📊 statistics

Causal Discovery via Statistical Power (CDSP)

The paper introduces Causal Discovery via Statistical Power (CDSP), a novel statistical inference framework that leverages effect-size asymmetry to estimate causal direction with uncertainty quantification, demonstrating improved robustness and reduced false discovery rates compared to existing methods.

Original authors: Shreya Prakash, Fan Xia, Elena A. Erosheva

Published 2026-05-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Shreya Prakash, Fan Xia, Elena A. Erosheva

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 figure out which of two people, Alice and Bob, is the boss and which is the employee. You can't ask them directly (no experiments allowed), so you only have a stack of old logs showing their daily activities.

Most detectives in this field use a specific rulebook: "If Alice's actions look like a perfect, straight-line cause of Bob's actions, then Alice is the boss." But this rulebook has a flaw: if the real relationship is slightly messy or curved (not a perfect straight line), the rulebook gets confused and often picks the wrong person as the boss. It also can't tell you how sure it is about its guess.

This paper introduces a new detective method called CDSP (Causal Discovery via Statistical Power). Instead of just checking if the rulebook fits, CDSP asks a different question: "Which story is easier to prove wrong?"

Here is how it works, using simple analogies:

1. The "Two-Way Test"

Imagine you have two theories:

Theory A: Alice causes Bob (Alice $\to$ Bob).
Theory B: Bob causes Alice (Bob $\to$ Alice).

To test these, the method tries to build a model for each direction.

If Theory A is true, the "noise" (the random stuff that happens to Bob that Alice didn't cause) should be totally unrelated to Alice.
If Theory B is true, the "noise" (random stuff happening to Alice) should be unrelated to Bob.

2. The "Effect Size" Analogy: The Weak vs. The Strong

The paper introduces a concept called Effect-Size Asymmetry. Think of this like trying to hear a whisper in a noisy room.

The Correct Direction: If you guess the right direction (Alice $\to$ Bob), the "noise" is mostly just the natural randomness of the world. It's a weak signal of a problem. It's like a faint whisper; it's hard to prove that something is "wrong" because the data looks mostly clean.
The Wrong Direction: If you guess the wrong direction (Bob $\to$ Alice), you are forcing a square peg into a round hole. The math breaks down in two ways at once: the noise isn't random, and the shape of the relationship is wrong. This creates a loud, strong signal of a problem. It's like a siren; it's very easy to prove this story is wrong.

The CDSP Insight: The method assumes that the wrong direction will always look "more broken" (have a stronger signal of error) than the right direction. Therefore, the direction that is harder to prove wrong is likely the correct one.

3. The "Confidence Score" (Uncertainty Quantification)

Old methods would just point a finger and say, "Alice is the boss!" without saying how sure they were. CDSP is different. It runs thousands of simulations (like a detective running the same case over and over in their head) to calculate a Directional Support Probability.

If the score is 0.99, it's like the detective saying, "I am 99% sure Alice is the boss because the 'Bob is the boss' story falls apart immediately."
If the score is 0.55, it's like saying, "I think Alice is the boss, but the evidence is shaky. The 'Bob is the boss' story isn't totally broken yet, so I'm not fully confident."

This helps users know when to trust the result and when to be skeptical.

4. What the Tests Showed

The authors tested this new method against the old "rulebook" method (called LiNGAM) using two types of tests:

The Simulation Lab: They created fake data where the relationship was slightly messy (not a perfect straight line).
- The old method (LiNGAM) got confused immediately and started guessing the wrong direction 100% of the time as soon as the data got slightly messy.
- The new method (CDSP) kept guessing correctly even when the data was messy, only starting to struggle when the mess became extreme.
The Real-World Case Files: They tested it on 100 real-world pairs of data (like "Temperature vs. Ozone" or "Population vs. Food Consumption").
- The old method was wrong about 56% of the time.
- The new method was wrong only 38% of the time.
- Crucially, when the old method was very confident (giving a high score) but wrong, CDSP was often able to say, "Wait, I'm not so sure," or correctly identify the direction where the old method failed.

Summary

CDSP is a new way to figure out cause-and-effect from observational data. Instead of blindly trusting a rigid mathematical rule, it compares how "broken" the story looks in both directions. It assumes the wrong story will look much more broken than the right one.

It also gives you a confidence score, telling you not just what the answer is, but how reliable that answer is, especially when the real world doesn't fit perfectly into neat mathematical boxes. The paper claims this makes it more robust and less likely to make false discoveries than current popular methods.

Technical Summary: Causal Discovery via Statistical Power (CDSP)

Problem Statement
Inferring causal direction from observational data is a fundamental challenge in scientific disciplines. While functional causal discovery methods utilize structural asymmetries (e.g., independence between cause and noise) to identify directionality, they typically rely on strong modeling assumptions (e.g., correct functional form specification, acyclicity, no unobserved confounding). A critical limitation of existing approaches is their lack of robust statistical inference and uncertainty quantification, particularly when these assumptions are violated. Traditional methods often fail to provide reliable directional evidence under mild to moderate model misspecification, and existing uncertainty quantification techniques (such as bootstrap stability or confidence sets) may be misleading if the underlying working models are misspecified.

