Louis Anthony “Tony” Cox: Drilling Into Politically Sensitive Risk Issues

Louis Anthony Cox, or "Tony" to members of SRA, is best known as Editor-in-Chief of the journal Risk Analysis since 2013, and as an Area Editor for mathematical modeling prior to that, from 2008 to 2013. Dr. Cox holds a PhD in Risk Analysis (1986) and an SM in Operations Research (1985), both from MIT; an AB from Harvard University (1978); and is a graduate of the Stanford Executive Program (1993). A long-time resident of Denver, Colorado, Tony has taught in the departments of mathematics, computer science, bioinformatics, and business analytics at the University of Colorado over the past 40 years. In 2020, he helped modernize the curriculum in Business Analytics to include new courses on Causal Analytics, Decision Analysis, Evaluative Analytics, Game Theory, and Artificial Intelligence for Business. Tony Cox is President of Cox Associates, a Denver-based company focusing on public and occupational health, safety, and environmental risk analysis; epidemiology; policy analytics; data science; and operations research. Since 1986, Cox Associates' analysts and scientists have applied epidemiological, risk analysis, and operations research models to measurably improve health and environmental risk assessment and decision-making for public and private sector clients. In 2006, Cox Associates was inducted into the Edelman Academy of the Institute for Operations Research and Management Science (INFORMS), recognizing outstanding real-world achievements in the practice of operations research and the management sciences. In 2012, Dr. Cox was elected to the National Academy of Engineering (NAE), "for applications of operations research and risk analysis to significant national problems." Along with his EIC role with Risk Analysis, Dr. Cox is Area Editor for real-world applications for the Journal of Heuristics. He has a long list of awards for research accomplishments, including multiple best paper awards, and he has over a dozen US patents on applications of artificial intelligence, signal processing, statistics, and operations research. We have deliberately focused this profile on several of his efforts to apply risk and decision science thinking and tools to politically sensitive issues. Cox's efforts to better understand these challenges have been welcomed by some and criticized by others. Our role is to try to present his views on these cases and thereby illustrate why risk analysis is such an interesting and fruitful endeavor. I have been fascinated by the possibilities and limitations of rational decision-making ever since. When I started my undergraduate work, I wanted to be a mathematical economist. I had taken an inspiring introductory economics course in high school and had come away with the impression that science and technology were more advanced than our understanding of how to use them wisely together to promote human flourishing. I hoped to apply mathematically rigorous thinking about decision-making and optimization to practical problems of cooperation, collective choice, productivity, and growth. I became interested in building causal models that would clarify the probabilities of consequences for different decision options and help to make clear and explicit where rational beliefs about probabilities came from based on available – usually incomplete – understanding of relevant causal mechanisms and processes. It seemed to me that decision analysis and game theory models usually just assumed some relationship between choices and outcome probabilities without justifying the assumed causal models. I wanted to understand how those models could be justified, so that even people with different subjective preferences for outcomes could agree on the objective factual aspects of a decision situation by agreeing on the outcome probabilities implied by different choices. I thought that decision analysis and game theory were seriously incomplete. Looking back from my perspective now, I would say that my college self was thirsting for a discipline of risk analysis that would help to determine the objective causal relationships among events, choices, and consequences because that seemed like an important hole in the theories of rational decision-making that I was studying. This course was a revelation to me. It complimented the much simpler perspectives on decision optimization that I was used to in decision analysis and game theory and got me thinking about the time consistency of beliefs, goals, and policies. None of the possible alternative non-causal explanations was consistent with the data, so by the process of elimination, we concluded that the suppression effect was a real causal effect. Charles made me co-author of his book, Beyond Probation: Juvenile Corrections and the Chronic Delinquent which was featured on The Today Show in 1979. I was 22, and this was my first book. It was a real thrill, of course, but it also introduced me to the challenges of causal analysis of data, as opposed to simply creating mathematical causal models. I've been fascinated by causal inference and modeling with observational data ever since. I consider such causal analysis to be fundamental for a lot of applied health risk assessment and epidemiology. I worked on an EPRI [Electric Power Research Institute] project quantifying the public and occupational health and safety risks from the same amount of energy produced using different fuel cycles. That led to my first solo publication, in