It’s quite easy to understand what we mean if we say an event causes something to happen. For example, drinking too much alcohol and then driving a car can cause road traffic accidents. That is easy to understand because we have everyday knowledge of the effects of alcohol on driving. When we are doing research, however, we are starting from a position of limited knowledge about the subject we are studying (see the The Scientific Method). So how can we go about discovering if something causes something else. One way is to look for a correlation, in other words determining whether two things change in step with one another.

A famous example is the debate around global warming and carbon dioxide emissions. The correlation between global carbon dioxide levels and global warming can be seen here http://www.ncdc.noaa.gov/paleo/globalwarming/temperature-change.html

An example from everyday life.

Suppose we were interested in whether the increased use of computer games is a cause of obesity, because playing games cause people to lead less active life styles. We could look to see if there is data available for computer games sales for the past twenty years and the proportion of people who are obese. If we saw that games sales and obesity increased around the same time, could say we have identified a correlation.

But that wouldn’t necessarily mean that computer games cause obesity. There are several reasons why this correlation might exist even if playing computer games doesn’t cause obesity.

1. The correlation might have occurred by pure chance. Sound unlikely? Suppose you notice that the amount of junk mail delivered varies from week to week. Now suppose you measure as many other weekly events as you can: how many times you saw a cat; how many times it rained; how many times you had to stop at a red light while driving; how many days you had toast for breakfast and so on. If you look at enough of these events, you are almost certain to find one that has the same pattern as the changes in junk mail delivery. In other words, if you look at enough things, correlations inevitably exist by pure chance. Look at this website http://www.tylervigen.com/spurious-correlations it lists hundreds of crazy correlations that have occurred by chance, but one event does not cause the other.
2. The correlations might occur because of “confounding factors”. Here is an example to explain that. It has been suggested that there is a correlation between sharks attacking humans and ice cream consumption. No-one is suggesting that the sharks are chasing people who are swimming while eating ice cream. However, people tend to eat more ice cream in the summer and it’s in the summer that people tend to go swimming in the sea increasing the chance of shark attacks (compared to being on dry land). In other words the hot weather “creates” a correlation between ice cream consumption and shark attacks even though one does not cause the other.
3. It’s difficult to infer the order of events from a correlation. Going back to our obesity and computer games example, we initially proposed that playing computer games leads to obesity. But isn’t the opposite also possible – that obesity, brought on by other factors, lead to people following less active hobbies such as playing computer games?
4. Some correlations are “inevitable”. There is a striking correlation between marriage and divorce. All people who get divorced have been married!

To summarize this, here is a diagram which shows potential explanations why a correlation might be observed between two events, labelled A and B.

So hopefully you can see that a correlation is only a starting point for investigating something by designing an appropriate experiment.

So how could we test if computer games cause obesity? We could encourage people to play computer games every day for several months, and see if they become obese. Then we could take away their computer games and see if they lose weight. This is called a prospective trial – because we start with the hypothesis that computer games cause obesity and directly test what happens in the future when we carry out a specific intervention (eg making people play with their computers). This is much more powerful than looking at historical data, because we can repeat it as often as we like to rule out the correlation occurring by chance or because of confounding factors.

Now if you really want to think about this, remember one behaviour could reinforce the other. Playing computer games could promote obesity, which in turn leads to playing more computer games, more obesity and so on. Biological scientists would call this positive feedback.