Causal machine learning is a family of methods that go beyond detecting patterns in data to estimating cause-and-effect relationships. Standard machine learning excels at prediction: given a patient’s medical history, what outcome is likely? Causal ML asks a harder question: if we change something, what happens next? That shift from prediction to intervention is what makes it valuable for decisions in medicine, business, and public policy, where acting on a mere correlation can backfire badly.
Why Standard Machine Learning Falls Short
A conventional deep learning model trained to classify images might learn to associate a grassy background with “cow” because most cow photos in the training data happen to have grass. The model performs well on similar photos but fails when it encounters a cow on a beach or in a barn. These background cues are spurious features: they correlate with the label in the training set but have nothing to do with what actually makes a cow a cow. Deep neural networks are prone to latching onto such features, and the problem worsens when models are deployed in safety-critical settings where the real-world data distribution differs from the training data.1PubMed Central. Beyond Distribution Shift: Spurious Features Through the Lens of Training Dynamics
The phenomenon is sometimes called the “Clever Hans” effect, named after a horse that appeared to solve math problems but was actually reading its handler’s body language. Modern models do something similar: they pick up on non-essential features of the input, like texture, background objects, or formatting artifacts, and treat those as if they were meaningful signals. When the data shifts even slightly, the model’s accuracy can collapse because the spurious correlation no longer holds.2arXiv. The Clever Hans Mirage: A Comprehensive Survey on Spurious Correlations in Machine Learning
This is not just an academic curiosity. A hiring algorithm might learn that applicants from certain zip codes tend to get hired and treat geography as a predictive feature, even though location has no causal link to job performance. A medical diagnostic tool might learn that the presence of a particular imaging device’s watermark correlates with disease prevalence, not because the device detects disease better, but because it is used more often in hospitals that see sicker patients. Causal ML’s core promise is to help models distinguish features that actually drive outcomes from features that merely happen to co-occur with them.
How Causal ML Thinks Differently
At the heart of causal ML is a simple but powerful idea: before running any algorithm, you need a model of how the world works. One of the most common tools for building that model is the directed acyclic graph, or DAG. A DAG is a diagram where arrows connect variables to show which ones are believed to influence which. If you think smoking causes lung cancer and also causes yellow teeth, the arrows run from smoking to both outcomes. Yellow teeth and lung cancer will be correlated in the data, but the DAG makes clear that teeth don’t cause cancer; they share a common cause.
DAGs are useful because they make hidden assumptions visible. They help researchers figure out which variables to account for and which to leave alone when estimating a causal effect. For example, adjusting for a common cause of treatment and outcome (a confounder) is usually necessary to get an unbiased estimate. But adjusting for the wrong variable can actually introduce bias where none existed. DAGs provide a formal way to reason through these decisions, using criteria like the “backdoor criterion” to identify which variables need adjustment.3PubMed Central. Directed acyclic graphs for clinical research: a tutorial When confounders are unmeasured, DAGs can also help researchers understand the direction and likely size of the resulting bias.4PubMed Central. Causal directed acyclic graphs and the direction of unmeasured confounding bias
A complementary approach comes from the potential outcomes framework. Instead of drawing diagrams, this framework asks: for a given patient, what would their outcome be if they received treatment, and what would it be if they did not? Only one of those outcomes is ever observed; the other is a counterfactual. The core challenge of causal inference is estimating that missing counterfactual. Both the DAG approach and the potential outcomes framework converge on the same goal: separating genuine causal effects from misleading associations.5Journal of Economic Literature. Potential Outcome and Directed Acyclic Graph Approaches to Causality: Relevance for Empirical Practice in Economics
Discovering Causal Structure From Data
Drawing a DAG requires domain knowledge: a clinician might know that age affects both treatment choice and disease progression, so they draw those arrows accordingly. But what happens when you don’t know the causal structure in advance? That is the problem causal discovery algorithms try to solve. These methods take observational data and attempt to reverse-engineer the causal relationships among variables, producing a candidate DAG or a set of plausible DAGs.6Journal of Data Science and Statistics. Causal Discovery for Observational Sciences Using Supervised Machine Learning
Causal discovery is harder than it sounds. From data alone, it is often impossible to pin down the direction of every arrow. If two variables are correlated and no third variable explains why, the data might be consistent with A causing B, B causing A, or both being caused by something unmeasured. Discovery algorithms can narrow the possibilities but rarely produce a single definitive answer without additional assumptions or experiments. Still, they are valuable for generating hypotheses, especially in complex systems with dozens or hundreds of interacting variables where human intuition struggles.
