A mathematical preprint reports that some covariance-based denominators used in causal estimators can be exactly self-normalizing under Gaussian sampling: their first-order sampling noise is proportional to the size of the denominator itself, and a standardized version is constant whenever it is defined.
The distinction matters near a weak denominator, meaning one approaching zero. The paper’s first-order analysis says that regime produces a Fieller-type ratio limit and a non-Gaussian Wald limit, warning that ordinary Wald inference may not describe uncertainty accurately there.
A classification of covariance formulas
The self-normalizing class consists of products of powers of nested covariance volumes. Under Gaussian sampling, the paper says these products have an exact factorization into independent chi-squared variables, giving them a specific sampling law as well as the algebraic normalization property.
The paper proves the converse completely in two dimensions. In three dimensions, it characterizes the degree-at-most-two factor class as having the same nested-flag structure, but leaves a mixed branch unresolved.
In practical terms, the results point to a split: formulas built from nested covariance volumes have the self-normalizing property, while slope-type formulas can enter the weak-denominator regime and may need direction-specific diagnostics or robust inversion.
What the simulations showed
The simulations drew known-mean sample covariances directly from Wishart distributions, using samples of 1000 observations and 100000 replications in each cell. In the front-door setting, the standardized denominator was 100.0 in every cell. As mediator residual variance fell across the study’s grid, robust standard deviation rose from 0.038 to 0.696, while Wald interval coverage stayed between 0.949 and 0.950.
For the stated front-door reduced form, with a nonzero treatment–mediator coefficient, the studentized statistic remained asymptotically standard normal along the allowed mediator-residual-variance boundary sequences.
The proximal simulations showed a different warning sign. As the variance parameter vA decreased, a naive marginal diagnostic rose from 179.7 to 624.6 even as the relevant partial diagnostic weakened. In the two weakest cells, Wald coverage fell to 0.930 and 0.931, while inversion coverage remained between 0.951 and 0.953.
A data audit points to the diagnostic problem
The study also audited the SUPPORT dataset, which included 5735 critically ill patients. After residualization, the analysis used rank 19 and 5716 residual degrees of freedom, and assessed sampling variability with 2000 patient-level nonparametric bootstrap replications that repeated imputation, coding and residualization.
The candidate proxy pairs produced sharply different signals. For pafi1/ph1, the naive statistic was 173.4, compared with 3.2 for the partial statistic and 2.3 for the sandwich statistic. For paco21/ph1, the corresponding figures were 16.4, 1980.0 and 315.2.
The SUPPORT exercise is a descriptive audit and cannot test whether the candidate proxies satisfy the assumptions required for proximal causal estimation.
What remains unresolved
The evidence combines mathematical identities, Gaussian simulations and an observational data audit. It does not establish clinical effects, and the simulation findings are specific to the mechanisms and parameter grids used in the study.
The paper leaves the mixed three-variable branch open. It also does not show that self-normalization makes an entire associated estimator regular or that front-door estimation remains precise as mediator residual variance vanishes.
The document identifies itself as arXiv:2608.20223v1, dated 20 Aug 2026. Python scripts for the numerical and real-data experiments are stated to be available, and the author reports support from JSPS KAKENHI Grant Number 25K24203.
Paper data and sources
Original title: Self-Normalizing Denominators in Rational Causal Estimation
Authors: Shu Tamano
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text