Implications of non-independence between sessions in Multi-session models

123 views
Skip to first unread message

mirjam....@gmail.com

unread,
Dec 4, 2023, 7:08:57 PM12/4/23
to secr
Hello Murray and everyone, 

When using multi-session models to estimate density for multiple years, but on species where there is considerable overlap in individuals between years (i.e., sessions are not independent) - what are the main caveats?

Murray mentions in the multi-session documentation that the density estimate errors will be too narrow for each year. 

Wouldn't one expect other issues? Would it also impact estimates of detection covariates, density covariates, etc? Would IC based model selection be biased? 

I'm finding several papers that published density trends for large carnivores that don't even mention this concern. 

Any insights and pointers would be much appreciated. 

Best, 

Mirjam 

Murray Efford

unread,
Dec 4, 2023, 8:59:13 PM12/4/23
to secr
Hi Mirjam

Interesting. I think the problem can be overstated, but it is there, and could use some serious attention (have you tried simulation?). Repeated sampling of the same AC while assuming independence results in overdispersion. Overdispersion inflates true sampling variance and reduces CI coverage, but has negligible effect on estimates themselves (does not cause bias). IC-based model selection is affected in principle, but should we really be taking that so seriously anyway?

Thinking aloud... The overdispersion we are talking about affects the 'n' component of the sampling variance of density (number of individuals detected), not the 'a' component (detection). Sampling the same population over and over is equivalent to sampling a highly clustered population. At worst (no turnover, no movement between sessions, high detection rate, low variance in a-hat), the true 'n' variance component might be inflated by the number of sessions (#AC per cluster).

Other approaches: try an open model (not my preferred solution because there are big overheads and the open model itself may be wrong); join multi-session data as single session (obvious risk of bias from turnover within duration of study, but maybe informative on variance).

Murray

mirjam....@gmail.com

unread,
Dec 7, 2023, 11:32:02 PM12/7/23
to secr
Thank you, Murray. 
This is really helpful. I'll try some simulations and also joining the data as a single session, just to see what happens. I still have to do some more reading to make sure I fully understand your comments and the implications for my analysis, but am a bit relieved so far as it appears that my main concerns are not necessarily warranted. We plan to use open pop models for the actual density estimation, which we have done in the past, so we may disagree here on preferred solutions :). 

Best, 
Mirjam

Murray Efford

unread,
Dec 10, 2023, 4:43:37 PM12/10/23
to secr
Well, I really don't like open models for density estimation, so expect some push back if I get to review your paper! The problem of temporal overdispersion does not arise if you use Horvitz-Thompson (derived) density estimates.That is a nice out: you get to fit a pooled (multi-year) model for estimating detection parameters, and AIC comparisons of such models are unaffected by temporal overdispersion. Annual CI for density are based on the number detected in each year, rather than the inflated number (summed over years). Of course this does not allow any model for density, specifically a trend model, which may be the whole point of the analysis.
Murray

mirjam....@gmail.com

unread,
Dec 11, 2023, 7:30:35 PM12/11/23
to secr
Interesting. Horvitz-Thompson estimates wouldn't allow for an inhomogeneous density model, though, either, right? Which is the whole point of the analysis, and necessary -  our study area is too large and too inhomogeneous to not use density models when estimating density. We're not as much interested in temporal trends, as the study didn't last long enough to detect temp trends unless they're incredibly drastic. 

So what do you suggest researchers do who cannot ignore spatial effects on density in their final models, but still need to sample large areas, usually over several years (to answer other questions, too - can't just count them) to get density estimates because their study species are wide ranging and also long lived but hard to detect? 

Best, 
Mirjam

Murray Efford

unread,
Dec 11, 2023, 8:17:10 PM12/11/23
to secr
Collect lots of data, learn Nimble, and don't expect too much.
Reply all
Reply to author
Forward
0 new messages