mvgam: Fit a Bayesian Dynamic GAM to Univariate or Multivariate Time...

In mvgam: Multivariate (Dynamic) Generalized Additive Models

Fit a Bayesian Dynamic GAM to Univariate or Multivariate Time Series

Description

This function estimates the posterior distribution for Generalised Additive Models (GAMs) that can include smooth spline functions, specified in the GAM formula, as well as latent temporal processes, specified by trend_model.

Further modelling options include State-Space representations to allow covariates and dynamic processes to occur on the latent 'State' level while also capturing observation-level effects. Prior specifications are flexible and explicitly encourage users to apply prior distributions that actually reflect their beliefs.

In addition, model fits can easily be assessed and compared with posterior predictive checks, forecast comparisons and leave-one-out / leave-future-out cross-validation.

Usage

mvgam(
  formula,
  trend_formula,
  knots,
  trend_knots,
  trend_model = "None",
  noncentred = FALSE,
  family = poisson(),
  share_obs_params = FALSE,
  data,
  newdata,
  use_lv = FALSE,
  n_lv,
  trend_map,
  priors,
  run_model = TRUE,
  prior_simulation = FALSE,
  residuals = TRUE,
  return_model_data = FALSE,
  backend = getOption("brms.backend", "cmdstanr"),
  algorithm = getOption("brms.algorithm", "sampling"),
  control = list(max_treedepth = 10, adapt_delta = 0.8),
  chains = 4,
  burnin = 500,
  samples = 500,
  thin = 1,
  parallel = TRUE,
  threads = 1,
  save_all_pars = FALSE,
  silent = 1,
  autoformat = TRUE,
  refit = FALSE,
  lfo = FALSE,
  ...
)

