bbsBayes

CRAN_Status_Badge

This README file provides an overview of the functionality that can be accomplished with ‘bbsBayes’. It is intended to provide enough information for users to perform, at the very least, replications of status and trend estimates from the Canadian Wildlife Service and/or United States Geological Survey. However, it provides more in-depth and advanced examples for subsetting data, custom regional summaries, and more.

Additional resources: Introductory bbsBayes Workshop Journal Article with worked example (preprint)

Overview

bbsBayes is a package to perform hierarchical Bayesian analysis of North American Breeding Bird Survey (BBS) data. ‘bbsBayes’ will run a full model analysis for one or more species that you choose, or you can take more control and specify how the data should be stratified, prepared for JAGS, or modelled.

Installation

Option 1: Stable release from CRAN (currently v2.3.8.2020)

# To install v2.3.8.2020 from CRAN:
install.packages("bbsBayes")

Option 2: Less-stable development version

# To install the development version from GitHub:
install.packages("devtools")
library(devtools)
devtools::install_github("BrandonEdwards/bbsBayes")

Basic Status and Trend Analyses

bbsBayes provides functions for every stage of Breeding Bird Survey data analysis.

Data Retrieval

You can download BBS data by running fetch_bbs_data. This will save it to a package-specific directory on your computer. You must agree to the terms and conditions of the data usage before downloading. You only need run this function once for each annual update of the BBS database.

fetch_bbs_data()

Data Preparation

Stratification

Stratification plays an important role in trend analysis. Use the stratify() function for this job. Set the argument by to stratify by the following options:

stratified_data <- stratify(by = "bbs_cws")

Jags Data

JAGS models require the data to be sent as a list depending on how the model is set up. prepare_jags_data subsets the stratified data based on species and wrangles relevent data to use for JAGS models.

jags_data <- prepare_jags_data(stratified_data, 
                               species_to_run = "Barn Swallow",
                                 min_max_route_years = 5,
                                 model = "gamye",
                                 heavy_tailed = T)

Note: This can take a very long time to run

MCMC

Once the data has been prepared for JAGS, the model can be run. The following will run MCMC with default number of iterations. Note that this step usually takes a long time (e.g., 6-12 hours, or even days depending on the species, model). If multiple cores are available, the processing time is reduced with the argument parallel = TRUE.

mod <- run_model(jags_data = jags_data)

Alternatively, you can set how many iterations, burn-in steps, or adapt steps to use, and whether to run chains in parallel

jags_mod <- run_model(jags_data = jags_data,
                               n_saved_steps = 1000,
                               n_burnin = 10000,
                               n_chains = 3,
                               n_thin = 10,
                               parallel = FALSE,
                          parameters_to_save = c("n", "n3", "nu", "B.X", "beta.X", "strata", "sdbeta", "sdX"),
                          modules = NULL)

The run_model function generates a large list (object jagsUI) that includes the posterior draws, convergence information, data, etc.

Convergence

The run_model() function will send a warning if Gelman-Rubin Rhat cross-chain convergence criterion is > 1.1 for any of the monitored parameters. Re-running the model with a longer burn-in and/or more posterior iterations or greater thinning rates may improve convergence. The seriousness of these convergence failures is something the user must interpret for themselves. In some cases some parameters of the model may not be separately estimable, but if there is no direct inference drawn from those separate parameters, their convergence may not be necessary. If all or the vast majority of the n parameters have converged (e.g., you’re receiving this warning message for other monitored parameters), then inference on population trajectories and trends from the model are reliable.

jags_mod$n.eff #shows the effective sample size for each monitored parameter
jags_mod$Rhat # shows the Rhat values for each monitored parameter

If important monitored parameters have not converged, we recommend inspecting the model diagnostics with the package ggmcmc.

install.packages("ggmcmc")
S <- ggmcmc::ggs(jags_mod$samples,family = "B.X") #samples object is an mcmc.list object
ggmcmc::ggmcmc(S,family = "B.X") ## this will output a pdf with a series of plots useful for assessing convergence. Be warned this function will be overwhelmed if trying to handle all of the n values from a BBS analysis of a broad-ranged species

Alternatively, the shinystan package has some wonderful interactive tools for better understanding convergence issues with MCMC output.

install.packages("shinystan")
my_sso <- shinystan::launch_shinystan(shinystan::as.shinystan(jags_mod$samples, model_name = "My_tricky_model"))

bbsBayes also includes a function to help re-start an MCMC chain, so that you avoid having to wait for an additional burn-in period.

