When we observe a strong correlation between two variables—such as a firm's digital maturity and its innovation output—our natural inclination is to draw a straight causal arrow from cause to effect. However, in complex economic, corporate, and social systems, causality is rarely a one-way street. Instead, variables frequently exist in a state of mutual influence, creating a "chicken-and-egg" dilemma known as bidirectional causality.
This circularity introduces endogeneity, a fundamental econometric challenge where an explanatory variable is correlated with the model's error term. When endogeneity is left unaddressed, traditional statistical conclusions collapse, transforming apparent causal relationships into mere descriptive correlations.
The Three Archetypes of Circular Causality
To build intuition for why endogeneity occurs, we can categorize these causal challenges into three distinct archetypes that recur across empirical literature.
1. The Resource Loop (Affordability vs. Causation)
In corporate governance and strategy, virtuous organizational traits consistently correlate with high performance. But does the trait cause the performance, or does the performance fund the trait?
For example, Yuan Yuan and Jingze Li note in Research on the Promotion of High-Quality Development of Manufacturing Enterprises in Tai'an City by Digital Productivity that while digital integration drives industrial upgrading, high-performing firms inherently possess more surplus capital to invest in expensive digital infrastructure.
We see this exact loop in the evaluation of Environmental, Social, and Governance (ESG) metrics. As Ye Wang argues in Analyzing the Influence of Corporate Environmental, Social, and Governance Ratings on Corporate Performance, it is difficult to isolate whether superior ESG ratings drive financial performance or if highly profitable firms simply have the luxury to allocate capital toward sustainability initiatives. This pattern repeats when researchers attempt to link firm performance to board gender diversity (Jiabo Fan et al.) or sophisticated human resource management (Peer-Olaf Siebers et al.): successful organizations are simply better positioned to afford and maintain progressive corporate structures.
2. The Selection and Targeting Trap
In macroeconomics and public policy, resources and interventions are rarely distributed at random; they are deliberately targeted based on existing conditions. This targeting destroys our ability to easily see cause and effect.
Consider law enforcement deployment. Douglas Newball-Ramírez et al. point out that police presence is naturally concentrated in high-crime neighborhoods. A naive regression might show a positive correlation between police numbers and crime rates, falsely implying that law enforcement causes illegal activity, when in reality, the high-activity environment drew the resources to itself.
A parallel issue distorts international development data. Sim Yee Lau and Pui Mun Yin observe that the relationship between foreign aid and economic growth is deeply obscured because low-growth environments naturally attract higher volumes of international aid, confounding efforts to measure the true catalytic effect of that capital.
3. The Shared Underlying Driver (Omitted Variable Bias)
Sometimes, two variables do not influence each other at all. Instead, they appear correlated because both are downstream products of a hidden, unobserved third variable.
In Study on the effect of ESG ratings of companies on financial restatement, Chenchen Yuan and Jinfeng Chen explain that a strong correlation exists between high ESG scores and a reduction in fraudulent financial restatements. However, they caution that an inherently ethical, transparent corporate culture will naturally prioritize both accurate accounting and environmental responsibility. Without a way to measure and control for this latent "culture," attributing the improved accounting quality directly to ESG mandates is an unverified leap.
Why Traditional Models Fail
To understand why bidirectional causality invalidates standard statistical tools like Ordinary Least Squares (OLS) regressions, we have to look at how these models handle error.
An OLS model assumes a clean, one-way flow of information:
$$\text{Outcome } (Y) = \beta \cdot \text{Cause } (X) + \epsilon \text{ (Unobserved Factors)}$$
For this model to calculate an unbiased weight ($\beta$), it relies on a strict mathematical assumption: the cause ($X$) must be completely independent of the unobserved factors ($\epsilon$).
When causality is bidirectional, this assumption breaks down through a process called simultaneity bias. If $X$ causes $Y$, but $Y$ also causes $X$, then any random shock to $Y$ (which lives inside $\epsilon$) immediately feeds back into $X$. Because $X$ and $\epsilon$ are now moving together, the regression model can no longer distinguish between the variance caused by $X$ and the variance caused by the background noise. The model will typically overestimate the strength of the relationship, and in severe cases, can even flip the mathematical sign of the true effect.
One-Way Causality (Valid OLS):
[Cause X] ----------> [Outcome Y (+ Error)]
Bidirectional Causality (Endogeneity Loop - Invalid OLS):
[Cause X] ----------> [Outcome Y]
^ |
| v
+-- (Feedback Loop) ---+ <--- Shocks to Y corrupt X
Chihiro Shimizu demonstrates this breakdown in urban economics within Do Urban Amenities or Walkability Drive Housing Rent?. High-income residents attract premium neighborhood amenities (like cafes and parks), but those premium amenities simultaneously drive up property values. A standard regression cannot untangle this loop, leaving policymakers unable to determine whether investing in neighborhood walkability will actively raise property values or if it is simply tracking pre-existing demographic wealth.
Untangling the Knot: Methodological Solutions
To break the endogeneity loop, researchers must isolate variation in the explanatory variable ($X$) that is completely untainted by the outcome ($Y$) or any unobserved background factors.
Natural Experiments and Exogenous Shocks
The cleanest way to isolate causality is to find an external event—a "shock"—that forces a change in the cause ($X$) without being influenced by the system itself.
Jian Huang and Han Yu utilized this strategy to study how shareholder activism impacts corporate risk-taking. Because activists selectively target volatile firms (a classic selection bias), a direct correlation is misleading. To bypass this, the authors used the passage of the Sarbanes-Oxley Act (SOX) as a regulatory natural experiment. Because SOX was an external mandate imposed uniformly by the government, it created an exogenous shift in corporate governance, allowing the authors to cleanly isolate how regulatory pressure alters management behavior without the interference of corporate feedback loops.
Instrumental Variables
When natural experiments are unavailable, researchers look for an Instrumental Variable (IV). An instrument is a third variable ($Z$) that satisfies two strict conditions:
- It is strongly correlated with the cause ($X$).
- It has no direct relationship with the outcome ($Y$), affecting it only through its influence on $X$.
Instrumental Variable Architecture:
[Instrument Z] ---> [Cause X] ---> [Outcome Y]
| ^
+------ (No Direct Path) ---------+
By using only the portion of $X$'s variance that is driven by the instrument $Z$, researchers can effectively strip away the endogenous feedback loop. For instance, Yong Duan and Binkai Chen deploy instrumental variables to untangle the complex, bidirectional relationship between geographical distance and the structural depth of global value chains, ensuring that the measured trade sensitivities are truly causal.
Dynamic Panel Data and GMM Modeling
In longitudinal studies where variables interact continuously over time, researchers often use the System Generalized Method of Moments (GMM).
Instead of looking for an external instrument, GMM uses the system's own history. It transforms the data by taking mathematical differences across time periods and uses lagged historical values of the endogenous variables (e.g., a firm's digital spending from two years ago) as instruments for current changes. Because the past cannot be caused by the present, these historical lags provide a clean, exogenous baseline.
Yining Liu and Rudan Zhao successfully deployed a system GMM model to study how agricultural insurance shapes industrial structures in China. Because insurance demand and industrial evolution continuously alter one another over time, using historical lags allowed the authors to isolate the structural impacts of insurance depth while neutralizing the simultaneous feedback loops inherent in macroeconomic development.
Key Foundational Work
The direction of causality between financial development and economic growth (2003) — contributed 3 concept elements to this direction.