Causal Forests and Personalized Treatment Effects
One of the most practically useful tools in causal ML is the causal forest. In medicine and social science, researchers often want to know not just whether a treatment works on average, but for whom it works best. A drug might reduce blood pressure by five points on average across all patients, but it might work much better for older patients and barely at all for younger ones. Causal forests are designed to detect exactly this kind of variation, known as treatment effect heterogeneity, in a flexible way that doesn’t require the researcher to guess which subgroups matter in advance.7Research & Politics. Estimating and evaluating treatment effect heterogeneity: A causal forests approach
A case study applying causal forests to randomized trial data illustrates both the potential and the challenges. Researchers looking at patient outcomes found that the algorithm identified serum potassium as a key variable splitting the population into groups with markedly different treatment effects. Patients with potassium levels above a certain threshold showed a large negative effect from treatment, while those below it showed a modest benefit. The difference in outcomes between these subgroups was substantial enough to have clinical implications.8PubMed Central. Application of causal forests to randomised controlled trial data to identify heterogeneous treatment effects: a case study That kind of finding can guide decisions about who should receive a treatment and who might be harmed by it.
Healthcare and Precision Medicine
Healthcare is where the stakes of getting causation wrong are highest, and it is also where causal ML has attracted the most attention. The goal of precision medicine is to determine the best treatment for an individual patient rather than relying on population averages. Causal ML contributes by estimating individualized treatment effects: what would happen to this specific patient under treatment A versus treatment B?9PubMed Central. Causal machine learning for healthcare and precision medicine
These methods can work with both clinical trial data and real-world data from electronic health records and clinical registries. Real-world data is attractive because it is cheap and plentiful, and it reflects the messy diversity of actual patient populations. But because treatments in real-world data are not randomly assigned, confounding bias is always a concern. A patient who receives an aggressive treatment may be sicker to begin with, making the treatment look less effective than it actually is. Causal ML methods are specifically designed to handle this kind of bias, though they require careful attention to assumptions. Using them carelessly with observational data can produce biased or incorrect predictions.10PubMed. Causal machine learning for predicting treatment outcomes
Business and Policy Applications
Outside of healthcare, causal ML is gaining traction in marketing and customer management. A classic example is customer retention. Traditional “propensity” models predict which customers are likely to leave. But knowing who will churn is not the same as knowing who will respond to a retention offer. Some customers who are likely to churn will leave regardless of any intervention. Others would have stayed even without the offer, making any outreach a waste of money. The interesting group is the “persuadables”: customers who will stay only if contacted.
Uplift models, a type of causal ML, try to estimate the treatment effect of a retention intervention on each individual customer. Recent work applying meta-learner uplift models found that targeting retention efforts based on these models reduced churn more effectively than traditional propensity-based targeting.11International Journal of Market Research. Using Meta-Learners and Propensity Score Matching to Optimize Customer Retention The same logic applies to personalized pricing, ad targeting, and policy interventions like job training programs: the question is not just “what works” but “for whom does it work.”
Building Models That Generalize
One of the more ambitious promises of causal ML is building models that transfer well to new environments. A model trained on hospital data from one country should ideally work in another country where patient demographics differ. Standard ML often fails at this because it relies on statistical patterns that may be specific to the training data. Causal approaches try to identify the features that have a stable, invariant relationship with the outcome across different settings, and ignore the features whose relationship with the outcome is environment-specific.
The reasoning is that causes tend to produce their effects reliably across contexts, while spurious correlations tend to shift. Many domain generalization methods now employ causal reasoning to find these invariant features.12arXiv. Domain Generalization — A Causal Perspective Some researchers have developed methods that can provably recover the direct causes of the target variable and provide formal generalization guarantees even in nonlinear settings.13arXiv. Nonlinear Invariant Risk Minimization: A Causal Approach Others focus on handling shifts in the distribution of invariant features themselves, acknowledging that even “stable” features can change in prevalence across environments without breaking the causal link to the outcome.14TMLR. Weighted Risk Invariance: Domain Generalization under Invariant Feature Shift
Fairness and Explainability
Causal thinking also changes how we evaluate whether a model is fair. Standard fairness metrics check whether a model’s predictions differ across protected groups, such as by race or gender. But those metrics cannot distinguish between a model that is biased and one that reflects real differences in the underlying data. Causal methods can ask a sharper question: would this person’s prediction change if only their protected attribute were different, holding everything else causally upstream constant? If the answer is yes, the model has a fairness problem.
Counterfactual explanations are one practical tool here. For each prediction, you can ask what minimal change to the input would flip the outcome. If these explanations systematically involve different attributes for protected and unprotected groups, that signals the model is relying on proxies for the sensitive attribute, even if the attribute itself is not directly used as an input. This kind of implicit bias is hard to detect with traditional fairness audits but shows up clearly through causal analysis.15PubMed Central. PreCoF: counterfactual explanations for fairness
Where Causal ML Hits Its Limits
Causal ML is not a magic fix, and its limitations are important to understand. The most fundamental issue is that causal inference from observational data requires assumptions that cannot be fully tested. The two biggest assumptions are:
- No unmeasured confounding: All common causes of the treatment and the outcome are observed and accounted for. In practice, this is almost never guaranteed. If an important confounder is missing from the data, causal estimates will be biased, and the model has no way to tell you this from the data alone.
