Correlation vs Causation

Two things can move together without one causing the other. A correlation tells you that knowing one variable helps you predict the other. It doesn’t tell you why. Before treating a correlation as cause and effect, rule out the other explanations.

Four reasons two things can be correlated

1. A third factor drives both. Ice cream sales and drownings rise and fall together. Ice cream doesn’t cause drowning. Hot weather sends more people to buy ice cream and more people into the water. The hidden factor is called a confounder, and it is the most common reason correlations mislead.

2. The cause runs the other way. Cities with more police officers tend to have more crime. More police don’t cause crime. Cities with more crime hire more police.

3. Chance. Search through enough pairs of variables and some will line up by accident. Over a decade, US cheese consumption tracked the number of people who died tangled in their bedsheets remarkably closely. With few data points and many comparisons, impressive-looking correlations are easy to find.

4. A real causal link. Smoking and lung cancer are correlated because smoking causes lung cancer. This is sometimes the answer, but you need more than the correlation to show it.

A real example: chocolate and Nobel prizes

In 2012 the New England Journal of Medicine published a short note showing that countries which eat more chocolate win more Nobel prizes per head, with a correlation of about 0.79. It was written partly tongue in cheek. Wealth is the obvious confounder: rich countries buy more chocolate and spend more on universities and research. Nobody thinks eating chocolate makes a country smarter.

How causation is established

  • Randomised experiments. If people are assigned to treatment or control at random, confounders are shared evenly between the groups on average. A difference in outcomes can then be put down to the treatment. This is the strongest evidence.
  • When an experiment isn’t possible (you can’t randomly assign people to smoke), researchers look for several kinds of evidence together: the cause comes before the effect, more exposure gives a stronger effect, the link holds across different populations and studies, there is a plausible mechanism, and known confounders have been measured and adjusted for.

Questions to ask about any correlation

  1. Could something else cause both?
  2. Could the effect be causing the “cause”?
  3. How many relationships did the researchers look at before finding this one?
  4. Was anything randomly assigned?
  5. How strong is the correlation, and how many data points is it based on?

On the last question: r = 0.3 is statistically significant with 50 pairs, but x then explains only 9% of the variation in y. A significant correlation can still be a weak one. The correlation coefficient calculator reports r, r², the p-value and a confidence interval, and the critical value table shows how strong a correlation must be to count as significant for your sample size.

Check yourself

Students who eat breakfast get better grades. Name one confounder.

Family circumstances are a likely one. Households with more time and money may be more likely to provide breakfast and to support schoolwork in other ways. Sleep habits are another.

People who use a meditation app report less stress. Give a reverse-causation explanation.

People who are already less stressed may be more likely to stick with a meditation app, while very stressed people give up on it.