Simulate complex data from explicit data-generating processes in R
simDAG is an R package for generating synthetic data from directed
acyclic graphs (DAGs). Users specify how variables are generated and how
they depend on one another, and may then simulate datasets from the
resulting data-generating process (DGP).
Beyond ordinary cross-sectional data, simDAG supports complex
longitudinal processes, time-to-event outcomes, censoring, recurrent and
competing events, and dependencies between individuals through networks.
It additionally implements an enhanced formula interface, allowing
specifications of arbitrary interactions, non-linear effects,
dependencies on categorical variables and random effects. The
development version also supports calibrating conditional model
parameters to achieve user-defined target characteristics, including
marginal effects.
Install the current CRAN release:
install.packages("simDAG")
To install the development version from GitHub, including the new
calibrate_dag() functionality:
install.packages("remotes")
# remotes::install_github("RobinDenz1/simDAG")
install.packages("simDAG", repos = c('https://robindenz1.r-universe.dev', 'https://cloud.r-project.org'))
Suppose we want to generate data with a continuous covariate $X$, a binary treatment $A$, and a binary outcome $Y$. Both treatment and outcome depend on $X$, and the outcome also depends on treatment.
library(simDAG)
dag <- empty_dag() +
node("X", type="rnorm", mean=0, sd=1) +
node("A", type="binomial",
formula= ~ -0.5 + X * 0.6) +
node("Y", type= "binomial",
formula= ~ -1 + A * 0.7 + X * 0.4)
set.seed(43)
data <- sim_from_dag(dag, n_sim=10000)
head(data)
#> X A Y
#> <num> <lgcl> <lgcl>
#> 1: -0.03751376 FALSE TRUE
#> 2: -1.57460441 FALSE FALSE
#> 3: -0.48596752 TRUE FALSE
#> 4: 0.46518623 TRUE TRUE
#> 5: -0.90409807 FALSE TRUE
#> 6: -0.27743280 TRUE FALSE
The DAG below represents the data-generating process: $X$ influences both treatment ($A$) and outcome ($Y$), while $A$ also influences $Y$.
Each node is generated according to its specified distribution or conditional model. The order of the nodes and their dependencies define the data-generating process. The same approach extends to much more complex DGPs, including custom generation functions and longitudinal processes.
| Use case | Main functionality |
|---|---|
| Cross-sectional data from a DAG | sim_from_dag() |
| Longitudinal data with time-varying variables | sim_discrete_time() |
| Continuous-time event-driven data | sim_discrete_event() |
| Data with dependencies between individuals |
network() functionality |
| Data with custom conditional models | Custom node() types and generation functions |
| Data with prespecified marginal characteristics |
calibrate_dag() (development version) |
The package implements many node types for user convenience, including
linear-, logistic-, Poisson- and Cox-, regression models, among many
others. See node() for an overview.
Generate data from static DAGs containing continuous, binary, categorical, count, and time-to-event variables. Nodes can be specified using conditional distributions, regression models or custom functions.
Learn more: Simulating Data from a Known DAG
Use sim_discrete_time() to generate complex longitudinal data with
time-varying covariates, time-dependent effects, censoring, recurrent
events, and competing events.
For suitable models, sim_discrete_event() offers an alternative based
on event-driven simulation in continuous time. It can be considerably
faster than discrete-time simulation, although it supports a narrower
range of time-dependent processes.
Learn more: Discrete-time simulation · Discrete-event simulation
simDAG also supports dependencies between individuals through static
and dynamic network structures, allowing the generated data to
incorporate network-based relationships.
Learn more: Simulating Data from a DAG with Network Dependencies
When defining a DGP through conditional models, the resulting marginal characteristics are not always obvious. For example, specifying the coefficients of a logistic regression model does not directly specify the marginal prevalence of its outcome or the marginal causal effect of treatment.
The calibrate_dag() function addresses this problem by allowing users
to leave selected parameters unspecified and determine them numerically
so that user-defined targets are achieved as closely as possible.
For example, suppose we want an outcome prevalence of 30%, while retaining the specified relationships between treatment, a confounder, and the outcome.
ldag <- lazy_dag(
empty_dag() +
node("X", type="rnorm", mean=0, sd=1) +
node("A", type="binomial",
formula= ~ -0.5 + X * 0.8) +
node("Y", type= "binomial",
formula= ~ param(intercept_Y) +
A * 0.7 + X * 0.4)
)
targets <- list(
prevalence_Y = list(
target = 0.30,
eval_fun = function(data) mean(data$Y)
)
)
params <- list(
intercept_Y = list(
start = 0,
range = c(-5, 5)
)
)
fit <- calibrate_dag(
lazy_dag = ldag,
targets = targets,
params = params,
n_sim_run = 5000,
n_sim_eval = 20000,
seed = 42
)
summary(fit)
#> ────────────────────────────────────────────────────────────
#> Calibrated DAG
#> ────────────────────────────────────────────────────────────
#> Parameters (1)
#> parameter estimate start lower upper parscale
#> intercept_Y -1.135 0 -5 5 NA
#>
#> Targets (1)
#> target desired achieved error scaled_error
#> prevalence_Y 0.3 0.3069 0.00685 0.00685
#>
#> Optimization
#> Method: uniroot
#> Convergence: successful
#> Root residual: 0
#> Iterations: 8
#> Simulations: 5,000 optimization; 20,000 final evaluation
#> ────────────────────────────────────────────────────────────
The result includes the calibrated parameters, the achieved target
values, and optimization diagnostics. The resulting fully specified DAG
is available as fit$dag and can be used with the usual simulation
functions.
Calibration is numerical and approximate. Inspect the achieved target values rather than relying solely on optimizer convergence.
Learn more: Calibrating a DAG to Match Target Characteristics
If you are unsure which simulation approach fits your problem, the introductory vignette and cookbook are good places to start.
If you use simDAG in your research, please cite:
Denz, R. and Timmesfeld, N. (2026). Simulating Complex Cross-Sectional and Longitudinal Data using the simDAG R Package. Journal of Statistical Software, 116(2). https://doi.org/10.18637/jss.v116.i02
Bug reports, questions about unexpected behavior, and feature requests are highly welcome. Please open an issue on GitHub.
simDAG is released under the GNU General Public License, version 3 or
later (GPL-3).