Arguments

formula	A formula object specifying the GAM observation model formula. These are exactly like the formula for a GLM except that smooth terms, s(), te(), ti(), t2(), as well as time-varying dynamic() terms, nonparametric gp() terms and offsets using offset(), can be added to the right hand side to specify that the linear predictor depends on smooth functions of predictors (or linear functionals of these). In nmix() family models, the formula is used to set up a linear predictor for the detection probability. Details of the formula syntax used by mvgam can be found in mvgam_formulae
trend_formula	An optional formula object specifying the GAM process model formula. If supplied, a linear predictor will be modelled for the latent trends to capture process model evolution separately from the observation model. Important notes: Should not have a response variable specified on the left-hand side (e.g., ~ season + s(year)) Use trend instead of series for effects that vary across time series Only available for RW(), AR() and VAR() trend models In nmix() family models, sets up linear predictor for latent abundance Consider dropping one intercept using - 1 convention to avoid estimation challenges
knots	An optional list containing user specified knot values for basis construction. For most bases the user simply supplies the knots to be used, which must match up with the k value supplied. Different terms can use different numbers of knots, unless they share a covariate.
trend_knots	As for knots above, this is an optional list of knot values for smooth functions within the trend_formula.
trend_model	character or function specifying the time series dynamics for the latent trend. Available options: None: No latent trend component (GAM component only, like gam) ZMVN or ZMVN(): Zero-Mean Multivariate Normal (Stan only) 'RW' or RW(): Random Walk 'AR1', 'AR2', 'AR3' or AR(p = 1, 2, 3): Autoregressive models 'CAR1' or CAR(p = 1): Continuous-time AR (Ornstein–Uhlenbeck process) 'VAR1' or VAR(): Vector Autoregressive (Stan only) 'PWlogistic', 'PWlinear' or PW(): Piecewise trends (Stan only) 'GP' or GP(): Gaussian Process with squared exponential kernel (Stan only) Additional features: Moving average and/or correlated process error terms available for most types (e.g., RW(cor = TRUE) for multivariate Random Walk) Hierarchical correlations possible for structured data See mvgam_trends for details and ZMVN() for examples
noncentred	logical. Use non-centred parameterisation for autoregressive trend models? Can improve efficiency by avoiding degeneracies in latent dynamic random effects estimation. Benefits vary by model - highly informative data may perform worse with this option. Available for RW(), AR(), CAR(), or trend = 'None' with trend_formula. Not available for moving average or correlated error models.
family	family specifying the exponential observation family for the series. Supported families: gaussian(): Real-valued data betar(): Proportional data on ⁠(0,1)⁠ lognormal(): Non-negative real-valued data student_t(): Real-valued data Gamma(): Non-negative real-valued data bernoulli(): Binary data poisson(): Count data (default) nb(): Overdispersed count data binomial(): Count data with imperfect detection when number of trials is known (use cbind() to bind observations and trials) beta_binomial(): As binomial() but allows for overdispersion nmix(): Count data with imperfect detection when number of trials is unknown (State-Space N-Mixture model with Poisson latent states and Binomial observations) See mvgam_families for more details.
share_obs_params	logical. If TRUE and the family has additional family-specific observation parameters (e.g., variance components, dispersion parameters), these will be shared across all outcome variables. Useful when multiple outcomes share properties. Default is FALSE.
data	A dataframe or list containing the model response variable and covariates required by the GAM formula and optional trend_formula. Required columns for most models: series: A factor index of the series IDs (number of levels should equal number of unique series labels) time: numeric or integer index of time points. For most dynamic trend types, time should be measured in discrete, regularly spaced intervals (i.e., c(1, 2, 3, ...)). Irregular spacing is allowed for trend_model = CAR(1), but zero intervals are adjusted to 1e-12 to prevent sampling errors. Special cases: Models with hierarchical temporal correlation (e.g., AR(gr = region, subgr = species)) should NOT include a series identifier Models without temporal dynamics (trend_model = 'None' or trend_model = ZMVN()) don't require a time variable
newdata	Optional dataframe or list of test data containing the same variables as in data. If included, observations in variable y will be set to NA when fitting the model so that posterior simulations can be obtained.
use_lv	logical. If TRUE, use dynamic factors to estimate series' latent trends in a reduced dimension format. Only available for RW(), AR() and GP() trend models. Default is FALSE. See lv_correlations for examples.
n_lv	integer specifying the number of latent dynamic factors to use if use_lv == TRUE. Cannot exceed n_series. Default is min(2, floor(n_series / 2)).
trend_map	Optional data.frame specifying which series should depend on which latent trends. Enables multiple series to depend on the same latent trend process with different observation processes. Required structure: Column series: Single unique entry for each series (matching factor levels in data) Column trend: Integer values indicating which trend each series depends on Notes: Sets up latent factor model by enabling use_lv = TRUE Process model intercept is NOT automatically suppressed Not yet supported for continuous time models (CAR())
priors	An optional data.frame with prior definitions or, preferably, a vector of brmsprior objects (see prior()). See get_mvgam_priors() and Details for more information.
run_model	logical. If FALSE, the model is not fitted but instead the function returns the model file and the data/initial values needed to fit the model outside of mvgam.
prior_simulation	logical. If TRUE, no observations are fed to the model, and instead simulations from prior distributions are returned.
residuals	logical. Whether to compute series-level randomized quantile residuals. Default is TRUE. Set to FALSE to save time and reduce object size (can add later using add_residuals).
return_model_data	logical. If TRUE, the list of data needed to fit the model is returned, along with initial values for smooth and AR parameters, once the model is fitted. Helpful for users who wish to modify the model file to add other stochastic elements. Default is FALSE unless run_model == FALSE.
backend	Character string naming the package for Stan model fitting. Options are "cmdstanr" (default) or "rstan". Can be set globally via "brms.backend" option. See https://mc-stan.org/rstan/ and https://mc-stan.org/cmdstanr/ for details.
algorithm	Character string naming the estimation approach: "sampling": MCMC (default) "meanfield": Variational inference with factorized normal distributions "fullrank": Variational inference with multivariate normal distribution "laplace": Laplace approximation (cmdstanr only) "pathfinder": Pathfinder algorithm (cmdstanr only) Can be set globally via "brms.algorithm" option. Limited testing suggests "meanfield" performs best among non-MCMC approximations for dynamic GAMs.
control	Named list for controlling sampler behaviour. Valid elements include max_treedepth, adapt_delta and init.
chains	integer specifying the number of parallel chains for the model. Ignored for variational inference algorithms.
burnin	integer specifying the number of warmup iterations to tune sampling algorithms. Ignored for variational inference algorithms.
samples	integer specifying the number of post-warmup iterations for sampling the posterior distribution.
thin	Thinning interval for monitors. Ignored for variational inference algorithms.
parallel	logical specifying whether to use multiple cores for parallel MCMC simulation. If TRUE, uses min(c(chains, parallel::detectCores() - 1)) cores.
threads	integer. Experimental option for within-chain parallelisation in Stan using reduce_sum. Recommended only for experienced Stan users with slow models. Currently works for all families except nmix() and when using Cmdstan backend.
save_all_pars	logical. Save draws from all variables defined in Stan's parameters block. Default is FALSE.
silent	Verbosity level between 0 and 2. If 1 (default), most informational messages are suppressed. If 2, even more messages are suppressed. Sampling progress is still printed - set refresh = 0 to disable. For backend = "rstan", also set open_progress = FALSE to prevent additional progress bars.
autoformat	logical. Use stanc parser to automatically format Stan code and check for deprecations. For development purposes - leave as TRUE.
refit	logical. Indicates whether this is a refit called using update.mvgam(). Users should leave as FALSE.
lfo	logical. Indicates whether this is part of lfo_cv.mvgam call. Returns lighter model version for speed. Users should leave as FALSE.
...	Further arguments passed to Stan: For backend = "rstan": passed to sampling() or vb() For backend = "cmdstanr": passed to cmdstanr::sample, cmdstanr::variational, cmdstanr::laplace or cmdstanr::pathfinder methods

