📈 economics

Identification and Inference for Causal Effects in Extremes under General Conditions

This paper develops identification and inference methods for causal effects in extreme events by analyzing the Causal Tail Coefficient within linear structural models, demonstrating how heterogeneous tail indices and heavy-tailed confounders influence causal structure identification and proposing adjusted estimators and tests to recover causal relations that average-based methods often miss.

Original authors: Lisa Leimenstoll, Melanie Schienle

Published 2026-08-25

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Lisa Leimenstoll, Melanie Schienle

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

In the world of data science, researchers have long relied on a simple principle: to understand how one thing affects another, look at the average. If you want to know if rain causes traffic jams, you might check how much slower cars move on a typical rainy day compared to a sunny one. This approach works well for everyday fluctuations, but it often fails when the stakes are highest. The most dangerous events in finance, climate, and infrastructure—market crashes, catastrophic floods, or systemic grid failures—do not happen on average. They happen in the extremes, in the rare, violent outliers where the usual rules of cause and effect can break down or hide entirely. For decades, statisticians struggled to untangle these extreme relationships because the tools they used were designed for the middle of the data, not the fringes. They could tell you how variables moved together in general, but they could not reliably say which one was pushing the other when everything was falling apart.

This is the problem tackled by a new study from researchers at the Karlsruhe Institute of Technology and Heidelberg Institute for Theoretical Studies. They developed a method to trace the path of extreme events through a system, asking a specific question: when a massive shock hits one part of a network, does it trigger a massive shock in another? To do this, they focused on a concept called the "causal tail coefficient." Imagine two variables, like rainfall and river levels. If a huge rainstorm consistently leads to a massive flood, there is a causal link. But if both are simply reacting to a third, hidden factor—like a massive weather system affecting the whole region—the link is an illusion. The researchers found that by looking specifically at the behavior of the most extreme values, they could distinguish between a true cause-and-effect chain and a coincidence caused by a hidden influence. Their work is particularly powerful because it handles real-world messiness: it works even when the two variables behave differently at the extremes, such as when one variable has a "heavier" tail, meaning it produces more frequent or more severe outliers than the other.

The researchers tested their theory using a framework where variables are connected in a specific direction, like a chain of dominoes. They discovered that the way these variables behave at the very edge of the distribution holds a secret key to identifying the direction of causality. If one variable has a "heavier" tail than the other—meaning it experiences more extreme spikes—their method can often tell you which one is the driver and which one is the follower. In cases where the tails are similar, the method still works, but it requires looking for a specific pattern in how the extremes align. Crucially, they also addressed the problem of hidden confounders. If a third, unobserved factor is causing both variables to spike, standard methods often mistake this for a direct cause. The team showed that if this hidden factor is also extreme (heavy-tailed), it can completely mask the true relationship. However, they devised a way to correct for this if researchers have a proxy—a measurable stand-in—for that hidden factor. By adjusting their calculations to account for this proxy, they could peel back the layer of confusion and reveal the true causal arrow, even in the presence of heavy-tailed noise.

To prove their method worked, the team ran thousands of computer simulations, creating artificial worlds with different types of extreme behaviors and hidden influences. They found that their new tests could correctly identify causal directions in situations where older, standard methods failed completely. For instance, in scenarios where a causal link existed only in the extreme tails of the data, traditional tools saw nothing, while the new method spotted the connection clearly. They also tested how well the method performed when the "tails" of the data were different, confirming that these differences actually helped them pinpoint the cause rather than confusing the issue. The simulations showed that the method remains reliable even when the data is noisy or when the sample size is not massive, provided the researchers choose the right threshold for what counts as an "extreme" event.

The researchers then took their tools out of the simulation lab and applied them to real-world data, starting with the Swiss railway system. They wanted to know if extreme precipitation caused extreme train delays. While it seems obvious that rain delays trains, standard statistical models looking at average data often struggle to prove this because delays happen for many reasons. By focusing only on the most extreme rain events and the most severe delays, the researchers found a clear, one-way causal link: heavy rain directly caused significant delays on the Zurich-to-Bern line. The method also revealed that the impact of rain accumulated over three hours was even more significant than the rain falling at a single moment, a nuance that average-based models missed.

Next, they turned to the rivers of Germany, examining the relationship between daily rainfall and river discharge on the Danube and the Main rivers. Here, the challenge was more complex. The data suggested that while rain caused the rivers to swell, there was also a hidden factor at play, likely related to the upstream catchment areas. The researchers' method detected this hidden influence, which they confirmed by checking the "tail indices" of the data. Once they adjusted their analysis to account for the upstream water flow as a proxy for the hidden factor, the true causal link between rain and river levels became sharp and undeniable. They also explored how long it took for rain to affect the river, finding that while the connection was strong after one day, the causal signal in the extremes faded much faster than the average signal did, disappearing after just a few days. This provided a more precise picture of flood dynamics than previously available.

