <jmmartelo at ...> wrote:

[snip]
 
# For you the year factor will have only two levels and that is too few
# to model the effect of year as a random effect.  When you incorporate
# a random-effects term in a model you end up estimating a variance
# component instead of trying to estimate coefficients in a linear model
# expression directly.  Having only two levels of year will not allow
# for a precise estimate of a variance component.  In fact, it will be a
# horribly imprecise estimate.

There are no hard and fast rules of how many levels are required to be
able to estimate a variance component but fewer than 5 is too few and
more than 10 is adequate.  I have used as few as 6 levels but that was
on nicely balanced data from a designed experiment.  Observational
data that is highly unbalanced requires more care.

jinx (snap):

http://separatedbyacommonlanguage.blogspot.ca/2006/10/jinx-and-snap.html

(making
   gmane
     happy
       with
         more
           stuff)