Chapter 4 The model

Below is a representation of the full model used in eDNAjoint, including all model variations. Note that inclusion of catchability coefficients, \(q_k\) (Eq. 2), and the regression with site-level covariates, \(\alpha\) (Eq. 4), are optional in implementation with eDNAjoint.

A reduced version of the joint model without these variations is also described in Keller et al. 2022.

4.1 Model description

The observed count, Y, of a species at site, i, in traditional survey sample, j, of gear type, k, is drawn from either

1. a negative binomial distribution with expected species catch rate, \(\mu_{i,k}\), and an overdispersion parameter, \(\phi\) (Equation 1.1)
2. a poisson distribution with expected species catch rate, \(\mu_{i,k}\) (Equation 1.2).

A third option allows for repeated continuous observations, Y, of a species at site, i, in traditional survey sample, j, of gear type, k.

3. a gamma distribution with shape parameter, \(\alpha_{mu}\) and rate parameter, \(\beta_{mu}\). The expected species catch rate, \(\mu_{i,k}\) is equal to \(\frac{\alpha_{mu}}{\beta_{mu}}\).

\[\begin{equation} \tag{Eq. 1.1} Y_{i,j,k} \sim NegativeBinomial(\mu_{i,k}, \phi) \end{equation}\]

\[\begin{equation} \tag{Eq. 1.2} Y_{i,j,k} \sim Poisson(\mu_{i,k}) \end{equation}\]

\[\begin{equation} \tag{Eq. 1.3} Y_{i,j,k} \sim Gamma(\alpha_{mu,i,k}, \beta_{mu,i,k}) \end{equation}\]

Catchability coefficients, \(q_k\), scale the catch rates of multiple gear types relative to gear type 1 (Equation 2).

\[\begin{equation} \tag{Eq. 2} \mu_{i,k} = q_k * \mu_{i,1} \end{equation}\]

The probability of a true positive eDNA detection, \(p_{11}\), at site i, is a function of expected species catch rate, \(\mu_{i,1}\) and scaling coefficient \(\beta_i\) (Equation 3).

\[\begin{equation} \tag{Eq. 3} p_{11,i} = \frac{\mu_{i,1}}{\mu_{i,1} + e^{\beta_i}} \end{equation}\]

The scaling coefficient \(\beta_i\) relates the sensitivity of eDNA sampling to the expected species catch rate and is a function of site-level covariate coefficients, \(\alpha_n\) and site-level covariate data, \(A_{i,n}\) (Equation 4).

\[\begin{equation} \tag{Eq. 4} \beta_i = A_{i,n}^{T} \cdot \alpha_n \end{equation}\]

The total probability of eDNA detection at site i, \(p_i\), is the sum of the probability of a true positive eDNA detection at site i, \(p_{11,i}\), and the probability of a false positive eDNA detection, \(p_{10}\) (Equation 5).

\[\begin{equation} \tag{Eq. 5} p_i = p_{11,i} + p_{10} \end{equation}\]

The number of positive quantitative PCR (qPCR) eDNA detections, K, out of the number of trials, N, in eDNA water sample m at site i is drawn from a binomial distribution, with a probability of success on a single trial, \(p_i\). (Equation 6).

\[\begin{equation} \tag{Eq. 6} K_{i,m} \sim Binomial(N_{i,m}, p_i) \end{equation}\]

Three informative prior distributions are included in the model for parameters, \(p_{10}\), \(\alpha_n\), and \(\phi\) (if a negative binomial distribution is used to describe the traditional survey observations, Eq. 1.2). See below for more details.

\[\begin{equation} p_{10} \sim Beta(\alpha, \beta) \end{equation}\] \[\begin{equation} \phi \sim Gamma(\alpha, \beta) \end{equation}\] \[\begin{equation} \alpha_n ~ Normal(0,10) \end{equation}\]

4.2 Bayesian modeling: Stan

The models that can be run with eDNAjoint use Bayesian inference for parameter estimation. The models are specified in the probabilistic programming language, Stan, which uses Hamiltonian Monte Carlo to obtain posterior simulations. For this reason, all the models fit using eDNAjoint are of the ‘stanfit’ class and can be analyzed and manipulated with functions in the rstan package, in addition to the functions outlined above.

The code for the models written in Stan can be found in this folder of the package Github repo.

4.3 Priors

Three non-uniform priors are used in the model. First, there is an informative prior distribution for the false positive probability of eDNA detection, \(p_{10}\), which is used for parameter identifiability. A beta distribution is used for the \(p_{10}\) prior with two parameters: alpha and beta. Second, an informative prior distribution for the overdispersion parameter, \(\phi\), in the negative binomial distribution for overdispersed count observations (Eq. 1.2). In eDNAjoint, these parameters can be user-specified. The default specification for the \(p_{10}\) prior is beta(1,20) (mean: 0.048, var: 0.045), and the default specification for the \(\phi\) prior is gamma(0.25,0.25) (mean: 1, var: 4). Additionally, a normally distributed shrinkage prior is used for \(\alpha_n\), which serves a similar role to regularization.