Model averaging

[simone77]

Hi,

I have run my analyses and did not find any model whose w i >0.9.
I would like to model average the estimates of the parameters of interest and, after, in case the parameters estimates were time varying, to calculate the mean value and their CI by using the formula indicated from Rémi Choquet in another post.

1. I am not sure if, given time varying parameters are best supported in the models with respect to the constant parameters, one could argument that it is not correct to calculate the mean value. To me it would make sense as far as one states that time variation is best supported and because the CI of that estimate would be affected by that variation.
Any conceptual problem with this procedure?

2. In MARK (I first learned to use it) model averaging is done by a specific tool implemented in the software. I haven't found something analogous in E-SURGE, how can I do this?

Thanks in advance for any help,

Simone

[simone77]

I have realized that the second question is too unspecific.

2. In MARK (I first learned to use it) model averaging is done by a specific tool implemented in the software. I haven't found something analogous in E-SURGE, how can I do this?

I am used to use the easy tool in MARK for that but I know the formula beyond this as specified in Burnham and Anderson (2002) and have found another paper (Johnson and Omland 2004) where the formulas to be used for model averaging are resumed this way (can't improve the image quality):


I think it should not be difficult to prepare a sheet in excel and run those formulas, I am just concerned about what to do when my parameter of interest is time dependent in some model/s and constant in some other model/s.
I have read somewhere that if the parameter does not exist in one of the models, you should simply mark a zero for that parameter in that model.
Said that, I reformulate the question:
2. How should I proceed if the parameter of interest is time varying in one model and constant in another? I may have two goals: (i) estimate the parameter as a single value and (ii) estimate the parameter as time varying.

I am still interested in some opinion on the first question of the above post.
Thank you again.

Simone

[CHOQUET]

1) there is no tool to do model averaging automatically in E-SURGE.

2) Model averaging is well defined in the Bayesian framework. This is more tedious in a frequentist
framework.

See however the book of Claeskens and Hjort "Model selection and model averaging"

Doing model averaging on parameters will not change the order of the error that you do
if you take only the best model. Concerning the test of hypothesis, you must select at the beginning
a set of model with clear hypothesis. Then make the sum of weigths of models containing the hypothesis H0 against the sum of weigths of models with H1.