Re: Using NPMR to compare multiple hypotheses: Is Bayes factor appropriate?

[br...@salal.us]

Pablo, the way I think of this is as a sample size issue. With a large sample size, you can detect week effects. This is no different between nonparametric statistics and parametric statistics. So if you have a large sample size and small effect size, you can still do something with it (i.e. publish it), but it will make a less compelling story than if you have a stronger effect size.

Although in early (Beta) versions of HyperNiche we referred to "logB" as Bayes factors, it was pointed out early on that this isn't strictly appropriate, and I agree. If you have presence-absence responses, I recommend comparing models with likelihood ratios: either logB or differences between logB from competing models.

The other component to this is doing randomization tests. Those help you decide if the effect size (however small) is larger than expected by chance alone. Again, if the sample size is large then the randomization test will tend to show that even small effect sizes are significant.

Good luck,
Bruce McCune


>-----Original Message-----
>From: Pablo Bacon [mailto:pabloand...@gmail.com]
>Sent: Monday, August 6, 2012 09:46 AM
>To: hyper...@googlegroups.com
>Subject: Using NPMR to compare multiple hypotheses: Is Bayes factor appropriate?
>
>Hey all,
>
>First, I want to say that reading these posts (and the ones from the PC-ORD
>group) has been extremely insightful.
>
>I am fairly new to NPMR/ HyperNiche and still exploring some of the
>applications. I am wanting to test multiple hypotheses regarding land-use
>land cover variables at multiple scales on the presence/absence of an
>aquatic beetle. The models are based on select lulc variables and have
>very low xr2 values. This may be due to other factors we did not
>quantitatively address in initial distributional surveys such as
>presence/absence of predators and competitors, indirect effects, etc. in
>addition to the environment and all interactions. However, the focus of my
>analyses are aimed at testing multiple hypotheses about these select
>variables at multiple spatial scales and how they interact (for example, at
>80% forest cover at the local riparian scale the watershed scale
>urbanization effect seems to have less of an effect on p/a of the beetle).
>
>So here is the question: Is it appropriate to use this approach to compare
>multiple hypotheses using the Bayes factor despite poor model fit/xr2?
>
>Thank you for your time,
>
>Pablo Andres Bacon
>

[Pablo Bacon]

Thank you so much for the quick resonse!  Yes, I have a terrible habit of saying Baye's factor when I do indeed mean logB.  I'm sure I will have more questions as I travel down the path of NPMR!

Take care,

 

Pablo Andres Bacon