[Paper Review] Event Conditional Correlation: Or How Non-Linear Linear Dependence Can Be

P-A. G. Maugis|arXiv (Cornell University)|Jan 6, 2014

Financial Risk and Volatility Modeling21 references3 citations

TL;DR

This paper introduces event conditional correlation (ECC), a new dependence parameter that estimates the correlation between two variables conditional on a specific event, correcting for sampling bias in partial samples. It proposes two consistent, asymptotically normal estimators—one for direct estimation of ECC and another for counterfactual estimation using different sampling regimes—demonstrating that dependence structures can be highly non-linear even under Gaussian assumptions, with critical implications for financial risk modeling during crises.

ABSTRACT

Entries of datasets are often collected only if an event occurred: taking a survey, enrolling in an experiment and so forth. However, such partial samples bias classical correlation estimators. Here we show how to correct for such sampling effects through two complementary estimators of event conditional correlation: the correlation of two random variables conditional on a given event. First, we provide under minimal assumptions proof of consistency and asymptotic normality for the proposed estimators. Then, through synthetic examples, we show that these estimators behave well in small-sample and yield powerful methodologies for non-linear regression as well as dependence testing. Finally, by using the two estimators in tandem, we explore counterfactual dependence regimes in a financial dataset. By so doing we show that the contagion which took place during the 2007--2011 financial crisis cannot be explained solely by increased financial risk.

Motivation & Objective

• Address the bias introduced in classical correlation estimators when data are collected only under specific conditions (e.g., only high-performing students or during market crises).
• Develop a formal framework for event conditional correlation (ECC) as a dependence parameter that captures conditional linear dependence given an event A with P(A) > 0.
• Establish theoretical properties—consistency and asymptotic normality—of ECC estimators under minimal regularity assumptions.
• Enable counterfactual analysis of dependence regimes by combining two estimators to infer ECC under one sampling condition from data collected under another.
• Demonstrate the method’s utility in non-linear regression, independence testing, and financial network topology analysis under varying risk regimes.

Proposed method

• Proposes a new dependence parameter: ρXY|A, the correlation of X and Y conditional on event A, defined via conditional expectations and variances.
• Derives an admissible estimator for ρXY|A using the Delta Method, based on consistent estimators of underlying parameters (ρXY, ρXZ, ρYZ, δ), under minimal assumptions.
• Introduces a second estimator that allows estimation of ρXY|A from an A′-sample (data collected under a different event A′), enabling counterfactual inference.
• Uses a piecewise-linear approximation framework to model conditional variances and covariances, particularly in the presence of conditioning on a thresholded variable Z.
• Applies the Delta Method to derive asymptotic normality of the ECC estimator, ensuring statistical inference is valid under regularity conditions.
• Employs simulation-based validation (using R and rgl) to demonstrate estimator performance in small samples and non-Gaussian settings.

Experimental results

Research questions

• RQ1 How can classical correlation estimators be biased when data are collected only under specific conditions (e.g., high Z values), and what is the true dependence structure under such sampling?
• RQ2 Can a consistent and asymptotically normal estimator be constructed for event conditional correlation ρXY|A under minimal assumptions?
• RQ3 To what extent can dependence parameters estimated under one sampling regime (e.g., A′-sample) be used to infer parameters under a different regime (e.g., A)?
• RQ4 How non-linear is the behavior of conditional correlation even in Gaussian models, and what does this imply for modeling dependence in high-risk scenarios?
• RQ5 Can event conditional correlation improve non-linear regression, independence testing, and financial network analysis under partial sampling?

Key findings

• Event conditional correlation can exhibit markedly different values from unconditional correlation—e.g., ρXY|Z>zc can be substantially higher than ρXY even in trivariate Gaussian models.
• The proposed estimators are consistent and asymptotically normal under minimal regularity conditions, enabling valid statistical inference.
• In synthetic examples, the estimators perform well even in small samples, supporting their use in practical applications with limited data.
• The method enables counterfactual analysis: by combining two estimators, one can infer ρXY|A from an A′-sample, allowing exploration of hypothetical dependence regimes.
• In a financial dataset, the study shows that the increased dependence (contagion) during the 2007–2011 crisis cannot be fully explained by increased market volatility alone, implying additional structural or systemic factors.
• Conditioning on high volatility (Z > zc) significantly alters the eigenvectors of the covariance matrix, indicating that the underlying dependence structure is non-trivially reconfigured during crises.