The Journey of Cornfield’s Inequalities and How They Shaped Modern Sensitivity Analysis

Abstract

Cornfield et al.’s 1959 paper introduced a seminal framework for evaluating the impact of unmeasured confounding by quantifying the strength an unobserved factor would need to overturn an observed association. This commentary revisits Cornfield’s original inequalities, clarifying their interpretation, and highlighting the reverse-reasoning principle that underlies modern sensitivity analysis. We summarize key methodological extensions developed by later researchers–ranging from conditional independence formulations and multiplicative bounds to the sharpened limits that led to the E-value, and illustrate how these refinements broadened the applicability of Cornfield’s logic. Finally, we discuss the continued influence of Cornfield-type reasoning across epidemiology and related fields, where sensitivity analysis has become an essential component of study design and causal interpretation. More than six decades later, Cornfield’s insight remains a foundational pillar of causal inference, guiding researchers in assessing the robustness of empirical findings to unmeasured confounding.

1. Introduction

Cornfield’s 1959 paper is widely recognized as a pioneering contribution to sensitivity analysis, introducing the idea that observed associations in observational studies may be vulnerable to unmeasured confounding. Their work articulated a simple but powerful philosophical question: how strong would an unobserved confounder need to be to overturn the observed association? This reasoning eventually came to be formalized as Cornfield’s inequalities. Although the original analysis applied only to binary outcomes and did not account for sampling variability, the conceptual clarity of the result became the foundation upon which numerous modern sensitivity analysis methods were built. The paper also emphasized the critical distinction between association and causation, influencing thinking across epidemiology, statistics, and the broader scientific community.

Stuart (2010) noted that Cornfield’s work directly inspired Rosenbaum and Rubin’s influential framework in the 1980s, which quantified how strong correlations with a hypothetical unmeasured covariate would need to be–both with treatment assignment and the outcome–to eliminate the observed treatment effect (Rosenbaum and Rubin, 1983; Rosenbaum, 2010). This line of work became known as Rosenbaum-style sensitivity analysis. More recently, Cornfield-type logic was extended in the 2010s with the development of the E-value by VanderWeele and Ding, providing an accessible, ready-to-use metric for quantifying the minimum strength of unmeasured confounding needed to explain away an association (Ding and VanderWeele, 2016; VanderWeele and Ding, 2017). Over time, Cornfield’s conceptual insight has gradually shaped the field into what we now refer to as modern sensitivity analysis.

A rich commentary literature has reflected on the historical and methodological significance of Cornfield’s contribution. Holland (1986), in his widely cited review on statistics and causal inference, viewed Cornfield’s inequality as providing an interpretable bound on the influence of unmeasured factors. Greenhouse (2009) offered an alternative derivation of the inequality, yielding additional intuition for general readers and reinforcing its foundational role in epidemiology and causal reasoning. Greenland (1996) documented the methodological extensions developed by successors, including external adjustment techniques and basic tools for bias sensitivity analysis, while Greenland (2012) further emphasized risk ratios and “risk relativism” as key measures for sensitivity assessment. VanderWeele (2022) compared Cornfield’s conditions with subsequent developments to position the E-value as a natural modern successor. More recently, Ding’s textbook, A First Course in Causal Inference, provided a comprehensive overview of unmeasured confounding, tracing a detailed lineage from Cornfield-style reasoning to contemporary E-value formulations (Ding, 2024).

In this commentary, we trace the intellectual journey of Cornfield’s inequality and its enduring influence on sensitivity analysis. First, we revisit the original two inequalities and summarize their role in motivating subsequent extensions. Second, we discuss how Cornfield’s reverse-reasoning principle—framing robustness in terms of the strength of confounding required to negate an association—shaped later methodological work and continues to inform modern applications. Throughout, we refer collectively to “Cornfield’s work,” though we fully acknowledge the contributions of all authors of Cornfield et al.’s landmark paper and their collective role in establishing the foundations of sensitivity analysis.

2. The Elegance of Cornfield’s Inequalities

2.1. Revisiting the Original Argument

We revisit the results now known as Cornfield’s inequalities, originally presented in Appendices A and B of Cornfield, Haenszel, Hammond, Lilienfeld, Shimkin, and Wynder (1959). For complete derivations, we refer the reader to the original paper. Here, we outline the foundational setup and key quantities.

