RWE Demo: Real vs. Synthetic Causal Analysis

The question

Does statin use reduce all-cause mortality?

Statins are among the most widely prescribed drugs in the world, and randomized trials support a mortality benefit for many patient groups. But most real-world statin use is observational: clinicians decide who gets prescribed a statin based on perceived cardiovascular risk, not a coin flip. That is exactly the setup where confounding by indication can distort a naive comparison, in either direction.

Study variables

Treatment	`statin`	reported taking a statin in the past month (yes/no)
Outcome	`death`	assumed deceased per the NCHS mortality linkage (through Dec 31, 2019)
Confounders	demographics + clinical measures	age, sex, race, poverty-income ratio, BMI, blood pressure, diabetes, smoking, total cholesterol

Cohorts

Real (NHANES analytic): 54,091 patients, 12 variables
Synthetic (Aindo-generated): 54,091 patients, 12 variables
NHANES 1999-2018, ten survey cycles, linked to the NCHS public-use mortality file (follow-up through Dec 31, 2019).

The data

How is the Aindo synthetic data generated?

Aindo generates synthetic patient records by learning the statistical and causal structure of a real, de-identified cohort, then sampling an entirely new synthetic cohort from what it learned: patients that never existed, with no one-to-one link back to a real person. This page loads the real NHANES analytic dataset side by side with a synthetic cohort Aindo generated from it.

1. Learn

A generative AI model is trained directly on the real, de-identified cohort. Because it learns the whole dataset jointly, not one column at a time, it captures not just each variable's own distribution, but how variables relate to one another (e.g. that statin use, age, blood pressure, and cholesterol move together). That joint structure is exactly the confounding structure a causal analysis like this one depends on.

2. Generate

Once training is done, the real data is set aside, and the model samples a brand-new synthetic cohort from what it learned: synthetic patients, not modified or resampled real ones, so no synthetic record traces back to a specific real person. This is Aindo's Synthesis mode, and it's what makes the output "void of personal information" while remaining statistically realistic, supporting GDPR-, EHDS-, and AI-Act-aligned data sharing.

3. Validate

Before a synthetic dataset ships, Aindo's platform reports a similarity score and a privacy score for the generation run, backed by univariate and bivariate analysis.

In code, the shape of it looks roughly like this (illustrative, not the actual Aindo API):

schema = {
    "statin": "categorical",
    "age": "numeric",
    "sex": "categorical",
    "total_cholesterol": "numeric",
    # ... one entry per column, with its real data type
}

model = AindoSyntheticDataModel(schema)
# Learn
model.fit(real_cohort)
# Generate
synthetic_cohort = model.generate(n_samples=<number of samples requested>)

Fidelity checks

Does the synthetic data look like the real data?

We perform three checks:

Boxplot comparison for the continuous confounders, do individual distributions match?
Prevalence comparison for the binary variables (statin, death, sex, diabetes, smoker), same question, for variables a boxplot doesn't apply to.
Pairwise correlation agreement across every variable, including the categorical ones (race and the binary variables above). Do variables relate to each other the same way in both datasets? This is the one that actually matters for a causal analysis: two datasets can have identical marginals while differing completely in their confounding structure.

Continuous confounders: boxplot comparison

Similarly-shaped boxes (median, IQR, and whiskers) mean the real and synthetic distributions agree.

Age

Poverty-income ratio

BMI

Systolic BP

Diastolic BP

Total cholesterol

Interpretation. Boxplot comparison. Every continuous confounder's real and synthetic boxes (median, IQR, and whiskers) line up closely, meaning the synthetic data's marginal distributions match the real data's.

Binary variables: prevalence

A boxplot doesn't apply to a two-valued variable: comparing prevalence is the equivalent check.

Interpretation. Binary prevalence. Rates agree closely between real and synthetic: statin 16.2% vs 16.6%, diabetes 12.2% vs 11.8%, smoker 45.3% vs 43.5%, sex (male) 48.0% vs 46.8%. death shows the largest gap, 15.8% vs 13.4%.

Pairwise correlation agreement

Every pairwise correlation among treatment, outcome, and confounders (categoricals one-hot encoded), real vs. synthetic, 120 pairs in total.

Agreement

Mean absolute difference: 0.0105
Worst-pair difference: 0.0619
Variable pairs compared: 120

Interpretation. Pairwise correlations. Across all 120 variable pairs (including the categorical ones, one-hot encoded), the mean absolute difference between real and synthetic correlations is ≈0.0105, and the single worst pair differs by only ≈0.0619. Each point is a correlation between two variables: the real-data correlation on the x-axis, the synthetic-data correlation on the y-axis.

Step 1

The naive comparison

Now we proceed with the main causal question posed in the beginning: does statin use reduce all-cause mortality?

An unadjusted comparison of mortality rates between statin users and non-users.

crude difference = Pr(death | statin) − Pr(death | no statin)

Run on both datasets to compare the two directly.

