Package 'bscm'

Description

Implements the synthetic control method of Abadie, Diamond, and Hainmueller (2010) within a Bayesian framework, enabling straightforward uncertainty quantification of treatment effects and other quantities of interest. Posterior sampling is performed using Markov chain Monte Carlo via Stan.

bscm supports time-varying covariates with potentially time-varying effects, single treated unit as well as multiple treated units possibly with staggered treatment adoption.

Output of bscm() is a bscmfit object with methods for model assessment, comparison, and selection based on placebo studies, cross-validation, and posterior predictive checks.

Author(s)

Maintainer: Jouni Helske [email protected] (ORCID)

References

Abadie A, Diamond A, and Hainmueller J (2010). Synthetic Control Methods for Comparative Case Studies: Estimating the Effect of California’s Tobacco Control Program. Journal of the American Statistical Association, 105(490), 493–505, https://doi.org/10.1198/jasa.2009.ap08746.

Stan Development Team (2025). RStan: the R interface to Stan. R package version 2.32.7. https://mc-stan.org.

See Also

Useful links:

• https://github.com/helske/bscm

• Report bugs at https://github.com/helske/bscm/issues

Description

Return posterior draws from BSCM fit as a draws object

Usage

## S3 method for class 'bscmfit'
as_draws(x, parameters = NULL, inc_warmup = FALSE, include = TRUE, ...)

Arguments

Value

An object of class draws_array containing the posterior draws of the specified parameters.

See Also

posterior::as_draws() for converting the output to other formats, and posterior::summarise_draws() for computing posterior summaries of the draws.

Examples

head(as_draws(fit_single_treated, parameters = c("alpha", "sigma")))

Description

Returns a data.frame representation of the posterior sample of the model parameters. For samples from posterior predictive distribution, see posterior_epred() and posterior_predict(). For donor weights, use donor_weights().

Usage

## S3 method for class 'bscmfit'
as.data.frame(
  x,
  row.names = NULL,
  optional = FALSE,
  parameters = NULL,
  include = TRUE,
  ...
)

Arguments

Value

A data.frame containing model parameters in a wide format.

See Also

as_draws.bscmfit().

Examples

head(as.data.frame(fit_single_treated, parameters = c("alpha", "sigma")))

Description

Pre- and post-treatment average effects of a Bayesian synthetic control model

Usage

average_treatment_effect(x, ...)

## S3 method for class 'bscmfit'
average_treatment_effect(
  x,
  average = TRUE,
  summary = TRUE,
  probs = c(0.025, 0.975),
  ...
)

Arguments

Value

A tibble of posterior summaries (summary = TRUE) or posterior draws (summary = FALSE) in long format.

Examples

average_treatment_effect(fit_single_treated)

Description

Function bscm estimates a Bayesian synthetic control models of varying types.

Usage

bscm(
  formula,
  data,
  treatment = "treatment",
  time = "time",
  unit = "id",
  error = "iid",
  priors = NULL,
  prior_only = FALSE,
  mcmc_diagnostics = TRUE,
  save_data = TRUE,
  compute_predictions = TRUE,
  ...
)

Arguments

Details

To define formula in case of no predictors, use outcome ~ 1 or outcome ~ 0. In the former case, as well by the default when using predictors (e.g., outcome ~ x + z), the model includes intercept term for each treated unit. Intercept can be omitted by using 0 in the RHS, e.g., outcome ~ 0 or outcome ~ 0 + x + z (equivalently, you can use -1 in place of 0). Formula should not contain the variable defining the treatment, which is defined separately using the argument treatment. In case the variable is present also in the formula, it is automatically removed.

To specify predictors with time-varying coefficients in formula, wrap them in tv(), e.g., outcome ~ x + tv(~ z, df = 10, type = "rw1") defines a model where x has a time-constant coefficient and z has a time-varying coefficient following penalized cubic spline with 10 spline basis functions and random walk prior on the spline coefficients. Terms inside tv() are automatically also included in the time-constant part of the model, since the time-varying coefficients are defined to have zero mean in the pre-treatment period (in case of multiple treated units, minimum period). Note that a common time-varying intercept is omitted as it would cancel out in the linear predictor.

