www.phidot.org

by **lenasmarques** » Mon Feb 27, 2017 4:00 pm

Hello all:

I have two independent datasets of voles’ detection data, one from live-trapping and the other from non-invasive genetics of fecal samples. The main objective of this study was to compare both methods to evaluate the use of noninvasive genetics to estimate population densities of small mammals.

Both datasets were analyzed with secr independently, the trap data using a model for multi-catch traps and the noninvasive genetics data using polygon detectors with detections modeled as counts, both with exponential detection functions.

Other than comparing the density estimates obtained from each method, i would also like to compare detectability parameters, especially g0, but i am not confident if this comparison can be made in such a direct manner (as in detectability is more than 10 times higher for noninvasive genetics (g0 = 0.07) than for trapping data (g0 = 0.0036)). If this straightforward comparison cannot be made, is there any way to make the detection parameters comparable between both methods?

Thanks in advance,
Helena

by **murray.efford** » Tue Feb 28, 2017 5:52 am

Hello Helena
For myself, I would be very uneasy about making a direct comparison of g0 - what would it mean? At least comparison of sigma could be construed as a comparison of trap-revealed and faecal-revealed home ranges (subject to qualification because neither behaviour - defaecation or tendency to enter a trap - is necessarily uniform across the home range). Comparisons of g0 (or better lambda0, using detectfn = 'HHN') to my mind only make sense when combined with sigma. You could devise an overall measure of detection efficiency (expected no. captures per unit effort, a function of lambda0 and sigma together) and compare that between methods. I can see that standardising 'effort' would be hard for two such radically different methods (and that is part of the problem in direct comparison of g0).
I hope this is some help - I know it stops short of giving the formula for expected number of captures - I may do that later if you want to go down that path.
Murray

by **lenasmarques** » Tue Mar 21, 2017 12:59 am

Hello Murray:

Thank you for your speedy response. It was indeed helpful, at least to be aware of what I can’t do. I’ve been doing some re-reading of the bibliography, and it really doesn’t make much sense to directly compare g0 between both methods.

The initial idea was really to be able to compare detection efficiency. Our final models give us higher vole densities with the noninvasive genetics dataset. We think that’s because noninvasive genetics is more efficient in detecting voles, and it would be nice to have some metrics to confirm this. Anyway, this is not essential for us at the moment. We can just discuss our results in a different perspective.

But I have another question. You mentioned detection function HHN. I’ve used the exponential detection function in the models for the polygon detectors, because AICc of the basic model was lower when using the exponential in comparison with using the hazard exponential. Do you still advise to use the hazard exponential in this case? Or is it better to stick with the exponential? And if I continue to use the exponential, g0 still gives me a detection probability, or the expected number of captures?

Thanks,
Helena

(sorry for the long delay of my response)

by **murray.efford** » Mon Mar 27, 2017 5:13 am

Sorry for the delay...

On your second paragraph -- a higher detection rate from genetic sampling is _not_ sufficient to explain a higher estimate of density. However there are related explanations that revolve around heterogeneity (differential catchability of different individuals). Most simply: if some population subclass is drastically undersampled by live trapping then it can drag down the whole estimate. There was evidence of this from pitfall trapping of voles by Charley Krebs' group way back, as you probably know. Faecal DNA seems much less susceptible to this sort of bias (depending on where you get it from - maybe young voles are less likely to use latrines). In nonspatial capture-recapture there is also a link between heterogeneity and detection rate - if all animals have high detection rates there is less scope for heterogeneity - high mean(p) implies low CV(p).

I have a personal preference for using the detection functions expressed as hazard (HHN, HEX etc.) (difficult to explain, but hazard is additive while probability is not). I am very surprised you got a better AICc from HEX than EX. The implementation of polygon detectors was a bit confused on exactly the point you mention (g0 vs lambda0), but I have cleaned that up: if you can wait for the next release secr 3.0 in a few weeks it would be worth re-running some models as a check. But I still think HHN, HEX etc. is the cleanest formulation.

Cheers
Murray

by **lenasmarques** » Tue Mar 28, 2017 10:48 pm

Hi again:

The best model for the trapping data includes a behavioral response and a time trend on detectability. Initially when I was beginning to analyze the data I ran models that included individual variables, like sex and age class, but it seems that there is no heterogeneity in catchability. Faecal DNA, on the other hand, is susceptible to other variables, that are included in the best model for the noninvasive genetics (genotyping success and the area of each searched habitat patch). So, I think that the heterogeneity in catchability is possibly not the reason for the lower density estimates obtained with live-trapping. I´ll post the print of these models in the end so you can have a look if you want.

The difference in AICc was high (12) between HEX and EX. I’ll post the print of these models also. I must discuss with my advisers the need to re-run the models using HEX instead of EX. There are more than 100 models that need to be re-run, if this is the case, and at least one month of computation time, with several models running parallel (some of these models took almost a week to run). At this point, when the paper is almost ready to submit, we need to evaluate if these extra time and effort is worthwhile.