Finally, the team looked at the volatile relationship between the S&P 500 stock index and Bitcoin. Financial markets are notorious for their extreme swings, and the question of whether stock market crashes drive cryptocurrency crashes, or vice versa, has been a subject of intense debate. Previous studies using average data had produced conflicting results. The researchers applied their extreme-event method to the left tail of the data—the side representing massive losses. They found that when the stock market suffered extreme negative returns, it triggered extreme negative returns in Bitcoin, but only after they accounted for broader market volatility as a hidden confounder. Without this adjustment, the signal was muddled, but with it, a clear causal path emerged from the stock market to the cryptocurrency. This suggests that during times of severe market stress, the traditional stock market acts as a leading indicator for the crypto market, a finding that could be vital for investors and regulators trying to understand systemic risk.

The study concludes that looking at the extremes is not just a niche exercise but a necessary tool for understanding how the world works when it is under the most pressure. By developing a way to separate true causal chains from coincidental correlations in the tails of distributions, the researchers have provided a new lens for economists, climatologists, and risk managers. Their work demonstrates that the most extreme events often carry the clearest signals of cause and effect, provided one knows how to read them. The method does not require perfect data or simple systems; it thrives in the messy, heavy-tailed reality of the natural and financial worlds, offering a way to see the true structure of causality when everything else is in chaos.

Technical Summary: Identification and Inference for Causal Effects in Extremes Under General Conditions

1. Problem Statement

The identification of causal effects from observational data is a central objective in econometrics. However, existing identification theories and methods predominantly focus on average effects or rely on information from the bulk of the distribution. In many critical economic and environmental applications—such as financial crises, systemic risk, and climate disasters—the phenomena of interest occur during rare and extreme events. In these settings, the causal mechanisms are often activated only by unusually large shocks, and impacts are most severe in the tails.

Current methods for causal inference in extremes, particularly those based on the Causal Tail Coefficient (CTC), typically rely on restrictive assumptions, most notably that all variables in the system exhibit comparable tail behavior (i.e., identical tail indices). In empirical reality, economic and environmental variables often display heterogeneous tail indices, and unobserved confounders may possess heavy tails that distort extremal dependence. This paper addresses the gap in identifying causal relations in extremes when variables exhibit heterogeneous tail indices and when heavy-tailed confounders may be present.

2. Methodology

Structural Framework

The authors analyze linear structural causal models (LSCM) defined on directed acyclic graphs (DAGs). The data-generating process is given by:
$X_j := \sum_{k \in pa(j,G)} \beta_{kj}X_k + \epsilon_j$
where the innovation terms $\epsilon_j$ are mutually independent and regularly varying with potentially different tail indices $\alpha_j$ . Regular variation is defined such that $P(\epsilon_j > x) \sim \ell(x)x^{-\alpha_j}$ as $x \to \infty$ .

The Causal Tail Coefficient (CTC)

The core measure of causal dependence is the Causal Tail Coefficient (CTC), defined for two regularly varying variables $X_i$ and $X_j$ as:
$\Gamma_{X_i \to X_j} := \lim_{u \to 1^-} E[F_j(X_j) \mid F_i(X_i) > u]$
Intuitively, this measures the expected rank of $X_j$ conditional on $X_i$ being in the extreme upper tail. Under standard conditions with equal tails, $\Gamma \to 1$ implies a causal link $X_i \to X_j$ , while $\Gamma \to 0.5$ implies independence.

Key Theoretical Extensions

The paper extends the CTC framework to general conditions:

Heterogeneous Tail Indices: The authors derive the limiting behavior of the CTC when $\alpha_i \neq \alpha_j$ . They show that differences in tail behavior provide additional identifying information. Specifically, if $\alpha_1 > \alpha_2$ (meaning $X_1$ has a lighter tail than $X_2$ ), the CTC asymmetry $\Delta = \Gamma_{X_1 \to X_2} - \Gamma_{X_2 \to X_1}$ can uniquely identify the direction $X_1 \to X_2$ even when standard equal-tail assumptions fail.
Confounders: The paper distinguishes between light-tailed and heavy-tailed unobserved confounders ( $H$ $H$ ).
- Light-tailed confounders ( $\alpha_H > \max(\alpha_1, \alpha_2)$ ) are asymptotically negligible for tail identification.
- Heavy-tailed confounders ( $\alpha_H \leq \max(\alpha_1, \alpha_2)$ ) can induce extremal dependence patterns indistinguishable from direct causal effects, leading to misidentification.
Adjusted CTC: To recover identification in the presence of heavy-tailed confounders, the authors propose using an adjusted CTC ( $\Gamma_{X_i \to X_j | H}$ ) that conditions on an observed proxy vector $H$ of the confounder. This involves fitting a Generalized Pareto Distribution (GPD) to the exceedances with scale parameters modeled as functions of the proxy.

