score.test: Score Test for SECR Models

In secr: Spatially Explicit Capture-Recapture

Score Test for SECR Models

Description

Compute score tests comparing a fitted model and a more general alternative model.

Usage

score.test(secr, ..., betaindex = NULL, trace = FALSE, ncores = NULL, .relStep = 0.001,
    minAbsPar = 0.1)

score.table(object, ..., sort = TRUE, dmax = 10)

Arguments

secr	fitted secr model
...	one or more alternative models OR a fitted secr model
trace	logical. If TRUE then output one-line summary at each evaluation of the likelihood
ncores	integer number of threads for parallel processing
.relStep	see fdHess
minAbsPar	see fdHess
betaindex	vector of indices mapping fitted values to parameters in the alternative model
object	score.test object or list of such objects
sort	logical for whether output rows should be in descending order of AIC
dmax	threshold of dAIC for inclusion in model set

Details

Score tests allow fast model selection (e.g. Catchpole & Morgan 1996). Only the simpler model need be fitted. This implementation uses the observed information matrix, which may sometimes mislead (Morgan et al. 2007). The gradient and second derivative of the likelihood function are evaluated numerically at the point in the parameter space of the second model corresponding to the fit of the first model. This operation uses the function fdHess of the nlme package; the likelihood must be evaluated several times, but many fewer times than would be needed to fit the model. The score statistic is an approximation to the likelihood ratio; this allows the difference in AIC to be estimated.

Covariates are inferred from components of the reference model secr. If the new models require additional covariates these may usually be added to the respective component of secr.

Mapping of parameters between the fitted and alternative models sometimes requires user intervention via the betaindex argument. For example betaindex = c(1,2,4) is the correct mapping when comparing the null model (D\sim{~}1, g0\sim{~}1, sigma\sim{~}1) to one with a behavioural effect on g0 (D\sim{~}1, g0\sim{~}b, sigma\sim{~}1).

The arguments .relStep and minAbsPar control the numerical gradient calculation and are passed directly to fdHess. More investigation is needed to determine optimal settings.

score.table summarises one or more score tests in the form of a model comparison table. The ... argument here allows the inclusion of additional score test objects (note the meaning differs from score.test). Approximate AIC values are used to compute relative AIC model weights for all models within dmax AIC units of the best model.

If ncores = NULL then the existing value from the environment variable RCPP_PARALLEL_NUM_THREADS is used (see setNumThreads).

Value

An object of class ‘score.test’ that inherits from ‘htest’, a list with components

statistic	the value the chi-squared test statistic (score statistic)
parameter	degrees of freedom of the approximate chi-squared distribution of the test statistic (difference in number of parameters H0, H1)
p.value	probability of test statistic assuming chi-square distribution
method	a character string indicating the type of test performed
data.name	character string with null hypothesis, alternative hypothesis and arguments to function call from fit of H0
H0	simpler model
np0	number of parameters in simpler model
H1	alternative model
H1.beta	coefficients of alternative model
AIC	Akaike's information criterion, approximated from score statistic
AICc	AIC with small-sample adjustment of Hurvich & Tsai 1989

If ... defines several alternative models then a list of score.test objects is returned.

The output from score.table is a dataframe with one row per model, including the reference model.

Note

This implementation is experimental. The AIC values, and values derived from them, are approximations that may differ considerably from AIC values obtained by fitting and comparing the respective models. Use of the observed information matrix may not be optimal.

Weights were based on AICc rather than AIC prior to version 5.0.0.

References

Catchpole, E. A. and Morgan, B. J. T. (1996) Model selection of ring-recovery models using score tests. Biometrics 52, 664–672.

Hurvich, C. M. and Tsai, C. L. (1989) Regression and time series model selection in small samples. Biometrika 76, 297–307.

McCrea, R. S. and Morgan, B. J. T. (2011) Multistate mark-recapture model selection using score tests. Biometrics 67, 234–241.

Morgan, B. J. T., Palmer, K. J. and Ridout, M. S. (2007) Negative score test statistic. American statistician 61, 285–288.

See Also

AIC, LR.test

Examples

## Not run: 
    AIC (secrdemo.0, secrdemo.b)
    st <- score.test (secrdemo.0, g0 ~ b)
    st
    score.table(st)

    ## adding a time covariate to separate occasions (1,2) from (3,4,5)
    secrdemo.0$timecov <- data.frame(t2 = factor(c(1,1,2,2,2)))
    st2 <- score.test (secrdemo.0, g0 ~ t2)
    score.table(st,st2)

## End(Not run)

secr documentation built on Nov. 4, 2024, 9:06 a.m.