I'm using an IT-approach and multi-model inference with some count data.

I have used model averaging to obtain parameter estimates from several GLMMs with Poisson-lognormal errors (Poisson family plus observation-level random effect). Model-averaged parameter estimates are as follows:

(Intercept) 1.9522
landuse -2.0318
treatment -0.8861
hedge -0.6943
sun -0.5191
landuse:treatment 1.0529
bound.area -0.2389

I am now trying to work out point estimates from these.

Landuse is a two level factor (a and p). Treatment is a two level factor (AF and MC). Hedge, sun and bound.area are all continuous variables.

So the intercept is treatment = AF and landuse = a. I am interested in working out point estimates for:

  1. treatment = AF when landuse = a
  2. treatment = AF when landuse = p
  3. treatment = MC when landuse = a
  4. treatment = MC when landuse = p

I have Grueber et al’s 2011 paper on MMI but I am not sure whether I have correctly managed to work out how to apply their methods (detailed in the paper's appendix) to my parameter estimates. My predictors were standardised to a mean of 0 and SD of 0.5, so I understand that I have to substitute standardized predictors and then solve the model for the required estimate at the mean (i.e. 0) of all the other predictors.

So for example for a point estimate of treatment=AF when landuse=a, can I simply do exp(intercept)? [Which is the same as exp(1.9522 – 2.0318*0 - 0.8861*0 - 0.6943*0 - 0.5191*0 + 1.0529*0*0 – 0.2389*0).]

But how do I work it out for any of the other combinations? Would treatment=AF and landuse=p be worked out like this? : exp(1.9522 – 2.0318 - 0.8861*0 - 0.6943*0 - 0.5191*0 + 1.0529 – 0.2389*0)

I have a feeling I’m missing something / have misunderstood...

Finally, am I right in thinking that I can’t work out a treatment effect by itself (as it’s included in an interaction)?

Any help much appreciated.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.