The Fundamental Problem of Causal Inference: Why the Counterfactual is Always Missing

The binding constraint of measuring any effect is that you must compare what happened against what would have happened if you had made a different choice. To know if a new marketing campaign drove a 15% increase in sales, or if a 200-milligram dose of ibuprofen cured a headache, you need to observe the exact same subject, at the exact same moment, under both conditions. That condition never holds. Time only moves in one direction, and reality only plays out once. You either take the pill or you do not. The alternative reality—the counterfactual—is permanently erased the moment a choice is made.[1][4]

This dilemma is not a mere philosophical puzzle; it is a hard mathematical limit known as the 'fundamental problem of causal inference.' Coined by statistician Paul W. Holland in a 1986 paper, the phrase describes the impossibility of observing a unit-level causal effect. Because one of the potential outcomes is always missing, causation is fundamentally a missing-data problem. Every algorithm, every A/B test, and every machine learning model attempting to measure cause and effect is ultimately trying to fill in a blank that reality refuses to provide.[2]

The technology industry frequently markets 'causal AI' and 'causal machine learning' as tools that can finally untangle correlation from causation. Vendors promise that their platforms can ingest datasets with 100,000 or 1,000,000 rows and output the true drivers of customer behavior or disease progression. But these systems do not possess a magic lens into alternate timelines. They are bound by the exact same mathematical constraints that Jerzy Neyman first formalized in his 1923 master's thesis and Donald Rubin expanded in 1974 with the Rubin Causal Model.[1][5]

Under the Rubin Causal Model, every subject has two potential outcomes: Y(1) if they receive the treatment, and Y(0) if they do not. The causal effect is simply Y(1) minus Y(0). But because we can only observe one of those states, calculating the true individual effect is impossible. 'At least 50% of the information needed to calculate the [individual causal effect] is missing,' notes statistician Stef van Buuren. Instead, statisticians and data scientists shift their focus from the individual to the group, attempting to calculate the Average Treatment Effect (ATE).[3]

To calculate an average effect, researchers must make a massive, untestable assumption: that the group that received the treatment is perfectly comparable to the group that did not. In a randomized controlled trial (RCT)—the gold standard of causal inference—this assumption is enforced by the study design. By randomly assigning the treatment to, say, 5,000 participants while giving a placebo to another 5,000, researchers ensure that the only systematic difference between the two groups is the drug itself. The control group serves as a proxy for the missing counterfactuals of the treated group.[1][4]

But most of the data collected by modern enterprises and governments is observational, not experimental. A company knows which 25,000 users clicked an ad, but it did not randomly assign the ad. A hospital knows which 400 patients took a drug, but the patients who sought out the drug might be sicker, wealthier, or 10 years older on average than those who did not. When data scientists apply causal inference algorithms to observational data, they are attempting to mathematically simulate a randomized trial after the fact.[4]

This is where the hype outpaces the capability. To extract causation from observational data, algorithms rely on assumptions like 'ignorability' or 'unconfoundedness.' These terms mean that the model assumes it has measured and accounted for every single variable that could influence both the treatment and the outcome. If a hidden variable exists—say, a patient's unrecorded genetic predisposition that increases their risk by 30%—the model's estimate of the counterfactual is wrong, and its 'causal' conclusion is just a dressed-up correlation.[2][5]

Structural Causal Models (SCMs), pioneered by Judea Pearl in the 1990s, offer a different approach by using directed acyclic graphs to map out the assumed relationships between variables. But SCMs also require the researcher to explicitly state the causal structure before the data is analyzed. The algorithm does not discover the causal rules from scratch; it merely computes the effects based on the human's structural assumptions. If the human draws the wrong map, the machine learning model will confidently output the wrong causal effect.[1][5]

The fundamental problem of causal inference cannot be solved by adding more compute power or larger datasets. A dataset with 10 billion rows of observational data still lacks the missing counterfactuals. As causal machine learning becomes a standard feature in enterprise software, the risk is that organizations will mistake sophisticated statistical adjustments for true causal certainty.[4][5]

This limitation is why causal inference remains as much a scientific discipline as a computational one. The algorithms are powerful, but they are entirely dependent on the quality of the study design and the validity of the human assumptions. When a vendor claims their AI can discover causality from raw data with 99% accuracy, they are obscuring the fact that the machine is simply executing a statistical solution to a scientific problem.[5]

The missing counterfactual forces a permanent humility on data science. We can measure associations with perfect precision, but causation will always require a leap of faith across the gap of what we cannot observe. The next time a model predicts exactly what would have happened if a different decision had been made, remember that it is not reporting a fact. The math can estimate what might have happened, but the true individual effect remains locked in a reality that never occurred.[5]