Details

Dynamic GAMs are useful when we wish to predict future values from time series that show temporal dependence but we do not want to rely on extrapolating from a smooth term (which can sometimes lead to unpredictable and unrealistic behaviours). In addition, smooths can often try to wiggle excessively to capture any autocorrelation that is present in a time series, which exacerbates the problem of forecasting ahead.

As GAMs are very naturally viewed through a Bayesian lens, and we often must model time series that show complex distributional features and missing data, parameters for mvgam models are estimated in a Bayesian framework using Markov Chain Monte Carlo by default.

Getting Started Resources:

• General overview: vignette("mvgam_overview") and vignette("data_in_mvgam")

• Full list of vignettes: vignette(package = "mvgam")

• Real-world examples: mvgam_use_cases

• Quick reference: mvgam cheatsheet

Value

A list object of class mvgam containing model output, the text representation of the model file, the mgcv model output (for easily generating simulations at unsampled covariate values), Dunn-Smyth residuals for each series and key information needed for other functions in the package. See mvgam-class for details. Use methods(class = "mvgam") for an overview on available methods.

Model Specification Details

Formula Syntax: Details of the formula syntax used by mvgam can be found in mvgam_formulae. Note that it is possible to supply an empty formula where there are no predictors or intercepts in the observation model (i.e. y ~ 0 or y ~ -1). In this case, an intercept-only observation model will be set up but the intercept coefficient will be fixed at zero. This can be handy if you wish to fit pure State-Space models where the variation in the dynamic trend controls the average expectation, and/or where intercepts are non-identifiable (as in piecewise trends).

Families and Link Functions: Details of families supported by mvgam can be found in mvgam_families.

Trend Models: Details of latent error process models supported by mvgam can be found in mvgam_trends.

Prior Specifications

Default priors for intercepts and any variance parameters are chosen to be vaguely informative, but these should always be checked by the user. Prior distributions for most important model parameters can be altered (see get_mvgam_priors() for details). Note that latent trends are estimated on the link scale so choose priors accordingly.

However more control over the model specification can be accomplished by setting run_model = FALSE and then editing the model code (found in the model_file slot in the returned object) before running the model using either rstan or cmdstanr. This is encouraged for complex modelling tasks.

Important: No priors are formally checked to ensure they are in the right syntax so it is up to the user to ensure these are correct.

Model Components

Random Effects: For any smooth terms using the random effect basis (smooth.construct.re.smooth.spec), a non-centred parameterisation is automatically employed to avoid degeneracies that are common in hierarchical models. Note however that centred versions may perform better for series that are particularly informative, so as with any foray into Bayesian modelling, it is worth building an understanding of the model's assumptions and limitations by following a principled workflow. Also note that models are parameterised using drop.unused.levels = FALSE in jagam to ensure predictions can be made for all levels of the supplied factor variable.

