A drug trial. The control group's outcomes form a bell curve centered at 50. The treatment group's outcomes form two peaks — one at 30, one at 70 — centered at the same mean of 50. The average treatment effect is zero. Every standard statistical test says the drug does nothing.
But the drug split the population in half. Some patients improved dramatically; others declined equally. The treatment didn't shift the distribution. It changed its shape from one hill to two. This is the failure case that motivates topological causal inference (arXiv:2603.14169): when the effect is structural rather than locational, mean-based estimands are blind.
Note (2026-11-07): the abstract proves that topological analogues of CATE and ATE are identifiable up to an explicit error bound under approximate topological ignorability, notes that a marginal persistence-diagram effect is not identified from conditional ignorability alone, and supports this with one synthetic experiment. The essay says identification is proved without the error-bound qualification; its numbers (peaks at 30 and 70 around a mean of 50) are an illustration of mine, and the claim that only topological tools see bifurcations is my own.
Persistent homology detects these shape changes. It counts connected components, loops, and voids in the distribution across multiple scales, producing a summary (the persistence diagram) that captures distributional geometry. The topological causal effect measures whether treatment changes the number or persistence of these features.
The complication: persistent homology doesn't commute with covariate adjustment. You can't just take the topological summary of the marginal distribution and call it causal, because confounders create spurious topological features. The paper proves identification under a conditional version of the ignorability assumption — you have to compute topological effects within strata, not marginally.
The through-claim: the tools you use to measure an effect determine which effects you can see. Mean-based tools see shifts. Variance-based tools see spreads. Only topological tools see bifurcations. The absence of evidence was the evidence of an absent tool.