Changelog • crywolf

crywolf 1.2.0

fp_occu(random_fp = TRUE) adds a site random intercept to the logit of the false-positive rate, for rates that vary among sites in ways the covariates do not capture; its prior scale is fp_priors(fp_site_sd = ). It is off by default: it slows the fits and mixes less well, and it does not replace covariates when false positives follow an environmental gradient (see ?fp_occu). With the option off, the draws are identical to those of crywolf 1.1.0.
fp_calibrate_fit() checks the calibration of candidate models on data simulated from a fitted model, which keep the design of the real data (sites, covariates, occasions, recorders, and verification).
simulate() method for fp_occu fits, from the posterior means or from random posterior draws.
The summary of a calibration prints every occupancy slope, not only occ_x.

The likelihood is summed over sites with Stan’s reduce_sum(), and the model is compiled with threading, so fp_occu(threads_per_chain = ) can split each chain among several cores. On the curiango data (97,806 automated occasions) a fit with four chains of four threads each took 16.5 minutes instead of 70, and each gradient evaluation was 2.7 times faster than with the previous model in a short benchmark on the same computer. The log density and its gradient are unchanged up to floating-point rounding, and the automated occasions use y * lp - n * log1p_exp(lp), an equivalent form of the binomial log-likelihood with half the transcendental functions. Executables compiled by earlier versions are rebuilt once.

First release, accompanying the simulation study of Provete et al. (fpcalib research compendium, v0.1.0, doi:10.5281/zenodo.23379100).

fp_occu() fits the multiple-methods false-positive occupancy model of Miller et al. (2011) in Stan, with any number of occupancy, detection, and false-positive covariates, an optional recorder random intercept, and configurable priors (fp_priors()). fp = NULL gives the standard occupancy model.
Observation-confirmation design of Chambert et al. (2015): argument confirmed of fp_data() and frac_confirmed of simulate_fp().
Methods print, summary, coef, predict, loo, and as_draws_df.
fp_ppc(): posterior predictive check with a Bayesian p-value.

simulate_fp() generates homogeneous, unstructured, and structured false-positive regimes; with the same seed it returns the same datasets as the simulator of the fpcalib compendium.
fp_calibrate() and fp_verification_curve() report the coverage and bias of credible intervals for a design, with a plot() method and a progress bar that shows the time remaining (needs ‘progressr’). They fit the recorder random intercept only when the simulated design has one, which avoids divergent transitions.

curiango dataset: acoustic monitoring of the Common Pauraque in the Pantanal (Piatti et al. 2026).
Articles: introduction, simulation, calibration, a case study with the curiango data, coming from unmarked::occuFP(), and the Stan model.