FairCauseSyn: Causally Fair Synthetic Data

Updated 15 April 2026

FairCauseSyn is a methodology that generates synthetic health data by enforcing explicit path-specific causal fairness, reducing bias from protected attributes.
It leverages large language models and iterative prompt refinement to produce tabular data matching real data fairness metrics within a 10% threshold.
Benchmark results indicate that when paired with a causally fair predictor, FairCauseSyn reduces direct bias by about 70%, advancing equity in health research.

FairCauseSyn is a methodology for synthetic data generation with explicit path-specific causal fairness constraints, leveraging LLMs as core generative components. It is developed specifically for health applications where equity and the attenuation of bias related to protected attributes (e.g., sex, race) are critical and existing approaches relying on generative adversarial networks (GANs) or standard LLMs enforce weaker, often merely statistical or counterfactual, fairness guarantees. FairCauseSyn is the first framework to generate tabular synthetic health data that matches real data on causal fairness metrics to within 10%, and when coupled with a causally fair predictor, reduces bias associated with protected attributes by approximately 70% compared to real data (Nagesh et al., 23 Jun 2025). This approach marks a fundamental advance in the generation of equitable, high-fidelity synthetic data for clinical and public health research.

1. Motivation and Problem Setting

The motivation behind FairCauseSyn stems from the widespread use of synthetic data as a means to conduct downstream machine learning on sensitive or limited-access datasets, particularly in healthcare, without the privacy risks associated with real patient records. However, synthetic data generators trained naively can propagate or exacerbate existing biases against protected groups. Previous synthetic data frameworks have primarily targeted statistical or counterfactual forms of fairness (e.g., demographic parity, equalized odds, counterfactual fairness), and their deployment in health contexts has been limited. Causal fairness is a stricter standard: it requires the preservation (or correction) of causal structures that could otherwise transmit unfair advantage or discrimination through direct, indirect, or spurious causal paths. In health applications, failure to combat confounding and mediated discrimination can lead to clinically relevant inequities. FairCauseSyn addresses these issues by enforcing path-specific causal-fairness constraints in the LLM-augmented generation loop, a design not employed in previous tabular data generators (Nagesh et al., 23 Jun 2025).

2. Formalization of Causal Fairness Constraints

FairCauseSyn operationalizes causal fairness using the Standard Fairness Model (SFM). Let $X$ denote a binary protected attribute (such as sex), $Z$ observed confounders (e.g., age), $W$ mediators (clinical variables), and $Y$ a binary outcome (e.g., survival). The system defines a structural causal model (SCM) $\langle V, U, \mathcal{F}, P(U)\rangle$ with $V = \{X, Z, W, Y\}$ observed and $U$ latent noise variables. Structural equations specify each variable as $v_i \gets f_i(\text{pa}(v_i), u_i)$ , with a causal graph encoding relations such as $X \to W$ , $Z \to W$ , $Z$ 0, etc. Causal influence from $Z$ 1 to $Z$ 2 is decomposed into path-specific effects (PSEs):

Direct Effect (DE): Difference in $Z$ 3 if $Z$ 4 is set to $Z$ 5 versus $Z$ 6, but mediators $Z$ 7 are held at their value under $Z$ 8.
Indirect Effect (IE): Effect of altering $Z$ 9 from their value under $W$ 0 to $W$ 1 while keeping $W$ 2 fixed.
Spurious Effect (SE): Captures bias due to pathways from $W$ 3 to $W$ 4 not mediated by $W$ 5 or confounded.
Total Variation (TV): Aggregate deviation encompassing DE, IE, and SE.

Causal fairness is attained if $W$ 6, $W$ 7, and $W$ 8 are all approximately zero, indicating the synthetic data do not transmit unfair causal effects from $W$ 9 to $Y$ 0 via any path (Nagesh et al., 23 Jun 2025).

3. LLM-Augmented Synthetic Data Generation Architecture

FairCauseSyn employs a distinctive LLM-orchestrated architecture in lieu of GAN or diffusion models. Key components include:

Data curation: Sampling representative batches from preprocessed real tabular data.
Prompt construction: Employing schema-guided prompts, in-context examples, and a self-consistency mechanism to induce the LLM to generate rows faithful to the original data schema.
Constraint evaluation: Computing causal-fairness metrics (DE, IE, SE, TV) for each batch of generated data.
Iterative refinement: If any fairness metric violates a prespecified threshold, prompts are modified—penalizing unfair generations or supplying additional exemplars—and the LLM is queried again.

This approach does not require LLM fine-tuning; iterative prompt refinement suffices to guide the model toward both marginal fidelity and causal fairness. Each refinement step effectively acts as lightweight “prompt-tuning” (Nagesh et al., 23 Jun 2025).

