This page compares a full observational causal-inference workflow, the kind used for real-world-evidence (RWE) analyses, run on real patient data against the same workflow run on two synthetic counterparts: a classical sequential-regression baseline and Aindo-generated synthetic data. We want to test not only whether synthetic data can preserve the statistical properties of the real cohort, but also its underlying causal mechanisms, and whether a deep generative model does a better job of this than a classical statistical synthesizer.

The example: a published NHANES cohort examining statin use and all-cause mortality. Confounding by indication is present here: clinicians preferentially prescribe statins to older, higher-risk patients, to the point that a naive, unadjusted comparison shows the opposite of the effect reported in the literature. This makes it a useful test case: it requires the analysis to correctly identify and adjust for confounding, not just reproduce a simple average.

The analysis comprises four steps: a naive comparison and three causal-adjustment methods (G-computation, IPTW, TMLE). For each, we run the identical analysis on the real data and on both synthetic datasets, and report all three results side by side.

Data. NHANES 1999-2018, ten survey cycles, linked to the NCHS public-use mortality file (follow-up through Dec 31, 2019).