recommended maximum number of parameters

[markmiller]

Recently someone informed me that Burnham and Anderson (2002) recommended, as a general rule, that the maximum number of structural parameters in a model = (N / 10).

I was using a band-recovery model. I think N was meant to be the number of birds banded. Although maybe N was meant to be the number of dead birds recovered.

I spent a few hours searching through Burnham and Anderson's (2002) book 'Model Selection and Multimodel Inference' without locating this rule. I also searched the internet for other references regarding model over-fitting and AIC without finding this rule mentioned.

Could someone tell me where I can find this rule in Burnham and Anderson (2002)? It kind of seems counter to my understanding of AIC to specify such a rule. In other words, I thought AIC was supposed to handle the issue of over-fitting for you. However, there is a very good chance I simply do not understand the nuances.

Sorry if this is not an appropriate post for this forum. It is not specific to MARK, but seems clearly related.

[ehileman]

Recently someone informed me that Burnham and Anderson (2002) recommended, as a general rule, that the maximum number of structural parameters in a model = (N / 10).

I was using a band-recovery model. I think N was meant to be the number of birds banded. Although maybe N was meant to be the number of dead birds recovered.

I spent a few hours searching through Burnham and Anderson's (2002) book 'Model Selection and Multimodel Inference' without locating this rule. I also searched the internet for other references regarding model over-fitting and AIC without finding this rule mentioned.

Could someone tell me where I can find this rule in Burnham and Anderson (2002)?

See p. 245 of Burnham and Anderson (2002).
Cheers,
Eric

[markmiller]

Thank you, Eric.

I guess I am still confused, Burnham and Anderson (2002:245) write that AICc correctly adjusts for small sample size or large K. It kind of seems to suggest that the rule is not necessary if AICc is used, as I did.

Although, perhaps the rule is best used regardless of whether AIC or AICc is used?

Mark

[ehileman]

I was using a band-recovery model. I think N was meant to be the number of birds banded. Although maybe N was meant to be the number of dead birds recovered.

The n/10 rule primarily applies to regression type models. In capture-recapture models, n can apply to the number of unique animals or the total number of captures.

I guess I am still confused, Burnham and Anderson (2002:245) write that AICc correctly adjusts for small sample size or large K. It kind of seems to suggest that the rule is not necessary if AICc is used, as I did.

Although, perhaps the rule is best used regardless of whether AIC or AICc is used?

Others may want to chime in here, but I'll take a first stab at it. In general you shouldn't fit models that use a sample size of fewer than 10 for a given parameter regardless of if you are using AIC or AICc. Unless you are collecting data in a controlled laboratory setting, the data are likely to be too noisy to yield anything meaningful (e.g., detectable differences between groups) or worse, they could give you spurious results. Hope this helps! -Eric

[markmiller]

I guess I am hoping for clarification. I am analyzing band recovery data. Suppose 200 birds were banded and released and 20 of those birds later were recovered dead.

With regards to the n / 10 rule, is n = 20 (the number of birds recovered dead), or 200 (the number of birds released) or 220 (the number of birds recovered dead + the number of birds released).

I appreciate Eric's response. Although, I am not really clear on the answer. For example, is the number of unique animals = 200? Is the total number of captures = 220?

The reason I am asking is that a reviewer insists n = 20 in this example, meaning the most complex model I can create is phi(.)r(.) using the Seber parameterization. This is preventing me from including any covariates in the analysis. So, I am hoping for clarification.

Sorry to be posting the same question twice.

[Eurycea]

I believe your effective sample size is the total number of releases/banded birds. MARK will even tell you your effective sample size, it is in the model output file as well as in the Adjustments drop-down menu to change the effective sample size (the number it defaults to). As for rules of thumb, they are just that. If your data is marginal for certain models, then you will have wide SE's on certain parameter estimates. Your band-recovery rate will have an impact on your ability to run more complex models, so perhaps that is what the reviewer is getting at. But to say you can only run a single model because you recaptured 20 birds, as opposed to 30, seems just silly to me. But I'm just a caveman.