the new journal Risk Analysis: "Artifactual Uncertainty in Risk Analysis" in 1982 (Cox 1982). That work won an award from ADL and gave me a chance to make a presentation to Chauncey Starr, then the president of EPRI. He really liked the work and encouraged my interest in risk analysis. I was hooked by this new field. One of the EPRI project managers, Paolo Ricci, became a lifelong friend and frequent collaborator in the early days of risk analysis. My boss at ADL, Joseph Fiksel, helped cofound the New England Chapter of the new Society for Risk Analysis, of which I became Secretary, and encouraged me to learn all I could about risk analysis and to work on all sorts of risk analysis projects at ADL. My Master's thesis on "Attribution of Risk in the Presence of Joint Causes" applied game-theoretic axioms and theorems to the problem of defining the fraction of blame for a jointly caused adverse outcome that should be attributed to each of the factors that increased its probability. This thesis, supervised by my kind, deeply insightful mentor and advisor Al Drake, who was the co-director of the Operations Research Center at MIT at the time, led to some papers in the journals Management Science and Risk Analysis that ended up being quite influential in the part of mathematical, epidemiology that deals with joint causation (Cox 1984, 1985). My 1984 paper in Risk Analysis on "Probability of Causation and the Attributable Proportion Risk" is still one of my favorites. Then in each set of activities that could lead to success, one should first try the available activity with the highest failure probability per unit cost so as to fail as early as possible if the set of activities could not all be completed. On the other hand, if there were also multiple different ways, or different sets of activities, whose completion would achieve success, then the set of activities that collectively had the highest probability of success per unit cost should be tried first. This basic idea had many applications – for example, to optimal sequential testing and repair in reliability systems (Cox and Qiu 1994; Cox, Chiu, et al. 1996) and to pursuit of goals in AI systems with uncertainty about which tasks can be completed, or with the ability to collect information or get questions answered before deciding what to do (Cox 1990a). Another application was figuring out how an intelligent attacker might try to compromise a target asset – either a physical asset or an information asset – through an attack consisting of a sequence of activities with different costs and success probabilities. Tony led an artificial intelligence project for ADL with the Air Force Computer Security Program Office, developing an "Automated Threat Assessment Methodology" (ATAM) that applied the same basic ideas as the R&D planning project to calculate the optimal attack strategies of intelligent attackers and to optimize defenses against them (Cox 1990b). I was annoyed by the widespread use of simplistic risk scoring methods and risk matrices that were clearly inconsistent with how a rational person or AI should or would do things That led me to write papers trying to clarify exactly what was wrong with such approaches and what ideas we could adopt from decision analysis, AI, and operations research to improve the modeling of optimal intelligent attacks and defenses against them. One of my favorites is "What's wrong with risk matrices?" which was published in Risk Analysis in 2008 (Cox 2008). This type of thinking became influential with a NAS committee I was on that focused on Methodological Improvements to the Department of Homeland Security's Biological Agent Risk Analysis (National Academies of Sciences, Engineering, and Medicine 2007) (See https://nap.nationalacademies.org/read/11836/chapter/2#2). This work helped to ignite a lot of game-theoretic and multi-level modeling by Vicki Bier, Seth Guikema, Jun Zhuang, and their students and many others that I think is now widely considered foundational for adversarial risk assessment. I have known Tony Cox for more than 35 years, since we studied with the same advisor in graduate school. Tony is perhaps the closest to a true polymath I have had the pleasure of knowing. I consider myself widely read, but am constantly astounded at the contributions Tony is able to make in integrating and applying the latest research results from a variety of fields, ranging from technical methods of correcting for multiple comparisons in statistics, to synthesizing the vast and complex literature on behavioral game theory, to what is known about the effects of diesel particulates on human health. Many practitioners (and even researchers) tend to stick to tried and true techniques, continuing to return to basic models and methods even when later developments in the field might be more appropriate for the problem at hand. By contrast, Tony is not only a voracious reader, keeping track of the latest developments in fields ranging from behavioral game theory to theories of carcinogenesis; he is also insightful in understanding where new methods or results could be useful in applications, and fearless in piloting such methods even in fields where they have not yet been applied. I have regularly coauthored with Tony (e.g., a book chapter on probabilistic risk analysis in my 2017 book on Risk in Extreme Environments), largely because I know he will inspire me to address interesting innovations and applications in a wider variety of fields than I might otherwise