Course contentsShow
← Data Science
Lesson 1444 of 2,14536. Causal Inference BasicsFree lesson

Selection Bias and Treatment Assignment

Learn how self-selection into treatment groups creates systematic differences that confound causal estimates.

Selection Bias and Treatment Assignment

You'll learn how self-selection into treatment groups creates systematic differences that confound your ability to estimate true causal effects.

The Core Problem

Selection bias occurs when individuals aren't randomly assigned to treatment and control groups, but instead choose (or are chosen for) their group based on characteristics that also influence the outcome. This creates systematic differences between groups that existed before treatment, making it impossible to tell whether outcome differences are due to the treatment or pre-existing differences.

A Real-World Example

Imagine studying whether a job training program increases earnings. People who voluntarily enroll may already be more motivated, better educated, or have stronger work histories than those who don't. If you compare their future earnings to non-participants, you can't tell if higher earnings came from the training or from those pre-existing traits.

Why It Matters for Causal Inference

In observational studies—where you don't control treatment assignment—selection bias is everywhere. Unlike randomized controlled trials (RCTs), where randomization ensures groups are balanced on all characteristics (measured and unmeasured), observational data reflects real-world selection processes. These processes often correlate with both treatment choice and outcomes, creating confounding that obscures causal relationships.

Selection bias can occur through:

  • Self-selection: individuals choose treatment based on expected benefits
  • Administrative selection: gatekeepers assign treatment based on need or eligibility
  • Survival bias: only certain types complete treatment and remain observable

Key Takeaway: Selection bias arises when treatment assignment depends on characteristics that also affect outcomes, creating pre-existing differences between groups that masquerade as treatment effects. Recognizing this pattern is the first step toward using methods that can adjust for it.