Causal Reasoning and the Ethical Foundation of Good Policy Decisions
Public policy has many so-called wicked problems, situations with no single right answer, competing stakeholders, and outcomes that ripple across time and populations. Decision intelligence (DI) is an emerging approach to these problems that combines data science, behavioral science, and management theory to help institutions make better decisions, especially repeated decisions.1 One of DI's most important contributions is that it insists that good decisions rest on causal understanding, not just statistical association. I assert that this insistence is what makes a decision-making process ethical, and below I outline the conditions required to support this claim.
Identification and Estimation
To understand why (not just how) an intervention affects an outcome, we actually need two separate things, and it frequently happens that one exists without the other.
Identification is the process by which we know why a quantity we can compute from data actually causes the outcome we care about. We might compute the difference in outcomes between a treated group and an untreated group with similar characteristics, but does that difference signify that the treatment caused the outcomes, or does it just reflect a preexisting difference between the two groups? A randomized controlled trial (RCT) solves this problem by design (assuming we control for the correct variables). However, RCTs are often impossible in policy settings: we can't randomly assign which frogs live in polluted ponds or which cattle herds face which predators. In these situations, identification has to come from causal inference, building a model (often visualized as a causal diagram, a map of which variables influence which) of the full network of factors involved so we know exactly what needs to be statistically conditioned.2,3
Estimation is a separate task of using a defined causal structure to calculate outcomes accurately, including how outcomes differ across subgroups or individuals, especially when relationships are nonlinear or the data is high-dimensional. This is where machine learning shines.
The crucial point is that associative ML methods (neural networks, random forests, gradient boosting, and so on) are excellent at estimation but do nothing to solve identification. A model that is fit to associate covariates with outcomes can pick up spurious correlation from confounding just as easily as it can pick up a real, causal difference in how a treatment affects different people.4 Statistical sophistication is not a substitute for causal reasoning. This is another way to think about "garbage in, garbage out" when referring to machine learning model predictions. Decision intelligence, because it intentionally focuses first on identification and then on which ML models may provide robust estimation within a causal system, addresses both parts.
Similar Approaches
While DI presents one solution to ensure that we can understand how outcomes are driven by treatments, it is not the only approach that can work. Structured decision making (SDM) is a widely used alternative in conservation and public policy, and it sharpens what causal reasoning is and isn't doing.5 In its simplest form, SDM aggregates stakeholder objectives to choose among alternatives, but doesn't itself generate causal understanding. Instead, it consumes predicted consequences rather than estimating them. In more advanced forms, however, SDM builds on influence diagrams6 (the extension of a causal diagram to add nodes for choices and stakeholder values), which generate exactly the causal structure needed to estimate differing effects of a treatment on sub-populations. Sophisticated versions of SDM therefore address both identification and estimation. The useful additional point that SDM makes is that knowing how a policy's effect differs across sub-populations is a separate question from whose benefit gets weighed against whose harm. Causal accuracy tells us the size and direction of an effect on a given group; it doesn't tell us how to trade one group's outcome against another. That doesn't change the core requirement, though: whatever decision framework a policymaker uses, correctly identifying and estimating how an effect varies across individuals and sub-populations remains the step that separates an ethical decision from an average-case estimate.
The Ecological Fallacy
But why is causal inference crucial to ethical decision-making? The ecological fallacy points to this answer. The ecological fallacy is the mistaken assumption that an effect measured on average across a group applies equally to every member of that group (or, conversely, that an effect observed in a few individuals will scale up to the whole population). 7 Picture a pond full of frogs: a treatment that helps one frog might not help all frogs, because frogs differ: sex, age, sun exposure, or prior exposure to pollutants. Apply a treatment across an entire pond based on one frog's result and you may harm frogs for whom the chemical interacts badly with higher UV exposure. The average effect and the individual effect are different quantities, and conflating them is the ecological fallacy.Now imagine deciding whether to kill a wolf blamed for recent cattle deaths. This is a life-or-death decision about one specific animal. Before that decision is made, we should require strong evidence that this wolf caused the deaths and that non-lethal interventions won't work as well, not just that wolves in general are responsible for livestock losses. Or, consider a healthcare analogy: should an insurer approve an expensive cancer drug for one patient? An average effect across a clinical trial population tells you the drug helps on average, it does not tell you it will help this patient, whose age, genetics, and comorbidities may put them far from that average. This is the basis of personalized medicine, an emerging discipline to address exactly this problem. Making high-stakes, individual decisions based only on group-level correlations risks avoidable harm to individuals with differing needs.