consider. The word "causality" is commonly applied to several distinct concepts, including attributing effects to factors with which they are associated; identifying the mechanisms by which changes in exposures change response probabilities; and predicting how outcome probabilities would change if we changed what we are doing. It is that third concept, sometimes called interventional causality (or, by some philosophers, manipulative causation) that I think is most relevant for most decision analysis and risk management applications. To me, data that provides no evidence that changing what we do changes outcome probabilities provides no evidence of a detectable interventional causal relationship between them. For example, if reducing high ambient levels of particulate air pollution in Ireland by more than 50% is observed to have no detectable effect on mortality rates over the following decades compared to areas that did not have that reduction, then these data provide no rational basis for a belief or prediction that reducing air pollution will create significant reductions in mortality rates (Cox 2024). The associations between air pollution and mortality may still satisfy other criteria for various types of causality, such as Bradford-Hill considerations for associative or attributive causation, but in my opinion, what matters for figuring out how to take causally effective actions to improve public health is just interventional causation. On the positive side, if statistical tests reject the null hypothesis of conditional independence between outcomes and actions, then there is reason to hope that changing outcomes can indeed change outcome probabilities. Technical methods such as Partial Dependence Plots (PDPs) and Individual Conditional Expectation (ICE) plots can then help to quantify the dependence of outcome probabilities and expected values on controllable inputs if all other inputs are held fixed (Cox 2023b). This opens the door to quantitatively informed decisions about what to do to make preferred outcomes, such as longer and healthier lives, more likely. This seems to me to be a much more promising approach to predicting and estimating effects of interventions than "causal" approaches based on subjective judgements about statistical associations. Developing mathematical models that could reproduce the main quantitative features of observed exposure-response data, such as their relative insensitivity to exposure concentrations above a certain level, and absence of an apparent increase in risk a sufficiently low concentrations, required considering molecular biological mechanisms in enough detail to identify major feedback loops and identify their mathematical stability properties (Cox 2011a). This could explain apparent epidemiological and experimental puzzles about why repeated high concentrations could be disproportionately dangerous, while total duration of exposure often mattered less. The key insight was to think of risk (e.g., transformed stem cells) as accumulating at a relatively high rate once chronic inflammation sets in, but at a much lower rate before that. The timing of the transition between these two regimes then drives lifetime risks, with the other details of exposure histories being much less important. I was thrilled when the NLRP3 inflammasome was discovered and the molecular biological details of its activation threshold and roles in the onset of chronic inflammation were elucidated, as these causal biological mechanisms seemed to me to fit what the mathematical modeling required of the biology and implied about the dynamics of exposure-response relationships (Cox 2019b, 2018). This dynamic understanding of dose and response is consistent with experimental exposure-response data in animals exposed to volatile organic chemicals such as isoprene or benzene. I had the opportunity to design some exposure-response experiments for these chemicals in the 1990s, and it was very clear that cumulative exposure was an inadequate basis for predicting biological responses, since time patterns of exposure played such a crucial role that larger cumulative exposures could easily create lower risks than smaller cumulative exposures if they were more spread out over time (Cox, Bird, et al. 1996). It seems to me now that understanding threshold-like nonlinearities in the dynamic responses of cells to toxic insults, including the activation thresholds for NLRP3 inflammasomes, helps to make sense of a great deal of observational data, and that it implies that our causal models of exposure-response relationships should be updated to take into account the importance of time patterns of high-concentration exposures. One of my most important and controversial advisory roles was as Chair of the US EPA Clean Air Scientific Advisory Committee (CASAC) from 2018 to 2020. The CASAC provides independent scientific advice to the EPA on National Ambient Air Quality Standards (NAAQS), reviewing the latest research on air pollutants and their health effects. It independently evaluates the strength of evidence, critiques EPA risk assessments, and makes recommendations on whether NAAQS should be revised, retained, or updated based on new scientific evidence. In that role, I argued that the study of air pollution health risks should adhere to the same scientific principles used in other areas of science, rather than relying on weight-of-evidence approaches advocated by the EPA and other CASACs that emphasize consensus expert judgments over