Changelog • drmr

drmr 0.5.0

New vignette on “algorithms” detailing the methods that can be used to draw samples from the posterior.
summary method for the output of the ages_edens output.
draws method added to the package.
marg method to .
Fixed a bug in the sdm.stan AR(1) term.
Fixed duplicated random effects in drm.stan
Improved performance for drm.stan and sdm.stan
Super class for DRMs and SDMs, so print, summary, and draws methods are simplified.
A class pred_drmr was created for the output of the predict and fitted methods. It facilitates joining those outputs to the observed data.
We had some functions called patches sites and vice-versa. Now everything is standardized using “site” for patches/sites (I know, it does sound confusing).
elpd functions temporarily unavailable

drmr 0.4.0

predict and fitted methods now have an argument called type, which allows to compute predictions based on the posterior predictive distribution (default and corresponds to old behavior), expected value of the posterior predictive distribution (i.e., its theoretical mean), and latent density (disconsidering observation error).
Removing loglogistic distribution for now (need to do some work on the parametrization)
Major bug in the fitted/predict functions fixed.
Some improvements to documentation.
Fixed bug in how we deal with zeros and NAs.
A function called summary_age_densities to process the output of ages_edens is now available.
Fixed (again) the lambda_drm function (to extract age specific densities) and changed its name to ages_edens (stands for “age-specific expected age densities”.
log_lik and fitted methods are now defined for adrm and sdm objects.
Calculations for log-likelihood improved for easier maintenance.
predict_drm and predict_sdm are now methods.
Another increase in performance for the predict (and elpd) function. Especially relevant for larger datasets.
print and summary methods are now available for adrm and sdm objects.
Different inference algorithms are now alowed through the algorithm argument. A new algo_args argument is also introduced for specific inputs regarding the algorithms. For example, iter_warmup (specifying the number of warmup iterations), iter_sampling (number of sampling iterations), chains (number of MCMC chains), parallel_chains (number of chains to run in parallel), adapt_delta among others are supposed to be input as elements of a list passed as algo_args parameter.

drmr 0.3.0

New classes (and validators) created for the output of the fit_drm and fit_sdm functions. The classes are called adrm and sdm, respectively.
An update method to facilitate model refitting is introduced.

drmr 0.2.3

Bug fixed when computing the log-lik from a truncated normal model.
The elpd method has been included with the package and works with the outputs of fit_drm and fit_sdm. This method is helpful in calculating out-of-sample ELPD to perform leave-future-out cross validation (e.g., see https://mc-stan.org/loo/articles/loo2-lfo.html )
The predict_* functions have been modified so they only return the predicted values at the new sites/times. Before this update, those functions also returned other quantities such as mu_proj (the expected value of the “non-zero” part of the density), rho_proj (probability of non-zero density), and zt (the AR random effect, if its switch is turned on).
Improved the code to facilitate manteinance of the likelihood functions
Fixing lambda_gq.
Creating fitted and log_likelihood functions.

drmr 0.2.2

Ensuring there are no NAs in the covariates, site, or time.
Added support for missingness on y.
Functions for vectorzing the zero-inflated densities were moved to R (as opposed to Stan). Therefore, the indexes of the “zeros” in the response variables become an input from R to Stan.
Spatial and iid random effects also included in the SDM functions.
Typos in the “Parameters, priors, and toggles” vignette were fixed.
instantiate version must be at least 0.2.3.9002, so we can use #include for functions.

drmr 0.2.1

A new toggle called rho_mu introduced. Its default value is 1 and stands for explicitly relating the probability of observing a 0 (rho) to the latent density (mu). See Yee 2014 and references therein.
A new vignette with a comprehensive list of the model parameters and their priors is available.
A new vignette detailing the initialization procedures is also included.

drmr 0.2.0

init now admits real numbers too.
init_data: default now is to initialize the age-classes for all patches ranging from .9 to .01.
Allowing for choosing which process is correlated in time;
- The variable raw became w_t;
- tau became sigma_t
- The flag time_ar became ar_re (stands for AR random effect). Now, ar_re can assume three values:
  - “none”, the default indicating no AR random effects.
  - “rec” (AR for recruitment)
  - “surv” (AR for survival)
  - “dens” (AR for density)
Unstructured random effects.
- The following variables are added to the model: z_i patch specific random effect and sigma_i the SD of the iid random effect.
- Similarly to ar_re, the unstructured random effects can be enabled through the flat iid_re, which also admits the following entries:
  - “none” (default) indicating no IID random effects.
  - “rec” (IID for recruitment)
  - “surv” (IID for survival)
  - “dens” (iID for density)
ICAR random effects
- The following variables are added to the model: w_s, z_s patch specific random effect and sigma_s the approx marginal SD of the spatial random effect.
- Similarly to the two structures mentioned above, the flag sp_re admits the following values: “none” (default), “rec” (recruitment), “surv” (survival), and “dens” (density).
Functions for data simulation were temporarily removed.
Fixing init_data
Fixing bugs regarding the popdyn initialization.