### if jags_mod has failed to converge...
new_initials <- get_final_values(jags_mod)

jags_mod2 <- run_model(jags_data = jags_data,
                               n_saved_steps = 1000,
                               n_burnin = 0,
                               n_chains = 3,
                               n_thin = 10,
                               parallel = FALSE,
                               inits = new_initials,
                          parameters_to_save = c("n", "n3", "nu", "B.X", "beta.X", "strata", "sdbeta", "sdX"),
                          modules = NULL)

Model Predictions

There are a number of tools available to summarize and visualize the posterior predictions from the model. ### Annual Indices of Abundance and Population Trajectories The main monitored parameters are the annual indices of relative abundance within a stratum (i.e., parameters “n[strata,year]”). The time-series of these annual indices form the estimated population trajectories.

indices <- generate_indices(jags_mod = jags_mod,
                            jags_data = jags_data)

By default, this function generates estimates for the continent (i.e., survey-wide) and for the individual strata. However, the user can also select summaries for composite regions (regions made up of collections of strata), such as countries, provinces/states, Bird Conservation Regions, etc. For display, the posterior medians are used for annual indices (instead of the posterior means) due to the asymetrical distributions caused by the log-linear retransformation.

indices <- generate_indices(jags_mod = jags_mod,
                            jags_data = jags_data,
                            regions = c("continental",
                            "national",
                            "prov_state",
                            "stratum"))
                            #also "bcr", "bcr_by_country"

Population trends can be calculated from the series of annual indices of abundance. The trends are expressed as geometric mean rates of change (%/year) between two points in time. \(Trend = (\frac {n[Minyear]}{n[Maxyear]})^{(1/(Maxyear-Minyear))}\)

trends <- generate_trends(indices = indices,
                          Min_year = 1970,
                          Max_year = 2018)

The generate_trends function returns a dataframe with 1 row for each unit of the region-types requested in the generate_indices function (i.e., 1 for each stratum, 1 continental, etc.). The dataframe has at least 27 columns that report useful information related to each trend, including the start and end year of the trend, lists of included strata, total number of routes, number of strata, mean observed counts, and estimates of the % change in the population between the start and end years.

The generate_trends function includes some other arguments that allow the user to adjust the quantiles used to summarize uncertainty (e.g., interquartile range of the trend estiamtes, or the 67% CIs), as well as include additional calculations, such as the probability a population has declined (or increased) by > X%.

trends <- generate_trends(indices = indices,
                          Min_year = 1970,
                          Max_year = 2018,
                          prob_decrease = c(0,25,30,50),
                          prob_increase = c(0,33,100))

Visualizing Predictions

Population Trajectories

Generate plots of the population trajectories through time. The plot_indices() function produces a list of ggplot figures that can be combined into a single pdf file, or printed to individual devices.

tp = plot_indices(indices = indices,
                         species = "Barn Swallow",
                         add_observed_means = TRUE)
# pdf(file = "Barn Swallow Trajectories.pdf")
# print(tp)
# dev.off()
print(tp[[1]])

print(tp[[2]]) 

etc.

Trend Maps

The trends can be mapped to produce strata maps coloured by species population trends.

mp = generate_map(trends,
                  select = TRUE,
                  stratify_by = "bbs_cws",
                  species = "Barn Swallow")
print(mp)

Geofacet Trajectories

For stratifications that can be compiled by political regions (i.e., bbs_cws, bbs_usgs, or state), the function geofacet_plot will generate a ggplot object that plots the state and province level population trajectories in facets arranged in an approximately geographic arrangement. These plots offer a concise, range-wide summary of a species’ population status and trends.

  gf <- geofacet_plot(indices_list = indices,
                     select = T,
                     stratify_by = "bbs_cws",
                     multiple = T,
                     trends = trends,
                     slope = F,
                     species = "Barn Swallow")
  
  #png("BARS_geofacet.png",width = 1500, height = 750,res = 150)
  print(gf)
  #dev.off()

EXAMPLE - Replicating the CWS status and trend estimates (2018 version onwards)

The CWS analysis, as of the 2018 BBS data-version, uses the GAMYE model. It also monitors two estimates of the population trajectory: * one for visualizing the trajectory that includes the annual fluctuations estimated by the year-effects “n” * and another for calculation trends using a trajectory that removes the annual fluctuations around the smooth “n3”.

