A new arXiv preprint reports that inverse-probability weighting performed close to its target in simulated case-control data, while broad-radius propensity-score matching was less reliable. In an ovarian-cancer application, estimates for regular aspirin use were all below 1, but the results varied with the method and the reported intervals.
A question about case-control data
The paper asks whether propensity scores fitted in controls can estimate population relative effects in case-control designs when outcome prevalence or sampling probabilities are unknown. It adapts inverse-probability-of-treatment weighting (IPTW), a one-step efficient doubly robust estimator and a two-stage propensity-score caliper matching procedure, and proposes diagnostic checks for overlap.
The simulations rewarded careful weighting
In the case-control simulations, each analytic dataset contained a random sample of 1,000 cases and 4,000 controls. In one scenario, the target marginal risk ratio among the treated (mRRT) was 0.93. Across 500 replications, IPTW had negligible bias and 95.6% coverage—the percentage of 95% intervals that contained the target—while matching with both radii set at 0.05 had 82% coverage.
Another scenario targeted a marginal risk ratio (mRR) of 0.77. IPTW and one-step estimation had coverage of 93.6% and 94.6%, respectively. Logistic regression showed bias of 0.063 and 80.6% coverage. In a nonlinear scenario, using machine learning for the supporting statistical models reduced bias and increased coverage relative to parametric models, although it could increase variance.
Matching was highly sensitive to its settings
Results also depended on the matching settings. In the test-negative simulation, IPTW had bias of 0.033 and 89.67% coverage. Matching with a first-stage radius of 0.05 had bias above 0.24 and coverage below 30%, while tighter radii improved performance.
An ovarian-cancer analysis gave a mixed signal
To apply the approach, the authors used PROVAQ, a Montreal population-based case-control study conducted from 2011 to 2016. It recruited women aged 18 to 79 and included 498 ovarian-cancer cases and 908 controls. Cases participated at a reported rate of 78%, compared with 56% for controls; controls were frequency-matched by five-year age strata and region.
IPTW and propensity-score matching were used for mRRT. For mRR, the researchers compared logistic regression, IPTW and a one-step doubly robust estimator, using a control-fitted model that included demographic, reproductive, lifestyle and medical covariates.
All three PROVAQ mRRT estimates for regular aspirin use were below 1: 0.753 with IPTW, 0.777 with IPTW SL0.001 and 0.758 with matching. Their reported 95% confidence intervals were [0.474, 1.007], [0.469, 0.942] and [0.460, 0.990], respectively.
For mRR, the parametric logistic, IPTW and one-step estimates were around 0.77 to 0.78. Estimates based on Super Learner after truncation were roughly 0.71 to 0.72, and only the IPTW confidence intervals excluded the null. Because the application was an observational case-control study, these estimates describe an association in the dataset and do not establish that aspirin caused a reduction in ovarian-cancer risk.
Why the result remains uncertain
The study is a version 1 preprint built around specified simulations and one case-control application. Its results show how the proposed estimators behaved under the tested conditions, but the ovarian-cancer finding remains sensitive to the statistical method used and should not be read as a definitive human causal result.
Paper data and sources
Original title: Causal inference via propensity scores for case-control studies
Authors: Yan Liu, Anita Koushik, Philippe Boileau et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text