unmarked modelnonparboot.RdResamples sites (with replacement) from an unmarkedFit object and
re-fits the model, to generate new non-parametric bootstrap samples of estimates.
These samples can then be used by vcov in order to get bootstrap-based
estimates of standard errors.
# S4 method for class 'unmarkedFit'
nonparboot(object, B = 1, keepOldSamples = TRUE, ...)Calling nonparboot on an unmarkedFit returns the original
unmarkedFit, with the bootstrap samples added on.
Then subsequent calls to vcov with the argument
method="nonparboot" will use these bootstrap samples.
Additionally, standard errors of derived estimates from either
linearComb or backTransform can be
instructed to use bootstrap samples by providing the argument
method = "nonparboot".
For all fitting functions, only sites are re-sampled. In previous versions of
unmarked both sites and occasions
of occu and occuRN fits were re-sampled, but now
only sites are re-sampled for these fit types to match the other fit types.
Note: by default, if the model fit to a new resampled dataset results in an
error or a convergence failure, nonparboot will continually
re-sample and re-fit until this is no longer the case. This could
result in the function hanging.
data(ovendata)
ovenFrame <- unmarkedFrameMPois(ovendata.list$data,
siteCovs=as.data.frame(scale(ovendata.list$covariates[,-1])), type = "removal")
(fm <- multinomPois(~ 1 ~ ufc + trba, ovenFrame))
#>
#> Call:
#> multinomPois(formula = ~1 ~ ufc + trba, data = ovenFrame)
#>
#> Abundance (log-scale):
#> Estimate SE z P(>|z|)
#> (Intercept) 0.102 0.119 0.864 0.388
#> ufc 0.100 0.126 0.794 0.427
#> trba -0.171 0.135 -1.262 0.207
#>
#> Detection (logit-scale):
#> Estimate SE z P(>|z|)
#> 0.288 0.233 1.24 0.217
#>
#> AIC: 326.1387
#> Number of sites: 70
#>
fm <- nonparboot(fm, B = 20) # should use larger B in real life.
#>
| | 0 % ~calculating
|+++ | 5 % ~00s
|+++++ | 10% ~00s
|++++++++ | 15% ~00s
|++++++++++ | 20% ~00s
|+++++++++++++ | 25% ~00s
|+++++++++++++++ | 30% ~00s
|++++++++++++++++++ | 35% ~00s
|++++++++++++++++++++ | 40% ~00s
|+++++++++++++++++++++++ | 45% ~00s
|+++++++++++++++++++++++++ | 50% ~00s
|++++++++++++++++++++++++++++ | 55% ~00s
|++++++++++++++++++++++++++++++ | 60% ~00s
|+++++++++++++++++++++++++++++++++ | 65% ~00s
|+++++++++++++++++++++++++++++++++++ | 70% ~00s
|++++++++++++++++++++++++++++++++++++++ | 75% ~00s
|++++++++++++++++++++++++++++++++++++++++ | 80% ~00s
|+++++++++++++++++++++++++++++++++++++++++++ | 85% ~00s
|+++++++++++++++++++++++++++++++++++++++++++++ | 90% ~00s
|++++++++++++++++++++++++++++++++++++++++++++++++ | 95% ~00s
|++++++++++++++++++++++++++++++++++++++++++++++++++| 100% elapsed=00s
vcov(fm, method = "hessian")
#> lambda(Int) lambda(ufc) lambda(trba) p(Int)
#> lambda(Int) 0.014052472 -1.228179e-03 2.364962e-03 -4.323007e-03
#> lambda(ufc) -0.001228179 1.594197e-02 8.057614e-03 -3.200697e-11
#> lambda(trba) 0.002364962 8.057614e-03 1.831812e-02 -3.459628e-11
#> p(Int) -0.004323007 -3.200697e-11 -3.459628e-11 5.415986e-02
vcov(fm, method = "nonparboot")
#> lambda(Int) lambda(ufc) lambda(trba) p(Int)
#> lambda(Int) 0.028845631 -0.001084086 0.0049178635 -0.0157796398
#> lambda(ufc) -0.001084086 0.013770757 0.0110038361 -0.0046697094
#> lambda(trba) 0.004917864 0.011003836 0.0186061311 0.0009194527
#> p(Int) -0.015779640 -0.004669709 0.0009194527 0.0948140633
avg.abundance <- backTransform(linearComb(fm, type = "state", coefficients = c(1, 0, 0)))
## Bootstrap sample information propagates through to derived quantities.
vcov(avg.abundance, method = "hessian")
#> [,1]
#> [1,] 0.01724521
vcov(avg.abundance, method = "nonparboot")
#> [,1]
#> [1,] 0.03539939
SE(avg.abundance, method = "nonparboot")
#> [1] 0.1881473