Multi-methods with missing methods at some sites

[AlexD]

I'm planning an analysis following the new Multi-methods extension of Nichols et al 2008.

5 different methods were being compared at up to 54 sites.
But some of the detection sites did not have the full complement of methods, e.g. cameras were only used at 24 of the sites.

I can't work out how the 'missing data' would be handled, if at all!

Can I use the full 54 sites in this analysis - with totally "missing data" for some methods at some sites?
Or should I only use the 24 sites which had all methods? (and treat the remainder as separate, single-method studies)

Thanks for any input!

[jhines]

Alex,

There should be no problem using all 54 sites, using '-' for a missing value. The general model should run, but you might be restricted on what you can estimate if you build a model which tries to estimate something based on data from the missing sites.

Cheers,

Jim

[Bird Counter]

Hello,

I have a couple of follow-up questions.

If there was more than one survey method was used, but Method #1 was less effective than Method #2 (resulting in p<0.30 on sites where only #1 was used), does it make sense to include sites where only Method #1 was used? The conditional probability of a particular site not being occupied should be inflated (and uncertainty higher) on sites where detectability is low.

What are the overall consequences of this if the objective is to estimate psi-hat? Does the ratio of sites only surveyed with Method #1 to those surveyed with both methods matter? For example if 30% of sites were only surveyed with Method 1 versus 10% of sites?


Thanks much.

[jhines]

Alex,

If detection prob for each method is estimated separately, the more effective method will have a higher p, and uncertainty will be lower for sites where that method was used. For sites where only the ineffective method was used, uncertainty will be higher. The degree of uncertainty will depend on the values of the p's, but the estimate of occupancy should still be OK (unbiased).

Jim

[Bird Counter]

Oh good, I wanted to make sure I was thinking about this correctly.

Thank you for your reply.

[AlexD]

Thanks for all your help Jim!

The follow-up questions were useful, though not asked by me (Alex).

My objective is somewhat different to "Bird Counter";
At the moment I am simply comparing the methods in a pilot study - the psi estimate is unimportant for now.
I will then use the p estimates to run simulations of the main event (sampling 1 million hectares!)


I have done the analysis with the "-" for missing methods, and it all seems to run fine, but I'm slightly concerned that I'm doing naughty things with that missing data.

The model with most support is the one I was kinda expecting - attractant methods have a declining p over time, while the passive method is constant. There are also clear differences in p for all 5 methods.

Can I model this sort of more complex trend info with the missing methods? When would I start getting into dangerous territory?

Thanks - I'd really appreciate any help with this!