Look at this graph. It’s incredible! As iPhone sales increase from 2007 to 2010, so did the number of deaths caused by people falling down the stairs. They even increased at the same rate. It’s crazy – iPhones cause people to die falling down the stairs!!
For this and other hilarious graphs of silly correlations see tylervigen.com
How obviously ridiculous this is. Most people could logically deduce that just because iPhone sales and deaths from falling down the stairs “correlate” that one does not “cause” the other. There are lots of other examples (and you can create your own silly correlations) from Spurious Correlations or on other news sites here or here. And in these cases, you can laugh realizing that just because these two things happen together does not mean that one causes the other.
But even though you’ve probably said that “correlation doesn’t imply causation” or heard someone say it, in science (and science news reporting) the difference is critical to tease out. Why? Understanding if something in science is a cause of the effect you’re seeing can
- Prevent something harmful from causing damage. For example knowing that smoking is a cause of lung cancer resulted in public health efforts to help people quit smoking.
- Treat a the cause of a disease or fix the cause of the problem. For example, knowing that the h.pylori bacteria causes gastritis and ulcers provides a method for treating ulcers by killing the h.pylori. Or if you know that factory waste being dumped into a river or lake causes animal life to die or stop procreating, you can work towards stopping the dumping to save wildlife.
- Prepare for diseases, outbreaks of disease, or natural disasters. If you know that earthquakes cause tsunamis, then a warning system can be developed to save people in the tsunami zone.
- Plan ahead and discuss the possible outcomes from a particular action. When you know that lack of water in a drought causes dry forest conditions leading to forest fires, you can plan to have greater funding available for fighting these fires in a particularly dry year.
“Cancer smoking lung cancer correlation from NIH” by Sakurambo – Vectorized version of Image:Cancer smoking lung cancer correlation from NIH.png, originally published on the nih.gov website. The source page has been deleted, but an archived copy is still accessible.Own work, created in Adobe Illustrator. Licensed under Public Domain via Commons
It’s just as important though to tease out when something doesn’t cause an effect – and unfortunately many false claims and pseudoscience is based on taking correlations and touting them as causes. So how do we figure this out?
In science, a lot of this is determined experimentally and statistically. In statistics (of which I do not claim to be an expert, but see the links below for more details), the strength of the relationship can be calculated and the stronger the more likely that one causes the other. The cause/effect relationship should also be tested experimentally, if possible, and the experiment should be repeated to see if the same results are obtained every time. Without experimental or repeatable experimental results, the relationship is less likely to be causal. Another interesting measurement is to look at the time frame – if the action takes place months, days, or years apart from the effect, you have to consider whether this would make sense or not. In the case of smoking and lung cancer, the separation of the two events by years makes sense, but in other cases it may not. Which also brings up the point of looking at the relationship and thinking about whether or not it makes sense or if a mechanism can be found for the cause and effect relationship. For example, we know that smoking causes DNA mutations and inflammation which is one of the mechanisms that leads to lung cancer. Alternatively, looking at the iPhone and dying from falling down the stairs example, it’s difficult to find a mechanism that could explain this relationship.
A great description of what to look for comes from the book club book that I’ll be talking about on Thursday, “Bad Science: Quacks, Hacks, and Big Pharma Flacks” by Ben Goldacre with a quote describing evidence-based medicine:
it needs to be a strong association, which is consistent, and specific to the thing you are studying, where the putative cause comes before the supposed effect in time; ideally there should be a biological gradient, such as a dose-response effect; it should be consistent or at least not completely at odds with what is already known (because extraordinary claims require extraordinary evidence); and it should be biologically plausible
Overall, it does come down to the data and some common sense. If there isn’t any data to support the relationship, you might just be looking at correlation and can confidently holler “iPhones do not kill people!”
To better understand the differences between correlation and causation and the math that can show which you are looking at, check out the Kahn Academy course. Read more about this topic from Stats with Cats Blog
For more Sci Snippets, click here.