empirically falsifiable and testable causal claims that can be independently tested and verified using the abundant data that have been collected in the United States and elsewhere over the past 50 years. I noted that many influential current studies on air pollution and mortality fail to meet basic scientific standards of sound causal inference such as fully controlling for key measured confounders before interpreting associations causally. I pushed for doing interventional causal analyses with available data to see whether such interpretations were consistent with the data (Cox 2019c). This was a very unpopular and controversial position. It led the Los Angeles Times and others to brand me as a fringe scientist and ideologue. However, I have long felt that the best response to such characterizations is to continue to try to do sound and useful work, hoping that if it is sound and useful enough to give demonstrably more accurate predictions, it will eventually be adopted. I have also served on many advisory committees for the US EPA and other agencies and have been involved in advising international and national authorities about policies for microbial hazards and food safety. Almost all of these committees have been a lot of fun. I have learned a lot from hearing about how different people think and deliberate about assessing and managing a variety of hazards. I love being an expert witness because I find that it is an excellent discipline for forcing precise and simple expression of important principles. I find myself explaining probability, statistics, and risk analysis concepts to people without specialized training but with a lot of practical interest in understanding what can be concluded and why. I think the discipline of communicating key ideas clearly to nonspecialists is extremely helpful in understanding what really matters in our field and why it is important and useful to people who need trustworthy conclusions and understandable explanations. Not dismissing the historic models and methods used to estimate dose–response relationships, Tony calls for building models that contain mechanistic information. Tony and I have spent considerable time together trying to dissuade others from too-simplistic probabilistic risk analysis based on unknowable probabilities that puts our nation at risk. In many cases, he takes to task established assessments using "currently accepted methods" and (with good grace) devastates these. Our most important work may have been classified, as was most of the scary bioterror data we analyzed together. Tony has a deep-seated sense of professional propriety. Tony refers with justified familial pride to "Uncle Archie," Archibald Cox, who as a US special prosecutor was fired by President Nixon as part of the infamous "Saturday Night Massacre" for refusing to keep damning Nixon recordings secret. In my experience with Tony over the years, honor is all. AI and machine learning (AI-ML) methods have also done a lot in the past 20 years to provide practical computational tools for identifying individual-level causal relationships from data. This has been one of my recent research interests. AI-ML has also been transforming the theory of practice of applied optimal decision-making under uncertainty, providing new algorithms and methods for optimizing exploration-exploitation trade-offs and learning how to manage complex uncertain systems (Cox 2021). The journal Risk Analysis recently pioneered an experimental AI system that gives opinions on whether draft papers or proposals are likely to be good fits to the journal based on the guidance we give authors, and that also explains its reasoning to authors to help them craft papers that will be as useful as possible to readers (See the tool at http://screener.riskanalysis.cloud). In addition to his research and advisory work, Tony Cox holds over a dozen US patents in areas spanning artificial intelligence, voice technology, and network optimization. Several of these pioneering inventions from the 1990s anticipated today's voice-based AI assistants by decades. For example, his patent on "a linguistically competent dialogue with a computerized service representative" described a system for helping users schedule appointments or get service information via natural voice interaction—years before Siri or Alexa existed. Other patents applied AI, optimization, and risk analysis to telecommunication network design, enabling regional phone carriers to deliver the same quality of service at far lower cost. One such system saved US WEST—then serving 14 states—over $100 million by optimizing network planning and installation. These patents reflect Cox's broader interest in practical, data-informed solutions that help systems work more efficiently and responsively, especially when uncertainty and complexity would otherwise make good decision-making difficult. Risk analysts can and do draw on methods of economics, econometrics, and political economy; decision analysis, decision science, and decision optimization; applied probability and statistics; causal modeling; risk psychology; AI-ML; reliability engineering; toxicology, biostatistics, and bioinformatics; and public health and epidemiology – basically, any tools that might be useful in addressing questions of risk assessment and risk perception, risk management, and risk communication and governance. So, I would advise people coming into the field to pay most attention to the fascinating and challenging questions that need to be