Methodology
The authors introduce Causal Discovery via Statistical Power (CDSP), a statistical inference framework that reframes causal direction estimation as a problem of directional detectability. The methodology is built on the following theoretical and procedural pillars:

Hypothesis Testing Framework: The approach formulates causal discovery as a pair of hypothesis tests for two variables $X$ and $Y$ . For the direction $X \to Y$ , the null hypothesis $H_0^Y$ posits that the predictor $X$ is independent of the residual $\eta$ (where $Y = m(X) + \eta$ ) and that the model class is correctly specified. The reverse direction $Y \to X$ is tested similarly.
Effect-Size Asymmetry Assumption: The core theoretical contribution is the Effect-Size Asymmetry Assumption. Unlike traditional asymmetry assumptions which rely on the population test statistic being zero in the correct direction and non-zero in the reverse, CDSP posits that the standardized deviation from the null (the effect size) is larger in the incorrect direction than in the correct direction.
- In the correct direction, a null violation arises solely from model misspecification (if any).
- In the incorrect direction, a null violation arises from both the inherent dependence between predictor and residual and model misspecification.
- The assumption states that the incorrect direction deviates more strongly from its null, making it easier to reject.
Directional Power and Detectability: The authors define Directional Power as the probability that the procedure correctly favors the true causal direction. They show that under the Effect-Size Asymmetry Assumption, this probability exceeds the probability of favoring the incorrect direction.
Estimation Procedure (Algorithm 1):
- Point Estimation: The method estimates the Directional Detectability Index ( $I_{X,n}$ and $I_{Y,n}$ ) for both directions. These indices are defined as the difference between the estimated population test statistic (e.g., Hilbert-Schmidt Independence Criterion, HSIC) and the critical value, standardized by the asymptotic standard deviation.
- Decision Rule: The direction with the larger detectability index is selected as the estimated causal direction.
- Uncertainty Quantification: The framework provides a Directional Support Probability ( $\hat{P}_{CDSP}$ ), estimated via nonparametric bootstrapping. This value represents the proportion of bootstrap samples where the estimated detectability index for the chosen direction exceeds that of the alternative. This serves as a measure of confidence, analogous to the Area Under the Curve (AUC), ranging from "little or no separation" (~0.5) to "very strong support" ( $\ge 0.9$ ).

Key Contributions

Theoretical Framework: The paper establishes a link between causal discovery and statistical power, introducing the Effect-Size Asymmetry Assumption as a condition for reliable direction detection even under model misspecification.
Robustness: Theoretical results (Theorem 1) demonstrate that if the effect-size asymmetry holds, the probability of correctly identifying the direction converges to 1 as sample size increases.
Uncertainty Quantification: CDSP provides a formal, interpretable measure of directional support ( $\hat{P}_{CDSP}$ ) that accounts for sampling variability and model misspecification, addressing a gap in current functional causal discovery literature.
Consistency: The authors prove the consistency of both the directional estimate and the bootstrap-based support probability under standard regularity conditions.

Results
The paper evaluates CDSP through simulations and real-data analysis:

Simulations (Linear Model Misspecification):
- Data was generated with varying degrees of nonlinearity ( $d \in \{1, 1.2, \dots, 1.5, 3\}$ ).
- Robustness: CDSP maintained high accuracy (100% for $d \le 1.3$ , 88% for $d=1.4$ ) under mild to moderate misspecification. In contrast, DirectLiNGAM failed abruptly (0% accuracy) starting at $d=1.25$ .
- Sample Size: Increasing sample size improved CDSP accuracy in moderate misspecification regimes, stabilizing the ordering of detectability indices.
- Severe Misspecification: Under severe misspecification ( $d=3$ ), the effect-size asymmetry assumption broke down, leading to incorrect inference by both methods, highlighting the limits of the approach.
Real Data (Tübingen Cause–Effect Pairs):
- Full Dataset (100 pairs): CDSP achieved a True Discovery Rate (TDR) of 62% and a False Discovery Rate (FDR) of 38%, compared to LiNGAM's 44% TDR and 56% FDR. CDSP reduced the FDR by approximately 18% relative to LiNGAM.
- Approximately Linear Subset (34 pairs): Performance improved for both methods, but CDSP maintained a significant advantage (20.6% FDR vs. 38.2% for LiNGAM).
- Uncertainty Calibration: Unlike LiNGAM's bootstrap rates, which assigned "very strong" support to many pairs with high error rates, CDSP's $\hat{P}_{CDSP}$ showed a clear correlation between support strength and accuracy (lower FDR in higher support categories).

Significance and Claims
The paper claims that CDSP offers a statistically principled basis for causal direction estimation that is robust to model misspecification and provides meaningful uncertainty quantification. By shifting the focus from strict identifiability under idealized assumptions to directional detectability via statistical power, the framework allows researchers to determine when data contains sufficient information to favor one causal direction over another.

The authors emphasize that CDSP is most informative under mild to moderate model misspecification, a regime where traditional methods often fail silently. The framework does not claim to solve causal discovery under all conditions (e.g., severe misspecification or unobserved confounding) but provides a diagnostic tool to assess the reliability of directional evidence. The paper concludes that CDSP represents a first step toward a power-motivated statistical framework, encouraging the use of such methods to learn when data provides sufficient evidence for causal inference, even when assumptions are not perfectly met.