Once again, thank you so much for your help.

Helena

Best model for the live-trapping data

Code: Select all: > print(Model58b.V3, deriv = TRUE) secr.fit(capthist = CaptHistA.V3, model = list(D ~ 1, g0 ~ T + B, sigma ~ B), mask = MaskA.V3, CL = FALSE, detectfn = 2, start = c(2.8063528, -5.6620229, 0.1937767, 2.7593377, 1.8213741, -0.1933307), method = "Nelder-Mead", verify = TRUE, trace = TRUE) secr 2.8.2, 09:57:21 01 Jul 2014 Detector type single Detector number 371 Average spacing 4.794238 m x-range 521534.6 522328 m y-range 4178202 4179022 m Usage range by occasion 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 min 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 max 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 N animals : 31 N detections : 81 N occasions : 20 Mask area : 11.1104 ha Model : D~1 g0~T + B sigma~B Fixed (real) : none Detection fn : exponential Distribution : poisson N parameters : 6 Log likelihood : -372.8012 AIC : 757.6023 AICc : 761.1023 Beta parameters (coefficients) beta SE.beta lcl ucl D 2.8083653 0.189386184 2.4371752 3.1795554 g0 -5.6287007 0.015227904 -5.6585468 -5.5988545 g0.T 0.1916259 0.006550623 0.1787869 0.2044649 g0.BTRUE 2.7079777 0.017825394 2.6730406 2.7429149 sigma 1.8426776 0.010155018 1.8227741 1.8625810 sigma.BTRUE -0.2197838 0.013723621 -0.2466816 -0.1928860 Variance-covariance matrix of beta parameters D g0 g0.T g0.BTRUE sigma sigma.BTRUE D 0.0358671266 -1.394867e-04 -3.642875e-04 1.416620e-04 -1.035854e-04 1.092309e-04 g0 -0.0001394867 2.318891e-04 -1.002422e-06 -2.318122e-04 -1.625975e-07 3.123691e-07 g0.T -0.0003642875 -1.002422e-06 4.291066e-05 7.311885e-07 -3.942257e-06 3.252029e-06 g0.BTRUE 0.0001416620 -2.318122e-04 7.311885e-07 3.177447e-04 2.490738e-07 -5.308419e-07 sigma -0.0001035854 -1.625975e-07 -3.942257e-06 2.490738e-07 1.031244e-04 -1.029646e-04 sigma.BTRUE 0.0001092309 3.123691e-07 3.252029e-06 -5.308419e-07 -1.029646e-04 1.883378e-04 Fitted (real) parameters evaluated at base levels of covariates link estimate SE.estimate lcl ucl D log 16.582788101 3.168923e+00 11.44067739 24.036064629 g0 logit 0.003580376 5.432641e-05 0.00347546 0.003688447 sigma log 6.313420226 6.411455e-02 6.18900352 6.440338060

Best model for noninvasive genetics

Code: Select all: > print(ModelR52.V4b, deriv = TRUE) secr.fit(capthist = CaptXY1.V4, model = list(D ~ 1, g0 ~ sa + log(tarea_m2), sigma ~ 1), mask = Mask1.V4, CL = FALSE, detectfn = 2, start = c(3.22372542, 4.64804125, 0.04209972, -0.95883076, 1.63094581), method = "Nelder-Mead", verify = TRUE, trace = TRUE) secr 2.8.2, 19:01:18 20 Dez 2014 Detector type polygon Number vertices 265 Number polygons 21 Total area 9.867063 ha x-range 521529.8 522347.7 m y-range 4178147 4179094 m N animals : 65 N detections : 115 N occasions : 4 Count model : Poisson Mask area : 11.1104 ha Model : D~1 g0~sa + log(tarea_m2) sigma~1 Fixed (real) : none Detection fn : exponential Distribution : poisson N parameters : 5 Log likelihood : -204.8343 AIC : 419.6686 AICc : 420.6856 Beta parameters (coefficients) beta SE.beta lcl ucl D 3.22638369 0.155871514 2.92088114 3.53188624 g0 4.65592872 1.501422374 1.71319494 7.59866250 g0.sa 0.04207808 0.006365284 0.02960236 0.05455381 g0.log(tarea_m2) -0.95982951 0.182463477 -1.31745135 -0.60220767 sigma 1.63177974 0.114332966 1.40769125 1.85586824 Variance-covariance matrix of beta parameters D g0 g0.sa g0.log(tarea_m2) sigma D 0.0242959287 0.039416865 2.658278e-04 -7.811774e-03 -1.743334e-04 g0 0.0394168650 2.254269144 -2.288354e-03 -2.667201e-01 3.028020e-02 g0.sa 0.0002658278 -0.002288354 4.051684e-05 3.523713e-05 8.706982e-06 g0.log(tarea_m2) -0.0078117741 -0.266720089 3.523713e-05 3.329292e-02 -3.309909e-03 sigma -0.0001743334 0.030280196 8.706982e-06 -3.309909e-03 1.307203e-02 Fitted (real) parameters evaluated at base levels of covariates link estimate SE.estimate lcl ucl D log 25.18840301 3.95012304 18.55763209 34.1883945 g0 log 0.07241937 0.02625803 0.03636664 0.1442137 sigma log 5.11296640 0.58649624 4.08650977 6.3972502

