IP models in DENSITY vs SECR for low populations

[mgreg]

I'm just learning SECR while analyzing an older dataset of mice captured with sherman traps which I hope to combare to more current data already collected with multitraps. In this area seasonality has a large impact, with density peaking in late summer and often a siginificant (up to 90%) die-off in the winter. Less than ten individuals may be recorded for some winter sessions on a 7x7 grid over 5 nights. IP models in Program DENSITY 4.4 give me what I consider to be reasonable estimates and SE's. The interesting thing is that dispite replicating the DENSITY model exactly with ip.secr in R, R refuses to run the model (or any other combination of parameters), giving either of the following error messages depending on the parameters I try:

Error in lm.fit(x, y, offset = offset, singular.ok = singular.ok, ...) :
> NA/NaN/Inf in foreign function call (arg 4)

or

> Error in matrix(unlist(temp), nrow = 3) : 'data' must be of a vector
> type In addition: Warning messages:
> 1: In mean.default(unlist(lapply(w, dbarx)), na.rm = T) :
> argument is not numeric or logical: returning NA
> 2: In mean.default(unlist(lapply(w, dbarx)), na.rm = T) :
> argument is not numeric or logical: returning NA
> 3: In mean.default(unlist(lapply(w, dbarx)), na.rm = T) :
> argument is not numeric or logical: returning NA

This is the general model I’ve been running, and I've tried mutiple combinations of parameter variations, initial starting values, etc.

ip.secr(myCH, model = g0~1, predictorfn = pfn, predictortype = "null", start = NULL, buffer = 200, CVmax = 1, detectfn = 0, boxsize = 0.4, centre = 3, min.nsim = 100, max.nsim = 2000, var.nsim = 5000,
maxbox = 1000, trace = TRUE)

The Program Density 4.4 parameters that work are:
Parameterisasion: Probability
Population dispersion: Poisson
Initial value method: AUTO
Non-target disturbance : 0.0 (0 %)
Random generator: Intrinsic Seed 987654321
Home range statistic: Sqrt(pooled spatial variance) RPSV
Parameter transformation : Odds(p-hat)
Design - Phase 1: Full design (3 centre points) Size ± (20,20,20) %, Min repl = 100, Max repl = 2000,
CV = 1.00%, Simulations for variance : N = 1000

Any ideas as to why the model may work in DENSITY but not SECR? If ip.secr cannot handle such few captures, any opinion as to the best alternative - IP in DENSITY? ML in SECR? Bayesian methods?

Any opinions or assistance would be greatly appreciated

Thanks!

If it helps, here is a typical winter trapping session

5 18 1 F5
5 7 1 E1
5 13 1 E5
5 61 1 A7
5 36 2 G4
5 25 2 G6
5 24 2 G7
5 7 2 E2
5 11 2 E3
5 18 2 E6
5 13 2 E7
5 47 2 C1
5 26 2 A2
5 18 3 F7
5 36 3 F4
5 7 3 C1
5 11 3 C7
5 47 3 B1
5 36 4 G5
5 25 4 G7
5 19 4 E7
5 11 4 D5
5 18 -4 C1
5 13 4 B7
5 7 4 B1
5 26 4 A2
5 47 4 A5
5 24 5 G5
5 36 5 G6
5 25 5 G7
5 13 5 E7
5 47 5 A1
5 26 5 A5

[murray.efford]

The ip.secr code in 'secr' is indeed less robust with small samples than Density. I can reproduce this with your example (thanks for your detailed report). It seems that 'secr' simply fails when a simulation generates no recaptures, whereas Density tries again. I hadn't realised this and will try to fix it in the next version. In the meantime, use Density if you need ip.secr.

The reason this has taken so long to be revealed is that most of us no longer bother with IP SECR (the ip.secr code in 'secr' was added for demonstration purposes). ML SECR is so much more powerful, especially with the extensions in 'secr'. The one exception might be if you really want an unbiased estimate of the g0 parameter and your traps are quite saturated. Presumably this is not the case in your study, at least at times of low capture rate.

Some of the benefits of ML SECR are especially relevant with sparse data. It is easy to build models with pooled estimates of detection parameters (g0, sigma), and to compare these with more general models by AIC. In this way you can obtain density or population size estimates in seasons when there are too few captures for standalone estimates (on the assumption that it is legitimate to pool detection parameters across the relevant samples, of course).

Murray