Estimation and Inference

Estimators: Non-parametric estimators for the CTC are constructed using empirical distribution functions and order statistics. An adjusted estimator is derived for the confounder-controlled case using a parametric pre-step (GPD fitting).
Asymptotic Theory: The paper establishes the asymptotic normality of the estimators and the difference statistic $\hat{\Delta}$ . The convergence rates depend on the tail indices and the choice of the tuning parameter $k$ (the number of top order statistics used).
Testing Strategy:
- Causality-Test: Tests $H_0: \beta_{12} = 0$ against $H_1: \beta_{12} \neq 0$ using the asymmetry $\hat{\Delta}$ .
- Confounder-Test: Tests for the presence of a heavy-tailed confounder by checking if $\Gamma_{X_2 \to X_1}$ deviates significantly from 0.5 when a causal link is suspected.
- Tail Index-Test: A pre-test to determine if tail indices are equal or heterogeneous, utilizing a modified Hill estimator.

3. Key Contributions

Identification under Heterogeneity: The paper demonstrates that heterogeneous tail indices are not merely a technical complication but a source of identifying information. It derives theoretical results showing how the ordering of tail indices ( $\alpha_1 > \alpha_2$ vs. $\alpha_1 < \alpha_2$ ) fundamentally alters the limiting behavior of the CTC, allowing for causal direction identification in scenarios where equal-tail theories fail.
Handling Heavy-Tailed Confounding: The authors provide a rigorous characterization of how heavy-tailed confounders distort extremal dependence. They show that while light-tailed confounders are asymptotically negligible, heavy-tailed ones can mimic causal effects. They propose a solution via the adjusted CTC, recovering identification when suitable proxy information is available.
Inference Procedures: The paper develops a complete inferential framework, including asymptotic distributions for estimators, hypothesis tests for causal direction and confounding, and practical guidance on tuning parameter selection ( $k$ ) via simulation.

4. Results

Simulation Evidence

Extensive simulations validate the finite-sample performance of the proposed methods:

Convergence: The estimators converge to their theoretical limits, though the rate depends on the tail indices and the underlying distribution family (e.g., Student's $t$ vs. Pareto).
Robustness: The Causality-Test successfully detects causal links in the tails even when the relationship is non-linear in the bulk or when tail indices differ.
Benchmarking: The proposed CTC-based methods outperform classical LiNGAM (Linear Non-Gaussian Acyclic Model) algorithms in tail-specific scenarios. LiNGAM often fails to detect causal effects confined to the extremes, particularly when tail indices are heterogeneous.
Confounder Detection: The Confounder-Test effectively identifies the presence of heavy-tailed confounders, and the adjusted CTC successfully recovers the true causal direction in these cases.

Empirical Applications

The methodology is applied to three distinct datasets:

Precipitation and Train Delays (Switzerland): The study confirms a causal effect of extreme precipitation on train delays. The tail tests detect this relationship even though standard methods focusing on the bulk of the distribution (LiNGAM) failed to find a significant link for current precipitation.
Precipitation and River Flows (Germany): Analyzing the Danube and Main rivers, the authors identify causal links from precipitation to river discharge. The method successfully accounts for upstream catchment areas as confounders and investigates temporal lags, finding that the causal effect diminishes more rapidly in the extreme tail than in the full distribution.
S&P 500 and Bitcoin Returns: In a setting where the causal direction is ambiguous, the analysis reveals a causal impact of extreme negative S&P 500 returns on Bitcoin returns after adjusting for market volatility (VIX) and broader market movements (MSCI Europe) as heavy-tailed confounders. This relationship was not detectable by standard LiNGAM on the full dataset.

5. Significance and Claims

The paper claims that its framework provides a robust tool for uncovering causal relations in extremes that remain undetected by approaches targeting average dependence. Key significance points include:

General Conditions: The method relaxes the restrictive assumption of equal tail indices, making it applicable to a wider range of empirical economic and environmental data where variables naturally exhibit different tail behaviors.
Confounder Handling: The ability to detect and adjust for heavy-tailed confounders is a critical advancement, as such confounders can otherwise lead to spurious causal conclusions in extreme value analysis.
Complementarity: The proposed tail tests complement existing structural causal models. While methods like LiNGAM are effective for global dependencies, the CTC-based approach is specifically designed for scenarios where causal mechanisms are activated primarily by extreme shocks.
Practical Utility: The simulation studies and empirical applications demonstrate that the methods are feasible for finite samples and provide actionable insights for risk management, climate science, and financial stability, particularly in identifying the propagation of extreme events.

The authors note limitations, such as the reliance on linear structural models (though noting non-linearity in the bulk may still be linear in the tail) and the current focus on independent observations, suggesting future work on time-series dependencies and non-linear structural models.