Both the time-constant and time-varying regression part is assumed to apply for all treated and donor units with common coefficients. If you want to apply a covariate only for treated units, just set the covariate values to zero for donors. This approach can be also used to vary behaviour of intercept: Defining formula such as y ~ 0 + intercept + x, where intercept is a name of constant column in the data, will define common intercept for all units, instead of unit-specific intercepts (e.g., fixed effects).

When model contains covariates $X$, their effect is subtracted from donors, i.e., for treated unit $i$, $y_i \sim N(\alpha_i + X_i\beta + Z^\ast\omega_i, \sigma_i^2)$, where column $j$ of $Z^\ast$ is $z_j - (\alpha_j + X_j \beta)$, i.e., the original donor outcome vector $z_j$ minus the donor-specific intercept $\alpha_j$ and the effect of covariates $X_j \beta$ for that donor. Note that $\beta$ is common across donors and the treated units.

Model can contain missing values in the outcome variable of the treated, but not of the donors, nor in the covariates. Missing outcomes are automatically imputed during MCMC sampling under a missing at random (MAR) assumption.

Value

An object of class bscmfit.

See Also

summary.bscmfit(), as_draws.bscmfit(), rstan::sampling().

Examples

# skip diagnostics and use small number of iterations for CRAN checks
fit <- bscm(
  y ~ 1, single_treated, "treatment", "time", "id",
  priors = list(omega = dirichlet_pr(0.5)),
  chains = 1, cores = 1, refresh = 0, iter = 1000,
  mcmc_diagnostics = FALSE
)
fit

Description

The functions normal_pr(), student_pr(), exponential_pr(), gamma_pr(), beta_pr(), half_normal_pr(), dirichlet_pr(), and logistic_normal_pr() specify prior distributions for parameters of the Bayesian synthetic control model estimated by bscm().

Usage

normal_pr(location, scale)

student_pr(df, location, scale)

gamma_pr(shape, rate)

exponential_pr(rate)

beta_pr(shape1, shape2)

half_normal_pr(scale)

dirichlet_pr(concentration)

logistic_normal_pr(scale)

Arguments

Details

For parameters that can take any real value (alpha, beta), currently supported prior distributions are normal_pr(location, scale) and student_pr(df, location, scale). Note however that for the intercept the prior corresponds to a value of the intercept when donors and predictors corresponding to time-constant coefficients are centered.

For parameters constrained to be positive (sigma, kappa), supported distributions are exponential_pr(rate), gamma_pr(shape, rate), and half_normal_pr(scale).

For autoregressive coefficients of the residuals (rho) the prior corresponds to transformation 0.5 * (1 + rho), so that the support is (0, 1). Only supported prior distribution for this is beta_pr(shape1, shape2).

For the donor weight vector omega, supported priors are dirichlet_pr(concentration) and logistic_normal_pr(scale). The (symmetric) Dirichlet prior with concentration $\alpha_\omega$ defined as $\omega \sim \text{Dirichlet}(\alpha_\omega, \ldots,\alpha_\omega)$. Values $\alpha_\omega < 1$ concentrate weight on few donors, while $\alpha_\omega = 1$ is uniform over the simplex, and $\alpha_\omega > 1$ pulls weights toward the center of the simplex. Logistic normal with with scale $\sigma_\alpha$ is defined as $\omega = \text{softmax}(\eta)$, where $\eta \sim N(0, \sigma_\alpha^2 I)$ constrained to sum to zero. Larger scale induces more concentrated (sparser) weights.

Value

A bscm_prior object.

See Also

bscm()

Examples

normal_pr(c(0, 0), c(1, 2))
student_pr(df = 4, location = 0, scale = 2)
exponential_pr(rate = c(1, 0.1))
gamma_pr(2, 1)
beta_pr(2, 2)
half_normal_pr(1)
dirichlet_pr(1)
logistic_normal_pr(2)

Description

This function is automatically called at the end of bscm() to check that the output can be trusted in terms of convergence of the MCMC sampling. Checks consists of the common diagnostics of Hamiltonian Monte Carlo variant used by Stan, as well as the Rhat values and effective sample sizes of model parameters. See rstan::check_hmc_diagnostics() and posterior::default_convergence_measures() for details on the definitions of these.

Usage

check_mcmc_diagnostics(x, ...)