answered, and to remain very flexible, open-minded, and willing to learn any tools that can help to answer them. Since the early 1990s, I have felt strongly that risk analysis involves some of the most interesting and important questions in our modern world – both scientific questions, such as how exposures affect health; and societal ones, such as how analysis and deliberation can and should be used to improve collective decisions in a modern democratic society. I think we in the field are blessed to be able to draw on the profound insights and methods of many other disciplines to better understand how decisions and interventions affect outcome probabilities in an uncertain world – and then helping to explain and justify those recommendations. Finally, we return to Tony Cox's role as EIC of Risk Analysis for more than a decade, role that allows him to help steer the field toward the ambitious goals he describes in the previous paragraph. Warner North (Greenberg and Lowrie 2017) comments on Dr. Cox's role as editor. Tony Cox has done a remarkable job piloting the Society's main journal, and I have been privileged to be his wingman for decision sciences. Other areas have more linkage to a specific discipline, while decision sciences cover theory and applications spread over a multitude of technical disciplines, some relatively uncharted territory for our Society. I've enjoyed the challenges Tony has given me, and I've enjoyed even more the highly collegial relationship we have maintained over many manuscripts, including oddities that Tony considered potential fits for Risk Analysis and sent to me for a decision. Tony continually strives to maintain and enhance both the quality and impact of Risk Analysis, which he affectionately calls "our favorite journal." He has a wise and gentle touch interacting not just with editors, but reviewers and authors, making them feel listened to and supported, and always demonstrating sound, logical reasoning in his management decisions, sprinkled with kindness. He treats the journal as his pride and joy, and he has been instrumental in keeping the journal at the forefront of scientific literature in the risk field. We leave final words to Arnold Barnett and Seth Guikema. Thinking about what Tony Cox means to risk analysis, I found myself realizing that Tony Cox actually is risk analysis now, to the extent that any one person can embody a discipline. I first met Tony in my MIT probability course, where he consistently asked deep questions and had great insights about how to answer them. Part of what makes Tony great is what he does not do. In fields like mathematics or economics, antagonism often develops between theoreticians and applied participants: the theorists see the applied people as intellectually shallow, while the applied people see the theorists as pointlessly abstract. But Tony simply transcends this divide: he knows that both groups must succeed for risk analysis to thrive, and he contributes mightily via his research and his example to each one. There is Tony's uncanny ability to express his ideas with great clarity, which goes a long way to explaining why he is such an "influencer" on urgent matters of public policy. Seth Guikema is a Professor in the Departments of Civil and Environmental Engineering and Industrial & Operations Engineering at the University of Michigan. Seth served as President of SRA in 2020, and is Area Editor for natural hazards and risk analysis (formerly mathematical modeling) for Risk Analysis: I consider Tony a role model of how to think deeply about risk, and I am honored to have him as a colleague. Tony is truly one of the giants of the field of risk analysis. His deep thinking across so many areas of the field, from fundamental to applied, has been critical to advancing the field, and his ability to ask concise, insightful questions pushes us all to be deeper thinkers. Cox, L. A., Jr. 1985. "A New Measure of Attributable Risk for Public Health Applications." Management Science 31, no. 7: 800–813. https://doi.org/10.1287/mnsc.31.7.800. Cox, L. A., Jr. 2001. Risk Analysis: Foundations, Models and Methods. International Series in Operations Research & Management Science, Vol. 45. Springer. Cox, L. A., Jr. 2008. ʻʻWhat's Wrong With Risk Matrices? Risk Analysis 28, no. 2: 497–512. https://doi.org/10.1111/j.1539-6924.2008.01030.x. Cox, L. A., Jr. 2011. "A Causal Model of Chronic Obstructive Pulmonary Disease (COPD) Risk." Risk Analysis 31, no. 1: 38–62. https://doi.org/10.1111/j.1539-6924.2010.01487.x. Cox, L. A., Jr. 2012. "Confronting Deep Uncertainties in Risk Analysis." Risk Analysis 32, no. 10: 1607–1629. https://doi.org/10.1111/j.1539-6924.2012.01792.x. Cox, L. A., Jr. 2013. Improving Risk Analysis. Springer. Cox, L. A., Jr. 2020. "Implications of Nonlinearity, Confounding, and Interactions for Estimating Exposure Concentration-Response Functions in Quantitative Risk Analysis." Environmental Research 187: 109638. https://doi.org/10.1016/j.envres.2020.109638. Cox, L. A., Jr. 2020. "Answerable and Unanswerable Questions in Risk Analysis With Open-World Novelty." Risk Analysis 40, no. S1: 2144-2177. https://doi.org/10.1111/risa.13553. Cox, L. A., Jr. 2021. "Thinking About Causation: A Thought Experiment With Dominos." Global Epidemiology 3: 100064. https://doi.org/10.1016/j.gloepi.2021.100064. Cox, L. A., Jr. 2024. "Objective Causal Predictions From Observational Data." Critical Reviews in Toxicology 54, no. 10: 895–924. https://doi.org/10.1080/10408444.2024.2399856.