Null model for polygon detectors, using EX

Code: Select all: > print(ModelR1.V4, deriv = TRUE) secr.fit(capthist = CaptXY1.V4, model = list(D ~ 1, g0 ~ 1, sigma ~ 1), mask = Mask1.V4, CL = FALSE, detectfn = 2, start = NULL, method = "BFGS", verify = TRUE, trace = TRUE) secr 2.10.4, 22:41:08 12 Mar 2017 Detector type polygon Number vertices 265 Number polygons 21 Total area 9.867063 ha x-range 521529.8 522347.7 m y-range 4178147 4179094 m N animals : 65 N detections : 115 N occasions : 4 Count model : Poisson Mask area : 11.1104 ha Model : D~1 g0~1 sigma~1 Fixed (real) : none Detection fn : exponential Distribution : poisson N parameters : 3 Log likelihood : -243.6708 AIC : 493.3416 AICc : 493.7351 Beta parameters (coefficients) beta SE.beta lcl ucl D 2.4809079 0.1391105 2.208256 2.7535594 g0 -0.9366488 0.1337347 -1.198764 -0.6745336 sigma 1.5666668 0.1027948 1.365193 1.7681410 Variance-covariance matrix of beta parameters D g0 sigma D 0.0193517297 -0.008350741 -0.0004722569 g0 -0.0083507409 0.017884971 0.0027955625 sigma -0.0004722569 0.002795562 0.0105667740 Fitted (real) parameters evaluated at base levels of covariates link estimate SE.estimate lcl ucl D log 11.9521105 1.6707404 9.0998353 15.698410 g0 log 0.3919391 0.0526511 0.3015667 0.509394 sigma log 4.7906535 0.4937581 3.9164777 5.859949

Null model for polygon detectors, using HEX

Code: Select all: > print(ModelR1.V4Z, deriv = TRUE) secr.fit(capthist = CaptXY1.V4, model = list(D ~ 1, g0 ~ 1, sigma ~ 1), mask = Mask1.V4, CL = FALSE, detectfn = 16, start = NULL, method = "BFGS", verify = TRUE, trace = TRUE) secr 2.10.4, 05:26:57 08 Mar 2017 Detector type polygon Number vertices 265 Number polygons 21 Total area 9.867063 ha x-range 521529.8 522347.7 m y-range 4178147 4179094 m N animals : 65 N detections : 115 N occasions : 4 Count model : Poisson Mask area : 11.1104 ha Model : D~1 lambda0~1 sigma~1 Fixed (real) : none Detection fn : hazard exponential Distribution : poisson N parameters : 3 Log likelihood : -249.7629 AIC : 505.5259 AICc : 505.9193 Beta parameters (coefficients) beta SE.beta lcl ucl D 2.531826 0.1418696 2.253766 2.809885 lambda0 -1.086550 0.1299047 -1.341158 -0.831941 sigma 1.526036 0.1170496 1.296623 1.755449 Variance-covariance matrix of beta parameters D lambda0 sigma D 0.0201269873 -0.008856176 -0.0002039613 lambda0 -0.0088561763 0.016875236 0.0025374188 sigma -0.0002039613 0.002537419 0.0137006034 Fitted (real) parameters evaluated at base levels of covariates link estimate SE.estimate lcl ucl D log 12.5764453 1.79323091 9.5235369 16.6080077 lambda0 log 0.3373786 0.04401262 0.2615426 0.4352037 sigma log 4.5999069 0.54026658 3.6569267 5.7860451

by **murray.efford** » Wed Mar 29, 2017 2:38 am

Hi

It's unseemly to get into an argument about your model set and data (you know them better than me), but I have my doubts about the wisdom of running 100 models - perhaps we could discuss some of those choices offline. I agree it is not smart to re-fit all those models.

I really would like to understand the reason for the AIC difference between EX and HEX. Could you possibly send me the R objects CaptXY1.V4 and Mask1.V4 in confidence? Possibly

Code: Select all: save(CaptXY1.V4 , Mask1.V4, file = 'dataformurray.RData')

Murray

www.phidot.org

Detectability parameters for polygon detectors

Detectability parameters for polygon detectors

Re: Detectability parameters for polygon detectors

Re: Detectability parameters for polygon detectors

Re: Detectability parameters for polygon detectors

Re: Detectability parameters for polygon detectors

Re: Detectability parameters for polygon detectors

Who is online