dNmixture: dNmixture distribution for use in 'nimble' models

dNmixtureR Documentation

dNmixture distribution for use in nimble models

Description

dNmixture_s and dNmixture_v provide Poisson-Binomial mixture distributions of abundance ("N-mixture") for use in nimble models. Overdispersion alternatives using the negative binomial distribution (for the abundance submodel) and the beta binomial distribution (for the detection submodel) are also provided.

Usage

dNmixture_v(x, lambda, prob, Nmin = -1, Nmax = -1, len, log = 0)

dNmixture_s(x, lambda, prob, Nmin = -1, Nmax = -1, len, log = 0)

rNmixture_v(n, lambda, prob, Nmin = -1, Nmax = -1, len)

rNmixture_s(n, lambda, prob, Nmin = -1, Nmax = -1, len)

dNmixture_BNB_v(x, lambda, theta, prob, Nmin = -1, Nmax = -1, len, log = 0)

dNmixture_BNB_s(x, lambda, theta, prob, Nmin = -1, Nmax = -1, len, log = 0)

dNmixture_BNB_oneObs(x, lambda, theta, prob, Nmin = -1, Nmax = -1, log = 0)

dNmixture_BBP_v(x, lambda, prob, s, Nmin = -1, Nmax = -1, len, log = 0)

dNmixture_BBP_s(x, lambda, prob, s, Nmin = -1, Nmax = -1, len, log = 0)

dNmixture_BBP_oneObs(x, lambda, prob, s, Nmin = -1, Nmax = -1, log = 0)

dNmixture_BBNB_v(x, lambda, theta, prob, s, Nmin = -1, Nmax = -1, len, log = 0)

dNmixture_BBNB_s(x, lambda, theta, prob, s, Nmin = -1, Nmax = -1, len, log = 0)

dNmixture_BBNB_oneObs(x, lambda, theta, prob, s, Nmin = -1, Nmax = -1, log = 0)

rNmixture_BNB_v(n, lambda, theta, prob, Nmin = -1, Nmax = -1, len)

rNmixture_BNB_s(n, lambda, theta, prob, Nmin = -1, Nmax = -1, len)

rNmixture_BNB_oneObs(n, lambda, theta, prob, Nmin = -1, Nmax = -1)

rNmixture_BBP_v(n, lambda, prob, s, Nmin = -1, Nmax = -1, len)

rNmixture_BBP_s(n, lambda, prob, s, Nmin = -1, Nmax = -1, len)

rNmixture_BBP_oneObs(n, lambda, prob, s, Nmin = -1, Nmax = -1)

rNmixture_BBNB_v(n, lambda, theta, prob, s, Nmin = -1, Nmax = -1, len)

rNmixture_BBNB_s(n, lambda, theta, prob, s, Nmin = -1, Nmax = -1, len)

rNmixture_BBNB_oneObs(n, lambda, theta, prob, s, Nmin = -1, Nmax = -1)

Arguments

x

vector of integer counts from a series of sampling occasions.

lambda

expected value of the Poisson distribution of true abundance

prob

detection probability (scalar for dNmixture_s, vector for dNmixture_v).

Nmin

minimum abundance to sum over for the mixture probability. Set to -1 to select automatically (not available for beta binomial variations; see Details).

Nmax

maximum abundance to sum over for the mixture probability. Set to -1 to select automatically (not available for beta binomial variations; see Details).

len

The length of the x vector

log

TRUE or 1 to return log probability. FALSE or 0 to return probability.

n

number of random draws, each returning a vector of length len. Currently only n = 1 is supported, but the argument exists for standardization of "r" functions.

theta

abundance overdispersion parameter required for negative binomial (*NB) N-mixture models. The negative binomial is parameterized such that variance of x is lambda^2 * theta + lambda

s

detection overdispersion parameter required for beta binomial (BB*) N-mixture models. The beta binomial is parameterized such that variance of x is V(x) = N * prob * (1-prob) * (N + s) / (s + 1)

Details

These nimbleFunctions provide distributions that can be used directly in R or in nimble hierarchical models (via nimbleCode and nimbleModel).

An N-mixture model defines a distribution for multiple counts (typically of animals, typically made at a sequence of visits to the same site). The latent number of animals available to be counted, N, follows a Poisson or negative binomial distribution. Each count, x[i] for visit i, follows a binomial or beta-binomial distribution. The N-mixture distributions calculate the marginal probability of observed counts by summing over the range of latent abundance values.

The basic N-mixture model uses Poisson latent abundance with mean lambda and binomial observed counts with size (number of trials) N and probability of success (being counted) prob[i]. This distribution is available in two forms, dNmixture_s and dNmixture_v. With dNmixture_s, detection probability is a scalar, independent of visit, so prob[i] should be replaced with prob above. With dNmixture_v, detection probability is a vector, with one element for each visit, as written above.

We also provide three important variations on the traditional N-mixture model: dNmixture_BNB, dNmixture_BBP, and dNmixture_BBNB. These distributions allow you to replace the Poisson (P) abundance distribution with the negative binomial (NB) and the binomial (B) detection distribution with the beta binomial (BB).

Binomial-negative binomial: BNB N-mixture models use a binomial distribution for detection and a negative binomial distribution for abundance with scalar overdispersion parameter theta (0-Inf). We parameterize such that the variance of the negative binomial is lambda^2 * theta + lambda, so large theta indicates a large amount of overdisperison in abundance. The BNB is available in three suffixed forms: dNmixture_BNB_v is used if prob varies between observations, dNmixture_BNB_s is used if prob is scalar (constant across observations), and dNmixture_BNB_oneObs is used if only one observation is available at the site (so both x and prob are scalar).