## S3 method for class 'bscmfit'
check_mcmc_diagnostics(x, warn = TRUE, ...)

Arguments

Value

Invisibly returns a list containing the results of the check.

References

https://mc-stan.org/learn-stan/diagnostics-warnings.html

Examples

check_mcmc_diagnostics(fit_single_treated, warn = FALSE)

Description

Extract regression coefficients of a Bayesian synthetic control model

Usage

## S3 method for class 'bscmfit'
coef(object, type = NULL, summary = TRUE, probs = c(0.025, 0.975), ...)

Arguments

Value

A tibble of posterior summaries (summary = TRUE) or list of tibbles of posterior draws (summary = FALSE).

Examples

coef(fit_single_treated)
coef(fit_single_treated, type = "beta")

Description

Computes the posterior distribution of the covariate adjustment to the synthetic control for each predictor, defined as the regression coefficient multiplied by the difference between the predictor value of the treated unit and the corresponding weighted predictor value of the donor units.

Usage

covariate_adjustment(x, ...)

## S3 method for class 'bscmfit'
covariate_adjustment(
  x,
  plot = TRUE,
  probs = c(0.025, 0.975),
  alpha = 0.5,
  scales = "free_y",
  ...
)

Arguments

Details

For predictor $k$, treated unit $i$, and time point $t$, the covariate adjustment is

$b_{tk} \left( X_{y,itk} - \sum_j \omega_{ji} X_{z,jtk} \right),$

where $b_{tk} = \beta_k$ for predictors with a fixed coefficient and $b_{tk} = \beta_k + \gamma_{tk}$ for predictors with a time-varying coefficient.

If there are multiple treated units, the covariate adjustments are averaged over treated units within each posterior draw before computing posterior summaries.

Value

A data frame containing the posterior mean and requested posterior quantiles of the covariate adjustment for each predictor and time point. If plot = TRUE, the corresponding plot is printed as a side effect.

Description

For models with covariates, returns and optionally visualizes the covariate imbalances

$\delta_{t} = \sqrt{\frac{1}{K}\sum_{k=1}^K(x_{k,0,t} - \bar x_{k,0,t})^2},$

$t=1,\ldots,T$, where $\bar x_{k,0,t} = \sum_{j=1}^J \omega_j x_{k, j, t}$ and $x_{k,0,t}$ is the value of $k$th covariate of a treated unit at time t, and similarly for donors $j=1,\ldots,J$. This is computed separately for each treated unit in case of multiple treated units.

Usage

covariate_imbalance(x, ...)

## S3 method for class 'bscmfit'
covariate_imbalance(x, plot = TRUE, probs = c(0.025, 0.975), ...)

Arguments

Value

A data.frame of posterior summaries of synthetic covariate distances.

Examples

covariate_imbalance(fit_single_treated, plot = TRUE, probs = c(0.05, 0.95))

Description

Returns ranking of donors either using posterior means of donor weights from bscmfit object, or ranking from a vsel object returned by projpred::varsel() or projpred::cv_varsel() when applied to a bscmfit object.

Usage

donor_ranking(x)

Arguments

Value

Character vector of donor IDs in order of selection.

Description

Extract donor weights of a Bayesian synthetic control model

Usage

donor_weights(x, ...)

## S3 method for class 'bscmfit'
donor_weights(x, summary = TRUE, probs = c(0.025, 0.5, 0.975), ...)

Arguments

Value

A tibble of posterior summaries (summary = TRUE) or posterior draws (summary = FALSE) in long format.

Examples

donor_weights(fit_single_treated) |> head(5)

Description

Effective donors is defined as $1 / \sum_{j=1}^J \omega_j^2$, where $\omega_j$ is the donor weight of control unit $j$.

Usage

effective_donors(x, ...)

## S3 method for class 'bscmfit'
effective_donors(
  x,
  average = FALSE,
  summary = TRUE,
  probs = c(0.025, 0.975),
  ...
)

Arguments

Value

A data.frame of posterior summaries (summary = TRUE) or posterior draws (summary = FALSE) in long format.

