Collider Bias & Selection Bias

Sometimes two traits appear linked due to a form of selection bias called collider bias. This illusion happens when we only look at a group selected for a common outcome, making two unrelated causes seem correlated. For example, among famous actors, talent and good looks might seem negatively correlated because an abundance of either can lead to fame. We don't see the aspiring actors with neither. Try this: When you see a surprising trade-off, ask what selection process might be shaping the group you're looking at.

Selection bias happens when the group we are observing isn't representative of the larger population. Collider bias is a specific and tricky version of this.

Imagine two separate causes, like 'artistic talent' and 'personal turmoil,' both lead to a common outcome, like 'becoming a famous artist written about in biographies.' This common outcome is the 'collider' because the two causal paths collide into it.

When we select our sample by looking only at the collider (e.g., reading biographies of famous artists), we can create a spurious correlation. In this selected group, it may look like talent and turmoil are negatively correlated. An artist might achieve fame through immense talent despite a stable life, or through a fascinatingly troubled life despite modest talent. To be included in the 'famous artists with biographies' group, one often needs a high level of at least one of these attributes. Those low on both are rarely written about.

This pattern appears in many domains. In business, if a company only hires people with either a top-tier degree or amazing work experience, looking at their employees might suggest a negative correlation between the two. You don't see the candidates who had neither and were not hired.

Similarly, studies on hospitalized patients can be misleading. If both a behavior (like smoking) and a virus (like COVID-19) increase the chances of hospitalization (the collider), studying only hospitalized patients can create a false association between smoking and the virus's severity.

Being aware of collider bias helps you question apparent trade-offs. The world is full of highly-selected groups—award winners, elite athletes, startup founders, even your social media feed. The patterns you see within them may not hold true for the world at large.

Selection bias is a primary threat to the validity of findings in observational studies, particularly in epidemiology and the social sciences. The concept of a 'collider' was formally defined and popularized by Judea Pearl using a graphical framework called Directed Acyclic Graphs (DAGs), which provides clear rules for when conditioning on a variable will introduce bias.

• Cole et al. (2010) illustrate how conditioning on birthweight (a collider) can make maternal smoking appear protective among low-birthweight infants. [1]
• Elwert & Winship (2014) detail how conditioning on a collider is a key 'endogenous selection problem' in sociology that can distort research on topics like social inequality and educational returns. [2]
• Judea Pearl's work, especially in The Book of Why (2018), demonstrates how DAGs make it possible to visually identify colliders and determine whether a statistical adjustment for a variable will reduce bias or create it. [3]

While the concept of a collider is clear, identifying all relevant causal paths and potential colliders in a real-world system is extremely difficult. The causal structure of a problem is often unknown or debated, making it hard to be certain you've avoided the bias. Furthermore, sometimes conditioning on a collider is unavoidable (e.g., studying a disease that is only diagnosed in hospitals). In these cases, researchers must use advanced statistical methods to attempt to correct for the bias, but these methods rely on their own set of strong, untestable assumptions.