📊 statistics

Asymptotic emergence of statistically supported false causal interpretation under unmeasured confounding

This paper demonstrates that under unmeasured confounding, increasing sample sizes can paradoxically lead to greater statistical certainty in identifying false causal structures based on observable proxies, highlighting a fundamental disconnect between statistical confidence and the recovery of true causal mechanisms.

Original authors: Mark Louie F. Ramos

Published 2026-07-31

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Mark Louie F. Ramos

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but you can only see the clues left behind, not the culprit themselves. This is the world of observational causal inference, a branch of science where researchers try to figure out what causes what using data from the real world, rather than running controlled experiments. In this field, a "cause" is like a domino that knocks over another domino (the "effect"). However, there's a tricky rule: just because two things happen at the same time (they are "associated") doesn't mean one caused the other. Sometimes, a hidden third factor—a "confounder"—is secretly pushing both of them. For example, ice cream sales and shark attacks both go up in the summer, but ice cream doesn't cause shark attacks; the hidden cause is the hot weather. Scientists use special maps called "DAGs" (Directed Acyclic Graphs) to draw their theories about how these dominoes are connected, but these maps rely on big assumptions about what they can't see.

The big question everyone cares about is: Can we ever be sure we've found the real cause? As we collect more and more data, we get better at spotting patterns. We can say with high confidence that "A is linked to B." But does that mean A causes B? This paper tackles a scary possibility: what if we get so good at finding patterns that we become too confident about the wrong answer? It suggests that even with perfect math and massive amounts of data, if the real cause is invisible to us, we might end up building a very convincing, statistically rock-solid story about the wrong suspect.

The Case of the Invisible Puppeteer

Let's say you are trying to figure out why a certain plant in your garden is growing huge. You suspect it's the "Secret Super-Fertilizer" (let's call it X) that a mysterious gardener is secretly pouring on the soil. But here's the catch: X is invisible. You can't see it, measure it, or even know it exists. However, you can see other things happening in the garden. You notice that every time the plant grows, the garden hose is also spraying water (Z). In reality, the invisible fertilizer (X) is causing both the plant to grow and the hose to spray (maybe the fertilizer makes the soil thirsty, or the gardener uses the hose to mix the fertilizer).

Now, imagine you are a data scientist with a super-computer. You decide to study the garden. You don't know about the invisible fertilizer, so you only look at the things you can see: the plant size and the water spray. You gather data from 10 plants, then 100, then 10,000. As you get more data, your computer gets better and better at spotting the link between the water spray and the plant growth.

The paper argues that as your data gets bigger and your math gets smarter, you will eventually reach a point where you are 100% certain that the water spray causes the plant to grow. Your statistical tests will scream, "This is a real effect! It's not a fluke!" You will have found a "stable, data-driven" relationship. But here is the twist: You are wrong. The water didn't cause the growth; the invisible fertilizer did. The water was just a "proxy"—a stand-in that happened to move along with the real cause.

The Trap of "More Data"

The author, Mark Louie F. Ramos, shows that this isn't a mistake in your math or a result of "p-hacking" (fudging numbers to get a result). It's a fundamental trap of the situation.

Think of it like trying to find a ghost in a haunted house. You can't see the ghost (the real cause), but you can see the curtains moving and the lights flickering (the proxies). If you stand there for a long time with a high-speed camera, you will eventually prove with absolute certainty that "When the curtains move, the lights flicker." You will have a perfect, unshakeable statistical proof of this connection. But if you conclude that "The curtains caused the lights to flicker," you've missed the ghost entirely.

The paper proves that if you keep adding more variables to your study (more things you can see in the garden) and keep increasing your sample size (watching more plants), you will inevitably find these "perfect" links between the wrong things. The more data you have, the more confident you become in a story that is completely wrong about why things are happening.

Why This Matters (And What It's Not)

It is important to understand what this paper is not saying. It is not saying that scientists are bad at math, or that they are cheating by picking and choosing data. The paper explicitly rules out the idea that this is just a "Type 1 error" (a false alarm caused by bad luck or bad statistics). In fact, because the invisible cause is real and strong, these "wrong" results are actually more likely to be repeated than random noise. If you run the experiment again with a million plants, you will get the same "wrong" answer again and again. The association is real; the causation is the lie.

The paper also clarifies that this isn't a problem with the tools we use (like the t-test). Those tools are great at telling us if two things are linked. The problem is that we often forget that the tools only work if we first agree on a map of the world that includes the invisible things. If our map is missing the "invisible fertilizer," the tools will happily tell us the "water hose" is the hero, and they will do it with perfect confidence.

The Takeaway

The main finding is a bit of a bummer for those who think "Big Data" will solve all our mysteries. The paper shows that more data does not fix the problem of missing causes. Instead, it makes us more confident in the wrong answers.

The author concludes that we need to stop thinking that statistical analysis can "validate" our ideas about what causes what. Statistics can only tell us the size of an effect if we assume our map of the world is correct. The real work—figuring out what the invisible causes are—has to happen before we crunch the numbers, using our brains, theories, and prior knowledge, not just our calculators.

So, the next time you read a study that says "We found a rock-solid, statistically significant link between A and B," remember the invisible puppeteer. The data might be perfect, the math might be flawless, and the confidence might be sky-high, but if the real cause is hiding in the shadows, the whole story could be a very convincing illusion.

