What is interference, or spillover, in causal inference?
Interference is when one unit’s treatment changes another unit’s outcome, and on farms it happens all the time: pests, pollinators, spray drift, water and disease don’t stop at the hedge. Bt maize cut European corn borer numbers across whole regions, so growers of ordinary maize benefited too, and comparing Bt with non-Bt fields there would underestimate Bt’s effect.
Last time's quick check: only units just either side of the line are comparable. Further away, they differ in other ways too. Almost every method so far makes a quiet assumption: one farm's treatment doesn't change another farm's outcome. When that assumption breaks, statisticians call it interference. On farms, it breaks all the time.
Pests, pollinators, spray drift, water and disease don't stop at the hedge. In the US Corn Belt, farmers who planted Bt maize, bred to resist the European corn borer, didn't only protect their own fields. Across whole regions, borer numbers fell, and farmers growing ordinary maize benefited too. A 2010 study estimated that a large share of the economic benefit went to growers who never planted Bt maize.
Now imagine comparing Bt and non-Bt fields in that region. You'd underestimate Bt's effect, because the comparison fields were partly treated too. A benefit that leaks across a boundary is called a spillover. Spillovers aren't only biological. In New Zealand, a study asked whether joining a farmer group made dairy farmers more likely to adopt good nutrient management. Farmers talk to their neighbours, so the researchers built neighbours' choices into a spatial version of propensity-score matching.
Group members had 1.81 times the odds of adopting nutrient management compared with similar non-members. Odds, not probabilities, and unmeasured differences between farmers could still play a part. But leaving out the neighbours would have told the wrong story. So what do you do? First, draw it on the map: an arrow from one farm's treatment to its neighbour's outcome.
Then design around it. Randomise whole areas instead of single fields. Leave buffer strips between treated and untreated plots. Or measure the spillover directly, because sometimes it's the most valuable part of the answer. Next: two famous cases where observational data pointed the wrong way, and how to stop it happening again.
First draw the spillover on the map, as an arrow from one farm’s treatment to its neighbour’s outcome. Then randomise whole areas instead of single fields, leave buffer strips between treated and untreated plots, or measure the spillover directly, because sometimes it’s the most valuable part of the answer.
No. In New Zealand, a study asked whether joining a farmer group made dairy farmers more likely to adopt good nutrient management. Because farmers talk to their neighbours, the researchers built neighbours’ choices into a spatial version of propensity-score matching. Leaving the neighbours out would have told the wrong story.
Causal evidence is changing how farming is measured and rewarded. Leave your email and we will tell you when new films, courses and findings are published.
Prefer to read online? Vitagri also writes on Substack. See our privacy policy for how we use your email.