Observation Level Parameters: When more than one series is included in data and an observation family that contains more than one parameter is used, additional observation family parameters (i.e. phi for nb() or sigma for gaussian()) are by default estimated independently for each series. But if you wish for the series to share the same observation parameters, set share_obs_params = TRUE.

Model Diagnostics

Residuals: For each series, randomized quantile (i.e. Dunn-Smyth) residuals are calculated for inspecting model diagnostics. If the fitted model is appropriate then Dunn-Smyth residuals will be standard normal in distribution and no autocorrelation will be evident. When a particular observation is missing, the residual is calculated by comparing independent draws from the model's posterior distribution.

Computational Backend

Using Stan: mvgam is primarily designed to use Hamiltonian Monte Carlo for parameter estimation via the software Stan (using either the cmdstanr or rstan interface). There are great advantages when using Stan over Gibbs / Metropolis Hastings samplers, which includes the option to estimate nonlinear effects via Hilbert space approximate Gaussian Processes, the availability of a variety of inference algorithms (i.e. variational inference, laplacian inference etc...) and capabilities to enforce stationarity for complex Vector Autoregressions.

Because of the many advantages of Stan over JAGS, further development of the package will only be applied to Stan. This includes the planned addition of more response distributions, plans to handle zero-inflation, and plans to incorporate a greater variety of trend models. Users are strongly encouraged to opt for Stan over JAGS in any proceeding workflows.

Recommended Workflow

How to Start: The mvgam cheatsheet is a good starting place if you are just learning to use the package. It gives an overview of the package's key functions and objects, as well as providing a reasonable workflow that new users can follow.

Recommended Steps:

1. Data Preparation: Check that your data are in a suitable tidy format for mvgam modeling (see the data formatting vignette for guidance)

2. Data Exploration: Inspect features of the data using plot_mvgam_series. Now is also a good time to familiarise yourself with the package's example workflows that are detailed in the vignettes:

   • Getting started vignette

   • Shared latent states vignette

   • Time-varying effects vignette

   • State-Space models vignette

   • "Fitting N-mixture models in mvgam"

   • "Joint Species Distribution Models in mvgam"

   • "Incorporating time-varying seasonality in forecast models"

   • "Temporal autocorrelation in GAMs and the mvgam package"

3. Model Structure: Carefully think about how to structure linear predictor effects (i.e. smooth terms using s(), te() or ti(), GPs using gp(), dynamic time-varying effects using dynamic(), and parametric terms), latent temporal trend components (see mvgam_trends) and the appropriate observation family (see mvgam_families). Use get_mvgam_priors() to see default prior distributions for stochastic parameters.

4. Prior Specification: Change default priors using appropriate prior knowledge (see prior()). When using State-Space models with a trend_formula, pay particular attention to priors for any variance parameters such as process errors and observation errors. Default priors on these parameters are chosen to be vaguely informative and to avoid zero (using Inverse Gamma priors), but more informative priors will often help with model efficiency and convergence.

5. Model Fitting: Fit the model using either Hamiltonian Monte Carlo or an approximation algorithm (i.e. change the backend argument) and use summary.mvgam(), conditional_effects.mvgam(), mcmc_plot.mvgam(), pp_check.mvgam(), pairs.mvgam() and plot.mvgam() to inspect / interrogate the model.

6. Model Comparison: Update the model as needed and use loo_compare.mvgam() for in-sample model comparisons, or alternatively use forecast.mvgam(), lfo_cv.mvgam() and score.mvgam_forecast() to compare models based on out-of-sample forecasts (see the forecast evaluation vignette for guidance).

7. Inference and Prediction: When satisfied with the model structure, use predict.mvgam(), plot_predictions() and/or plot_slopes() for more targeted simulation-based inferences (see "How to interpret and report nonlinear effects from Generalized Additive Models" for some guidance on interpreting GAMs). For time series models, use hindcast.mvgam(), fitted.mvgam(), augment.mvgam() and forecast.mvgam() to inspect posterior hindcast / forecast distributions.

8. Documentation: Use how_to_cite() to obtain a scaffold methods section (with full references) to begin describing this model in scientific publications.

Author(s)

Nicholas J Clark

References

Nicholas J Clark & Konstans Wells (2023). Dynamic generalised additive models (DGAMs) for forecasting discrete ecological time series. Methods in Ecology and Evolution. 14:3, 771-784.