- Positivity: Every type of patient (defined by their characteristics) must have some chance of receiving each treatment. If certain patients never receive a particular treatment in the data, the model has nothing to learn from for that group. Violations of this assumption show up when, for example, a drug is contraindicated for patients with a specific condition, so no one in that subgroup ever receives it.
Positivity violations are more common than many practitioners realize, especially in high-dimensional data where combinations of characteristics can quickly produce subgroups with zero exposure to one treatment arm.16PubMed Central. Core concepts in pharmacoepidemiology: Violations of the positivity assumption in the causal analysis of observational data When models try to extrapolate into these regions of the data, their predictions become unreliable. Recent work on regularizing extrapolation highlights a difficult tradeoff: in regions with poor data overlap, you face competing sources of error from distributional imbalance, model misspecification, and statistical noise, and reducing one can increase the others.17PubMed Central. Regularizing Extrapolation in Causal Inference
Combining Experiments and Observational Data
One increasingly active area of research tries to get the best of both worlds by combining experimental and observational data. Randomized trials give unbiased causal estimates but are expensive, often small, and frequently conducted on narrow populations that don’t represent the broader public. Observational data from health records or insurance databases is cheap and covers real-world populations, but it suffers from confounding.18WIREs Computational Statistics. Methods for Combining Observational and Experimental Causal Estimates: A Review
Data fusion methods attempt to combine these sources in a principled way. For example, a trial might establish a causal effect in a narrow study population, and observational data can then be used to transport that estimate to a broader target population by adjusting for differences in patient characteristics between the trial participants and the population of interest.19PubMed Central. Causal Inference Methods for Combining Randomized Trials and Observational Studies: A Review The logic runs both directions: observational data can also refine trial-based estimates by providing information about effect modifiers that the trial was too small to detect.
Time Series and Temporal Challenges
Much of the causal ML literature deals with cross-sectional data: one snapshot of many individuals at one point in time. But many scientific and business questions involve time series, where the same system is measured repeatedly. Climate scientists want to know whether rising sea surface temperatures cause changes in rainfall patterns. Economists want to know whether a policy change caused a shift in employment. Causal inference in time series introduces unique difficulties that static methods cannot handle.
Contemporaneous causation is one: two variables may influence each other within the same time step, making it hard to assign a direction. Hidden confounders that evolve over time are another. And many real-world systems are non-stationary, meaning the causal relationships themselves can change. Dedicated methods for time-series causal inference are an active area of development, particularly in climate science and the Earth sciences, where the data are abundant but controlled experiments are impossible.20Nature Reviews Earth & Environment. Causal inference for time series
Causal Reasoning and Reinforcement Learning
Reinforcement learning trains agents to make sequences of decisions by trial and error. A robot learns to walk, a recommendation system learns which content to surface, a treatment algorithm learns when to adjust drug doses. Classical reinforcement learning struggles with explainability, robustness, and generalization. Integrating causal inference into reinforcement learning addresses these problems by giving the agent a model of how its actions affect the world, rather than just a record of what rewards followed what actions.21arXiv. Unifying Causal Reinforcement Learning: Survey, Taxonomy, Algorithms and Applications
One particularly tricky problem in this area is off-policy evaluation: estimating how a new decision policy would perform using data collected under a different policy. If a hospital collected data under its old treatment protocol, can you predict outcomes under a proposed new protocol without actually running it? When there are unmeasured confounders, standard methods can produce misleading estimates. Newer approaches use techniques borrowed from econometrics, such as fixed-effects models, to simultaneously learn about unmeasured confounders and estimate the value of new policies.22NeurIPS Proceedings. What Is Causal Machine Learning and Why Is It Important?
What Large Language Models Cannot Do (Yet)
With the explosion of large language models, a natural question is whether these systems can perform causal reasoning. The short answer is: not reliably. LLMs can produce text that looks like causal reasoning because they have absorbed patterns from scientific papers and textbooks in their training data. But they are fundamentally pattern-matching systems. When the training data contain incomplete or incorrect causal relationships, LLMs struggle with causal reasoning tasks, because they are replicating steps they have seen rather than reasoning from a causal model.23arXiv. A Survey on Enhancing Causal Reasoning Ability of Large Language Models
This does not mean LLMs are useless for causal tasks. They can help generate candidate causal graphs based on domain knowledge encoded in their training corpus, and they can assist with translating research questions into formal causal queries. But trusting an LLM to determine whether A causes B is a mistake. The model has no mechanism for distinguishing causation from correlation in its own learned representations. For now, causal reasoning remains something that needs to be built into the system architecture rather than something that emerges from scale alone. Researchers are actively exploring hybrid approaches that pair LLMs with formal causal inference tools, but this work is still in its early stages.