Improving model fit in proc glimmix, interpretation of chi-sq/df?

[bowerske]

Hi, I am analyzing data from an experiment testing whether insects from northern latitudes are more cold-hardy that those from southern latitudes. I have five large cages in each of three locations (north, mid, south) Each cage has 21 potted plants on which I placed 10 insects from north, mid, or southern populations( 21= 3 sources of insects x 7 months). Each month I remove one potted plant per source from each cage and count what proportion of insects survived (The variable prop is a calculated variable equal to live insects/found insects). The code that I think is the best so far is below. I have also tried binomial, which had worse fit statistics and neg binomial, which didn’t converge.

proc glimmix data=puncv.new plots=pearsonpanel (conditional marginal);

class Month Location Source Cage;

model prop= Month Location Source ;

random intercept /subject=Cage;

run;

When I run this, I get the below fit statistics and residual graphs? What is the generalized chi-sq/df= 0.05 mean about how the model fits? Is there a model that would produce better residual graphs (i.e. fit the data better?)

[SteveDenham]

Rather than analyzing the calculated proportion, would it make sense to analyze the actual live counts? I ask that because it looks like you "seeded" the pots with equal numbers at the beginning.. If that is the case, consider a model like:

proc glimmix data=puncv.new plots=pearsonpanel (conditional marginal) method=laplace;
class Month Location Source Cage;
model live_count= Month Location Source /dist=poisson;
random intercept /subject=Cage;
run;

From here, you could consider interactions between you main effects, or whether there are correlations over time.  And in the end, you can use the mean estimates on the original scale divided by the number introduced as a proportional survival.

SteveDenham

[bowerske]

Hi Steve,

I will try that; I didn't start off with that because even though I did seed 10 weevils in each pot, we didn't always recover 10. Does that matter?

Kristen

[SteveDenham]

Well, you could consider the total as an offset for a count variable.  Just a thought about this though, if you put in 10, and only recover 1, and it is alive, that doesn't give me a lot of confidence that if I could have recovered all 10, then all 10 would be alive. And that is the assumption if you plug this into a binomial (or proportion) analysis.  Perhaps better to just analyze the counts of found alive, and if available, the number found dead.

SteveDenham