4. Optimization and Constraint-Satisfaction Protocol

The optimization objective is composite: minimize both the $Y$ 1 distance on feature-wise statistics and penalize deviations from causal fairness. Explicitly, the objective is

$Y$ 2

with $Y$ 3 denoting vector differences in marginal statistics. In practical terms, sampled data are accepted only if all fairness metrics fall below strict thresholds (e.g., $Y$ 4, $Y$ 5, $Y$ 6), yielding a constraint-satisfaction loop in the generative process. Thresholds are set so that the final synthetic data metrics are within $Y$ 7 of real data (Nagesh et al., 23 Jun 2025).

5. Benchmarking: Experimental Design and Quantitative Results

FairCauseSyn was evaluated on the Heart-Failure Clinical Records dataset, using sex as the protected attribute, age and other clinical markers as confounders/mediators, and binary survival as outcome. Benchmark baselines included TVAE, CTGAN, TabDDPM (standard tabular GAN/diffusion generators), TabFairGAN, FairGAN/CFGAN (statistical/counterfactual GANs), Decaf (a causal GAN), and recent LLM-only frameworks with counterfactual constraints.

The table below summarizes causal-fairness metrics (mean ± SD, 5 seeds) for data-level and prediction-level evaluation using both unconstrained and causally fair predictors (FairAdapt).

Evaluation	Dataset	TV	DE	IE	SE
Data Fairness	Real	-0.0121±.0537	-0.0477±.0026	-0.0472±.0068	0.0116±.0556
	Synthetic	-0.0492±.0571	-0.0429±.0043	-0.0002±.0072	0.0064±.0580
No-fairness Model	Real	0.0130±.0458	0.0060±.0022	-0.0163±.0074	0.0093±.0477
	Synthetic	-0.0433±.0550	-0.0287±.0037	-0.0217±.0117	0.0363±.0552
Causally-fair Model	Real	0.0248±.0631	-0.0070±.0016	-0.0538±.0054	0.0219±.0637
	Synthetic	0.0003±.0568	-0.0020±.0030	0.0076±.0108	-0.0099±.0556

All fairness metrics for synthetic data are within ±10% of real data, and training a FairAdapt model on synthetic data reduces direct-effect bias by over 70% compared to real data (Nagesh et al., 23 Jun 2025).

6. Comparative Analysis and Methodological Innovations

FairCauseSyn’s explicit modeling of SFM and PSEs in the generative loop preserves the entire causal-fairness profile (DE, IE, SE), a property lacking in GAN-based and diffusion-based methods that primarily enforce marginal or counterfactual fairness. The LLM-driven generation and iterative prompt refinement provide higher controllability and avoid the instability observed when training deep models with competing fairness losses. Integrating FairCauseSyn data with a causally fair downstream learner yields a significant reduction in direct bias transmitted through the synthetic data. This approach offers a practical solution for producing high-fidelity, causally fair data that supports equitable algorithmic decision-making in health research (Nagesh et al., 23 Jun 2025).

7. Limitations and Prospects for Future Development

Despite marked improvements, the variance in spurious effects (SE) remains high, pointing to challenges in controlling confounder-mediator pathways using only prompt-based mechanisms. The absence of LLM fine-tuning constrains the expressive adaptation of fairness principles; emerging strategies such as parameter-efficient tuning or reinforcement learning–guided prompting may further advance causal fairness enforcement. Scalability to multiclass or continuous protected attributes, and handling more complex cohort structures, represent open frontiers. Bridging LLM-generation with explicit causal graphical modeling (e.g., causal Bayes nets) is anticipated to deliver stronger guarantees and more robust control against spurious associations (Nagesh et al., 23 Jun 2025).

8. Context within Causal-Fairness and Synthetic Data Methodologies

FairCauseSyn’s framework is situated at the intersection of recent advances in fairness-aware synthetic data generation, causal inference, and LLMs. In contrast to PreFair which enforces justifiable fairness on tree-structured graphical models with privacy guarantees via differential privacy, FairCauseSyn targets path-specific causal effects in health datasets using LLM-based synthesis (Pujol et al., 2022). Sensitivity analyses underscore the fragility of fairness metrics under real-world data biases, reinforcing the necessity for robust, causally grounded fairness constraints in the generation pipeline (Fawkes et al., 2024). The development of contrastive and path-specific fairness measures in the synthetic data context further motivates continued innovation beyond statistical and counterfactual notions (Chakraborti et al., 2019).

In summary, FairCauseSyn represents a substantive advance in LLM-augmented, causally fair tabular data generation, enabling both methodological rigor and practical utility for health equity research (Nagesh et al., 23 Jun 2025).