Clarifications on the pcountOpen model output

458 views
Skip to first unread message

Clem N

unread,
Feb 11, 2018, 7:55:18 PM2/11/18
to unmarked
Hello,

Being rather new to statistics and to this type of models, I find myself struggling to be sure about my model results and how to interpret them correctly. I hope you will forgive the elementary nature of my questions…

I am studying the temporal trends of several bird species. Here is an example of the model used for one species : 

### unmarked frame

> summary(umf_sp1)
unmarkedFrame
Object


25 sites
Maximum number of observations per site: 21
Mean number of observations per site: 19.12
Number of primary survey periods: 7
Number of secondary survey periods: 3
Sites with at least one detection: 25


Tabulation of y observations:
   
0    1    2    3    4    5    6 <NA>
 
156  163  111   30   14    3    1   47


Site-level covariates:
  w_area             w_height          w_dens      
 
Min.   :-1.28928   Min.   :-1.5452   Min.   :-1.94598  
 
1st Qu.:-0.88181   1st Qu.:-0.8470   1st Qu.:-0.49685  
 
Median :-0.01456   Median : 0.2234   Median :-0.08281  
 
Mean   : 0.00000   Mean   : 0.0000   Mean   : 0.00000  
 
3rd Qu.: 0.90019   3rd Qu.: 0.7819   3rd Qu.: 0.33123  
 
Max.   : 2.14665   Max.   : 1.8989   Max.   : 2.40143  


Observation-level covariates:
  dAs  
 
Min.   :-1.5017  
 
1st Qu.:-0.9049  
 
Median : 0.0009  
 
Mean   : 0.0000  
 
3rd Qu.: 0.9733  
 
Max.   : 1.7983  
 
NAs   :47      


Yearly-site-level covariates:
  dOs               rod               year        
 
Min.   :-2.2336   Min.   :-1.0154   Min.   :-1.1514  
 
1st Qu.:-0.7556   1st Qu.:-0.8070   1st Qu.:-0.8635  
 
Median : 0.4289   Median :-0.1817   Median :-0.2878  
 
Mean   : 0.0000   Mean   : 0.0000   Mean   : 0.0000  
 
3rd Qu.: 0.8048   3rd Qu.: 0.6519   3rd Qu.: 0.5757  
 
Max.   : 1.1796   Max.   : 3.3611   Max.   : 2.0149  
                   
NAs   :3      




### model

mod_sp1
<- pcountOpen(~1 + w_area + w_height, ~1 + rod + dOs, ~1, ~1 + w_area + w_dens + w_height + dAs + year, umf_sp1, mixture="P", K=40, dynamics="trend")





### model results

> summary(mod_sp1)
Abundance (log-scale):
             
Estimate    SE      z  P(>|z|)
(Intercept)     3.1471 0.207 15.190 4.09e-52
w_area        
-0.0395 0.200 -0.198 8.43e-01
w_height      
-0.0949 0.151 -0.630 5.29e-01


Growth Rate (log-scale):
           
Estimate     SE      z P(>|z|)
(Intercept)  -0.0521 0.0392 -1.327   0.184
rod          
0.0134 0.0204  0.656   0.512
dOs          
-0.0199 0.0228 -0.873   0.383


Detection (logit-scale):
               
Estimate     SE       z  P(>|z|)
(Intercept)     -2.7830 0.1448 -19.225 2.30e-82
w_area          
0.1300 0.2097   0.620 5.35e-01
w_dens          
0.0377 0.0709   0.532 5.95e-01
w_height        
0.1318 0.1699   0.776 4.38e-01
dAs              
0.2756 0.0663   4.157 3.22e-05
year            
0.3018 0.1305   2.312 2.08e-02




After that, I backtransformed my estimates with the linear combination method, for the intercept and the covariates, such as : 

### backtransformation

lambda_int      
<- backTransform(linearComb(mod_sp1, type="lambda", coefficients=c(1,0,0)))
lambda_w_area  
<- backTransform(linearComb(mod_sp1, type="lambda", coefficients=c(0,1,0)))
lambda_w_height
<- backTransform(linearComb(mod_sp1, type="lambda", coefficients=c(0,0,1)))




And so on for all the estimates, to obtain this table : 