Nicholas J Clark, SK Morgan Ernest, Henry Senyondo, Juniper Simonis, Ethan P White, Glenda M Yenni, KANK Karunarathna (2025). Beyond single-species models: leveraging multispecies forecasts to navigate the dynamics of ecological predictability. PeerJ. 13:e18929 https://doi.org/10.7717/peerj.18929

See Also

jagam(), gam(), gam.models, get_mvgam_priors(), jsdgam(), hindcast.mvgam(), forecast.mvgam(), predict.mvgam()

Examples

## Not run: 
# =============================================================================
# Basic Multi-Series Time Series Modeling
# =============================================================================

# Simulate three time series that have shared seasonal dynamics,
# independent AR(1) trends, and Poisson observations
set.seed(0)
dat <- sim_mvgam(
  T = 80,
  n_series = 3,
  mu = 2,
  trend_model = AR(p = 1),
  prop_missing = 0.1,
  prop_trend = 0.6
)

# Plot key summary statistics for a single series
plot_mvgam_series(data = dat$data_train, series = 1)

# Plot all series together
plot_mvgam_series(data = dat$data_train, series = "all")

# Formulate a model using Stan where series share a cyclic smooth for
# seasonality and each series has an independent AR1 temporal process.
# Note that 'noncentred = TRUE' will likely give performance gains.
# Set run_model = FALSE to inspect the returned objects
mod1 <- mvgam(
  formula = y ~ s(season, bs = "cc", k = 6),
  data = dat$data_train,
  trend_model = AR(),
  family = poisson(),
  noncentred = TRUE,
  run_model = FALSE
)

# View the model code in Stan language
stancode(mod1)

# View the data objects needed to fit the model in Stan
sdata1 <- standata(mod1)
str(sdata1)

# Now fit the model
mod1 <- mvgam(
  formula = y ~ s(season, bs = "cc", k = 6),
  data = dat$data_train,
  trend_model = AR(),
  family = poisson(),
  noncentred = TRUE,
  chains = 2,
  silent = 2
)

# Extract the model summary
summary(mod1)

# Plot the historical trend and hindcast distributions for one series
hc_trend <- hindcast(mod1, type = "trend")
plot(hc_trend)

hc_predicted <- hindcast(mod1, type = "response")
plot(hc_predicted)

# Residual diagnostics
plot(mod1, type = "residuals", series = 1)
resids <- residuals(mod1)
str(resids)

# Fitted values and residuals can be added directly to the training data
augment(mod1)

# Compute the forecast using covariate information in data_test
fc <- forecast(mod1, newdata = dat$data_test)
str(fc)
fc_summary <- summary(fc)
head(fc_summary, 12)
plot(fc)

# Plot the estimated seasonal smooth function
plot(mod1, type = "smooths")

# Plot estimated first derivatives of the smooth
plot(mod1, type = "smooths", derivatives = TRUE)

# Plot partial residuals of the smooth
plot(mod1, type = "smooths", residuals = TRUE)

# Plot posterior realisations for the smooth
plot(mod1, type = "smooths", realisations = TRUE)

# Plot conditional response predictions using marginaleffects
conditional_effects(mod1)
plot_predictions(mod1, condition = "season", points = 0.5)

# Generate posterior predictive checks using bayesplot
pp_check(mod1)

# Extract observation model beta coefficient draws as a data.frame
beta_draws_df <- as.data.frame(mod1, variable = "betas")
head(beta_draws_df)
str(beta_draws_df)

# Investigate model fit
mc.cores.def <- getOption("mc.cores")
options(mc.cores = 1)
loo(mod1)
options(mc.cores = mc.cores.def)


# =============================================================================
# Vector Autoregressive (VAR) Models
# =============================================================================

# Fit a model to the portal time series that uses a latent
# Vector Autoregression of order 1
mod <- mvgam(
  formula = captures ~ -1,
  trend_formula = ~ trend,
  trend_model = VAR(cor = TRUE),
  family = poisson(),
  data = portal_data,
  chains = 2,
  silent = 2
)

# Plot the autoregressive coefficient distributions;
# use 'dir = "v"' to arrange the order of facets correctly
mcmc_plot(
  mod,
  variable = 'A',
  regex = TRUE,
  type = 'hist',
  facet_args = list(dir = 'v')
)

