Statistical errors may taint as many as half of mouse studies
Seven years ago, Peter Kind, a neuroscientist at the University of Edinburgh in Scotland, found himself in an uncomfortable situation. He was reading a study about fragile X syndrome, a developmental condition characterized by severe intellectual disability and, often, autism. The paper had appeared in a high-profile journal, and the lead scientist was a reputable researcher — and a friend. So Kind was surprised when he noticed a potentially serious statistical flaw.
The research team had looked at 10 neurons from each of the 16 mice in the experiment, a practice that in itself was unproblematic. But in the statistical analysis, the researchers had analyzed each neuron as if it were an independent sample. That gave them 160 data points to work with, 10 times the number of mice in the experiment.
“The question is, are two neurons in the brain of the same animal truly independent data points? The answer is no,” Kind says. “The problem is that you are increasing your chance of getting a false positive.”
The more times an experiment is replicated, the more likely it is that an observed effect is not just a lucky roll of the dice. That’s why more animals (or people) means more reliable results. But in the fragile X study, the scientists had artificially inflated the number of replications — a practice known as ‘pseudoreplication.’
This practice makes it easier to reach the sweet spot of statistical significance, especially in studies involving small numbers of animals. But treating measurements taken from a single mouse as independent samples goes against a fundamental principle of statistics and can lead scientists to find effects that don’t actually exist.
Read more here at Spectrum.