Goodness-of-Fit Test
Simple Monte-Carlo goodness-of-fit tests for full-likelihood SECR models. The approach is to calculate a statistic from either the raw data or a fitted model, and to relate this to the distribution of the statistic under the original fitted model. The distribution is estimated by simulating data from the model, and possibly re-fitting the model to each simulated dataset.
The suitability of different test statistics has yet to be assessed.
secr.test(object, nsim = 99, statfn, fit = FALSE, seed = NULL, ncores = NULL, tracelevel = 1)
object |
a fitted secr model |
nsim |
integer number of replicates |
statfn |
function to compute a numeric vector of one or more statistics from a single-session ‘capthist’ object or from a fitted model (see Details) |
fit |
logical; if TRUE the model is re-fitted to each simulated dataset |
seed |
either NULL or an integer that will be used in a call to |
ncores |
integer number of threads for parallel processing |
tracelevel |
see |
The test statistic(s) may be computed either on a dataset or on a
fitted model, as determined by the argument fit. The single
argument expected by statfn should be either a ‘capthist’ object
(fit = FALSE) or an ‘secr’ object (fit = TRUE).
The default statistic when fit = FALSE is the proportion of
individuals observed on only one occasion, which is equivalent to
statfn = function(CH) c(f1 = sum(apply(abs(CH) > 0,1,sum) == 1) /
nrow(CH)). Repeat detections on one occasion at the same or different
detectors are not counted. The default statistic is therefore not
appropriate for some data, specifically from ‘count’ or ‘polygon’
detectors with few occasions or only one.
The default statistic when fit = TRUE is the deviance divided by
the residual degrees of freedom (i.e., statfn = function(object)
c(devdf = deviance(object) / df.residual(object))).
The reported probability (p) is the rank of the observed value in the
vector combining the observed value and simulated values, divided by
(nsim + 1). Ranks are computed with rank using the default
ties.method = "average".
Simulations take account of the usage attribute of detectors in the original capthist object, given that usage was defined and ignoreusage was not set.
Setting ncores = NULL uses the existing value from the environment variable
RCPP_PARALLEL_NUM_THREADS (see setNumThreads).
statfn may return a vector of statistics for each observed or
simulated dataset or model: then the value of each statistic will be
calculated for every simulated dataset, and summarised. If fit =
TRUE the vector of statistics may include both those computed on the
raw data (object$capthist) and the fitted model.
An object of class c('secrtest','list') with components:
object |
as input |
nsim |
as input |
statfn |
as input |
fit |
as input |
seed |
as input |
output |
list comprising the simulated values, the observed value(s), and estimated probabilities |
For multi-session input when fit = FALSE, ‘output’ is a list
in which each session provides one component.
Print and plot methods are provided for ‘secrtest’ objects.
simulate.secr is always used to simulate the raw data, but
simulate.secr does not work for all types of fitted
model. Models fitted by maximizing the likelihood conditional on
n (CL = TRUE in secr.fit) potentially include
individual covariates whose distribution in the population is
unknown. This precludes simulation, and conditional-likelihood models
in general are therefore not covered by secr.test.
Other exclusions include exotic non-binary behavioural responses (“bn", “bkn", “bkc", “Bkc" - but these are generally undocumented in any case).
If fit = TRUE then sim.secr is used.
At each simulation a new population is generated across the extent of the original mask. If the extent is unduly large then time will be wasted simulating the possibility of detection for many essentially undetectable animals. This is an argument for keeping the mask tight - large enough only to avoid mask-induced bias.
## Not run:
secr.test(secrdemo.0, nsim = 99)
secr.test(ovenbird.model.1, nsim = 20)
## example combining raw data summary and model fit
## assumes single-session
bothfn <- function(object) {
CH <- object$capthist
f1 <- sum(apply(abs(CH) > 0, 1, sum) == 1) / nrow(CH)
devdf <- deviance(object) / df.residual(object)
c(f1 = f1, devdf = devdf)
}
test <- secr.test (secrdemo.0, nsim = 19, statfn = bothfn, fit = TRUE)
test
plot(test, main = '')
## End(Not run)Please choose more modern alternatives, such as Google Chrome or Mozilla Firefox.