The full script to run the CWS analysis for the 2018 BBS data version is accessible here: https://github.com/AdamCSmithCWS/BBS_Summaries

species.eng = "Pacific Wren"

stratified_data <- stratify(by = "bbs_cws") #same as USGS but with BCR7 as one stratum and PEI and Nova Scotia combined into one stratum

jags_data <- prepare_jags_data(strat_data = stratified_data,
                                 species_to_run = species.eng,
                                 min_max_route_years = 5,
                                 model = "gamye", 
                                 heavy_tailed = T) #heavy-tailed version of gamye model
                                 
                                 
jags_mod <- run_model(jags_data = jags_data,
                               n_saved_steps = 2000,
                               n_burnin = 10000,
                               n_chains = 3,
                               n_thin = 20,
                               parallel = F, 
                          parameters_to_save = c("n","n3","nu","B.X","beta.X","strata","sdbeta","sdX"),
                          modules = NULL) 

# n and n3 allow for index and trend calculations, other parameters monitored to help assess convergence and for model testing 

EXAMPLE - Replicating (approximately) the earlier USGS status and trend estimates (2011 - 2017 data versions)

The USGS analysis, from 2011 through 2017, uses the SLOPE model. Future analyses from the USGS will likely use the first difference model (see, Link et al. 2017 https://doi.org/10.1650/CONDOR-17-1.1)

NOTE: the USGS analysis is not run using the bbsBayes package, and so this analysis may not exactly replicate the published version. However, any variations should be very minor.

species.eng = "Pacific Wren"

stratified_data <- stratify(by = "bbs_usgs") #BCR by province/state/territory intersections

jags_data <- prepare_jags_data(strat_data = stratified_data,
                                 species_to_run = species.eng,
                                 min_max_route_years = 1,
                                 model = "slope", 
                                 heavy_tailed = FALSE) #normal-tailed version of slope model
                                 
                                 
jags_mod <- run_model(jags_data = jags_data,
                               n_saved_steps = 2000,
                               n_burnin = 10000,
                               n_chains = 3,
                               n_thin = 20,
                               parallel = FALSE,
                               track_n = FALSE,
                               parameters_to_save = c("n2"), #more on this alternative annual index below
                          modules = NULL) 

Advanced options and customized models

Alternative Models

The package has (currently) four status and trend models that differ somewhat in the way they model the time-series of observations. The four model options are slope, gam, gamye, and firstdiff.

slope

The slope option estimates the time series as a log-linear regression with random year-effect terms that allow the trajectory to depart from the smooth regression line. It is the model used by the USGS and CWS to estimate bbs trends since 2011. The basic model was first described in 2002 (Link and Sauer 2002; https://doi.org/10.1890/0012-9658(2002)083[2832:AHAOPC]2.0.CO;2) and its application to the annual status and trend estimates is documented in Sauer and Link (2011; https://doi.org/10.1525/auk.2010.09220) and Smith et al. (2014; https://doi.org/10.22621/cfn.v128i2.1565).

    #stratified_data <- stratify(by = "bbs_usgs")
    
    #jags_data_slope <- prepare_jags_data(stratified_data, 
    #                           species_to_run = "American Kestrel",
    #                           min_max_route_years = 3,
    #                           model = "slope")

    #jags_mod_full_slope <- run_model(jags_data = jags_data)
                               
    slope_ind <- generate_indices(jags_mod = jags_mod_full_slope,
                                  jags_data = jags_data_slope,
                                  regions = c("continental"))
    slope_plot = plot_indices(indices = slope_ind,
                         species = "American Kestrel SLOPE",
                         add_observed_means = TRUE)
    #png("AMKE_Slope.png", width = 1500,height = 900,res = 150)
    print(slope_plot)
    #dev.off()
    

gam

The gam option models the time series as a semiparametric smooth using a Generalized Additive Model (GAM) structure. See https://github.com/AdamCSmithCWS/Smith_Edwards_GAM_BBS for more information (full publication coming soon)

    #stratified_data <- stratify(by = "bbs_usgs")
    