Beta-binomial-Poisson: BBP N-mixture uses a beta binomial distribution for detection probabilities and a Poisson distribution for abundance. The beta binomial distribution has scalar overdispersion parameter s (0-Inf). We parameterize such that the variance of the beta binomial is N * prob * (1-prob) * (N + s) / (s + 1), with greater s indicating less variance (greater-than-binomial relatedness between observations at the site) and s -> 0 indicating the binomial. The BBP is available in three suffixed forms: dNmixture_BBP_v is used if prob varies between observations, dNmixture_BBP_s is used if prob is scalar (constant across observations), and dNmixture_BBP_oneObs is used if only one observation is available at the site (so both x and prob are scalar).

Beta-binomial-negative-binomial: dNmixture_BBNB is available using a negative binomial abundance distribution and a beta binomial detection distribution. dNmixture_BBNB is available with _s, _v, and _oneObs suffixes as above and requires both arguments s and theta as parameterized above.

The distribution dNmixture_oneObs is not provided as the probability given by the traditional N-mixture distribution for length(x) = 1 is equivalent to dpois(prob * lambda).

For more explanation, see package vignette (vignette("Introduction_to_nimbleEcology")).

Compared to writing nimble models with a discrete latent state of abundance N and a separate scalar datum for each count, use of these distributions allows one to directly sum (marginalize) over the discrete latent state N and calculate the probability of all observations for a site jointly.

If one knows a reasonable range for summation over possible values of N, the start and end of the range can be provided as Nmin and Nmax. Otherwise one can set both to -1, in which case values for Nmin and Nmax will be chosen based on the 0.0001 and 0.9999 quantiles of the marginal distributions of each count, using the minimum over counts of the former and the maximum over counts of the latter.

The summation over N uses the efficient method given by Meehan et al. (2020, see Appendix B) for the basic Poisson-Binomial case, extended for the overdispersion cases in Goldstein and de Valpine (2022).

These are nimbleFunctions written in the format of user-defined distributions for NIMBLE's extension of the BUGS model language. More information can be found in the NIMBLE User Manual at https://r-nimble.org.

When using these distributions in a nimble model, the left-hand side will be used as x, and the user should not provide the log argument.

For example, in nimble model code,

observedCounts[i, 1:T] ~ dNmixture_v(lambda[i], prob[i, 1:T], Nmin, Nmax, T)

declares that the observedCounts[i, 1:T] (observed counts for site i, for example) vector follows an N-mixture distribution with parameters as indicated, assuming all the parameters have been declared elsewhere in the model. As above, lambda[i] is the mean of the abundance distribution at site i, prob[i, 1:T] is a vector of detection probabilities at site i, and T is the number of observation occasions. This will invoke (something like) the following call to dNmixture_v when nimble uses the model such as for MCMC:

dNmixture_v(observedCounts[i, 1:T], lambda[i], prob[i, 1:T], Nmin, Nmax, T, log = TRUE)

If an algorithm using a nimble model with this declaration needs to generate a random draw for observedCounts[1:T], it will make a similar invocation of rNmixture_v, with n = 1.

If the observation probabilities are visit-independent, one would use:

observedCounts[1:T] ~ dNmixture_s(observedCounts[i, 1:T], lambda[i], prob[i], Nmin, Nmax, T)

Value

For dNmixture_s and dNmixture_v: the probability (or likelihood) or log probability of observation vector x.

For rNmixture_s and rNmixture_v: a simulated detection history, x.

Notes for use with automatic differentiation

The N-mixture distributions are the only ones in nimbleEcology for which one must use different versions when AD support is needed. See dNmixtureAD.

Author(s)

Ben Goldstein, Lauren Ponisio, and Perry de Valpine

References

D. Turek, P. de Valpine and C. J. Paciorek. 2016. Efficient Markov chain Monte Carlo sampling for hierarchical hidden Markov models. Environmental and Ecological Statistics 23:549–564. DOI 10.1007/s10651-016-0353-z

Meehan, T. D., Michel, N. L., & Rue, H. 2020. Estimating Animal Abundance with N-Mixture Models Using the R—INLA Package for R. Journal of Statistical Software, 95(2). https://doi.org/10.18637/jss.v095.i02

Goldstein, B.R., and P. de Valpine. 2022. Comparing N-mixture Models and GLMMs for Relative Abundance Estimation in a Citizen Science Dataset. Scientific Reports 12: 12276. DOI:10.1038/s41598-022-16368-z

See Also

For occupancy models dealing with detection/nondetection data, see dOcc. For dynamic occupancy, see dDynOcc.

Examples

# Set up constants and initial values for defining the model
len <- 5 # length of dataset
dat <- c(1,2,0,1,5) # A vector of observations
lambda <- 10 # mean abundance
prob <- c(0.2, 0.3, 0.2, 0.1, 0.4) # A vector of detection probabilities

# Define code for a nimbleModel
 nc <- nimbleCode({
   x[1:5] ~ dNmixture_v(lambda, prob = prob[1:5],
                        Nmin = -1, Nmax = -1, len = 5)

   lambda ~ dunif(0, 1000)

   for (i in 1:5) {
     prob[i] ~ dunif(0, 1)
   }
 })

# Build the model
nmix <- nimbleModel(nc,
                    data = list(x = dat),
                    inits = list(lambda = lambda,
                                 prob = prob))
# Calculate log probability of data from the model
nmix$calculate()
# Use the model for a variety of other purposes...

nimbleEcology documentation built on June 27, 2024, 5:09 p.m.