### backtransformed estimates

lambda_int     lambda_w_area      lambda_w_height
23.26807342    0.96122565         0.90950410
         
gamma_int      gamma_rod          gamma_dOs
0.94927353     1.01349871         0.98030617

det_int        det_w_area         det_w_dens       det_w_height     det_dAs         det_year
0.05825069     0.53244820         0.50942617       0.53290999       0.56847207      0.57488290



So now I have a few questions/clarifications to be sure that I understood the meaning of these estimates properly : 

1) lambda intercept gives the mean abundance for the first year of study for the covariates fixed at their mean value (because they are scaled). So it is an average value for one hypothetical site, not over the whole "metapopulation"?

2) gamma intercept gives the finite rate of increase. Is it as an average over all sites? and over the whole study period?

3) the detection estimate seems to be very low (here p=5.8%, and it is one of the highest of all my species). Is it the mean probability of detection of an individual rather than the global p for the entire species? Is it also a mean value over the whole study period?

4) Probably a very basic algebra question : Is there a way to calculate the effect of a covariate on the detection estimate when they are backtransformed? It is easy for lambda and gamma, because their exp() backtransformation has a simple rule of exp(a+b) = exp(a) * exp(b), so what was additive now has to be simply multiplied to see the backtransformed effects on the parameter. But for detection, due to the more complex logit() / logistic() backtransformation, I cannot find a way to combine directly those backtransformed covariates…


Thanks in advance for any help!

Clem

Jim Baldwin

unread,
Feb 12, 2018, 2:11:53 PM2/12/18
to unma...@googlegroups.com
I think the main issue for the backtransformations looking odd is that you need to include the intercept always in all backtransformations when you want a prediction.

If you want the relative effect of changing a covariate by 1 unit (or some meaningful number of units) while keeping the other covariates constant, then that is another story.  Also, I would imagine that using "1 standardized unit" will usually make no sense as that depends on the the variability of the covariate in your data.  For example you'd probably want to plug in the associated standardized value for 1 year rather than the value 1.0.

The detection estimate of 5.8% is for a single visit.  For 3 visits the detection probability would be 1 - (1-0.05825069)^3 = 0.1647703.  The estimate of abundance at the average of the covariates is lambda_int = 23.26807342 but the maximum count for any one visit is 6 and most of the time it is 2 or less.  So does 5.8% still seem low?

I recognize that my response certainly doesn't answer all (or maybe even any) of your questions.

Jim


--
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+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

Jim Baldwin

unread,
Feb 13, 2018, 12:02:00 PM2/13/18
to unma...@googlegroups.com
Hopefully this is a more informative response.  When faced with estimated equations that deal with different facets of a predicted outcome, I need to draw a picture.  Below is one way to look at the effects of each predictor.  (Note that I only have the "standardized" predictors and not the predictors in their original units.  If you supply the necessary values to "unstandardize" the predictors, I can make the display more informative.)  I've attached the Wolfram CDF (computable document format) code and a screenshot below:

Inline image 1
Such displays can give you a "feeling" for the importance (or lack of importance) of each variable by moving the sliders.  (Categorical variables and confidence intervals can also be included.  Never go anywhere without some sort of measure of precision.)  The free CDF player can be downloaded at www.wolfram.com/cdf-player and works on Windows, Mac, and some flavors of Linux.  (I'm not associated with or promoting Wolfram products.  It just happens to be the software that I know how to use.  There's stuff in R-Studio that will accomplish the same thing.)  Oh...and it's a huge file to download: 2 or 3 gb but you only have to do it once.

Jim

Jim Baldwin

unread,
Feb 13, 2018, 12:03:30 PM2/13/18
to unma...@googlegroups.com
Sorry.  Forgot to include the CDF document.
pcountOpen example.cdf
Message has been deleted

Clem N

unread,
Feb 15, 2018, 6:46:35 PM2/15/18
to unmarked
Hi Jim,

Thank you so much for talking the time to answer! (And I'm sorry for such a late answer, life really knows how to get in the way sometimes…)
Your first message really gave me some valuable insight and a new perspective on my data, so that was a huge help.
I'll definitely look into Wolfram, it looks really interesting, thanks for the suggestion!
Reply all
Reply to author
Forward
0 new messages