My assertion that accurate causal understanding is required for ethical decision-making therefore rests on two claims: first, that individual cases are detectably and meaningfully different from one another, and second that individual lives matter enough that those differences should change our decisions. The first claim is exactly what identification and estimation accomplish together: a well-built causal model can capture the confounders and mechanisms that make one wolf, one frog, or one patient different from the group average. This allows us to estimate a heterogeneous treatment effect, an effect size specific to that individual or sub-population rather than a single average across everyone. That is the technical machinery underneath the ethical claim.
The second claim is a moral one, and it's supported from various different directions across ethical theory. Rights theorists assert that there is an intrinsic value of being what Tom Regan calls a "subject-of-a-life," a status that doesn't depend on usefulness to others. 8 Utilitarians argue that capacity to suffer provides rationale for protecting individual lives, arguing that suffering counts against an act regardless of whose it is.9 Kantians, working from within a framework built around rational agency, argue that any being with a good of its own must still be treated as an end and not merely a means. 10 Feminist care ethics grounds the claim not in abstract principle but in an ethic of attentiveness to the vulnerability of particular others. 11 While I won't relitigate these debates here, it's worth noting that this claim isn't the property of any one school of thought: individuals matter and harming them without adequate justification is wrong.
The Limits
My argument above has certain technical limits, which are important to note:• Identification is often incomplete. Causal discovery methods rely on assumptions, like the absence of unmeasured confounders, that can't be tested from data alone. In genuinely wicked problems, we may never achieve full identification, no matter how much rigor we apply.
• Data can encode injustice. A causal model trained on historical data reflects the process that generated that data, including any bias in it (for example, historical enforcement patterns feeding a "causal" model of risk). A causal diagram and any ML models associated with it are only as fair as the data and the model choices behind them.
• Estimates carry uncertainty. Heterogeneous treatment effects have confidence intervals and are not certainties. Ethical decision-making requires being honest about that uncertainty, not just presenting a point estimate as fact.
• Rigor has a cost. Full causal identification takes time, expertise, and data that policymakers facing urgent decisions may not have. Insisting on perfect causal rigor before acting can itself cause harm through delay.
Conclusion
Considering everything above, the strongest version of a "causal models lead to ethical decision-making" assertion is that causal reasoning, honestly bounded by its own limits, is a necessary but not sufficient condition for ethical policy decisions that affect individuals or sub-populations. It must be paired with an explicit account of whose values are being weighed, transparency about uncertainty, and realism about when full identification simply isn't achievable. In service to this assertion, a causal diagram is a powerful tool for taking individuals seriously, particularly when combined with the harder, values-laden work of deciding what we owe humans and other sentient beings.References:
Pratt, L., Bisson, C. & Warin, T. Bringing advanced technology to strategic decision-making: The Decision Intelligence/Data Science (DI/DS) Integration framework. Futures 152, 103217 (2023).
Imbens, G. W. & Rubin, D. B. Causal Inference for Statistics, Social, and Biomedical Sciences: An Introduction. (Cambridge University Press, Cambridge, 2015). doi:10.1017/CBO9781139025751.
Pearl, J. Causality: Models, Reasoning and Inference. (Cambridge University Press, Cambridge New York, NY Port Melbourne New Delhi Singapore, 2022).
Athey, S. & Imbens, G. Recursive partitioning for heterogeneous causal effects. Proc. Natl. Acad. Sci. 113, 7353–7360 (2016).
Gregory, R. et al. Structured Decision Making: A Practical Guide to Environmental Management Choices. (Wiley-Blackwell, Chichester, West Sussex Hoboken, N.J, 2012).
Howard, R. A. & Matheson, J. E. Readings on the Principles and Applications of Decision Analysis. (Sdg Decision Systems, Menlo Park, Calif, 1989).
Robinson, W. S. Ecological Correlations and the Behavior of Individuals. Am. Sociol. Rev. 15, 351–357 (1950).
Regan, T. The Case for Animal Rights. (University of California Press, Berkeley, 1985).
Singer, P. Animal Liberation: A New Ethics for Our Treatment of Animals. (Distributed by Random House, New York, 1975).
Korsgaard, C. M. Fellow Creatures: Our Obligations to the Other Animals. (Oxford University Press, Oxford, 2018).
Donovan, J. Animal Rights and Feminist Theory. Signs J. Women Cult. Soc. 15, 350–375 (1990).
Image by Andrea Bohl from Pixabay