Causal analysis of absolute and relative risk reductions

Björn Meder, Charley M. Wu, Felix G. Rebitschek

TL;DR

This paper reveals that the two most common ways of reporting medical trial results—Absolute Risk Reduction (ARR) and Relative Risk Reduction (RRR)—are not just different "flavors" of the same data, but represent fundamentally different causal models of the world. By mapping these metrics onto Causal Bayes Nets, the authors show that RRR is mathematically equivalent to "causal power," making it a far more robust tool for predicting how a treatment will perform in new populations with different baseline risks.

What Problem Does This Paper Address?

In medical research, we often ask: "How well does this treatment work?" The answer is usually given as an ARR (the simple difference in infection rates between groups) or an RRR (the percentage reduction in risk). While researchers have debated for decades which metric is "better" or more "honest," this debate has mostly focused on how to communicate risk to patients.

The authors identify a deeper, technical gap: transportability. If a vaccine trial in a low-risk population shows a certain benefit, how do we accurately predict its effect in a high-risk population? The paper addresses this by moving beyond descriptive statistics and treating these metrics as parameters in a generative causal model.

Methods

The researchers employ the framework of Causal Bayes Nets (CBNs) to analyze the underlying logic of risk metrics. Their approach involves:

  1. Formal Proof of Equivalence: They mathematically demonstrate that ARR is equivalent to the $\Delta P$ metric (a linear-additive model), while RRR is equivalent to causal power (a noisy-logical model, specifically a "noisy-AND-NOT" gate for preventive causes).
  2. Causal Graph Parameterization: They use a simple graph where a Cause ($C$, the treatment) and Background Factors ($A$, the baseline risk) both influence an Effect ($E$, the disease).
  3. Counterfactual Generalization: By fixing the "causal strength" parameter ($w_c$) derived from trial data and varying the "baseline risk" parameter ($w_a$), they simulate how each metric predicts outcomes in novel contexts.

Datasets

The paper validates its theoretical findings using real-world data from three major COVID-19 vaccine trials:

  • BNT162b2 (Pfizer-BioNTech)
  • mRNA-1273 (Moderna)
  • ChAdOx1 (AstraZeneca)

They also use three hypothetical scenarios (low, intermediate, and high baseline risk) to illustrate the "crossover points" where the two metrics diverge most sharply.

Key Insights and Ideas

  • The Hidden Assumption: Choosing a metric is an implicit choice of a causal mechanism. Using ARR assumes the treatment "subtracts" a fixed amount of risk regardless of the environment. Using RRR assumes the treatment "blocks" a proportion of the background risk.
  • The "Additive Trap": Because ARR assumes a linear-additive relationship, it can produce "impossible" results. For example, if a treatment has an ARR of 5% but is applied to a population with only a 2% baseline risk, the model predicts a -3% chance of getting the disease—a mathematical absurdity.
  • RRR as "Causal Power": RRR (Causal Power) is "better-behaved." It naturally stays within the [0, 1] probability range. It aligns with the biological intuition that a vaccine's "power" is its ability to prevent a certain percentage of potential infections, whether the virus is rare or rampant.
  • Divergence in Generalization: Both metrics agree perfectly on the data they were calculated from. However, as soon as the baseline risk changes (e.g., moving from a clinical trial to a real-world "hotspot"), their predictions diverge wildly.

Key Results

Using the Pfizer trial data (ARR $\approx$ 0.84%, RRR $\approx$ 95%):

  • In a high-risk scenario (10% baseline risk): The ARR model predicts the vaccine would only reduce risk to 9.16%. The RRR model predicts it would reduce risk to 0.5%.
  • In a "Human Challenge" scenario (100% exposure): ARR predicts a 99.16% infection rate, whereas RRR predicts only 5%.
  • The Verdict: RRR is a more "portable" metric. It captures a stable property of the treatment's preventive power, whereas ARR is highly sensitive to the specific baseline risk of the trial population.

Limitations and Future Work

  • Stability of Causal Strength: The authors assume the causal strength ($w_c$) remains constant across populations. In reality, factors like age, genetics, or viral variants might change the actual "power" of the treatment.
  • Binary Outcomes: The current analysis is limited to binary (yes/no) outcomes. Future work could extend this to continuous variables or time-to-event data.
  • Psychological Application: The authors suggest that these findings should be used to design better experiments in cognitive psychology to see if humans naturally think in "additive" or "proportional" causal terms when learning from data.
  • Risk Communication: While RRR is better for modeling, the authors acknowledge it can be misleading to laypeople if the baseline risk isn't also provided. They advocate for "Fact Boxes" that show both numbers to support informed decision-making.

Summary updated

Abstract

Any new medical innovation must first prove its benefits with reliable evidence from clinical trials. Evidence is frequently expressed using two metrics, summarizing treatment benefits based on either absolute risk reductions (ARRs) or relative risk reductions (RRRs). While both metrics are derived from the same data (e.g., observed frequencies of a disease in a treatment and control group), they implement conceptually and causally distinct ideas. Here, we analyze these risk reductions measures from a causal modeling perspective, revealing an isomorphic relationship to central measure of causal strength and causal Bayesian networks. First, we show that ARR is equivalent to the Delta P metric, while RRR is equivalent to the causal power metric, thus clarifying the implicit causal assumptions. Second, we show how this formal equivalence establishes a relationship with causal Bayesian networks theory, offering a basis for incorporating risk reduction metrics into a computational causal modeling framework. Drawing on these analyses, we demonstrate that under dynamically varying baseline risks, ARRs are inadequate when generalizing treatment effects to novel contexts. Specifically, the inherent assumption of a linear parameterization of the underlying causal graph leads to incorrect conclusions when generalizing to baseline risks differing from those from which the original effect was obtained. For instance, generalizing the effect of a vaccine to new contexts with different baseline risks (e.g., from one population to another). Our analyses highlight the shared principles underlying risk reduction metrics and measures of causal strength, emphasizing the potential for explicating causal structure and inference in epidemiological research.
BibTeX
@article{meder2025causal,
  author = {Björn Meder and Charley M. Wu and Felix G. Rebitschek},
  title = {{Causal analysis of absolute and relative risk reductions}},
  journal = {Journal of Mathematical Psychology},
  year = {2025},
  doi = {10.1016/j.jmp.2025.102942},
}
Figure