# Plot the process error variance-covariance matrix in the same way
mcmc_plot(
  mod,
  variable = 'Sigma',
  regex = TRUE,
  type = 'hist',
  facet_args = list(dir = 'v')
)

# Calculate Generalized Impulse Response Functions for each series
irfs <- irf(
  mod,
  h = 12,
  cumulative = FALSE
)

# Plot some of them
plot(irfs, series = 1)
plot(irfs, series = 2)

# Calculate forecast error variance decompositions for each series
fevds <- fevd(mod, h = 12)

# Plot median contributions to forecast error variance
plot(fevds)


# =============================================================================
# Dynamic Factor Models
# =============================================================================

# Now fit a model that uses two RW dynamic factors to model
# the temporal dynamics of the four rodent species
mod <- mvgam(
  captures ~ series,
  trend_model = RW(),
  use_lv = TRUE,
  n_lv = 2,
  data = portal_data,
  chains = 2,
  silent = 2
)

# Plot the factors
plot(mod, type = 'factors')

# Plot the hindcast distributions
hcs <- hindcast(mod)
plot(hcs, series = 1)
plot(hcs, series = 2)
plot(hcs, series = 3)
plot(hcs, series = 4)

# Use residual_cor() to calculate temporal correlations among the series
# based on the factor loadings
lvcors <- residual_cor(mod)
names(lvcors)
lvcors$cor

# For those correlations whose credible intervals did not include
# zero, plot them as a correlation matrix (all other correlations
# are shown as zero on this plot)
plot(lvcors, cluster = TRUE)


# =============================================================================
# Shared Latent Trends with Custom Trend Mapping
# =============================================================================

# Example of supplying a trend_map so that some series can share
# latent trend processes
sim <- sim_mvgam(n_series = 3)
mod_data <- sim$data_train

# Here, we specify only two latent trends; series 1 and 2 share a trend,
# while series 3 has its own unique latent trend
trend_map <- data.frame(
  series = unique(mod_data$series),
  trend = c(1, 1, 2)
)

# Fit the model using AR1 trends
mod <- mvgam(
  formula = y ~ s(season, bs = "cc", k = 6),
  trend_map = trend_map,
  trend_model = AR(),
  data = mod_data,
  return_model_data = TRUE,
  chains = 2,
  silent = 2
)

# The mapping matrix is now supplied as data to the model in the 'Z' element
mod$model_data$Z

# The first two series share an identical latent trend; the third is different
plot(residual_cor(mod))
plot(mod, type = "trend", series = 1)
plot(mod, type = "trend", series = 2)
plot(mod, type = "trend", series = 3)


# =============================================================================
# Time-Varying (Dynamic) Coefficients
# =============================================================================

# Example of how to use dynamic coefficients
# Simulate a time-varying coefficient for the effect of temperature
set.seed(123)
N <- 200
beta_temp <- vector(length = N)
beta_temp[1] <- 0.4
for (i in 2:N) {
  beta_temp[i] <- rnorm(1, mean = beta_temp[i - 1] - 0.0025, sd = 0.05)
}
plot(beta_temp)

# Simulate a covariate called 'temp'
temp <- rnorm(N, sd = 1)

# Simulate some noisy Gaussian observations
out <- rnorm(N,
  mean = 4 + beta_temp * temp,
  sd = 0.5
)

# Gather necessary data into a data.frame; split into training / testing
data <- data.frame(out, temp, time = seq_along(temp))
data_train <- data[1:180, ]
data_test <- data[181:200, ]

# Fit the model using the dynamic() function
mod <- mvgam(
  formula = out ~ dynamic(
    temp,
    scale = FALSE,
    k = 40
  ),
  family = gaussian(),
  data = data_train,
  newdata = data_test,
  chains = 2,
  silent = 2
)

# Inspect the model summary, forecast and time-varying coefficient distribution
summary(mod)
plot(mod, type = "smooths")
fc <- forecast(mod, newdata = data_test)
plot(fc)

# Propagating the smooth term shows how the coefficient is expected to evolve
plot_mvgam_smooth(mod, smooth = 1, newdata = data)
abline(v = 180, lty = "dashed", lwd = 2)
points(beta_temp, pch = 16)


# =============================================================================
# Working with Offset Terms
# =============================================================================