Examples

effective_donors(fit_single_treated)

Description

The object fit_single_treated contains a Bayesian synthetic control model estimated as

fit <- bscm(
  formula = y ~ x, data = single_treated, treatment = "treatment",
  chains = 2, cores = 1, refresh = 0, iter = 2000, warmup = 1000,
  control = list(adapt_delta = 0.8)
)

Format

A bscmfit object.

Examples

fit_single_treated
plot(fit_single_treated)

Description

Returns posterior draws (or summaries) of expected values y_mean of the posterior predictive distribution of the model.

Usage

## S3 method for class 'bscmfit'
fitted(object, summary = TRUE, probs = c(0.025, 0.975), ...)

Arguments

Value

A data.frame of posterior summaries (summary = TRUE) or posterior draws (summary = FALSE) in long format.

Examples

fitted(fit_single_treated) |> head()

Description

Get prior specifications used in the Bayesian synthetic control model

Usage

get_priors(x, ...)

## S3 method for class 'bscmfit'
get_priors(x, ...)

Arguments

Value

A named list of bscm_prior objects for the parameters of the model.

Description

Creates a refmodel object from bscmfit to be used with projpred package. This enables the usage of projpred::varsel() and projpred::cv_varsel() for donor selection. Note that for predictions with newdata, you need to call proj_predict_bscm() which is a wrapper of projpred::proj_predict() to which converts the data to the wide format used by projpred.

Usage

## S3 method for class 'bscmfit'
get_refmodel(object, ...)

Arguments

Details

Projection are based on only on the pretreatment period.

This function is experimental and currently only a single treated unit is supported.

For projpred integration, donors are treated as separate "predictors" in the formula. Currently only supported model is one without extra predictors. There are also other restrictions, namely lack of support for K-fold.

Value

An object of class refmodel.

See Also

projpred::varsel(), projpred::cv_varsel()

Examples

fit <- bscm(
  y ~ 1, treatment = "treatment", time = "time", unit = "id",
  data = single_treated, refresh = 0, chains = 2, cores = 2
)
refmodel <- get_refmodel(fit)
vs <- projpred::varsel(refmodel)
plot(vs, stats = c("elpd", "rmse"), alpha = 0.05)
# predictions using the projection:
predictions <- projpred::proj_predict(vs)
# posterior mean of the predictions
colMeans(predictions)

Description

Reconstructs the data list passed to Stan in bscm(). Used by methods such as loo() with reloo = TRUE.Requires the original data (save_data = TRUE).

Usage

get_standata(x, ...)

Arguments

Value

A named list suitable for passing to rstan::sampling().

Description

This function returns the output (stanfit object) from rstan::sampling().

Usage

get_stanfit(x, ...)

## Default S3 method:
get_stanfit(x, ...)

## S3 method for class 'bscmfit'
get_stanfit(x, ...)

Arguments

Value

Object of class stanfit.

Description

leave_donor_out() re-estimates the original model after omitting donors from the donor pool. By default, it removes one donor at a time. Setting cumulative = TRUE instead removes donors cumulatively following the chosen order, which can be used to assess how sensitive the results are to the most influential donors.

Usage

leave_donor_out(x, ...)

## S3 method for class 'bscmfit'
leave_donor_out(
  x,
  order = "descending",
  cumulative = FALSE,
  probs = c(0.025, 0.5, 0.975),
  ...
)

Arguments

Details

The donor order can always be supplied explicitly. For one-donor leave-out runs (cumulative = FALSE), that order controls the order of the returned runs. For cumulative leave-out runs (cumulative = TRUE), donors are removed from the ordered pool one by one. If no order is supplied, one-donor leave-out uses the original donor order, while cumulative leave-out orders donors by posterior mean donor weight in descending order.

For models with multiple treated units, automatic donor ranking is based on the average posterior mean donor weight across treated units.

Value

An object of class bscm_ldo with data frames effect, rmse, rmse_ratio, weights, and diagnostics, and a metadata list. The data frames contain posterior summaries for the original fit (step = 0) and for each leave-out run, identified by step, n_removed, and last_removed. The metadata list contains the omitted donors, whether donor omission was cumulative, summary probabilities, and model metadata needed for plotting. The result can be visualized with plot_weights() and plot_effects().