Crude odds ratio

Source	Odds ratio	95% CI
Real	2.428	[2.300, 2.563]
Synthetic	2.597	[2.455, 2.748]

Crude mortality rate difference

RealSynthetic

Interpretation. Both datasets show a higher crude mortality rate among statin users, with odds ratios above 1 (2.43 [2.30– 2.56] on the real dataset, 2.60 [2.46–2.75] on the synthetic): an unadjusted analysis would conclude statins increase mortality. This is inconsistent with the randomized-trial literature and is a signal of confounding, addressed in the next sections. Both real and synthetic data reproduce this pattern; the crude mortality rate difference estimates are also well aligned (0.1398 and 0.1342).

Why the naive estimate is wrong

Confounding by indication

Statins are not prescribed at random. Clinicians prescribe them to patients judged to be at higher cardiovascular risk. If statin users are systematically older and sicker, they have higher baseline mortality risk independent of the drug. The plot below shows the standardized mean difference (SMD) between statin users and non-users for each confounder, in the real data.

RealSynthetic

Dashed lines mark |SMD| = 0.1, the conventional imbalance threshold.

Interpretation. Several covariates exceed the |SMD| > 0.1 threshold, age most of all: statin users are, on average, roughly two decades older than non-users. Since age is a strong predictor of mortality, the naive comparison above largely reflects the age difference between groups rather than an effect of the drug. The synthetic data shows the same imbalance, consistent with its naive-comparison result above.

Step 2 · Outcome modeling

G-computation

The first technique used for dealing with these differences is G-computation. G-computation estimates an outcome model. This model learns the predictors of the outcome (death) and uses them to predict every participant's mortality risk in a counterfactual scenario where statin use is set to 1 and to 0 for everyone, then averages the difference. Here we use two different outcome models (a logistic regression and an XGBoost forest). Finally, we run the same analysis on both datasets and compare the results.

The confidence interval comes from a nonparametric bootstrap: we resample the dataset with replacement (same size as the original) 500 times, refit the outcome model from scratch on each resample, and take the 2.5th and 97.5th percentile of the resulting 500 estimates as the interval. The point estimate itself is not an average over these resamples. It is computed once, by fitting the model on the full original dataset; the bootstrap is only used to quantify the uncertainty around it.

Estimated treatment effect (risk difference)

RealSynthetic

Point estimates & 95% CI

Model	Source	ATE	95% CI
Logistic regression	Real	-0.0365	[-0.0422, -0.0317]
Logistic regression	Synthetic	-0.0340	[-0.0386, -0.0274]
XGBoost	Real	-0.0132	[-0.0169, -0.0070]
XGBoost	Synthetic	-0.0148	[-0.0187, -0.0060]

Interpretation. Once the outcome model adjusts for age and the other confounders, the estimated effect of statin use on mortality flips sign relative to the naive comparison, for both the real and synthetic data, model by model, with overlapping confidence intervals.

Step 3 · Propensity weighting

Inverse Probability of Treatment Weighting

Another approach is to model the treatment: the rationale is to learn the patterns that determine treatment assignment, then weight each individual by their probability of receiving the treatment they actually got. IPTW reweights each participant by the inverse probability of the statin use they actually received, so that in the reweighted pseudo-population, statin use no longer depends on the confounders. We then compare weighted mortality means between arms. As with G-computation, the confidence interval is a nonparametric bootstrap of 500 resamples of the dataset.

Estimated treatment effect (risk difference)

RealSynthetic

Point estimates & 95% CI

Model	Source	ATE	95% CI
Logistic regression	Real	0.0109	[0.0023, 0.0197]
Logistic regression	Synthetic	0.0050	[-0.0027, 0.0151]
XGBoost	Real	0.0547	[0.0469, 0.0647]
XGBoost	Synthetic	0.0497	[0.0437, 0.0632]

Interpretation. The propensity models still do not point in the reversed, beneficial direction on both datasets. This is a known limitation of IPTW: the estimate depends entirely on the propensity model, and a more flexible model can produce a small number of extreme weights that distort the weighted comparison; TMLE is designed to be robust to this. The synthetic data reproduces this same instability rather than smoothing it out.

Does weighting actually balance the groups?

As a direct check (weighting should make statin users and non-users look statistically alike on the measured confounders), we redo the SMD check on the real and synthetic data, weighted by the (stabilized) IPTW weight.

Real: unweighted vs. weighted

UnweightedWeighted (XGBoost)

Synthetic: unweighted vs. weighted

UnweightedWeighted (XGBoost)

Interpretation. After weighting, the |SMD| of most confounders drops well below the unweighted values. Nonetheless, IPTW does not completely solve the problem, as mentioned above. For example, propensity weighting does not make the two groups' ages fully comparable, in either the synthetic or the original dataset.

Step 4 · Doubly robust

Targeted Maximum Likelihood Estimation

TMLE combines G-computation and IPTW into a single doubly-robust estimate. It starts from the same two ingredients used above, an outcome model that predicts mortality risk and a propensity model that predicts treatment assignment, and adds a "targeting" step: a small, one-parameter adjustment to the outcome model's predictions, fit using the propensity scores, that nudges the plain G-computation estimate just enough to solve for the ATE directly. The result is doubly robust, meaning it is consistent if either the outcome model or the propensity model is correctly specified, so the estimate does not have to bet entirely on one nuisance model the way G-computation and IPTW each do on their own.