# Example showing how to incorporate an offset; simulate some count data
# with different means per series
set.seed(100)
dat <- sim_mvgam(
  prop_trend = 0,
  mu = c(0, 2, 2),
  seasonality = "hierarchical"
)

# Add offset terms to the training and testing data
dat$data_train$offset <- 0.5 * as.numeric(dat$data_train$series)
dat$data_test$offset <- 0.5 * as.numeric(dat$data_test$series)

# Fit a model that includes the offset in the linear predictor as well as
# hierarchical seasonal smooths
mod <- mvgam(
  formula = y ~ offset(offset) +
    s(series, bs = "re") +
    s(season, bs = "cc") +
    s(season, by = series, m = 1, k = 5),
  data = dat$data_train,
  chains = 2,
  silent = 2
)

# Inspect the model file to see the modification to the linear predictor (eta)
stancode(mod)

# Forecasts for the first two series will differ in magnitude
fc <- forecast(mod, newdata = dat$data_test)
plot(fc, series = 1, ylim = c(0, 75))
plot(fc, series = 2, ylim = c(0, 75))

# Changing the offset for the testing data should lead to changes in
# the forecast
dat$data_test$offset <- dat$data_test$offset - 2
fc <- forecast(mod, newdata = dat$data_test)
plot(fc)

# Relative Risks can be computed by fixing the offset to the same value
# for each series
dat$data_test$offset <- rep(1, NROW(dat$data_test))
preds_rr <- predict(mod,
  type = "link",
  newdata = dat$data_test,
  summary = FALSE
)
series1_inds <- which(dat$data_test$series == "series_1")
series2_inds <- which(dat$data_test$series == "series_2")

# Relative Risks are now more comparable among series
layout(matrix(1:2, ncol = 2))
plot(preds_rr[1, series1_inds],
  type = "l", col = "grey75",
  ylim = range(preds_rr),
  ylab = "Series1 Relative Risk", xlab = "Time"
)
for (i in 2:50) {
  lines(preds_rr[i, series1_inds], col = "grey75")
}

plot(preds_rr[1, series2_inds],
  type = "l", col = "darkred",
  ylim = range(preds_rr),
  ylab = "Series2 Relative Risk", xlab = "Time"
)
for (i in 2:50) {
  lines(preds_rr[i, series2_inds], col = "darkred")
}
layout(1)


# =============================================================================
# Binomial Family Models
# =============================================================================

# Example showcasing how cbind() is needed for Binomial observations
# Simulate two time series of Binomial trials
trials <- sample(c(20:25), 50, replace = TRUE)
x <- rnorm(50)
detprob1 <- plogis(-0.5 + 0.9 * x)
detprob2 <- plogis(-0.1 - 0.7 * x)
dat <- rbind(
  data.frame(
    y = rbinom(n = 50, size = trials, prob = detprob1),
    time = 1:50,
    series = "series1",
    x = x,
    ntrials = trials
  ),
  data.frame(
    y = rbinom(n = 50, size = trials, prob = detprob2),
    time = 1:50,
    series = "series2",
    x = x,
    ntrials = trials
  )
)
dat <- dplyr::mutate(dat, series = as.factor(series))
dat <- dplyr::arrange(dat, time, series)
plot_mvgam_series(data = dat, series = "all")

# Fit a model using the binomial() family; must specify observations
# and number of trials in the cbind() wrapper
mod <- mvgam(
  formula = cbind(y, ntrials) ~ series + s(x, by = series),
  family = binomial(),
  data = dat,
  chains = 2,
  silent = 2
)
summary(mod)
pp_check(mod,
  type = "bars_grouped",
  group = "series", ndraws = 50
)
pp_check(mod,
  type = "ecdf_overlay_grouped",
  group = "series", ndraws = 50
)
conditional_effects(mod, type = "link")

# To view predictions on the probability scale,
# use ntrials = 1 in datagrid()
plot_predictions(
  mod,
  by = c('x', 'series'),
  newdata = datagrid(
    x = runif(100, -2, 2),
    series = unique,
    ntrials = 1
  ),
  type = 'expected'
)

# Not needed for general use; cleans up connections for automated testing
closeAllConnections()

## End(Not run)

mvgam documentation built on Jan. 21, 2026, 9:07 a.m.