Evidence ratio

[mpwilson09]

Hello,

I'm working with avian point count data and using multi-state occupancy modeling to figure out what variables are influencing detection probabilities and occupancy for certain species. I was told about calculating an evidence ratio for each variable which from my understanding, is a way to understand how much that variable influences detection probability and occupancy. I pasted some of my results for Acadian flycatcher below;

Variable Sum Evidence ratio Influence
OBA 1.00 #DIV/0! Strong
OBSV 1.00 #DIV/0! Strong
SPH 0.378926782 0.610116119 None
TEMP 0.429776311 0.753697749 None

I have a couple questions regarding this method. First, is this a good overall way to describe variable influence on detection probability and occupancy? What is considered a strong vs. moderate vs. a weak influence...are there categories? When you get a sum of 1, does that mean the evidence ratio is infinity and how do you explain that?

I hope this makes sense, it is all very new to me. Please let me know if you have any questions and I appreciate your help!

Thanks,
Meredith

[jhines]

Yes, the evidence ration seems like a good way to describe how a variable influences detection probability or occupancy (in my opinion). A value of 1 means that all evidence from the model-set indicates the variable is an important factor for the estimation of the parameter.

Regarding, the definition of how strong or weak the evidence is, those are subjective terms which you can use as you like. Obviously, an evidence ration (ER) of 1.0 should be considered 'strong' (or 'very strong' or 'super-duper strong' ), and an ER=0 would be no evidence of an effect. Any ER > 0 indicates some evidence, but since we're dealing with estimates of AIC and model weights, I would call anything less than some value (eg. 0.1 = a number I just pulled out of...uh... thin air) 'virtually no evidence'. Above that value could be called 'weak', 'moderate' or 'strong'. I think an ER=0.5 might be considered 'moderate' as 1/2 of the model weight is for models with the effect of the variable and 1/2 without the effect. This is much higher than I would expect if the data were completely random with respect to the variable. I don't know where I would draw the lines between 'weak' and 'moderate' or 'moderate' and 'strong'.

Jim

[jhines]

I forgot to mention that I hope your model-set includes an equal number of similar models, with and without the covariate.

Jim