Description

Estimates the leave-future-out (LFO) expected log predictive density (ELPD) for bscmfit models. The LFO-CV is performed over the pre-treatment period: at each step t (from L to T_pre_min - 1), the model is evaluated on its ability to predict the next pre-treatment observation y[t+1] having been fitted on y[1:t]. For models with multiple treated units, all units must have at least L + 1 pre-treatment observations, and only the first min(T_pre) time points are used so the number of treated units stays constant throughout the evaluation period. In this case, a joint ELPD of all treated units is computed at each step.

Usage

lfo(x, ...)

## S3 method for class 'bscmfit'
lfo(x, L, exact = FALSE, k_threshold = 0.7, ...)

Arguments

Value

An object of class bscm_lfo, a list with components:

• ELPD: Total expected log predictive density.

• ELPD_SE: A crude approximation of standard error of ELPD, ignoring serial dependency of ELPD estimates.

• ELPDs: Vector of per-step ELPDs (length T_pre_min - L). Element k is the ELPD for predicting observations at L + k.

• ks: Pareto k values (length T_pre_min - L - 1, NULL for exact LFO).

• refits: Time indices at which the model was re-estimated.

• L: The value of L used.

• T_pre_min: The minimum pre-treatment period length across treated units.

• times: Vector of all unique time values from the original data.

• time: Name of the time variable.

• k_thres: The Pareto k threshold used.

References

Bürkner PC, Gabry J, and Vehtari A (2020). Approximate leave-future-out cross-validation for Bayesian time series models. Journal of Statistical Computation and Simulation, 90(14), 2499–2523, https://doi.org/10.1080/00949655.2020.1783262.

Examples

lfo_approx <- lfo(fit_single_treated, L = 10)
lfo_approx

lfo_exact <- lfo(fit_single_treated, L = 10, exact = TRUE)
lfo_exact

Description

Returns posterior draws of pointwise log-likelihoods of the treated units per pre-treatment time point

Usage

## S3 method for class 'bscmfit'
log_lik(object, ...)

Arguments

Value

A matrix where rows correspond to posterior draws and columns to observed (non-missing) time periods. Output columns are ordered by unit then time: all observed pre-treatment time points for the first treated unit, then the second, and so on.

Examples

log_lik(fit_single_treated) |> head()

Description

bayes_R2 computes the Bayesian $R^2$ measure of model fit for a Bayesian synthetic control model, while loo_R2 computes a leave-one-out adjusted version of the same quantity.

Usage

## S3 method for class 'bscmfit'
loo_R2(object, summary = TRUE, probs = c(0.025, 0.975), fixed_seed = TRUE, ...)

## S3 method for class 'bscmfit'
bayes_R2(object, summary = TRUE, probs = c(0.025, 0.975), ...)

Arguments

Value

data.frame of posterior summary of (LOO-adjusted) R-squared values.

References

Gelman A, Goodrich B, Gabry J, and Vehtari A (2019). R-squared for Bayesian regression models. The American Statistician. 73(3), 307–309, https://doi.org/10.1080/00031305.2018.1549100.

Examples

bayes_R2(fit_single_treated)

Description

Approximate leave-one-out (LOO) cross-validation for Bayesian synthetic control models

Usage

## S3 method for class 'bscmfit'
loo(x, r_eff = TRUE, reloo = FALSE, k_threshold = 0.7, ...)

Arguments

Value

An output from loo::loo().

References

Vehtari A, Gelman A, and Gabry J (2017). Practical Bayesian model evaluation using leave-one-out cross-validation and WAIC. Statistics and Computing. 27(5), 1413–1432, https://doi.org/10.1007/s11222-016-9696-4.

Examples

loo(fit_single_treated)

Description

This simulated data is generated based on a latent factor model with two factors, 40 time points and 53 units. For the first three units, a "treatment" $\tau = 2$ for the last 5 time points $t=0,\ldots,4$ is added to an outcome y. There are two time-varying covariates x and z, with latter one having time-varying effect on y. To see exactly how the data was generated, see data-raw folder on the Github repository of the package.

Format

A data frame with 1,550 rows and 5 columns:

• time: Time index from ranging from -30 to 9.

• id: Unit index variable 'ranging from 1 to 33.

• y: Outcome variable.

• x: Time-varying predictor.

• z: Time-varying predictor.

• treatment: Binary indicator variable where 1 corresponds to the treatment.