Unlike G-computation and IPTW, TMLE's confidence interval is not a bootstrap. The targeting step comes with a closed-form standard error derived from its efficient influence curve, computed directly from the fitted models on the full dataset, no resampling needed.

Estimated treatment effect (risk difference)

RealSynthetic

Point estimates & 95% CI

Model	Source	ATE	95% CI
Logistic regression	Real	-0.0271	[-0.0373, -0.0169]
Logistic regression	Synthetic	-0.0306	[-0.0393, -0.0220]
XGBoost	Real	-0.0187	[-0.0255, -0.0120]
XGBoost	Synthetic	-0.0207	[-0.0269, -0.0146]

Interpretation. TMLE reverses the naive conclusion for both nuisance-model choices, including XGBoost (the one case where IPTW alone did not reverse) on both real and synthetic data. This reflects TMLE's doubly-robust property: its XGBoost-based estimate uses information from the outcome model to correct for the XGBoost propensity model's poor behavior seen in the IPTW section. The doubly-robust estimate and its confidence interval are the ones we would report, and real and synthetic results agree closely.

Putting it together

Naive vs. causally-adjusted estimates, real vs. synthetic

The naive/crude estimate points one direction; every causally-adjusted method points the other. The synthetic data's results match the real data's at every step.

RealSynthetic

A further use

Aindo's model as a predictor, not just a generator

Everything above used classical ML (logistic regression, XGBoost) as the outcome and propensity models. As a further demonstration, separate from the real-vs-synthetic comparison above, the same deep model architecture behind Aindo's synthetic data generation can also be trained as a plain predictive model: not to generate data, but to predict it.

Here we take a model trained by Aindo directly on the real cohort (no synthetic data involved in this section) and use it as the outcome model for a G-computation-style ATE and, paired with a logistic-regression propensity model, a TMLE estimate.

Outcome model AUC, predicting death (real data)

Outcome model	AUC
Logistic regression	0.8751
XGBoost	0.8890
Aindo model	0.9100

ATE, XGBoost vs. Aindo model as the outcome model (real data)

Method	Outcome model	ATE	95% CI
G-computation	XGBoost	-0.0132	[-0.0169, -0.0070]
G-computation	Aindo model	-0.0118	[-0.0120, -0.0116]
TMLE	XGBoost	-0.0187	[-0.0255, -0.0120]
TMLE	Aindo model	-0.0124	[-0.0184, -0.0064]

Interpretation. Using Aindo's own model as the outcome model gives a G-computation-style ATE of -0.0118 [-0.0120, -0.0116] and a TMLE estimate of -0.0124 [-0.0184, -0.0064]. Both point in the same protective direction as every adjusted estimate above, and both confidence intervals overlap with the corresponding XGBoost-based estimates from earlier on this page; the G-computation CI here in fact sits entirely inside XGBoost's.

It also outperforms both classical outcome models on plain predictive performance: at discriminating who actually died, Aindo's model reaches an AUC of 0.91, ahead of XGBoost's 0.89 and logistic regression's 0.88. Aindo's modeling technology, in other words, isn't only useful for generating a realistic synthetic cohort: the same underlying model is a stronger nuisance model for the causal estimate than the classical alternatives used elsewhere in this analysis.

Note: this section uses only the real data.

Takeaway

Aindo's synthetic cohort reproduces the same causal conclusion

Takehome message: Aindo's synthetic cohort reproduces the same causal conclusion as the real data, at every stage of a full observational RWE pipeline, not just in its summary statistics.

Before any modeling, the synthetic data already matched the real data on individual variable distributions, binary prevalences, and the pairwise correlation structure across all confounders (the harder property to fake, and the one a causal analysis actually depends on).
The naive comparison was confounded by indication (higher-risk patients were preferentially prescribed statins) and pointed the wrong way (on both real and synthetic data: crude mortality rate 0.14 real vs. 0.13 synthetic). A real-world-evidence analysis has to correct for this; a naive synthetic-data check would not have caught it.
G-computation and TMLE corrected for the measured confounders and reversed that conclusion, in close agreement between the two datasets at every step (e.g. TMLE: -0.027 real vs. -0.031 synthetic for logistic regression, -0.019 vs. -0.021 for XGBoost). TMLE's doubly-robust estimate is the one we would report, and it lands in the same place whether it's computed on real patients or on Aindo's synthetic cohort. In all cases, point estimates from the synthetic data fall within the confidence intervals of the real data, and vice versa.

Put together, this is more than "the synthetic data looks similar"; it shows Aindo's synthetic data preserves the underlying causal mechanism, including the exact confounding-by-indication pattern this dataset was chosen to test. That is what makes it a credible substitute for the real patient data in this kind of analysis: an external collaborator, reviewer, or regulator could run this entire causal-inference workflow on the synthetic cohort alone, never touching an identifiable patient record, and reach the conclusion the real data supports, preserving both patient privacy and the validity of the inference.