Multi-method error

[NicolaG]

Hi!

I am having issues running a single-season, multi-method occupancy model in PRESENCE, with two methods. When I create a new project, if I enter 2 for the #methods/survey on the opening page, I get the error "Numerical convergence may not have been reached. Parameter estimates converged to approximately 3.19 significant digits." when I run the null model with only the detection history. I have tried to change the initial value (using Supple Initial Values under Options), but I continue to get the error. However, if I don't enter 2 for the #methods/survey when I'm creating the project and just leave it blank, but when I run the multi-method model, I say the number of methods is 2 when prompted, the null model runs fine, and I don't get an error.

Is it ok to run it the second way, or is there a way to fix the error if I run it the first way?

To provide a bit of background on my data. I have two methods, Sherman trapping and camera trapping, which were conducted over 69 transects. The detection history is formatted with Method1Day1, Method2Day1, Method1Day2, and Method2Day2 along the columns in my spreadsheet. I have a maximum of 7 days (repeated surveys) for the Sherman trapping and 12 days for the camera trapping. I have included columns for days 8-12 for the Sherman trapping with no data, which I put as "-" in each cell. The two methods were not done simultaneously at each transect but were done within a few months of each other. I have 11 transects where the interval was longer. I removed the camera trapping data for these transects and put the cells as no data. I would love to keep all the data, but I couldn't find a way to run a multi-season, multi-method model. There were three transects where the camera traps failed, so they also have no data for every day. I don't know if the number of cells with no data might be causing this issue. I have a total of 772 no data cells out of a matrix with 1,656 cells. I never had errors with the null model when I was running single-season models on each method separately. I'm not necessarily interested in looking at how effective the different methods were. Still, I want to include the detection history for both methods when I include site covariates, as they were caught on several different transects with only one method or the other.

I'm sorry, I'm rather new to occupancy modelling, so I hope that makes sense. I am also really struggling to find a tutorial on multi-method occupancy models.

I'd appreciate any help!

Thanks,

Nicola

[jhines]

Hi Nicola,

The message, "Numerical convergence may not have been reached" is a warning, not an error. It's telling you that the likelihood surface at the maximum likelihood value is somewhat "flat" meaning that the model estimates are not as precise as possible. However, my experience has been that with 3 or more significant digits in the likelihood, the estimates are accurate to 4 or more decimal digits. So, I wouldn't be concerned with the warning. The fact that it converges to the same result (and warning) with different initial values is an indication that the maximum likelihood value is not a local maximum of a surface with many maximum values.


Entering 2 when prompted should give the same result as when you entered 2 methods per survey in the initial setup. Are the estimates different? If you would like to send me (jhines@usgs.gov) a copy of your most recent project zipfile, I'd be happy to see why they differ.

The standard single-method model can compute estimates for detection for each method by building a model where every odd numbered survey has just an intercept in the design matrix and every even numbered survey has the intercept plus an additional beta for the effect of the 2nd method on detection. The reason you might want to run the multi-method model is that it's possible to break the detection process into the product of local availability in a survey (theta) and probability of detection, given availability (p). The standard (single-method) model estimates the product of those two terms.

Regardless of which of those two models you choose, it is important that species exposure to the two methods is the same for a survey. If the species was there during the Sherman trapping, but gone a few months later when the camera traps were activated, then the assumption of closure during a survey is violated, causing biased estimates.

Presence does have a multi-season, multi-method model, which requires input of the number of surveys per season. So, you'll need to decide how to define a "season", where occupancy status does not change within seasons. Otherwise, running that model should be similar to the single-season, multi-method model.

If you are familiar with the R development software, there is a package for it called, "OccupancyTuts" which is a tutorial for occupancy modeling using R. A collaborator and I developed the package for teaching an introductory occupancy modeling course. To install the tutorial in R, type:

install.packages('OccupancyTuts", repo="https://www.mbr-pwrc.usgs.gov/mbrCRAN")

[NicolaG]

Hi

Thank you for your response; I really appreciate it!

I will send over the project files.

In the version of PRESENCE (2.13.39) I’m using, the multi-method model is only an option under Run -> Analysis:single-season; there is no multi-method option under Analysis:multi-season. Is it labelled as another type of model or is this possibly due to the version of PRESENCE I’m using?

I am working my way through the OccupancyTuts to better understand occupancy models, although R is definitely not my strong suit. The tutorials have been great so far.

Thanks,

Nicola

[jhines]

Hi Nicola,

Thanks for sending the files. I've discovered that Presence was not using the value "2" that you entered when running the model. Instead, it used the value of the number of methods when you created the input pao object. Consequently, I have corrected this error and uploaded a new version of Presence. This is the reason that the first and third models you ran produced estimates of zero for the beta estimate of theta. With only one method, theta cannot be estimated, so it simply remained at the starting value. The model where you used the input pao with the number of methods set to 2 is correct and produced an estimate of ~25 for the beta associated with theta. This translates to a value of 1.0 for the probability of local occupancy, given overall occupancy.

As a side note, when a real parameter is near one of the boundaries (0 or 1), the associated beta estimate is either very large negative or positive value and the likelihood function is "flat" (ie., small changes in beta parameter values don't produce much change in the likelihood). In some of these cases, this "flatness" leads to more imprecise estimates, causing the warning about significant digits.

In this case (theta=1.0) the multi-method model is not needed since it is equivalent to the simple model without the theta parameter. However, other more complicated multi-method models may produce theta estimates < 1.0, so I wouldn't disregard the multi-method models.

[NicolaG]

Hi Jim,

Thank you for all your help.

Out of curiosity, if I run more complicated mixed-method models with site and survey covariates, and the theta is still 1.0, would it be better to just run it as a simple single-season model and include a covariate for the type of method?

Thanks,

Nicola

[jhines]

Yes, the resulting likelihood and other parameter estimates should be the same, but the precision should be better without having to estimate a parameter on the boundary. Also, modeling should be a bit easier, not having to deal with the theta parameters.

[NicolaG]

That's good to know. Thank you for all your help!

Nicola