N-Mixture for gibbon abundance: in need of feedback on results.

177 views
Skip to first unread message

camille c

unread,
Sep 15, 2014, 9:22:14 AM9/15/14
to unma...@googlegroups.com

Hello all,

I have applied the N-mixture Pcount model for gibbon survey data using the count of groups heard from 'listening posts' on 4 consecutive mornings at 30 different sites (one listening post in each of 30 consecutive grid cells of 4km2).
In order for each of my 'site' to be independent (e.g. no double counting of groups between two sites), I have included only the groups counted that were estimated to be located within 1 km-radius from the listening post (see map bellow). So my 'site' is arbitrarily defined as 1km-radius around each of my listening posts i.e. 3.14 km2 (so that I can then infer density estimate from my abundance per site estimate).

I have used some site covariates (elevation, ruggedness, observer, distance to main water source) and some survey covariates (temperature, humidity, windspeed, date)

I have run 13 different models. Here is a table that shows the results of all models:


model name

λ

±SE 

p

±SE 

k

AIC

AIC Δ

AIC wgt

Cum wgt

p Value

λ(.)p(Wind+Temp)

8.43

±1.97

0.27

±0.06

4

397.34

0.00

0.96

0.96

0.446

λ(distWater)p(Temp)

7.74

±1.89

0.29

±0.07

4

404.66

7.33

0.03

0.99

0.119

λ(.)p(Temp)

8.16

±1.92

0.28

±0.07

3

406.93

9.60

0.01

1.00

0.208

λ(distWater)p(Wind)

5.29

±0.85

0.43

±0.06

4

409.19

11.85

0.00

1.00

0.347

λ(distWater)p(.)

5.33

±0.89

0.43

±0.06

3

416.27

18.93

0.00

1.00

0.475

λ(.)p(Wind)

5.34

±0.79

0.45

±0.06

3

416.83

19.49

0.00

1.00

0.297

λ(.)p(Rain)

5.85

±1.02

0.40

±0.07

3

417.85

20.51

0.00

1.00

0.139

λ(.)p(Observer)

6.32

±1.16

0.37

±0.06

3

419.60

22.27

0.00

1.00

0.277

λ(Rugged)p(.)

5.37

±0.87

0.44

±0.06

3

420.70

23.36

0.00

1.00

0.248

λ(.)p(Humidity)

5.42

±0.83

0.44

±0.06

3

420.91

23.58

0.00

1.00

0.228

λ(.)p(Elev)

5.46

±0.83

0.44

±0.06

3

423.01

25.67

0.00

1.00

0.248

λ(.)p(.)

5.39

±0.83

0.44

±0.06

2

423.52

26.18

0.00

1.00

0.079

λ(Elev)p(.)

5.40

±0.84

0.44

±0.06

3

424.67

27.33

0.00

1.00

0.228

λ(.)p(Date)

5.38

±0.82

0.45

±0.06

3

425.46

28.12

0.00

1.00

0.129

λ: abundance; p: probability of detection; SE: standard error;  k: number of parameters estimated in the model; AIC: Alkaike Information Criterion = -2 (ln[maximum likelihood]) + 2 K; Δ AIC: best model’s AIC – AIC of model i; AIC wgt: weight of model i; cum wgt: cumulative models’ AIC wgt; p Value: p value of the goodness of fit test, using the Χ2 statistic with 100 bootstraps, p>0.05 indicates no evidence of lack of fit.



A few points I would like some feedback on:


1- What surprised me at first was the very low detection probability for the top ranking model (p=0.27), although in the field we could hear gibbons (at least one group, but mostly > one group) almost on any given day. Why would the probability of detection be so low?


2- The probability of detection between the different models varies quite a lot as well. Can the survey covariates used really influence that much estimates of p ? (note that wind and temperature are actually known to influence gibbon call frequencies)

3- There is a relatively high variation in abundance (
λ) estimates between the different models run. Does that make sense to have very these different estimates (or at least between the top-ranked ones and the lowest-ranked ones) between the different models run?


This is the first time this method is applied to gibbons so it is very interesting to test its potential for these species; I would really appreciate any comments on the results.


Thank you very much for your help,


Camille.

www.conservationlaos.com

Jeffrey Royle

unread,
Sep 15, 2014, 10:10:20 AM9/15/14
to unma...@googlegroups.com
Dear Camille,
 It makes sense that estimates of abundance should vary quite a bit once you start explaining heterogeneity in detection probability. That makes sense to me.
 
 As for "p" changing so much among models, the interpretation of that is based on the intercept of the model for logit(p) and how the covariates are defined. Are they categorical levels or continuous? Are they standardized?  If they are categorical ("factors") then the interpretation of p is at the baseline level (the summary output should make this clear). If they are continuous but not standardized then the interpretation is at the value "x = 0", so you should probably standardize each continuous covariate.
 
 Does p = 0.27 seem reasonable?  Keep in mind that this is INDIVIDUAL level p and not SURVEY level p.  For example, if there are N = 2 groups then the "survey level" p is :   1- (1-0.27)^2  = 0.47 or, in general,   1-(1-0.27)^N
which will be much higher.
 
regards,
andy
 

--
You received this message because you are subscribed to the Google Groups "unmarked" group.
To unsubscribe from this group and stop receiving emails from it, send an email to unmarked+u...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

Kery Marc

unread,
Sep 15, 2014, 10:52:00 AM9/15/14
to unma...@googlegroups.com

Hi Camille and Andy,

 

what is the counting unit ? Do you count the groups or really individuals ? If the former, then p is for the group.

 

Regards  --- Marc

camille c

unread,
Sep 16, 2014, 3:19:41 AM9/16/14
to unma...@googlegroups.com
Dear Andy and Marc,

Thank you for your replies.

The counting unit is gibbon group.

so if I understand correctly the detection probability p is the average detection for any single group (INDIVIDUAL level) but not the probability of detecting at least one group on any given day (SURVEY level)?...so because I had most of the time a detection of >1 group, my SURVEY detection should always be much higher than the INDIVIDUAL detection. If that's right, then it makes much more sense to me then because detection was very high (0.27 seemed too low...)

My covariates -- survey and site -- are continuous but they were standardized in Unmarked using following code:
oc <- obsCovs(Umf_gibbon)
oc[,1:6] <- scale(oc[,1:6])
obsCovs(Umf_gibbon) <- oc

sc <- siteCovs(Umf_gibbon)
sc[,1:7] <- scale(sc[,1:7])
siteCovs(Umf_gibbon) <- sc

So, the p change between model might just be what it is, because of the covariates having strong influence on detection?

Thank you so much for the help and comment, this is very useful.

Best,
Camille.

Jeffrey Royle

unread,
Sep 16, 2014, 7:04:51 AM9/16/14
to unma...@googlegroups.com
hi Camille,
 Yes your statements are correct.
 And, because covariates are scaled, I think the change in p among models is simply due to strong effects of the covariates.
regards,
andy

camille c

unread,
Sep 17, 2014, 4:38:40 AM9/17/14
to unma...@googlegroups.com
Thank you so much; Very helpful!
I think the N-mixture model is very applicable to gibbon surveys!

Best,
Camille.
<td width="67" nowrap valign="top" style="padding:0c
...
Reply all
Reply to author
Forward
0 new messages