FoMind
Suppose a survey finds that kids who attend after-school tutoring average 85 points, while kids who don't average 75. The conclusion seems obvious: tutoring adds 10 points.
Not so fast. Which families send their kids to tutoring? Usually families with higher incomes and parents who care a lot about schoolwork. Their kids might have more books, a quieter place to study, and more support, tutoring or not.
What is a confounder?
A confounder is a third factor that affects both the "cause" and the "effect." It makes two things look causally linked when some or all of the link actually comes from the third factor.
In the tutoring example, family background is the confounder: it affects both whether a kid attends tutoring and how well they do in school.
How to strip it out
- Compare like with like: only compare kids from similar families. If, within the same income bracket, tutored kids score just 2 points higher, the real effect of tutoring is closer to 2 points than 10.
- Randomized experiments: randomly decide who gets tutoring so family background is spread evenly across both groups. This is the most reliable method, but you rarely get one in daily life.
- Ask "who chooses X?": whenever people choose to do X themselves, suspect that they were different from everyone else to begin with.
Using this in forecasts
When forecasting the effect of a policy, product, or method, confounders are the easiest trap to fall into. "Housing prices rise faster in districts that get a new subway line" might simply reflect that subway lines are built first in areas that were already set to grow.
When writing your reasoning, try to list one or two possible confounders and ask: once you take them out, how much weight is left in your argument?
When you see "people who did X did better," first ask: were they different to begin with?