Abstract
In causal inference, the effect of an exposure $X$ on an outcome $Y$ can be distorted by conditioning on a common effect, known as a collider. While qualitative descriptions of collider bias are common, analytical solutions for binary variables—prevalent in social and epidemiological research—have been lacking. This paper provides a comprehensive analytical framework for calculating the exact magnitude and direction of collider bias in binary systems. By deriving precise formulas for various causal structures (V, Y, and M-structures), the authors identify the conditions that determine the sign of bias on covariance, risk difference, and odds ratio scales.
Introduction
Causal inference often focuses on confounding, where a common cause influences both treatment and outcome. However, collider bias presents a distinct and often more insidious challenge: it occurs when researchers condition on a "collider"—a variable that is a common effect of two other variables.
Imagine two independent traits, "Beauty" and "Talent," both of which help someone become a "Celebrity." If you only look at celebrities (conditioning on the collider), you might find a spurious negative correlation: if a celebrity isn't talented, they must be beautiful, and vice versa. This is the "explaining away" effect. While previous research often assumed Gaussian distributions, binary variables behave with greater complexity. This paper addresses this gap, providing rigorous mathematical foundations for binary collider bias.
Causal Structures and Methodology
The authors utilize Causal Directed Acyclic Graphs (DAGs) to categorize scenarios where collider bias arises. Three primary structures are analyzed:
- V-structure: $X$ and $Y$ directly influence a common collider $C$.
- Y-structure: $X$ and $Y$ influence $C$, and conditioning occurs on $C$’s child, $D$.
- M-structure: $X$ and $Y$ are influenced by independent causes ($A$ and $B$) that also influence a common collider $C$.
The research derives exact formulas for bias across three association measures: Covariance, Risk Difference (RD), and Odds Ratio (OR). Additionally, the authors analyze bias resulting from linear regression adjustment, demonstrating that it functions as a variance-weighted average of strata-specific biases.
Key Insights
The Interaction Rule
The sign and magnitude of bias in a V-structure are determined by how $X$ and $Y$ interact to influence the collider $C$. Specifically, if $X$ and $Y$ do not interact on the Risk Ratio scale, the bias on the Odds Ratio scale is zero.
The "Same-Sign" Heuristic
A general rule of thumb emerges: if $X$ and $Y$ both have positive effects on the collider, the induced bias is typically negative. This formalizes the "explaining away" phenomenon mathematically.
Structural Extensions
For complex structures like the M-bias, the total bias can be decomposed. The paper proves that the sign of the bias in an M-structure is the product of the "embedded" V-bias and the signs of the paths connecting the variables to the collider.
Key Results
The paper establishes several key theorems that allow researchers to move beyond qualitative warnings:
- Theorem 1 (V-bias): Provides fundamental formulas. On the covariance scale, bias is proportional to $(P_{c|00}P_{c|11} - P_{c|10}P_{c|01})$.
- Scale Invariance: The direction (sign) of bias is often consistent across RD, OR, and Covariance, provided the parents of the collider are marginally independent.
- Regression vs. Stratification: Bias from linear regression adjustment generally shares the same sign as bias from conditioning on a specific level of the collider, though the magnitude differs due to variance weighting.
Conclusion
This work provides a template for identifying spurious associations in binary data. By moving beyond qualitative heuristics, researchers can now anticipate the direction and potential magnitude of bias in common epidemiological and social science scenarios. Future work may extend these results to ordinal variables and scenarios where collider bias and confounding occur simultaneously.
Reference: Nguyen, T. Q., Dafoe, A., & Ogburn, E. L. (2019). The Magnitude and Direction of Collider Bias for Binary Variables. Epidemiologic Methods.
Summary updated