drmr 0.1.3

Modified examples.qmd to prevent errors due to pathfinder failing.
Updated github actions.
Same problematic behaviour that was affecting make_data had to be dealed with prior initialization and predictions.
Options for “data informed” phi priors through the floag phi_hat
- When family = "gamma" and a prior for phi is not provided, the default prior for this parameter is a Gamma with shape = 2 and rate = ybar * s2. This prior has its theoretical mean at the method of moments estimator of phi.
- When family = "lognormal" and a prior for phi is not provided, the default prior for this parameter is a Gamma having its theoretical mean at the MLE of phi.
fixed typo when defining log-logistic reparametrization

drmr 0.1.2.9

fixed weird behaviour with make_data

drmr 0.1.2

lambda_drm function added to recover the age-specific densities form a fit_drm call.
fixed a bug on the forecast.stan script.

drmr 0.1.1

Vignettes updated so R CMD check “works”.
The lambdas are no longer returned by the fit and predict functions. Returning those values imply a significant decrease in the computational performance, especially as increasing the sample size.
The AR term in the model was “corrected”.
Some bugs with prior initialization for the sdm were fixed.

drmr 0.1.0

Exporting fix_linbeta, max_quad_x, and int_score functions
Fixing check_between function (not exported; This functions is just a helper)
Changed prior on zeta and allowing for user to input the hyperparameters. Before, we had a standard normal prior on the logit of zeta. Now, we place a beta prior on zeta.
Included support for different types of population dynamics initialization
Fixed how pdf and random number generation from location-scale Student’s t distribution
rlang became a dependency.
Included functions for simulating data from the DRM model and prior predictive checks.
Trying to avoid overflow by making calculations on the log-scale whenever it’s possible.
Constraining alpha to $(0, 1)$ .
Prior on phi now is Gamma.
fit_drm and fit_sdm returns changed. Now, the element draws is called stanfit. In addition, there is an additional element to the returned list called formulas. The formulas elements aims at making the predict_* functions less error prone.
The QR parametrization toggles were completely removed.
The coef_* parameters were converted to beta_*. That is, coef_r now is beta_r; while coef_t now is beta_t, and so on.
The pr_logsd_r_* inputs converted to pr_ltau_*.
Functions to plot effect of covariates on recruitment, survival, or absence probability were included.
est_mort becomes est_surv (as it makes more sense with the text).
get-started and examples vignette updated.

drmr 0.0.24

Initial values for population dynamics have been fixed.
Vignettes to quarto
New between function based on data.table::between.
int_score function to calculate the interval score was also included. The interval score helps to assess interval predictions.
age_at_maturity is replaced by ages_movement. The former can take either a single integer indicating the age at which individuals start to move, or a vector with 0s for age-groups that do not move and 1s for age-groups that are allowed to move.

drmr 0.0.23

The prior for $\alpha$ has been modified. In particular, instead of a pcp prior, now we put a Beta prior on $(\alpha + 1) / 2$ . The hyperparameters of this Beta distribution are pr_alpha_a and pr_alpha_b, respectively.
New functions to initialize the parameters from the prior have been introduced. Now the init parmeter from the fit_sdm and fit_drm functions can take three possible values:
- “default”: the standard initialization in Stan (For details see Stan's documentation).
- “prior”: initialize the parameters using samples from their respective prior distributions;
- “pathfinder”: uses the Pathfinder algorithm to initialize the parameters.
Fixed some issues with documentation.

drmr 0.0.22

fit_drm and fit_sdm functions to make model fitting slightly simpler.
pr_phi_a and pr_phi_b become pr_phi_mu and pr_phi_sd. The prior for $\log(\phi)$ is a Student’s t with 3 degrees of freedom, mean pr_phi_mu and SD pr_phi_sd.
Selectivity in make_data was fixed, thanks to Mark. Before, it was not being used when users input it.

drmr 0.0.21

make_data_sdm function (analogous to make_data) created for SDM.
predict_sdm function (analogous to predict_drm) created for SDM.

drmr 0.0.2

Parameters in the code and documentation were properly matched.
p_error toggle becomes time_ar toggle (more appropriate).
predict_drm function created
the make_data function now has a family argument indicating the probability distribution assumed for the response (given all the model parameters and latent variables)

drmr 0.0.1

First version of the package containing pre-compiled code for fitting the DRM, making forecasts and calculating density weighted centroids.