Should truncation distance be the same when comparing different model fits (haz rate, half-normal)?

[Brianna Wright]

My question is about right truncation of detection functions. I know Buckland et al. 2001 recommend truncating at g(x)=0.10, which is what I am doing. However, many papers that use Distance sampling compare AIC values for different model types (half-normal, hazard rate etc.) to pick the best model, but they apply the same truncation distance to each of the models. What I'm wondering is, if setting the truncation distance to the distance where g(x)=0.10 depends on the detection function itself, shouldn't there be a different truncation distance for each model type? For instance, in my data g(x)=0.10 occurs at w=2600 m for the hazard rate model and w=4300 m for the half-normal model. Is it correct to truncate each by these two different values, and then compare the AIC values between models? Or do I have to truncate by the same value for both model types if I'm hoping to compare AIC's? If the latter is the case, then how do I choose w if it differs for g(x)=0.10 between the two models? Thanks for any assistance!

[Steve]

Brianna, yes, this is an issue - you cannot use AIC to select between models when truncation distance w differs.  It is unusual to have such a large difference between models.  Remember that g(x)=0.1 is just a guideline, and not the only guideline.  You want to choose a w to avoid having to fit the tail of the distribution, which can increase the number of parameters needed, and reduce precision and reliability.  Where 2 models differ by this much, I would tend to go for more truncation rather than less, suggesting something closer to 2600m than 4300m.  But I would try different values, and inspect histograms and goodness of fit stats, before selecting a single value to use across models.

Steve Buckland

On 22/01/2015 01:48, Brianna Wright wrote:

My question is about right truncation of detection functions. I know Buckland et al. 2001 recommend truncating at g(x)=0.10, which is what I am doing. However, many papers that use Distance sampling compare AIC values for different model types (half-normal, hazard rate etc.) to pick the best model, but they apply the same truncation distance to each of the models. What I'm wondering is, if setting the truncation distance to the distance where g(x)=0.10 depends on the detection function itself, shouldn't there be a different truncation distance for each model type? For instance, in my data g(x)=0.10 occurs at w=2600 m for the hazard rate model and w=4300 m for the half-normal model. Is it correct to truncate each by these two different values, and then compare the AIC values between models? Or do I have to truncate by the same value for both model types if I'm hoping to compare AIC's? If the latter is the case, then how do I choose w if it differs for g(x)=0.10 between the two models? Thanks for any assistance!

--
You received this message because you are subscribed to the Google Groups "distance-sampling" group.
To unsubscribe from this group and stop receiving emails from it, send an email to distance-sampl...@googlegroups.com.
To post to this group, send email to distance...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/distance-sampling/d8b60844-fc35-4060-b3d6-308306ad8735%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

-- 
____________________________________________________________________________
Stephen T. Buckland
CREEM, The Observatory, Buchanan Gdns, St Andrews KY16 9LZ, Scotland
Tel. +44 (0)1334 461841           Secretary  +44 (0)1334 461842
Fax  +44 (0)1334 461800           e-mail     st...@st-andrews.ac.uk

Professor of Statistics           http://www.st-and.ac.uk/
Co-Director, NCSE                 http://www.ncse.org.uk/

The University of St Andrews is a charity registered in Scotland:No SC013532
____________________________________________________________________________

[Brianna Wright]

Hi Steve,

Thanks very much for the clarification, this is very helpful.

 

Cheers,

- Brianna