Package 'syntheticdata'

Compare multiple synthesis methods

Description

Runs all three synthesis methods on the same data and returns a comparative validation table.

Usage

compare_methods(data, n = nrow(data), seed = NULL)

Arguments

data	A data frame of real data.
n	Number of synthetic records. Default: same as input.
seed	Random seed passed to synthesize().

Value

A method_comparison object (tibble) with columns: method, metric, value, interpretation.

References

Jordon J, et al. (2022). Synthetic Data – what, why and how? arXiv preprint arXiv:2205.03257. doi:10.48550/arXiv.2205.03257

Examples

set.seed(42)
real <- data.frame(x = rnorm(100), y = rnorm(100))
compare_methods(real, seed = 42)

---

Downstream model fidelity test

Description

Trains a predictive model on synthetic data and evaluates it on real data. Compares to a model trained on real data (gold standard). Measures whether synthetic data preserves predictive signal.

Usage

model_fidelity(x, outcome, predictors = NULL)

Arguments

x	A synthetic_data object.
outcome	Character. Name of the outcome column.
predictors	Character vector (optional). Predictor columns. Default: all other numeric columns.

Details

The real-data baseline uses in-sample evaluation (train and test on the same real data) to provide an upper bound on achievable performance. The synthetic-data model is also evaluated on real data, so the comparison reflects how well the synthetic data preserves predictive signal.

Value

A tibble with columns: train_data, metric, value. For binary outcomes the metric is AUC; for continuous outcomes it is R-squared.

References

Jordon J, et al. (2022). Synthetic Data – what, why and how? arXiv preprint arXiv:2205.03257. doi:10.48550/arXiv.2205.03257

Examples

set.seed(42)
real <- data.frame(
    x1 = rnorm(200), x2 = rnorm(200),
    y = rbinom(200, 1, 0.3))
syn <- synthesize(real, seed = 42)
model_fidelity(syn, outcome = "y")

---

Compute privacy risk metrics

Description

Evaluates re-identification risk of synthetic data through multiple privacy metrics: nearest-neighbor distance ratio, membership inference accuracy, and attribute disclosure risk.

Usage

privacy_risk(x, sensitive_cols = NULL)

Arguments

x	A synthetic_data object from synthesize.
sensitive_cols	Character vector (optional). Columns considered sensitive for attribute disclosure assessment.

Value

A privacy_assessment object (tibble) with columns: metric, value, risk_level.

References

Snoke J, et al. (2018). General and specific utility measures for synthetic data. Journal of the Royal Statistical Society A, 181(3):663–688. doi:10.1111/rssa.12358

Examples

set.seed(42)
real <- data.frame(age = rnorm(100, 65, 10),
                   sbp = rnorm(100, 130, 20))
syn <- synthesize(real, seed = 42)
privacy_risk(syn)

---

Generate synthetic data from a real dataset

Description

Creates a synthetic version of the input data that preserves marginal distributions and pairwise correlations while adding controlled noise for privacy protection.

Usage

synthesize(
  data,
  method = c("parametric", "bootstrap", "noise"),
  n = nrow(data),
  noise_level = 0.1,
  seed = NULL
)

Arguments

data	A data frame of real clinical data.
method	Synthesis method: "parametric" (default): fits Gaussian copula to continuous variables, multinomial to categorical. Fast, interpretable. "bootstrap": nonparametric resampling with optional noise. "noise": adds calibrated Laplace noise to each variable (differential privacy inspired).
n	Number of synthetic records. Default: same as input.
noise_level	For method = "noise": scale of Laplace noise relative to variable SD. Default 0.1.
seed	Random seed for reproducibility. If non-NULL, the global RNG state is saved before and restored after synthesis so that calling code is not affected.

Details

The parametric method uses a Gaussian copula approach: marginal distributions are estimated empirically and the joint dependence structure is captured via the correlation matrix of normal scores. This preserves both marginal shapes and pairwise associations while generating genuinely new observations.

Value

A synthetic_data object (list) with components: ⁠$synthetic⁠ (tibble of synthetic records), ⁠$real⁠ (tibble of the original data, retained for downstream validation), ⁠$method⁠, ⁠$n_original⁠, ⁠$n_synthetic⁠, ⁠$variables⁠.

References

Jordon J, et al. (2022). Synthetic Data – what, why and how? arXiv preprint arXiv:2205.03257. doi:10.48550/arXiv.2205.03257

Examples

set.seed(42)
real <- data.frame(
  age = rnorm(200, 65, 10),
  sbp = rnorm(200, 130, 20),
  sex = sample(c("M", "F"), 200, replace = TRUE),
  outcome = rbinom(200, 1, 0.3)
)
syn <- synthesize(real, method = "parametric", seed = 42)
syn

---

Validate synthetic data quality

Description

Computes utility and privacy metrics comparing synthetic data to the original real dataset.

Usage

validate_synthetic(
  x,
  metrics = c("distributional", "correlation", "discriminative", "privacy")
)

Arguments

x	A synthetic_data object from synthesize().
metrics	Character vector of metrics: "distributional": KS statistic per numeric variable. "correlation": Frobenius norm of correlation difference. "discriminative": AUC of real-vs-synthetic classifier. "privacy": nearest-neighbor distance ratio.

Details

Utility metrics assess how well the synthetic data preserves statistical properties. Privacy metrics assess the risk of re-identification.

Discriminative accuracy near 0.5 means the synthetic data is indistinguishable from real data. Privacy ratio > 1 means synthetic records are not closer to real records than real records are to each other.

Value

A synthetic_validation object (tibble) with columns: metric, value, interpretation.

References

Snoke J, et al. (2018). General and specific utility measures for synthetic data. Journal of the Royal Statistical Society A, 181(3):663–688. doi:10.1111/rssa.12358

Examples

set.seed(42)
real <- data.frame(age = rnorm(100, 65, 10), sbp = rnorm(100, 130, 20))
syn <- synthesize(real, seed = 42)
validate_synthetic(syn)

---