I'm having a problem with some models in PRESENCE not converging on the correct maximum likelihood values. I'm modeling frog calling survey data collected over 3 surveys at 42 locations. I'm using two site covariates: one describes the distance to the nearest dam and the other is the percentage of urban land cover within a buffer zone surrounding the site. Strangely, when considering two models with the exact same Psi parameterization, I get convergence with the model incorporating survey-specific detection, but fail to get results with constant detection (which, of course, has fewer parameters). For example, with the model Psi (Distance and Urban), p (t), which incorporates the additive effects of the two site covariates and survey-specific detection (a total of 6 parameters), I get the following output:
Custom Model:
Number of parameters = 6
Number of significant digits = 7.0
Number of function calls = 232
-2log(likelihood) = 77.9793
AIC = 89.979277
Model has been fit using the logistic link.
Untransformed Estimates of coefficients for covariates (Beta's)
==============================================================================
estimate std.error
A1
ccupancy psi 0.699522 (4.993891)A2
ccupancy psiDistance -0.839609 (3.209474)A3
ccupancy psi%_Urban -1.112380 (3.389385)B1 :detection p1 -2.527900 (1.537090)
B2 :detection p2 -1.052536 (1.742663)
B3 :detection p3 -1.751560 (1.563875)
Variance-Covariance Matrix of Untransformed estimates (Beta's):
A1 A2 A3 B1 B2 B3
A1 24.938947 -15.381930 -16.297533 -6.719803 -8.394042 -7.302120
A2 -15.381930 10.300725 10.592774 4.121308 5.148132 4.478448
A3 -16.297533 10.592774 11.487930 4.380191 5.471515 4.759765
B1 -6.719803 4.121308 4.380191 2.362647 2.276872 1.980690
B2 -8.394042 5.148132 5.471515 2.276872 3.036876 2.474179
B3 -7.302120 4.478448 4.759765 1.980690 2.474179 2.445704
The standard errors are large, but at least PRESENCE computes them. However, for the model Psi (Distance and Urban) p (.) (only 4 parameters), I get the following:
Custom Model:
Number of parameters = 4
Number of significant digits = 6.2
Number of function calls = 266
-2log(likelihood) = 78.6339
AIC = 86.633892
Model has been fit using the logistic link.
********************** WARNING ***************************
Variance-covariance matrix has not been computed successfully.
Ignore matrix values in the below output.
(sum=-0.000282)
--------------------------------------------------------------
Untransformed Estimates of coefficients for covariates (Beta's)
==============================================================================
estimate std.error
A1
ccupancy psi 328.618177 (-1.#IND00)A2
ccupancy psiDownDistance -271.797008 (-1.#IND00)A3
ccupancy psi%_Urban -312.655292 (-1.#IND00)B1 :detection p1 -1.956839 (3.374978)
Variance-Covariance Matrix of Untransformed estimates (Beta's):
A1 A2 A3 B1
A1 -0.000000 0.000000 0.000000 0.000000
A2 0.016780 -0.000000 -0.000000 -0.000000
A3 0.851691 -0.000000 -0.000000 0.000000
B1 0.050719 -0.000000 0.000000 11.390476
I've followed the advice I read on the forum and tried to set initial values myself, with no satisfactory result. In fact, the results listed above used the exact betas from the Psi (Distance and Urban), p (t) model as initial values (I guessed -2 as the constant detection value) and it still didn't converge. Are there any other strategies I can try to get the model to run properly? As I said, this particular example confused me because it was the model with fewer parameters that failed to converge. Any help will be greatly appreciated.
Evan