Abstract
We study causal inference under case-control and case-population sampling, focusing on the binary-outcome and binary-treatment case. The parameters of interest are causal relative and attributable risks defined via the potential outcome framework. It is shown that strong ignorability is not always as powerful as it is under random sampling and that certain monotonicity assumptions yield comparable results in terms of sharp identified intervals. Specifically, the usual odds ratio is shown to be a sharp identified upper bound on causal relative risk under the monotone treatment response and monotone treatment selection assumptions. We offer algorithms for inference on the causal parameters that are aggregated over the true population distribution of the covariates. We demonstrate the approach through three empirical examples: private school attendance in Pakistan, gang involvement in Brazil, and physician billing practices in the United States.
Introduction
In many fields, researchers seek to understand the causal impact of a treatment on a rare outcome. For instance, what is the effect of a specific educational intervention on entering an elite university, or the impact of social programs on preventing criminal gang involvement? In these scenarios, standard random sampling is often impractical; a random sample of thousands might yield only a handful of "cases" where the outcome occurred.
To address this, researchers frequently employ outcome-based sampling, such as case-control designs, where individuals who experienced the outcome (cases) are oversampled and compared to those who did not (controls). However, traditional causal inference tools, such as the "strong ignorability" (unconfoundedness) assumption, do not translate directly from random samples to outcome-based samples. This paper provides a robust framework for causal inference in these settings by moving away from point identification and toward sharp partial identification using intuitive monotonicity assumptions.
Methodology: Beyond Unconfoundedness
Standard causal inference often relies on the assumption that all confounders are observed. In observational data, this is rarely certain. Furthermore, in case-control studies, even assuming strong ignorability is often insufficient to point-identify parameters like Relative Risk (the ratio of probabilities) or Attributable Risk (the difference in probabilities) without knowing the exact prevalence of the outcome in the general population.
Monotonicity Assumptions
Instead of relying solely on unconfoundedness, this research utilizes two key monotonicity assumptions to bound the causal effects:
- Monotone Treatment Response (MTR): This assumes that the treatment is potentially beneficial (or at least neutral) but never harmful. For example, staying in school is assumed not to increase the likelihood of joining a gang.
- Monotone Treatment Selection (MTS): This assumes that those who select into the treatment group are naturally more likely to experience the positive outcome regardless of the treatment itself.
The Odds Ratio as a Sharp Bound
A central theoretical contribution of this work is the proof that under MTR and MTS assumptions, the standard Odds Ratio (OR)—a common measure of association in medical and social sciences—serves as the sharp upper bound for the Causal Relative Risk. This provides a formal causal justification for using the OR as a "guardrail" or conservative estimate of effect size when the true relative risk cannot be identified.
Implementation
The authors provide a "retrospective sieve logistic estimator" to aggregate these causal effects over the population distribution of covariates. To facilitate the use of these methods, the authors have released the ciccr R package, allowing researchers to calculate these bounds in their own case-control or case-population studies.
Empirical Applications
The framework was applied to three distinct datasets to demonstrate its versatility:
1. Private Education and University Entrance (Pakistan)
The study examined whether attending a private school increases the probability of entering a "Very Selective University" (VSU) in Pakistan. Using survey data from male students in Islamabad and Lahore, the analysis found that the causal relative risk of entering a VSU due to private schooling is bounded by an upper limit of 1.26. This suggests that while private schooling has a positive association, its actual causal impact on elite university entrance is likely modest.
2. Schooling and Gang Involvement (Brazil)
In Rio de Janeiro favelas, the researchers analyzed the link between staying in school and joining drug-trafficking gangs. By combining census data with a specialized gang member survey, the study found a massive potential impact. The relative risk of joining a gang for those not in school was bounded by an odds ratio of 15.01, highlighting the critical role of educational retention in crime prevention.
3. Physician Billing Practices (USA)
The researchers investigated whether physicians in small group practices (fewer than six people) were more likely to be flagged for high Medicare billing (over 100 hours per week). Analyzing a dataset of 78,165 physicians, the study found an upper bound relative risk of approximately 2.68, suggesting a significant correlation between practice size and billing intensity.
Discussion and Limitations
The paper highlights that Case-Population sampling (comparing cases to a general population sample) is often more informative and mathematically tractable for causal inference than traditional case-control sampling.
However, the framework has limitations. The bounds for Attributable Risk are sensitive to the assumed population prevalence ($p_0$); if $p_0$ is entirely unknown, the bounds can become quite wide. Additionally, the current methodology is restricted to binary treatments and outcomes. Finally, the validity of the results depends on the plausibility of the MTR and MTS assumptions in the specific context of the study.
Conclusion
By shifting the focus from elusive point estimates to sharp identified bounds, this research provides a rigorous way to interpret outcome-based studies. It demonstrates that the ubiquitous odds ratio is not just a convenient statistic, but a mathematically sound upper bound for causal effects under reasonable monotonicity assumptions.
Summary updated