The Fiction of the Isolated Subject
A standard randomized experiment assumes that treating one person doesn't affect anyone else. In the real world, this is rarely true. People talk, markets react, and diseases spread. This phenomenon, called **interference** or **spillover**, can cause us to drastically misinterpret an experiment's results.
Standard Assumption: No Interference
The outcome for each person depends only on their own treatment status.
Real World: Interference
The treatment given to one person "spills over" and affects their neighbors.
Note that Outcome C is different from Outcome B because of the spillover from the treated unit.
Experimental Designs for Measuring Spillovers
To properly measure spillover, we can't just randomize individuals. We need experimental designs that explicitly manipulate treatment at a group level. Explore the toolkit below.
Cluster-Randomized Trial
Randomize entire groups (e.g., villages, schools) to 100% treatment or 0% treatment. This is the simplest way to account for interference.
What it Measures: The **total effect** of the program (direct + spillover).
Limitation: Cannot separate the direct effect from the spillover effect.
Saturation Design (Two-Stage Randomization)
First, select clusters. Second, within each cluster, randomize the *proportion* of individuals who get the treatment (e.g., 0%, 33%, 66%, 100%).
What it Measures: This powerful design can disentangle the **direct effect**, the **spillover effect**, and the **total effect**.
Limitation: More complex and costly to implement.
Network-Based Randomization
If you have data on social networks (e.g., who is friends with whom), you can randomize treatment in a way that explicitly varies the number of treated peers each person has.
What it Measures: A granular estimate of spillover effects based on network distance (e.g., the effect of a treated friend vs. a friend-of-a-friend).
Limitation: Requires detailed, often expensive, pre-existing network data.
Interactive Simulation: A Deworming Program
Let's simulate a public health study inspired by the classic deworming experiments. We will use a saturation design to see how treating a certain percentage of children in a school affects the health of everyone. Click the buttons to change the treatment saturation level and observe the effects.
At 0% saturation, this is the baseline health score. No one is treated, so there are no effects.
Why Getting Spillovers Right Matters
Measuring spillover isn't just an academic exercise. It has profound implications for science and policy.
Estimating True Program Impact
Ignoring positive spillovers (like in the deworming case) leads to **underestimating** a program's total benefit. Ignoring negative spillovers (e.g., market competition from a job training program) leads to **overestimating** its benefits. In either case, policymakers get the wrong answer.
Predicting Scalability
The results from a small-scale pilot study may not hold when a program is scaled up. A saturation design helps predict these "general equilibrium" effects, providing a much more accurate forecast of how a program will perform when implemented universally.