Let U denote an unmeasured (or unobserved) binary confounder, E the exposure, and D the outcome (commonly a disease in public health studies). In the absence of unmeasured confounding, conventional analyses focus on the relative risk (RR) of the outcome under exposure, defined as

$$\begin{array}{c}{\text{RR}}_{\mathit{E}\mathit{D}}={\displaystyle \frac{\mathit{P}\mathrm{(}\mathit{D}=1\mid \mathit{E}=1\mathrm{)}}{\mathit{P}\mathrm{(}\mathit{D}=1\mid \mathit{E}=0\mathrm{)}}}.\end{array}$$

It is important to emphasize that the order of subscripts determines interpretation. The quantity ${\text{RR}}_{\mathit{E}\mathit{D}}$ is not interchangeable with ${\text{RR}}_{\mathit{D}\mathit{E}}$. The latter,

$$\begin{array}{c}{\text{RR}}_{\mathit{D}\mathit{E}}={\displaystyle \frac{\mathit{P}\mathrm{(}\mathit{E}=1\mid \mathit{D}=1\mathrm{)}}{\mathit{P}\mathrm{(}\mathit{E}=1\mid \mathit{D}=0\mathrm{)}}}\mathrm{,}\end{array}$$

captures the prevalence of exposure E among individuals with and without the outcome D. Although numerically related through Bayes’ rule under certain conditions, these two relative risks represent fundamentally different epidemiologic measures and must not be conflated.

Cornfield’s Inequality 1. If the exposure E and the unmeasured confounder U are positively associated, then the prevalence ratio of U across levels of E,

$$\begin{array}{c}{\text{RR}}_{\mathit{E}\mathit{U}}={\displaystyle \frac{\mathit{P}\mathrm{(}\mathit{U}=1\mid \mathit{E}=1\mathrm{)}}{\mathit{P}\mathrm{(}\mathit{U}=1\mid \mathit{E}=0\mathrm{)}}}\mathrm{,}\end{array}$$

must satisfy

$$\begin{array}{c}{\text{RR}}_{\mathit{E}\mathit{U}}\ge {\text{RR}}_{\mathit{E}\mathit{D}}.\end{array}$$

In many applications, a more scientifically meaningful quantity is the strength of the association between the unmeasured confounder and the outcome. This is captured by

$$\begin{array}{c}{\text{RR}}_{\mathit{U}\mathit{D}}={\displaystyle \frac{\mathit{P}\mathrm{(}\mathit{D}=1\mid \mathit{U}=1\mathrm{)}}{\mathit{P}\mathrm{(}\mathit{D}=1\mid \mathit{U}=0\mathrm{)}}}.\end{array}$$

Cornfield and colleagues showed that this quantity must also be large in order to explain away the observed exposure–outcome association. Their second inequality formalizes this requirement.

Cornfield’s Inequality 2. Among the unexposed ($\mathit{E}=0$), the relative risk relating U to D must satisfy

$$\begin{array}{c}{\text{RR}}_{\mathit{U}\mathit{D}\mid \mathit{E}=0}\ge {\text{RR}}_{\mathit{E}\mathit{D}}.\end{array}$$

The elegance of these inequalities lies in their simplicity and interpretability: both follow directly from elementary algebra and probability, yet they yield powerful constraints on how strong an unmeasured confounder must be to account for an observed association. Their clarity and accessibility inspired subsequent statisticians to extend these results to more general settings and to apply them across diverse scientific domains.

2.2. Extensions by Successors

Several subsequent works have refined Cornfield’s inequalities by introducing additional assumptions that sharpen the bounds or lead to more practically useful applications.

A common strengthening begins by assuming conditional independence of the exposure and the outcome given the unmeasured confounder, $\mathit{E}⫫\mathit{D}\mid \mathit{U}$. This assumption states that once U is taken into account, the exposure no longer affects the outcome; equivalently, U fully explains the observed association between E and D. Under this condition, one obtains the bound

$$\begin{array}{c}{\text{RR}}_{\mathit{U}\mathit{D}}\ge {\text{RR}}_{\mathit{E}\mathit{D}}\mathrm{,}\end{array}$$

highlighted in Leppälä (2023) in the context of extending Cornfield-type reasoning to sensitivity analysis of odds ratios.

