This is “Correlation and Causality”, section 17.23 from the book Theory and Applications of Microeconomics (v. 1.0). For details on it (including licensing), click here.

Has this book helped you? Consider passing it on:
Creative Commons supports free culture from music to education. Their licenses helped make this book available to you.
DonorsChoose.org helps people like you help teachers fund their classroom projects, from art supplies to books to calculators.

17.23 Correlation and Causality

Correlation is a statistical measure describing how two variables move together. In contrast, causality (or causation) goes deeper into the relationship between two variables by looking for cause and effect.

Correlation is a statistical property that summarizes the way in which two variables move either over time or across people (firms, governments, etc.). The concept of correlation is quite natural to us, as we often take note of how two variables interrelate. If you think back to high school, you probably have a sense of how your classmates did in terms of two measures of performance: grade point average (GPA) and results on a standardized college entrance exam (SAT). It is likely that classmates with high GPAs also had high scores on the SAT exam. In this instance, we would say that the GPA and SAT scores were positively correlated: looking across your classmates, when a person’s GPA score is higher than average, that person’s SAT score is likely to be higher than average as well.

As another example, consider the relationship between a household’s income and its expenditures on housing. If you conducted a survey across households, it is likely that you would find that richer households spend more on most goods and services, including housing. In this case, we would conclude that income and expenditures on housing are positively correlated.

When economists look at data for the whole economy, they often focus on a measure of how much is produced, which we call real gross domestic product (GDP), and the fraction of workers without jobs, called the unemployment rate. Over long periods of time, when GDP is above average (so that the economy is doing well), the unemployment rate is below average. In this case, GDP and the unemployment rate are negatively correlated, as they tend to move in opposite directions.

The fact that one variable is correlated with another does not inform us about whether one variable causes the other. Imagine yourself on an airplane in a relaxed mood, reading or listening to music. Suddenly, the pilot comes on the public address system and requests that you buckle your seat belt. Usually, such a request is followed by turbulence. This is a correlation: the announcement by the pilot is positively correlated with air turbulence. The correlation is, of course, not perfect, because sometime you hit some bumps without warning and sometimes the pilot’s announcement is not followed by turbulence.

But—obviously!—this doesn’t mean that we could solve the turbulence problem by turning off the public address system. The pilot’s announcement does not cause the turbulence. The turbulence is there whether the pilot announces it or not. In fact, the causality runs the other way. The turbulence causes the pilot’s announcement.

We noted earlier that real GDP and unemployment are negatively correlated. When real GDP is below average, as it is during a recession, the unemployment rate is typically above average. But what is the causality here? If unemployment caused recessions, we might be tempted to adopt a policy that makes unemployment illegal. For example, the government could fine firms if they lay off workers. This is not a good policy because we don’t think that low unemployment causes high real GDP. But nor do we necessarily think that high real GDP causes low unemployment. Instead, based on economic theory, we think there are other influences that affect both real GDP and unemployment.

More Formally

Suppose that you have N observations of two variables, x and y, where xi and yi are the values of these variables in observation I = 1,2,...,N. The mean of x, $μx$, is the sum over the values of x in the sample divided by N, and likewise for y.

$μx =x1 +x2 +...+xN ; μy =y1 +y2 +...+yN N .$

We can also calculate the variance and standard deviations of x and y. The calculation for the variance of x, denoted $σx2$ is

$σx2 =(x1 −μx )+(x2 −μx )+...(xN −μx ) N$

The standard deviation of x is the square root of $σx2$,

With these ingredients, the correlation of (x,y), denoted corr(x,y), is given by

$corr(x,y)=(x1 −μx )(y1 −μy )+(x2 −μx )(y2 −μy )+...(xN −μx )(yN −μy ) Nσx σy .$