Technical Summary: Asymptotic Emergence of Statistically Supported False Causal Interpretation under Unmeasured Confounding

Problem Statement
The paper addresses a critical limitation in observational causal inference: the tendency for data-driven causal discovery workflows to produce statistically robust but causally false interpretations when unmeasured confounding is present. While it is a foundational principle in causal theory that unobserved common causes prevent the recovery of true causal effects from observational data alone, the paper argues that the implications of this limitation for causal discovery are often understated. Specifically, as observational datasets grow larger and richer, standard statistical procedures may increasingly converge on stable, significant associations involving observable proxy variables, mistakenly interpreting these non-causal correlations as evidence of causal mechanisms.

Methodology and Setup
The author formalizes this phenomenon through a theoretical asymptotic framework involving a Directed Acyclic Graph (DAG):

True Structure: A true causal variable $X$ exists such that $X \to Y$ with a total effect $\rho$ , but $X$ is unobserved ( $X \notin O$ ).
Proxies: A sequence of observable variables $\{Z_j\}$ exists in the observed set $O$ . These variables are non-causal proxies of $X$ (i.e., $X \to Z_j$ ) but have no direct causal edge to the outcome ( $Z_j \not\to Y$ ). However, due to the correlation with $X$ , they exhibit non-zero covariance with both $X$ and $Y$ ( $Cov(X, Z_j) \neq 0$ and $Cov(Z_j, Y) \neq 0$ ).
Data Expansion: The analysis considers a sequence of increasing subsets of observable variables $\{O_k\}$ and increasing sample sizes $n$ .
Procedure: A causal discovery procedure $\mathcal{A}_{k,n}$ tests hypotheses of the form $H_0: \theta_i = 0$ (no association) versus $H_1: \theta_i \neq 0$ for focal variables $W_{i,k} \in O_k$ . The procedure utilizes consistent estimators for the association $\theta_i$ and standard hypothesis testing. Under the assumption of no unmeasured confounding between the focal variable and the outcome, a rejection of $H_0$ is interpreted as the variable causing $Y$ .

Key Results
The paper provides an asymptotic characterization demonstrating that:

Inevitable Inclusion: As the set of observed variables expands ( $k \to \infty$ ), the probability that the procedure includes at least one proxy variable $Z_j$ in its hypothesis tests approaches 1.
Non-Zero Association: For any such proxy $Z_j$ , the true association with the outcome is non-zero ( $\theta_i \neq 0$ ) due to the confounding path through $X$ .
Asymptotic Certainty: As the sample size $n \to \infty$ , the probability that the procedure correctly rejects the null hypothesis for these proxies approaches 1, driven by the consistency of the estimators and the non-zero true association.
False Causal Conclusion: Consequently, for sufficiently large $k$ and $n$ , the procedure will conclude with probability approaching 1 that a proxy variable $Z_j$ causes $Y$ ( $Z_j \to Y$ ), despite the fact that $Z_j$ is not a causal parent of $Y$ .

Key Contributions
The primary contribution is the explicit demonstration that statistical certainty and causal validity are distinct concepts in the presence of unmeasured confounding. The paper shows that:

Data Quantity Does Not Resolve Ambiguity: Increasing sample size and the richness of observed features do not necessarily lead to the identification of the true causal structure. Instead, they increase the statistical confidence in relationships involving observable proxies.
Replicability of False Causality: Unlike issues arising from $p$ -hacking or Type I error inflation (e.g., data dredging), these false causal conclusions are robust and replicable. Because the underlying associations are real (driven by the unobserved confounder), repeating the study with larger samples will consistently yield the same "statistically significant" but causally incorrect results.
Distinction of Assumptions: The work clarifies the separation between identification assumptions (e.g., the structure of the DAG, absence of unmeasured confounding) and statistical assumptions (e.g., normality, linearity). While statistical assumptions can often be validated or are robust to minor violations, identification assumptions are rigid and cannot be validated by the data itself.

Significance and Claims
The paper does not claim to introduce new methods for causal identification or to discourage the use of observational causal inference. Rather, its significance lies in underscoring a fundamental epistemological boundary:

Measurement vs. Validation: The purpose of observational causal inference is to measure the magnitude of an effect conditional on specific assumptions, not to validate those assumptions using the same data.
The Operative Question: The correct framing of research is not "Does $X$ cause $Y$ ?" but rather "Suppose $X$ causes $Y$ , what is the magnitude of that effect?"
Asymptotic Convergence to Error: The paper warns that strategies seeking to discover stable, data-driven DAGs from increasingly large datasets may asymptotically converge on explanations that accurately capture observed associations but incorrectly attribute causality to proxy variables rather than the unobserved mechanism.

The author concludes that researchers, reviewers, and readers must treat the assumptions underpinning causal models with at least as much scrutiny as the statistical analysis itself, recognizing that more data alone cannot resolve causal ambiguity when the true causal variable is unobserved.

The Case of the Invisible Puppeteer

The Trap of "More Data"

Why This Matters (And What It's Not)

The Takeaway

Technical Summary: Asymptotic Emergence of Statistically Supported False Causal Interpretation under Unmeasured Confounding

More like this