Overdispersion in point DS

14 views
Skip to first unread message

Sven Buchmann

unread,
Mar 18, 2022, 6:27:08 AMMar 18
to distance...@googlegroups.com

Good morning, 

I’m a master student currently working with data from a marmot point sampling survey. I’m following a forward selection procedure with a set of covariates. I would like to hear your opinion on different ways to perform model selection when there is overdispersion in the data. 

The alpine marmot occurs in family groups, meaning one should use “family” as sampling unit. In the field, however, it is not possible to distinguish the families. Therefore, only individuals that were closer than 20m were recorded as a group with corresponding group size. 

In a first analysis I naively treated every individual as independent, i.e., I duplicated rows with n animal in the group n times. As described in Howe et al. 2018 the AIC selected for an overcomplex model. When analysing the data with clusters, AIC still tends to select complex models, indicating that there is still some overdispersion present (as expected by the field method), but not as pronounced as in the naïve approach (i.e., there are several competing models of different complexities and with overdispersion in mind one can argue for a less complex model). I performed four model selection approaches to deal with the overdispersion. 

  • No clusters + model selection with AIC: model gets way to complex 
  • No clusters + QAIC: Results in a model with ~Temperature (3 levels). 
  • Clusters + AIC: Model with ~Temperature + Weather (3 levels). However, it is followed by a more complex model (dAIC = 0.86), indicating that there is still some remaining overdispersion. 
  • With clusters + QAIC: selects null model. Visual inspection and field experience show that some covariates have an influence, so the penalty might be too rigorous.

I would appreciate to hear your expertise on this issue, since I see positive and negative points in all model selection approaches, and I don’t have arguments to prefer one of these. 

Sincerely, 

Sven 


Eric Rexstad

unread,
Mar 18, 2022, 8:16:23 AMMar 18
to Sven Buchmann, distance...@googlegroups.com
Sven

I think mechanisms for coping with overdispersion is still an area of active research, but I'll offer some observations to which others might add more sophisticated opinions.

Overdispersion is unlikely to have a profound effect upon your estimates of marmot density.  Often (including the famous amakihi point transect data set in Marques et al. (2007)) covariates in the detection function have minor effects upon the density estimates.  Examine Figure 2 of Howe et al. (2018) to see the difference in point estimates of density produced using different model selection criteria are small relative to the width of the estimate confidence intervals. I consider it to be a third-order problem.

I suggest entertaining models in the candidate model set that contain no more than X parameters, where the value of X can be argued, but perhaps no greater than 3 or 4.

From: distance...@googlegroups.com <distance...@googlegroups.com> on behalf of Sven Buchmann <sven.b...@uzh.ch>
Sent: 18 March 2022 10:26
To: distance...@googlegroups.com <distance...@googlegroups.com>
Subject: {Suspected Spam} [distance-sampling] Overdispersion in point DS
 
--
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 view this discussion on the web visit https://groups.google.com/d/msgid/distance-sampling/GV0P278MB01627B5585B3CBE2945B0DA088139%40GV0P278MB0162.CHEP278.PROD.OUTLOOK.COM.

Sven Buchmann

unread,
Mar 18, 2022, 10:15:07 AMMar 18
to Eric Rexstad, distance...@googlegroups.com

Dear Eric, 

thank you for your answer!  

My density estimates show the same behaviour, for all models the confidence intervals contain the density estimates. Would you agree that analysing the data with clusters is preferable over an individual based analysis, even though it does not capture real family sizes? I would argue that this still eliminates some of the overdispersion. The model selection could then be done with AIC and your suggestion on constraining the number of parameters. 

All the best, 

Sven  



Von: Eric Rexstad <Eric.R...@st-andrews.ac.uk>
Gesendet: Freitag, 18. März 2022 13:16
An: Sven Buchmann <sven.b...@uzh.ch>; distance...@googlegroups.com <distance...@googlegroups.com>
Betreff: Re: {Suspected Spam} [distance-sampling] Overdispersion in point DS
 

Eric Rexstad

unread,
Mar 18, 2022, 10:31:32 AMMar 18
to Sven Buchmann, distance...@googlegroups.com
Sven

I'm reluctant to offer an opinion regarding the handling of groups with camera trapping.  I would suggest the opportunity for bias to arise from inaccurate group size measurement is greater than problems arising from overdispersion and model selection, but I have no evidence with which to support that claim.

Perhaps other readers will offer their opinions.

From: Sven Buchmann <sven.b...@uzh.ch>
Sent: 18 March 2022 14:14
To: Eric Rexstad <Eric.R...@st-andrews.ac.uk>; distance...@googlegroups.com <distance...@googlegroups.com>
Subject: AW: {Suspected Spam} [distance-sampling] Overdispersion in point DS
 

Sven Buchmann

unread,
Mar 18, 2022, 10:41:28 AMMar 18
to Eric Rexstad, distance-sampling

I'm sorry, there is a misunderstanding. The survey was a traditional one without camera traps. But still I suspect overdispersion in the data, because in the field we not able to determine to which family group an individual belonged and because AIC tends to select models with too many parameters if it is allowed to. 



Von: Eric Rexstad <Eric.R...@st-andrews.ac.uk>
Gesendet: Freitag, 18. März 2022 15:31

Eric Rexstad

unread,
Mar 18, 2022, 10:51:07 AMMar 18
to Sven Buchmann, distance-sampling
Sorry Sven, my mistake.

I've been thinking about camera trapping data and falsely presumed that was the source of the data you were describing. Narrow camera fields of view add further complications to the group size issue, but that is not the situation here.

If you cannot accurately assign individuals to groups, then your estimates of group size cannot be reliable.  This is a problem common for researchers estimating density of primates in jungle habitat.  Using detections of individuals rather than groups is the preferred solution to this problem.

See the following describing the situation with primates

Buckland, S. T., Plumptre, A. J., Thomas, L., & Rexstad, E. A. (2010). Design and analysis of line transect surveys for primates. International Journal of Primatology, 31(5), 833–847. https://doi.org/10.1007/s10764-010-9431-5


Sent: 18 March 2022 14:41
To: Eric Rexstad <Eric.R...@st-andrews.ac.uk>; distance-sampling <distance...@googlegroups.com>
Subject: {Suspected Spam} AW: {Suspected Spam} [distance-sampling] Overdispersion in point DS
 

Stephen Buckland

unread,
Mar 18, 2022, 10:52:13 AMMar 18
to Sven Buchmann, Eric Rexstad, distance-sampling

I agree with Eric – I would analyse individuals.  There is some discussion of the issue on pp75-76 and p127 of the 2001 distance sampling book.

 

Steve Buckland

 


Sent: 18 March 2022 14:41
To: Eric Rexstad <Eric.R...@st-andrews.ac.uk>; distance-sampling <distance...@googlegroups.com>

Reply all
Reply to author
Forward
0 new messages