R help - Dec 2013 - GLMM parameter estimates giving opposite trends

I apologize if this is a simple question.

I am running GLMM's using glmmML and model averaging with MuMIn. One of the
parameter estimates for a parameter (firefreq) in the best model is giving
a positive number, where in reality I know this to be a negative
correlation.
I have checked and double checked the data that has gone in and this is not
the issue. This is occurring for numerous variables in my models.

As far as I was aware the parameter estimate is indicative of the direction
of the relationship? Is there any reason why this model would give me
opposite trends?

Let me know if any other information would be useful in answering this
question.
Thank you in advance for any input.


This is the best model:
M17<-
glmmML(ldeli~bare+firefreq+canopy+treatment,cluster=season,family=poisson,data=ldeli)

Model-averaged coefficients:
                Estimate Std. Error z value Pr(>|z|)
(Intercept)   -84.248439  30.376197   2.774  0.00555 **
bare           -0.102111   0.023231   4.396 1.11e-05 ***
firefreq        3.370089   1.183091   2.849  0.00439 **
canopy         -0.013598   0.007420   1.832  0.06688 .
treatmentLU    87.939276  30.376750   2.895  0.00379 **
treatmentSM01  67.595184  24.477350   2.762  0.00575 **
treatmentSMWF  64.612540  23.322285   2.770  0.00560 **
treatmentT01   80.142838  28.030787   2.859  0.00425 **
treatmentT03   77.088813  26.836430   2.873  0.00407 **
treatmentTB    56.163472  19.744217   2.845  0.00445 **
treatmentWF    84.313036  29.214505   2.886  0.00390 **
-- 
_ _ _ _
Diana

	[[alternative HTML version deleted]]

Diana Virkki <d.virkki <at> griffith.edu.au> writes:
> 
> I apologize if this is a simple question.
> 
> I am running GLMM's using glmmML and model averaging with
>  MuMIn. One of the
> parameter estimates for a parameter (firefreq) in the
>  best model is giving
> a positive number, where in reality I know this to be a negative
> correlation.
> I have checked and double checked the data that has
> gone in and this is not
> the issue. This is occurring for numerous variables in my models.
> 
> As far as I was aware the parameter estimate is 
> indicative of the direction
> of the relationship? Is there any reason why this model would give me
> opposite trends?
  It's a little hard to guess without a reproducible example (see
http://tinyurl.com/reproducible-000), but one guess is that you have
one or more confounding variables
<http://en.wikipedia.org/wiki/Confounding> in your multivariate model;
that is, the _marginal_ effect of fire frequency is to decrease the
mean response, but the effect _conditional_ on all of the other
variables in the model is to increase it.  This phenomenon is most
common when the predictors are strongly correlated.

  Do you get a sensible sign when you fit a model with just the
focal parameter?

  Ben Bolker