A further refinement follows by assuming that the confounder is more prevalent among the exposed than among the unexposed, i.e., ${\text{RR}}_{\mathit{E}\mathit{U}}\mathrm{>}1$. This condition implies that

$$\begin{array}{c}\mathit{P}\mathrm{(}\mathit{U}=1\mid \mathit{E}=1\mathrm{)}\mathrm{>}\mathit{P}\mathrm{(}\mathit{U}=1\mid \mathit{E}=0\mathrm{)}\mathrm{,}\end{array}$$

so the unmeasured confounder is positively associated with the exposure. Combining this with Cornfield’s inequalities yields

$$\begin{array}{rl}{\text{RR}}_{\mathit{E}\mathit{U}}\times {\text{RR}}_{\mathit{U}\mathit{D}}& \ge {\text{RR}}_{\mathit{E}\mathit{D}}^{\text{obs}}.\end{array}$$

Rearranging provides a lower bound for the confounder–outcome association:

$$\begin{array}{r}{\text{RR}}_{\mathit{U}\mathit{D}}\ge {\displaystyle \frac{{\text{RR}}_{\mathit{E}\mathit{D}}^{\text{obs}}}{{\text{RR}}_{\mathit{E}\mathit{U}}}}\mathrm{,}\end{array}$$

where ${\text{RR}}_{\mathit{E}\mathit{D}}^{\text{obs}}$ denotes the observed relative risk. By varying the magnitude of ${\text{RR}}_{\mathit{E}\mathit{U}}$ from 1 upward, one obtains a range of possible values for ${\text{RR}}_{\mathit{U}\mathit{D}}$ that could explain the observed association. Costa (2025), for example, used this inequality within a Monte Carlo sensitivity analysis to evaluate whether unmeasured confounding alone could plausibly explain the observed effectiveness of COVID-19 vaccines.

A major conceptual advance came from VanderWeele and Ding, who sharpened the lower bound on the confounding strength ${\text{RR}}_{\mathit{U}\mathit{D}}$ Ding and VanderWeele (2016). Specifically, they established the inequality

$$\begin{array}{r}{\text{RR}}_{\mathit{U}\mathit{D}}\ge {\displaystyle \frac{{\text{RR}}_{\mathit{E}\mathit{U}}\times {\text{RR}}_{\mathit{E}\mathit{D}}^{\text{obs}}-{\text{RR}}_{\mathit{E}\mathit{D}}^{\text{obs}}}{{\text{RR}}_{\mathit{E}\mathit{U}}\times {\text{RR}}_{\mathit{E}\mathit{D}}^{\text{true}}-{\text{RR}}_{\mathit{E}\mathit{D}}^{\text{obs}}}}\mathrm{,}\end{array}$$

where ${\text{RR}}_{\mathit{E}\mathit{D}}^{\text{true}}$ denotes the true (confounding-adjusted) relative risk. This formulation laid the groundwork for the development of the E-value, which quantifies the minimum strength of unmeasured confounding required to reduce the true exposure–outcome relative risk to 1 (VanderWeele and Ding, 2017). Setting ${\text{RR}}_{\mathit{E}\mathit{D}}^{\text{true}}=1$ and adding an assumption that ${\text{RR}}_{\mathit{E}\mathit{U}}={\text{RR}}_{\mathit{U}\mathit{D}}$ yields the now-standard expression:

$$\begin{array}{r}{\text{RR}}_{\mathit{U}\mathit{D}}\ge {\text{RR}}_{\mathit{E}\mathit{D}}^{\text{obs}}+\sqrt{{\text{RR}}_{\mathit{E}\mathit{D}}^{\text{obs}}\mathrm{(}{\text{RR}}_{\mathit{E}\mathit{D}}^{\text{obs}}-1\mathrm{)}}=\text{E-value}.\end{array}$$

The E-value thus provides a convenient, interpretable, and widely applicable measure of the minimum confounding strength needed to explain an observed effect, representing a modern extension of Cornfield’s original intuition.