Examples

head(multiple_treated)
multiple_treated |>
  dplyr::mutate(treated = any(treatment), .by = id) |>
  ggplot2::ggplot(ggplot2::aes(time, y)) +
  ggplot2::geom_line(
    ggplot2::aes(group = id, colour = treated, alpha = treated)
  ) +
  ggplot2::scale_alpha_manual(values = c(0.3, 1)) +
  ggplot2::theme_bw()

Description

Return the number of chains of bscmfit object

Usage

## S3 method for class 'bscmfit'
nchains(x)

Arguments

Value

Number of Markov chains used in sampling as a single integer value.

Description

Return the number of posterior draws of a bscmfit object

Usage

## S3 method for class 'bscmfit'
ndraws(x)

Arguments

Value

Number of posterior draws as a single integer value.

Description

Estimates in-space and in-time placebo effects based on the model definition of previously estimated Bayesian synthetic control model.

Usage

placebo_effects(x, ...)

## S3 method for class 'bscmfit'
placebo_effects(x, type, L = NULL, probs = c(0.025, 0.975), ...)

Arguments

Details

For the in-space placebo (type = "donor"), original model is re-estimated using each donor as the treated unit in turn (omitting the original, true treated unit). The obtained effect estimates should be around zero, under the assumption that the treatment affected only the true treated unit.

For the in-time placebo (type = "time") we still estimate the treatment effect for the original treated, but move the start of the treatment from $L + 1$ to $T_{pre} + 1$, where $L$ is the minimum number of pre-treatment time points to use, and $T_{pre}$ is the true last pre-treatment time point. In all cases, the obtained treatment effects should fluctuate around zero for time points before the true treatment time, under the assumption of no anticipation effects.

In both cases, the output includes also the results of the input model.

Value

A list with data frames effect, rmse, rmse_ratio and diagnostics, and a metadata list. The data frames contain posterior summaries for each run, identified by a placebo column. For type = "donor", placebo is the treated unit name for the original fit and the donor name for each placebo run. For type = "time", placebo is the assumed treatment start time for each run. The metadata list contains the placebo type, summary probabilities, model setup, and the placebo labels used. The result can be visualized with plot_effects().

Description

Plots posterior means and posterior intervals for the model coefficients. For time-constant coefficients (type = "fixed"), a horizontal point-range plot is produced. For time-varying coefficients (type = "varying"), a ribbon-and-line plot over time is produced, faceted by parameter.

Usage

plot_coefs(x, ...)

## S3 method for class 'bscmfit'
plot_coefs(
  x,
  type = c("fixed", "varying"),
  probs = c(0.025, 0.975),
  combine = FALSE,
  alpha = 0.5,
  scales = "free_y",
  ...
)

Arguments

Value

A ggplot object.

Examples

plot_coefs(fit_single_treated)

Description

plot_effects() plots the posterior mean and posterior interval of the treatment effect over time. For the output of leave_donor_out() or placebo_effects(), it additionally overlays the posterior mean of the treatment effect from each leave-out or placebo fit.

Usage

plot_effects(x, ...)

## S3 method for class 'bscmfit'
plot_effects(x, probs = c(0.025, 0.975), unit = NULL, ...)

## S3 method for class 'bscm_ldo'
plot_effects(x, probs = NULL, ...)

## S3 method for class 'bscm_placebo_effects'
plot_effects(x, probs = NULL, ...)

Arguments

Details

For models with multiple treated units, plot_effects.bscmfit() defaults to plotting the average treatment effect over time since treatment. Supply a unit identifier to plot the effect for a specific treated unit instead.

Value

A ggplot object.

Examples

plot_effects(fit_single_treated)

Description

plot_residuals() plots the posterior mean and posterior interval of the pre-treatment residuals over time (type = "time"), or the posterior distribution of the residual autocorrelations up to a given number of lags (type = "autocorrelation"), again only for pre-treatment time points.

Usage

plot_residuals(x, ...)

## S3 method for class 'bscmfit'
plot_residuals(x, type = "time", probs = c(0.025, 0.975), max_lag = 5L, ...)

Arguments

Details

For models with multiple treated units, returns a named list of per-unit plots.

Value

A ggplot object, or a named list of ggplot objects when x is a bscmfit with multiple treated units.

