Construction workers in high-rise buildings face greater risk of death on the job than office workers do. By comparing wages in risky and less risky occupations, controlling for, education, experience, and other determinants of wages, economists can get some sense about what value people put on their own lives. Studies using this approach conclude that the value of a human life is about $10 million.

What does controlling for mean?

3 Answers 3


When you are using statistics to understand a relationship between two things, you have to be careful that a third thing doesn't cause a problem.

Suppose you want to see how human height is caused racial genetics. You measure the height of 100 randomly selected white people from around the world, and 100 randomly selected black people from around the world. You find that white people are taller on average. But then you "control for" diet by comparing the white people who ate very little with the black people who ate very little, and comparing the white people who ate a lot with the black people who ate a lot, and you discover that race has very little influence on height. It turns out most white people are from rich countries where people eat a lot and most black people are from Africa where people don't have as much food.

When doing statistical studies, you have to control for things that might mess up your results.


In your example

controlling for

means that a particular variable is kept the same or nearly the same when comparing other variables. For example, if you were to look at risky occupations in general, there will be some people with high levels of education and some with low. When looked at separately, those with low education may show different results than those with high education. By controlling for education, the level of education is taking out of the equation, for instance in a multivariable regression.


It depends on the context. In a broader sense, it also means data binning a variable you suspect is affecting the relationship between the variables you are interested in and include it into your hypothesized model. If for each bin of the controlled variable you observe the similar effect between the variables you’re interested in, then you can conclude that the controlled variable is not a confounding factor. Otherwise, it is. Theory from Pearl’s Structural Causal Modeling formalizes the possible scenarios of relationship between variables and a systematic method for knowing where spurious correlation affects the model regardless the data collected used to estimate it.

  • Categorising essentially continuous variables is not necessary (as another answer has already pointed out).
    – mdewey
    May 29, 2021 at 16:02
  • I agree, but as I said, "it also means"...which means "not necessary" May 30, 2021 at 17:14

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .