Model Averaging - General question

[constant survivor]

Hi all,
I have a basic question about model averaging.

Assume two simple models (just for demonstration; 5 Occasions):
1) phi(.) p(.)
2) phi(t) p(.)

Model 1 has lowest AICc value but to address model uncertainty, model averaging (for phi) is done.

If I got it right, then model averaging should return 4 estimates because all the estimates from model 2 should 'go into' the averaging procedure.

But what to report then? Would you report 4 estimates, even though model 1 has constant phi? Or would you calculate the mean of the weighted average estimates (not easy, I know)? How to 'pay tribute' to the fact, that the best model is one with constant phi??

Thank you and kind regards
Hannes

[cooch]

Hi all,
I have a basic question about model averaging.

Assume two simple models (just for demonstration; 5 Occasions):
1) phi(.) p(.)
2) phi(t) p(.)

Model 1 has lowest AICc value but to address model uncertainty, model averaging (for phi) is done.

If your best model is a 'dot' model, then either (i) your data are lousy, or (ii) your model set is lousy.

If I got it right, then model averaging should return 4 estimates because all the estimates from model 2 should 'go into' the averaging procedure.

Correct.

But what to report then? Would you report 4 estimates, even though model 1 has constant phi? Or would you calculate the mean of the weighted average estimates (not easy, I know)? How to 'pay tribute' to the fact, that the best model is one with constant phi??

An 'overall mean' based on only 4 estimates is not particularly informative. So, I'd suggest simply plotting the model-averaged estimates. If your 'dot model' has most of the AIC support, then the plot will fairly reflect that.

[constant survivor]

Hi cooch,
thanks for reply.

As I said: the models are just for demonstration purposes. They do not at all reflect my model set.
For me it's more about the general problem of time variant vs. time invariant models in the candidate list.

Other example:
What would you suggest if the candidate model set includes phi(.) models and phi(linear Trend) models and they are supposed to be averaged? If the trend model is the 'best' model - ok. The slope of the regression line maybe just gets lower. But what if vice versa - the dot model is best and the trend model is among those to average?


I guess there are a lot of other possible combinations and I was asking for some general advice on that (if such can be given).

Best
Hannes

[cooch]

If you're model-averaging, plot the model averaged estimates. Its that simple.