Causal Generative Model

Encodes the underlying Directed Acyclic Graph (DAG) of a system directly into the generative architecture. Unlike correlation-based generators, the model explicitly represents variables as functions of their direct causes and independent noise terms. This allows the model to distinguish between observational and interventional distributions, ensuring that generated data respects the invariant causal mechanisms rather than spurious statistical associations.

Leverages Judea Pearl's do-operator to simulate the effect of setting a variable to a specific value, cutting off its incoming causal edges. This enables the generation of data from a post-intervention distribution P(Y | do(X=x)) rather than the conditional distribution P(Y | X=x). This is critical for medical applications where one must simulate the outcome of administering a treatment, removing the confounding bias of who originally received it.

Computes answers to retrospective 'what if' questions at the individual unit level. The model uses abduction (inferring the latent noise variables for a specific factual observation), action (applying the do-operator to alter the causal mechanism), and prediction (computing the resulting outcome). This generates a synthetic twin of a patient showing what their outcome would have been under a different treatment plan, enabling personalized medicine simulations.

Automatically identifies and neutralizes hidden confounders that distort the relationship between independent and dependent variables. By modeling the joint distribution via the causal graph, the generator ensures that synthetic datasets do not propagate selection bias or Simpson's Paradox. This produces balanced datasets where treatment and control groups are statistically exchangeable, preventing downstream predictive models from learning spurious shortcuts.

Operates on the principle that the conditional distribution of an effect given its causes is an autonomous, modular mechanism that remains invariant to changes in other parts of the system. This modularity allows the model to recombine mechanisms to generate data for novel environments or patient populations not seen during training, providing robust out-of-distribution generalization for clinical trial simulations.

Enables the transfer of causal knowledge from a source population to a target population where only observational data exists. By encoding selection diagrams that identify where causal mechanisms differ between domains, the model generates synthetic data that accurately reflects the target population's characteristics. This is essential for adapting models trained on one hospital's data to another with different demographics.

The mathematical backbone of causal generative models. An SCM defines a system of equations X_i = f_i(PA_i, U_i) where PA_i represents the direct causes (parents) of variable X_i and U_i represents exogenous noise. Unlike correlation-based models, SCMs encode asymmetric causal direction and support the do-operator for simulating interventions. In synthetic data generation, the SCM provides the generative ordering—variables are sampled sequentially following the causal graph's topological sort, ensuring that interventional distributions P(Y | do(X=x)) are valid and confounding bias is structurally eliminated.

A formal rule system developed by Judea Pearl for deriving interventional distributions from observational data. Do-calculus provides three rules for transforming expressions containing the do(·) operator into standard probabilistic expressions that can be estimated from data. Key capabilities:

• Rule 1: Insertion/deletion of observations when variables are d-separated
• Rule 2: Exchanging actions for observations when backdoor paths are blocked
• Rule 3: Insertion/deletion of actions when no causal path exists In causal generative models, do-calculus validates that generated counterfactual samples respect the underlying causal structure and that simulated interventions produce logically consistent outcomes.

The process of answering 'what would have happened if' questions at the individual unit level. Counterfactual inference requires three steps:

1. Abduction: Infer the exogenous noise U for a specific observed case using the SCM
2. Action: Apply the do(·) operator to modify the causal structure
3. Prediction: Compute the outcome under the modified structure In synthetic patient data, counterfactual generation enables creating parallel clinical trajectories—for example, generating a synthetic patient's outcome both with and without a specific drug exposure, enabling direct treatment effect estimation without randomized trials.

A primary advantage of causal generative models over standard GANs or VAEs. Confounders are variables that influence both the treatment and outcome, creating spurious correlations. Causal models address this through:

• Backdoor adjustment: Conditioning on the confounder set that blocks all backdoor paths between treatment and outcome
• Front-door criterion: Using mediators when confounders are unmeasured
• Instrumental variable methods: Leveraging variables that affect treatment but not outcome directly Synthetic datasets generated causally preserve the true treatment effect signal while eliminating the confounding associations that would mislead downstream predictive models trained on the data.

Data-driven methods for learning causal graph structure from observational data when the true causal relationships are unknown. Key algorithm families:

• Constraint-based (PC algorithm, FCI): Use conditional independence tests to infer edges and orientations
• Score-based (GES, NOTEARS): Search over graph space optimizing a goodness-of-fit criterion
• Functional causal models (LiNGAM, ANM): Exploit distributional asymmetries to determine causal direction These algorithms provide the structural prior for causal generative models when domain expertise is incomplete, enabling automated construction of the SCM before synthetic data generation begins.

The population-level metric used to validate causal generative models. ATE is defined as E[Y(1) - Y(0)]—the expected difference in outcomes between treated and untreated states. In synthetic data evaluation:

• Generated ATE should match the true ATE embedded in the causal model
• Discrepancies indicate the generative model has failed to preserve the causal structure
• Conditional Average Treatment Effect (CATE) extends this to subgroups, testing whether heterogeneous treatment effects are faithfully reproduced ATE preservation is a stricter quality metric than distributional fidelity alone, ensuring synthetic data supports valid causal conclusions in downstream analyses.