    #jags_data_gam <- prepare_jags_data(stratified_data, 
    #                           species_to_run = "American Kestrel",
    #                           min_max_route_years = 3,
    #                           model = "gam")

    #jags_mod_full_gam <- run_model(jags_data = jags_data)
                               
    gam_ind <- generate_indices(jags_mod = jags_mod_full_gam,
                                jags_data = jags_data_gam,
                                regions = c("continental"))
    gam_plot = plot_indices(indices = gam_ind,
                         species = "American Kestrel GAM",
                         add_observed_means = TRUE)
    #png("AMKE_gam.png", width = 1500,height = 900,res = 150)
    print(gam_plot)
    #dev.off()
    

gamye

The gamye option includes the semiparametric smooth used in the gam option, but also includes random year-effect terms that track annual fluctuations around the smooth. This is the model that the Canadian Wildlife Service is now using for the annual status and trend estimates.

    #stratified_data <- stratify(by = "bbs_usgs")
    
    #jags_data_gamye <- prepare_jags_data(stratified_data, 
    #                           species_to_run = "American Kestrel",
    #                           min_max_route_years = 3,
    #                           model = "gamye")

    #jags_mod_full_gamye <- run_model(jags_data = jags_data)
                               
    gamye_ind <- generate_indices(jags_mod = jags_mod_full_gamye,
                                  jags_data = jags_data_gamye,
                                  regions = c("continental"))
    gamye_plot = plot_indices(indices = gamye_ind,
                         species = "American Kestrel GAMYE",
                         add_observed_means = TRUE)
    #png("AMKE_gamye.png", width = 1500,height = 900,res = 150)
    print(gamye_plot)
    #dev.off()
    

firstdiff

The firstdiff option models the time-series as a random-walk from the first year, so that the first-differences of the sequence of year-effects are random effects with mean = 0 and an estimated variance. This model has been described in Link et al. 2017 https://doi.org/10.1650/CONDOR-17-1.1

    #stratified_data <- stratify(by = "bbs_usgs")
    
    #jags_data_firstdiff <- prepare_jags_data(stratified_data, 
    #                           species_to_run = "American Kestrel",
    #                           min_max_route_years = 3,
    #                           model = "firstdiff")

    #jags_mod_full_firstdiff <- run_model(jags_data = jags_data)
                               
    firstdiff_ind <- generate_indices(jags_mod = jags_mod_full_firstdiff,
                                      jags_data = jags_data_firstdiff,
                                      regions = c("continental"))
    firstdiff_plot = plot_indices(indices = firstdiff_ind,
                         species = "American Kestrel FIRSTDIFF",
                         add_observed_means = TRUE)
    #png("AMKE_firstdiff.png", width = 1500,height = 900,res = 150)
    print(firstdiff_plot)
    #dev.off()
    

Alternate extra-Poisson error distributions

For all of the models, the BBS counts on a given route and year are modeled as Poisson variables with over-dispersion. The over-dispersion approach used here is to add a count-level random effect that adds extra variance to the unit variance:mean ratio of the Poisson. In the prepare_jags_data function, the user can choose between two distributions to model the extra-Poisson variance:

The heavy-tailed version is well supported for many species, particularly species that are sometimes observed in large groups. Note: the heavy-tailed version can require significantly more time to converge (~2-5 fold increase in processing time).

#stratified_data <- stratify(by = "bbs_usgs")
    
    #jags_data_firstdiff <- prepare_jags_data(stratified_data, 
    #                           species_to_run = "American Kestrel",
    #                           min_max_route_years = 3,
    #                           model = "firstdiff",
    #                           heavy_tailed = TRUE) 

    #jags_mod_full_firstdiff <- run_model(jags_data = jags_data)

In all the models, the default measure of the annual index of abundance (the yearly component of the population trajectory) is the derived parameter “n”. The run_model function monitors n by default, because it is these parameters that form the basis of the estimated population trajectories and trends.

Alternate retransformations

There are two ways of calculating these annual indices for each model. The two approaches differ in the way they calculate the retransformation from the log-scale model parameters to the count-scale predictions. The user can choose using the following arguments in run_model() and generate_indices().

mod <- run_model(... ,
                 parameters_to_save = "n",
                 ... )
indices <- generate_indices(... ,
                            alternate_n = "n",
                            ... )
mod <- run_model(... ,
                 parameters_to_save = "n2",
                 ... )
indices <- generate_indices(... ,
                            alternate_n = "n2",
                            ... )

The default approach parameters_to_save = c("n") slightly underestimates the uncertainty of the annual indices (slightly narrower CI width). However, we have chosen this approach as the default because:

For example, in the figures below, the predicted annual indices (blue line and CI-band) are much more similar to the observed average counts (grey dots) for the Default n approach.

