Rethinking Statistics and Causality: Why Mechanisms Cannot Be Inferred from Data Distributions

Across psychology, neuroscience, and applied economics, many findings fail to replicate or disappear outside the lab. Even without fraud, statistically “significant” effects often fail to match real-world behavior. These breakdowns are not discipline-specific—they reflect a deeper mismatch between how real systems generate phenomena and how statistical inference interprets the data left behind.

In practice, statistics and causal inference operate entirely within data space. But data are already a compressed projection of a much higher-dimensional world; most semantic structure is lost before analysis even begins. Once mechanisms are removed in this projection, no amount of geometry or factorization applied to the remaining traces can reconstruct how the system actually works.

Many failures stem from two assumptions: that alignment in data space reflects structure in the underlying system, and that probabilistic factorizations reveal the process that generated the outcomes. Yet these tools operate in a domain where the mechanism is no longer present, which is why they break even when used correctly.

Statistical and causal inference have become universal currencies of explanation across the sciences, particularly in domains where underlying mechanisms remain opaque. Their apparent rigor—spanning psychology, economics and biomedicine—rests on the assumption that patterns within data can reveal the processes that generate them. Yet persistent mismatches between empirical predictions and real-world behaviour expose a deeper limitation: mechanisms cannot be inferred from data distributions alone.

To address this limitation, we revisit the foundations of both paradigms, showing how statistical inference reduces explanation to geometric alignment, while causal inference, evolved from Bayes’ theorem and graphical models, extends this misstep by conflating probabilistic structure with causal truth. Both expose the same epistemic gap: data encode a lower-dimensional projection of structure, not the mechanism that generates it.

We argue that understanding the world follows two routes: one is data-driven, expanding models toward richer function classes to achieve high-precision prediction, as exemplified by modern deep learning; the other is mechanism-driven, proposing and testing structural hypotheses as in the physical sciences. A robust framework requires both: data-driven models for high-precision prediction, and mechanistic models for reconstructing how the world produces the data we observe.

These failures do not stay inside statistics. They propagate across every discipline that relies on data-space reasoning. Psychology, neuroscience, political science, and applied economics routinely draw conclusions from patterns that could never reveal how the system actually works. Entire literatures grow around correlations, regressions, and conditional probabilities whose quantities have no mechanism-level meaning.

Causal inference amplifies the problem. Its algebra multiplies probabilities into joint events that do not exist in real systems. Because observed variables are low-dimensional projections of a richer generative world, the true joint distribution is undefined—yet causal formulas fabricate it through factorization. The machinery operates on objects that have no semantic or generative counterpart in reality.

Machine learning and statistical learning theory inherit the same statistical ontology, but their actual behavior no longer matches their notation. Quantities like likelihoods, expected risk, factorizations, and losses survive as syntactic artifacts, even though the true system has no mechanism that corresponds to them. Models work through optimization heuristics and implementation patches, not through the semantics their formulas suggest. What runs in practice is not what the notation claims.

That is why the foundation must be rebuilt. As long as inference is performed in data space, every downstream field will continue scaling the same structural error — operating on quantities that do not correspond to the systems they aim to explain.