Although secr.fit is quite robust, it does not always
work. Inadequate data or an overambitious model occasionally cause
numerical problems in the algorithms used for fitting the model, or
problems of identifiability, as described for capture--recapture models
in general by Gimenez et al. (2004). Here are some tips that may help
you.
This page has largely been superceded by secr-troubleshooting.pdf.
Most likely you have infeasible starting values for the
parameters. try some alternatives, specifying them manually with the
start argument.
This usually means the model did not fit and the estimates should not be trusted. Extremely large variances or standard errors also indicate problems.
Try another maximization method (method = "Nelder-Mead"
is more robust than the default). The same maximum likelihood should
be found regardless of method, so AIC values are comparable across
methods.
Repeat the maximization with different starting values. You can use
secr.fit(..., start = last.model) where last.model is a
previously fitted secr object.
If you think the estimates are right but there was a problem
in computing the variances, try re-running secr.fit() with the
previous model as starting values (see preceding) and set
method = "none". This bypasses maximization and computes the
variances afresh using fdHess from nlme.
Try a finer mask (e.g., vary argument nx in
make.mask). Check that the extent of the mask matches your
data.
The maximization algorithms work poorly when the beta coefficients
are of wildly different magnitude. This may happen when using
covariates: ensure beta coefficients are similar (within a factor of
5--10 seems adequate, but this is not based on hard evidence) by scaling
any covariates you provide. This can be achieved by setting the
typsize argument of nlm or the parscale
control argument of optim.
Examine the model. Boundary values (e.g., g0 near 1.0) may give
problems. In the case of more complicated models you may gain insight by
fixing the value of a difficult-to-estimate parameter (argument
fixed).
See also the section `Potential problems' in secr-densitysurfaces.pdf.
This condition does not invariably indicate a failure of model fitting. Proceed with caution, checking as suggested in the preceding section.
A feature of the maximization algorithm used by default in nlm
is that it takes a large step in the parameter space early on in the
maximization. The step may be so large that it causes floating point
underflow or overflow in one or more real parameters. This can be
controlled by passing the `stepmax' argument of nlm in the
... argument of secr.fit (see first example). See also the
previous point about scaling of covariates.
This is a problem particularly when using individual covariates in a model fitted by maximizing the conditional likelihood. The memory required is then roughly proportional to the product of the number of individuals, the number of occasions, the number of detectors and the number of latent classes (for finite-mixture models). When maximizing the full-likelihood, substitute `number of groups' for `number of individuals'. [The limit is reached in external C used for the likelihood calculation, which uses the R function `R_alloc'.]
The mash function may be used to reduce the number of
detectors when the design uses many identical and independent
clusters. Otherwise, apply your ingenuity to simplify your model,
e.g., by casting `groups' as `sessions'. Memory is less often an issue
on 64-bit systems (see link below).
These models have known problems due to multimodality of the likelihood. See secr-finitemixtures.pdf.
Gimenez, O., Viallefont, A., Catchpole, E. A., Choquet, R. and Morgan, B. J. T. (2004) Methods for investigating parameter redundancy. Animal Biodiversity and Conservation 27, 561--572.
secr.fit, Memory-limits