Decomposing the population trajectories for two of the models

For two of the main model types "slope" and "gamye", users can choose two different ways to calculate trajectories and population trends. With these two model types, the population trajectories are composed of two largely independent components, a long-term smooth and the random annual fluctuations around that smooth. Because the two components are largely independent, the population trajectory can be decomposed.
The default approach is to include the annual fluctuations around the linear (slope) or GAM-smooth (gamye) components of the trajectories. These trend estimates are more comprehensive in that they include the full estimated trajectory, but they will vary more between subsequent years (e.g., more variability between a 1970-2017 trend and a 1970-2018 trend), because they include the effects of the annual fluctuations.

mod <- run_model(... ,
                 parameters_to_save = "n",
                 ... )
indices <- generate_indices(... ,
                            alternate_n = "n",
                            ... )

An alternative approach is to decompose the full trajectory and to exclude the annual fluctuations around the linear (slope) or smooth (gamye) components. In this case, the predicted trends will be much more stable between subsequent years. For the CWS status and trend analyses, the visualized population trajectories are calculated using the full trajectory, and the trend estimates are derived from the decomposed trajectory using only the smooth component.

mod <- run_model(... ,
                 parameters_to_save = c("n","n3"),
                 ... )
indices_visualize <- generate_indices(... ,
                                      alternate_n = "n",
                                      ... )
indices_trend_calculation <- generate_indices(... ,
                                              alternate_n = "n3",
                                              ... )

For example, the figure below (produced using a modified version of the standard plotting functions), shows the two kinds of trajectories for Pacific Wren from the 2018 CWS analysis. The light-blue trajectory is the visualized trajectory, including the yearly fluctuations. The orange trajectory is the one used for trend calculations, which includes only the GAM-smooth component. For the kinds of broad-scale status assessments that form the primary use of the published estimates of trends, this decomposition is a particularly useful feature of these two models.

The figure below provides another example of the benefits of removing the year-effect annual fluctuations when calculating trends.

Each point on the graph represents the 10-year trend estimate for Wood Thrush in Canada, ending in a given year (e.g., the points at 2015 represent the species national population trend from 2005-2015). The red and green points are the trend estimates from the default trend estimates derived from the full population trajectories for the gamye and slope models. The Blue points represent the trends calculated using the decomposed trajectory of the gamye model, including only the smooth component. When the annual fluctuations are included (SLOPE and GAMYE including Year Effects), the population trends surpass the IUCN trend-criterion, in some years (e.g., 2011) suggesting that if assessed in those years the species would be listed as Threatened (trend in the orange region). However, a more stable trend estimate from the decomposed trajectory (GAMYE - Smooth only in Blue) shows that the species is probably best thought of as in decline, but not surpassing the Threatened criterion.

Alternate Measures of Trend and Population Change

The generate_trends() function produces much more than just the trend estimates.

The default trend calculation is an interval-specific estimate of the geometric mean annual change in the population. \(Trend = (\frac {n[Minyear]}{n[Maxyear]})^{(1/(Maxyear-Minyear))}\) It relies on a comparison of the annual indices in the first and last years of the trend period to quantify the mean rate of population change. However, it ignores the pattern of change between the two end-points.

The user can choose an alternative estimate of change that is calculated by fitting a log-linear slope to the series of all annual indices between the two end-points (e.g., all 11 years in a 10-year trend from 2008-2018). The slope of this line could be expressed as an average annual percent change across the time-period of interest. If working with estimates derived from a model with strong annual fluctuations and for which no decomposition is possible (e.g., “firstdiff” model), this slope-based trend may be a more comprehensive measure of the average population change, that is less dependent on the particular end-point years. These slope trends can be added to the trend output table by setting the slope = TRUE argument in generate_trends(). The standard trends are still calculated, but additional columns are added that include the alternate estimates. NOTE: the generate_map() function can map slope trends as well with the same slope = TRUE argument.