3. Cornfield’s Reverse Reasoning in Sensitivity Analysis

In the 1950s, mainstream statistical practice emphasized error control through confidence intervals and hypothesis testing, alongside point estimation. Methodological innovation largely focused on reducing sampling error, improving prediction accuracy, and increasing statistical significance–all directed toward strengthening scientific conclusions. Against this prevailing mindset, Cornfield and colleagues introduced a strikingly different idea: rather than asking how strongly the data support a conclusion, they asked under what conditions the observed association could be overturned. This reverse-reasoning perspective–evaluating what magnitude of unmeasured confounding would be required to explain away an effect–became a conceptual cornerstone of modern sensitivity analysis. Although the term sensitivity analysis has since expanded far beyond Cornfield’s original formulation, the underlying spirit remains intact. Whether framed through bias functions, Rosenbaum-type models or E-values, the essential question is the same: under what plausible departures from ideal conditions would a causal conclusion no longer hold? This way of thinking continues to guide applied researchers in assessing the robustness of their findings.

Cornfield’s work also played a critical role in raising awareness of unmeasured confounding among applied medical researchers. Even when investigators could not directly address this limitation, his insight encouraged them to acknowledge potential unmeasured confounders as a meaningful source of uncertainty. As a result, discussions of residual confounding and calls for future research became routine in epidemiologic reporting for decades.

The influence of Cornfield’s inequalities has been profound. Many statisticians and applied researchers were inspired by the clarity and accessibility of his logic, prompting a range of methodological extensions to accommodate diverse research questions. Greenland (1996) provided a foundational synthesis of methods for sensitivity analysis of biases. Leppälä (2023) developed further refinements for sensitivity analysis of the odds ratio. Zhao, Small, and Bhattacharya (2019) extended Rosenbaum’s sensitivity framework using inverse probability weighting estimators, thereby integrating classical ideas with modern causal inference tools.

Applications of Cornfield’s inequality have appeared across a breadth of substantive research areas. Huesch (2013) applied Cornfield’s method to external adjustment problems. Costa (2025) used the inequality to quantify the limits of unmeasured confounding in applied settings. Hauben and Hung (2013) conducted an exploratory aggregate analysis of pneumothorax as a potential adverse drug event using Cornfield-type calculations. Chang et al. (2018) incorporated Cornfield-based sensitivity analysis in a visual assessment of relative risks for venous thrombosis, pulmonary embolism, and peripheral artery disease. These diverse examples demonstrate how the reverse-reasoning framework originally articulated by Cornfield continues to shape both methodological research and applied investigation.

4. Enduring Influence on Modern Causal Inference

Sensitivity analysis has become deeply embedded across a wide spectrum of scientific disciplines. In medical and pharmaceutical research, it is now common to see systematic reviews emphasizing the importance of sensitivity analyses in pharmacoepidemiology (Latour et al., 2025). Its influence extends even further; fields as diverse as criminal justice research have incorporated sensitivity analysis frameworks to interrogate the robustness of observational findings (Bingenheimer, Brennan, and Earls, 2005). This broad adoption reflects how central Cornfield’s original insight has become to empirical research that relies on non-experimental data.

Cornfield’s contribution imparted a lasting methodological mindset: associations should be treated with caution, and causal interpretations should be scrutinized for their vulnerability to alternative explanations. In many real-world settings, it is unlikely that a single risk factor directly and exclusively causes an outcome. Sensitivity analysis provides a structured way to ask whether a stronger, unmeasured factor could plausibly overturn a reported effect. This encourages researchers not only to quantify robustness but also to cultivate a habit of scientific skepticism–one of Cornfield’s enduring intellectual legacies.

Despite its recognized importance, the implementation of sensitivity analysis in applied research remains uneven. Many contemporary study protocols in public health and clinical sciences still do not require explicit sections devoted to sensitivity analysis, even for observational studies in which unmeasured confounding is a central threat. Ideally, sensitivity analysis should not be an optional addendum at the end of an analysis but rather a planned component of study design, integrated from the earliest stages. Treating sensitivity analysis as a routine step—much like specifying primary endpoints or confounder adjustment strategies—would further strengthen the credibility of observational evidence.

Cornfield’s reverse-reasoning framework thus continues to shape modern causal inference both conceptually and practically. Its influence is evident in the methodological tools widely used today and in the increasing expectation that researchers justify the robustness of their findings. As empirical science continues to rely heavily on observational data, Cornfield’s perspective remains not only relevant but indispensable.