Correlation does not imply causality is one of the central tenets of science. But most people don’t think like scientists (some interesting research has shown that even scientists usually don’t think like scientists when they’re not doing science). That makes it easy to pull the wool over people’s eyes.
The purpose of this article is to help you appreciate that correlation and causality often get mixed up (either deliberately or mistakenly), and to use that awareness to defend yourself against the wool-pullers—including yourself.
When you examine the marketing claims of providers of services such as driver training you’ll often come across phrases like, “Here’s proof that what we do works.” You’re then presented with a “case study” (which usually bears no resemblance to a proper, scientific, randomised control study) which implies that intervention A produced result B. In other words, A caused B. But did it?
Is causality real or implied?
Even with carefully-designed and methodically-implemented scientific studies it’s often very difficult to establish causality, although there may be statistically-significant correlations. And yet in the world of fleet safety (where proper scientific evaluation is only a little more common than hens’ teeth) service providers may offer advice or sell services on the basis of implied causality.
It’s important to appreciate that implied causality is not just a sneaky device employed by people who are trying to sell you something. It’s just as common to find their customers doing it. If you had just spent a big chunk of the company’s money on purchasing a “solution” to a problem, it’s highly likely that you would attribute any improvement in the situation to your wonderfully-wise purchasing decision.
This is what psychologists call confirmation bias: we tend to notice those things that confirm our decision more than those that disconfirm it. And we’re all much more prone to it than we realise.
The only way to guard against it is to be very methodical and rational in examining all the things that may have changed on the input side of a process before drawing any conclusions about what may have caused the output.
In which direction does causality act?
Only when there is one single change on the input side is it reasonable to assume that it may have caused the change in output (this, of course, is what you aim to do in a scientific experiment). And even then you have to be careful not to confuse the cart with the horse; there may be causality but can you be sure of its direction? Is there a clear-cut input and output?
As the famous psychologist William James put it: “You do not run from a bear because you are afraid of it, but rather become afraid of the bear because you run from it.” In many cases causality can operate in both directions, in a feedback loop.
What did what?
When a number of changes occur simultaneously, as is often the case in the real world, it’s difficult if not impossible to determine the relative influence of each one. So, if we take the example of a business that introduces a driver training programme and then sees an improvement in fleet safety, there will have been many factors that could have influenced this improvement but which are easily overlooked.
Prior to the delivery of any training there would have been communications within the business that basically conveyed the message: “We’re now taking driving seriously.” The safety record of the fleet and the performance of drivers may have found their way onto management agenda for the first time. Driving performance may have been introduced into employee reviews. And so on and so forth.
It’s easy to assume that it was the driver training and the driver training alone that produced the improvement because it was the most obvious and visible change, and it cost more than anything else (and so gives rise to strong confirmation bias).
Is it just coincidence?
Sometimes factors correlate but their correlation may only be coincidental with no causality at all. However, if you present the correlation as though it were causality in a serious-looking argument you may get away with it.
For example, by adopting the manner of a social historian I could perhaps persuade you that rising hemlines in women’s clothing is a precursor to economic collapse. Shorter dresses in the Roaring Twenties preceded the Great Depression of the 1930s. It happened again when the mini-skirts of the Swinging Sixties led to the strikes, fuel shortages and power cuts of the 1970s recession.
I’ll bet you never realised it was all Mary Quant’s and Twiggy’s fault.
…Or is it something else?
Now before you blurt out, “Yes, but…” I’ll get in first. I’ve just created a flippant, and unsubstantiated, causal relationship. A proper social historian would be able to construct a solid argument to show that the rising hemline was a manifestation of growing hedonism and risk-taking combined with a reduction in conservatism, including fiscal conservatism, which ultimately led to an economic collapse or reversal.
And that’s another common error of causality: it’s not that A (hemlines) caused B (economic recession) but that C (all that other stuff) caused both A and B. Whenever you see a change in output, before you attribute it to an “obvious” change in input always ask whether there could be another (perhaps hidden) factor driving both the observed changes.
Let’s have a road safety example of that. You might review the collision statistics for your user-chooser car fleet and find that two-door coupes have the highest crash rate—a commonly-found correlation. You could easily infer that driving this type of car causes more crashes and so decide to remove all two-door coupes from choice lists. But you find no subsequent reduction in the fleet’s overall crash rate.
The actual causality in this case lies with the drivers who were attracted to the two-door coupes (drivers who are more likely to be young and single or, at least, without kids). Deprived of their preferred choice of car they drive something else but their driver behaviour remains unchanged.
Be on your guard…
So the messages of this article are: be watchful for dodgy, implied causality and beware your own tendency towards confirmation bias. You’re much less likely to be persuaded by “proof” that a particular product or service will produce a good return on investment if you always remember…
Correlation does not imply causality