An excellent example of what statistical analysis can and cannot show:
Do mobile phone towers make people more likely to procreate? Could it be possible that mobile phone radiation somehow aids fertilisation, or maybe there’s just something romantic about a mobile phone transmitter mast protruding from the landscape?
These questions are our natural response to learning that variation in the number of mobile phone masts across the country exactly matches variation in the number of live births. For every extra mobile phone mast in an area, there are 17.6 more babies born above the national average.
This was discovered by taking the publicly available data on the number of mobile phone masts in each county across the United Kingdom and then matching it against the live birth data for the same counties. When a regression line is calculated it has a “correlation coefficient” (a measure of how good the match is) of 98.1 out of 100. To be “statistically significant” a pattern in a dataset needs to be less than 5% likely to be found in random data (known as a “p-value”), and the masts-births correlation only has a 0.00003% probability of occurring by chance.
Part of the problem is that our brains have evolved to detect patterns and relationships — even when they’re not really there:
Mobile phone masts, however, have absolutely no bearing on the number of births. There is no causal link between the masts and the births despite the strong correlation. Both the number of mobile phone transmitters and the number of live births are linked to a third, independent factor: the local population size. As the population of an area goes up, so do both the number of mobile phone users and the number people giving birth.
The problem is that our first instinct is to assume that a correlation means that one factor is causing the other. While this does not cause a problem when using pattern-spotting as an evolved survival tool, it does cause severe problems when assessing possible health scares based on a recently uncovered correlation. For the majority of cases, correlation does not indicate the presence of causality.
H/T to Maggie Koerth-Baker for the link.
