Saturday, 5 October 2013

Fuck statistics




Eagle rock gospel singers - What Paul did (2013)




A new paper published in PLOS Biology argues that neuroscientific research is compromised due to a significant bias in the reporting of animal studies, thus greatly exaggerating the impact of future therapy platforms in humans. While this is excellent news to anyone who wishes to put an end to animal testing in science, this study still holds several testimonies as to the various maladies inflicting contemporary science. First and foremost is the utter addiction to statistics. The reality in 20th/21st century experimental science is that if it doesn't have numbers than it's meaningless. It is completely irrelevant that statistical methods are changed frequently, even within the same paper, just so they'll deliver the desired result. It's also irrelevant that the statistical tools used are often wrong and do not fit the experimental design. It's all beside the point: put a table rife with numbers and augment it with a line-rich graph and you're done. As a corollary, statistical power is the king: scientists in medical and life sciences are not bothered by ethical, only financial and statistical. They have to come up with the largest N (group size) that will a) allow for a statistical power and b) be cheap. Unfortunately for most labs all over the world, particularly in neuroscience, where monkeys are considered to be the best animal model, b often trumps a, with authority.
A quasi solution that is all too popular is a methodological tool called meta-analysis: the pooling together of a number (preferably large, though more often than not, rather small) of published studies to create a larger N. to justify this seemingly scientifically absurd procedure (those are DIFFERENT papers, with different procedures, designs, subjects, and goals), authors present an elaborate scheme of criteria for the exclusion/inclusion of papers into their analysis.
A new meta-analysis conducted by researchers from Stanford takes this concept a step further, and performs a meta-analysis of 160 previously published other meta-analyses, covering a total of roughly a thousand research papers. The topic of interest was studies of potential treatments of various human neurological disorders such as MS, Alzheimer's, Parkinson's, stroke, and spinal cord injury. The aim of the paper was to test the validity of the statistical tools used and whether a suitable group size was used. To that end, the authors picked only the most precise study in each meta-analysis as an indication whether the expected number allowed for statistically significance.
The main finding, one that should seriously alarm any funding agency, public health policy makers, and the public in general, is that only eight (8!!!!) of the 160 papers covered can claim to have established valid statistical significance.
In addition, almost half of the studies suffered from the basic flaw of small N. bear in mind, those are published papers, i.e., they were approved by editors and peer reviewers.
As only 108 papers were deemed "somehow effective", the number of studies claiming statistical significance for positive results was double what it should be. Such a statistically horrendous skew should immediately raise various ethical concerns. However, the authors are quick to dismiss the possibility of a ubiquitous phenomenon of scientific fraud, and focus on two other explanations.
The first is that investigators selectively choose the statistical tool that provides them with the a priori set desired outcome (and that does not constitute fraud?).
The second reflects the immensely strong bias editors of "prestigious" journal have toward publishing positive results (and preferably novel ones). This creates a reality in which an astronomical amount of data (yes, science mostly produces negative results) never sees the light of day and discarded and never shared with the relevant scientific community. Thus, the bulk of trials and experiments simply cannot be included in any analysis.
The authors suggest that these biases are the culprit in the inappropriate promotion of treatments from animal studies into human clinical trials. I suggest that this papers offers an extremely rare honest glimpse into the way science is performed today. I have to reserve that: the way the BUSINESS of science is conducted today. Those are profoundly two different things, and we (as well as hundreds of thousands of innocent animals) suffer the consequences of this dissonance every day.

Tsilidis KK et al. 2013. Evaluation of excess significance bias in animal studies of neurological diseases. PLoS Biology. 11 (7): e1001609 (10.1371/journal.pbio.1001609)