Examples

plot_residuals(fit_single_treated)
plot_residuals(fit_single_treated, type = "autocorrelation", max_lag = 10L)

Description

plot_weights() visualizes posterior summaries of donor weights for a fitted Bayesian synthetic control model. When applied to the output of leave_donor_out(), it instead visualizes how donor weights or their ranks change across the leave-out runs.

Usage

plot_weights(x, ...)

## S3 method for class 'bscmfit'
plot_weights(
  x,
  point_estimate = "median",
  order = NULL,
  coverage = c(0.5, 0.95),
  linewidth = 1,
  point_size = 2,
  reverse = NULL,
  ...
)

## S3 method for class 'bscm_ldo'
plot_weights(
  x,
  type = "weight",
  point_estimate = "median",
  coverage = 0.95,
  linewidth = 1,
  point_size = 2,
  reverse = TRUE,
  ...
)

Arguments

Value

A ggplot object, or a named list of ggplot objects for models with multiple treated units.

Examples

plot_weights(fit_single_treated)

Description

Plots Pareto k values against the corresponding time variable values, with a horizontal line at the refitting threshold. Only available (and relevant) for PSIS-LFO (not exact LFO).

Usage

## S3 method for class 'bscm_lfo'
plot(x, ...)

Arguments

Value

A ggplot object, or NULL invisibly for exact LFO.

Description

A plot of the posterior mean and posterior interval of the treatment effect and synthetic control over time for single treated unit models. For models with multiple treated units, plots the average treatment effect over time since treatment across all treated units.

Usage

## S3 method for class 'bscmfit'
plot(x, probs = c(0.025, 0.975), ...)

Arguments

Value

A ggplot object

Examples

plot(fit_single_treated, probs = c(0.05, 0.95))

Description

Returns draws from the posterior distribution of the expected value of the Bayesian synthetic control model.

Usage

## S3 method for class 'bscmfit'
posterior_epred(object, ...)

Arguments

Value

A matrix where rows correspond to posterior draws and columns to time periods.

Description

Returns draws from the posterior distribution of the linear predictor of the Bayesian synthetic control model. Since the model uses a Gaussian likelihood with identity link, this is equivalent to posterior_epred.bscmfit().

Usage

## S3 method for class 'bscmfit'
posterior_linpred(object, transform = FALSE, ...)

Arguments

Value

A matrix where rows correspond to posterior draws and columns to time periods.

Description

Returns draws from the posterior predictive distribution of the Bayesian synthetic control. Note that this function does not simulate new realizations from this distribution, but rather returns the posterior draws computed during the model estimation.

Usage

## S3 method for class 'bscmfit'
posterior_predict(object, ...)

Arguments

Value

A matrix of posterior predictive draws of the synthetic control, where rows correspond to posterior draws and columns to time periods.

Description

Prints the summary of the leave-future-out cross-validation.

Usage

## S3 method for class 'bscm_lfo'
print(x, ...)

Arguments

Value

Returns x invisibly.

Description

Print method for bscmfit objects

Usage

## S3 method for class 'bscmfit'
print(x, ...)

## S3 method for class 'summary_bscmfit'
print(x, ...)

Arguments

Value

Returns x (invisibly).

Returns x (invisibly).

Description

Print MCMC Diagnostics

Usage

## S3 method for class 'bscmfit_diagnostics'
print(x, print_table = NULL, ...)

Arguments

Value

input x (invisibility).

Description

Convenience wrapper around projpred::proj_predict() that accepts data in the original long format used by bscm(). If newdata is supplied in long format, it is converted internally to the wide format for projpred.

Usage

proj_predict_bscm(object, newdata = NULL, ...)

Arguments

Value

The output of projpred::proj_predict().

Description

Autocorrelation function for residuals of BSCM

Usage

residual_acf(x, max_lag = 5L, probs = c(0.025, 0.975))

Arguments

Value

A data frame.

Examples

residual_acf(fit_single_treated, max_lag = 10)

Description

Returns posterior draws or summaries of residuals, defined as the difference between the observed outcome and the posterior expected value.

Usage

## S3 method for class 'bscmfit'
residuals(
  object,
  summary = TRUE,
  probs = c(0.025, 0.975),
  pretreatment_only = TRUE,
  ...
)

Arguments

Value

A data.frame of posterior summaries (summary = TRUE) or posterior draws (summary = FALSE) in long format.

