High P values and backtransform error in occupancy model output

[sbuitr...@gmail.com]

Hello everybody,

I am trying to run a single season occupancy model based on some data collected through camera traps and I have two questions that I hope someone could hep me to answer. 

1. After using the function dredge and selecting the best model based on AICc, I see that the estimates obtained for each of my parameters have very high values (way above the expected 0.05). I have seen that in some cases people do not pay attention to the P values in occupancy models (for example in chapter 10 of AHM1 they got very high P values but did not pay attention to those), but I would like to know the reason? Or if is there something else that I should check? Here is my model output:

 Occupancy (logit-scale):

                    Estimate   SE      z P(>|z|)
(Intercept)             38.4  221  0.174   0.862
Diversidad_Shannon     -34.3  217 -0.158   0.875
Distancia_Cultivos    -174.8 1079 -0.162   0.871
Distancia_Vias          49.9  222  0.225   0.822
Distancia_Incendios    -71.3  490 -0.146   0.884
NDVI                  -177.4 1170 -0.152   0.879
Altitud                124.7  716  0.174   0.862
Inclinacion            -64.0  529 -0.121   0.904

Detection (logit-scale):
            Estimate    SE     z  P(>|z|)
(Intercept)    -2.59 0.328 -7.88 3.32e-15
trigger        -1.93 0.709 -2.72 6.45e-03

My data has 63 sites and 35 surveys that represent 35 days of cameras being active. I checked with a null model that my detection probability is very low, this may be a reason for the high P values?

2. I have tried to backtrasform the results from the previous model, but even though I followed the steps in "Overview of Unmarked" (Fiske and Chandler, 2019), I keep getting the next error. 

Error in (function (cond)  :
  error in evaluating the argument 'obj' in selecting a method for function 'backTransform': ncol(coefficients) == length(est) is not TRUE

Thanks in advance to anyone who can help me with my doubts. I have to say that I am a little bit new with these models.

Kind regards, 

[Ken Kellner]

I wouldn't say that you should "expect" p-values below 0.05. It could be that none of your covariates affect occupancy probability, and thus all p > 0.05.

However the results you present indicate a problem with the model fit, likely some kind of failure in the optimization. You shouldn't be getting estimates and SEs with absolute values that large. This could be because of low detection probability, but I would guess it is more likely that you have problems with your occupancy covariates. Are the estimates more reasonable values when you fit the null model? I'd suggest fitting models iteratively, making them progressively more complex until you can identify which covariate(s) might be causing the problem.

It is hard to answer your question about backTransform without more information, e.g. additional code showing exactly what you are trying to run.

Ken

[Jim Baldwin]

Suppose the overall detection probability at a site is 1 (the best it can be).  Then your model is essentially performing logistic regression with 7 predictors and just 65 observations.  I'm probably more conservative than most in that I think that maybe only 2 or 3 covariates at a time should be considered with such a limited number of sites (as the measurement being just 0 or 1 is awfully limited information).  And I do understand the difficulty and high cost of getting even 65 sites surveyed.

So, in short, I'm just suggesting to take Ken's suggestion one step farther:  start with 1 or 2 covariates and then stop.

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+u...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/unmarked/175cff7e-f358-4f5a-ae1f-c757b5d9af12n%40googlegroups.com.

[sbuitr...@gmail.com]

Hello Ken and Jim,

I truly appreciated your answers. I followed your instructions and tried to start by a null model and the results seem much better. Nevertheless, I now have another question. I tried to check the GOF of my model using the Mackenzie and Bailey approach, but I am getting some measures of bad fit. I do believe that the reason of this is the detection history from one of my sites (in which I recorded a presence in every visit, something that was not even close to happen in the other places). Here are the results of my MB GOF for the null model:

MacKenzie and Bailey goodness-of-fit for single-season occupancy model

Pearson chi-square table:

        Cohort Observed Expected Chi-square
0000000      0       12    10.91       0.11
0010000      0        1     0.30       1.68
0100000      0        1     0.30       1.68
0100001      0        1     0.07      13.28
000000.      1       30    31.37       0.06
000001.      1        2     1.01       0.96
000010.      1        1     1.01       0.00
000011.      1        1     0.22       2.67
000100.      1        1     1.01       0.00
001000.      1        2     1.01       0.96
010000.      1        2     1.01       0.96
100000.      1        1     1.01       0.00
101001.      1        1     0.05      18.04
111111.      1        1     0.00    1817.75
00.000.      2        1     0.77       0.07
0.0000.      3        1     0.77       0.07
0..00..      4        1     0.84       0.03
.00000.      5        1     0.77       0.07
...0000      6        1     1.60       0.23
...1000      6        1     0.07      11.96

Chi-square statistic = 1879.478
Number of bootstrap samples = 1000
P-value = 0.006

Quantiles of bootstrapped statistics:
  0%  25%  50%  75% 100%
  34  127  188  300 9969

Estimate of c-hat = 6.77

Can any of you give me an advice or direction on how to proceed here? I already read the Mackenzie and Bailey paper and it seems  like this also happened in their example data with salamanders (some sites with unusual detection histories).

Thanks ahead of time for your consideration, any help will be much appreciated.