    #jags_data_firstdiff <- prepare_jags_data(stratified_data, 
    #                           species_to_run = "American Kestrel",
    #                           model = "firstdiff")

    #jags_mod_full_firstdiff <- run_model(jags_data = jags_data)
                               
    #firstdiff_ind <- generate_indices(jags_mod = jags_mod_full_firstdiff,
    #                                  jags_data = jags_data_firstdiff,
    #                                  regions = c("continental","stratum"))
    fd_slope_trends_08_18 <- generate_trends(indices = firstdiff_ind,
                                             Min_year = 2008,
                                             Max_year = 2018,
                                             slope = TRUE)
    generate_map(fd_slope_trends_0.8_18,
                 slope = TRUE,
                 stratify_by = "bbs_usgs")
    )

Percent Change and probability of change

The generate_trends() function also produces estimates of the overall percent-change in the population between the first and last years of the trend-period. This calculation is often easier to interpret than an average annual rate of change. These percent change estimates have associated uncertainty bounds, and so can be helpful for deriving statements such as “between 2008 and 2018, the population has declined by 20 percent, but that estimate is relatively uncertain and the true decline may be as little as 2 percent or as much as 50 percent”

In addition, the function can optionally calculate the posterior conditional probability that a population has changed by at least a certain amount, using the prob_decrease and prob_increase arguments. These values can be useful for deriving statements such as “our model suggests that there is a 95% probability that the species has increased (i.e., > 0% increase) and a 45 percent probability that the species has increased more than 2-fold (i.e., > 100% increase)”

    fd_slope_trends_08_18 <- generate_trends(indices = firstdiff_ind,
                                             Min_year = 2008,
                                             Max_year = 2018,
                                             slope = TRUE,
                                             prob_increase = c(0,100))
                                                       

Custom regional summaries

Yes, you can calculate the trend and trajectories for custom combinations of strata, such as the trends for Eastern and Western populations of Lincoln’s Sparrow.

    #stratification <- "bbs_cws"
    #strat_data <- stratify(by = stratification, sample_data = TRUE)
    #jags_data <- prepare_jags_data(strat_data, 
    #                           species_to_run = "Lincoln's Sparrow",
    #                           model = "gamye")

    #jags_mod <- run_model(jags_data = jags_data)

Assuming the above setup has been run. The user could then generate population trajectories using a customized grouping of the original strata.

First extract a dataframe that defines the original strata used in the analysis.

st_comp_regions <- get_composite_regions(strata_type = stratification)

The add a column to the dataframe that groups the original strata into the desired custom regions.

st_comp_regions$East_West <- ifelse(st_comp_regions$bcr %in% c(7,8,12:14,22:31),"East","West")

st_comp_regions can now be used as the dataframe input to the argument alt_region_names in generate_indices(), with “East_West” as the value for the argument regions. The relevant trends can be calculated using just the generate_trends() function.

east_west_indices <- generate_indices(jags_mod = jags_mod,
                                      jags_data = jags_data,
                                      alt_region_names = st_comp_regions,
                                      regions = "East_West")
east_west_trends <- generate_trends(indices = east_west_indices)

Exporting the JAGS model

You can easily export any of the bbsBayes models to a text file.

model_to_file(model = "slope",
              filename = "my_slope_model.txt")

Then, you can modify the model text (e.g., try a different prior) and run the modified model

run_model <- function(... ,
                      model_file_path = "my_modified_slope_model.txt",
                      ... )

Details coming soon…

Modifying the JAGS model and data

You can even export the bbsBayes model as text, and modify it to add in covariates. For example a GAM smooth to estimate the effect of the day of year on the observations, or an annual weather covariate, or… Then add the relevant covariate data to the jags_data object, and you’re off! We’ll add some more details and examples soon.

Comparing Models

Finally, bbsBayes can be used to run Bayesian cross-validations. For example, the get_final_values() function is useful to provide an efficient starting point for a cross-validation runs, without having to wait for another full burn-in period.

Paper that includes an example of how to implement a cross-validation using bbsBayes.

Pre-print: https://doi.org/10.1101/2020.03.26.010215 Supplement: DOI

NOTE: although bbsBayes includes functions to calculate WAIC, recent work has shown that WAIC performs very poorly with the BBS data (https://doi.org/10.1650/CONDOR-17-1.1). We recommend a k-fold cross-validation approach, as in the above zenodo archive.