Examples

residuals(fit_single_treated) |> head()

Description

Returns posterior draws (or summaries) of root mean squared errors (RMSEs) for the pre-treatment and post-treatment periods.

Usage

rmse(x, ...)

## S3 method for class 'bscmfit'
rmse(x, average = FALSE, summary = TRUE, probs = c(0.025, 0.975), ...)

Arguments

Value

A tibble of posterior summaries (summary = TRUE) or posterior draws (summary = FALSE).

Examples

rmse(fit_single_treated, probs = c(0.01, 0.5, 0.8))

Description

Returns posterior draws (or summaries) of the ratio of the post-treatment RMSE to the pre-treatment RMSE.

Usage

rmse_ratio(x, ...)

## S3 method for class 'bscmfit'
rmse_ratio(x, average = FALSE, summary = TRUE, probs = c(0.025, 0.975), ...)

Arguments

Value

A tibble of posterior summaries (summary = TRUE) or posterior draws (summary = FALSE).

Examples

rmse_ratio(fit_single_treated, probs = c(0.01, 0.5, 0.8))

Description

Extract residual standard deviations of a Bayesian synthetic control model

Usage

## S3 method for class 'bscmfit'
sigma(object, summary = TRUE, probs = c(0.025, 0.975), ...)

Arguments

Value

A data.frame of posterior summaries (summary = TRUE) or posterior draws (summary = FALSE) in long format.

Examples

sigma(fit_single_treated)

Description

This simulated data is generated based on a latent factor model with three factors, 50 time points and 51 units. For the first unit (id = 1), a "treatment" $\tau = 1 + t$ for the last 10 time points $t=0,\ldots,9$ is added to an outcome y. To see exactly how the data was generated, see data-raw folder on the Github repository of the package.

Format

A data frame with 2550 rows and 5 columns:

• time: Time index from ranging from -40 to 9.

• id: Unit index variable 'ranging from 1 to 31.

• y: Outcome variable.

• x: Time-varying predictor.

• treatment: Binary indicator variable where 1 corresponds to the treatment.

Examples

head(single_treated)
single_treated |>
  ggplot2::ggplot(ggplot2::aes(time, y, group = id)) +
  ggplot2::geom_line(alpha = 0.3) +
  ggplot2::geom_line(
    data = single_treated |> dplyr::filter(id == 1), colour = "tomato"
  ) +
  ggplot2::theme_bw()

Description

Generates posterior summary statistics for a Bayesian synthetic control model estimated with bscm(). Note that many (more) summaries are available via separate methods, e.g., donor_weights().

Usage

## S3 method for class 'bscmfit'
summary(object, probs = c(0.025, 0.975), ...)

Arguments

Value

A data frame with various posterior summaries of the model.

Description

Returns posterior draws (or summaries) of the pre- and post-treatment trajectories of treated units, i.e. draws y_repfrom the posterior predictive distribution of the model.

Usage

synthetic_control(x, ...)

## S3 method for class 'bscmfit'
synthetic_control(x, summary = TRUE, probs = c(0.025, 0.975), ...)

Arguments

Value

A data.frame of posterior summaries (summary = TRUE) or posterior draws (summary = FALSE) in long format.

Examples

synthetic_control(fit_single_treated) |> tail()

Description

Treatment effect estimates of a Bayesian synthetic control model

Usage

treatment_effect(x, ...)

## S3 method for class 'bscmfit'
treatment_effect(
  x,
  average = TRUE,
  summary = TRUE,
  probs = c(0.025, 0.975),
  ...
)

Arguments

Value

A tibble of posterior summaries (summary = TRUE) or posterior draws (summary = FALSE) in long format.

Examples

treatment_effect(fit_single_treated) |> tail()

Description

Function tv() validates its arguments and returns the necessary variables needed for the construct the time-varying coefficients in bscm(). You should not call tv() separately, instead it should be part of a formula, e.g., y ~ x + tv(~ z * w, df = 20).

Usage

tv(tv_formula, df = 10, type = "rw1", noncentered = TRUE)

Arguments

Value

Object of class tv_term (a list).