Over the last 20-30 years the use of fMRI in psychology and related disciplines has become increasingly fashionable and well used. Seemingly the ability to create easy to understand images depicting brain areas involved in such and such task, aglow with eye catching colours is simply too much for the lay person, the news reporter or indeed most cognitive scientists to ignore. Indeed, The fMRI machine has allowed us fast, non-intrusive access to the function of deep layers of the brain in a way hitherto barely accessible without autopsy or the lower resolution PET.
My own gut reaction to sensationalist fMRI headlines is a causal satisfaction, a gratifying A = B closed loop which appeases my inner scientist – something which seems to rarely happen in the misleading and statistically unsure worlds of psychology and brain science.
Indeed, I am not alone in this respect, the persuasive effect of neuroimaging is well documented (McCae and Castel 2008)- presenting physical evidence of abstract cognitive processes both allows us to gloss over system complexity and seemingly understand things much more certainly providing some sort of credible objective difference that we can easily see and measure with colourful images. This extends outside of the lab too – as researchers have shown the persuasive nature of the fMRI in courtroom judgement.
The question is whether fMRI deserves this reputation? A cautionary study by Bennet, Baird, Miller and Wolford (2010) might suggest otherwise.
In the study Bennet et al placed an Atlantic salmon (dead at the time of study) into the fMRI machine, the salmon was shown a perspective taking task which was later shown to a group of (live) humans. With an uncorrected analysis, it was shown that there were some active voxels in the Salmon’s brain cavity and spinal column.
So how did this happen? First we need to understand a little bit about functional magnetic resonance imaging.
In an fMRI brain image, colour represents activation. It recognises activation by first using a powerful magnetic field (the big white rings) to align the protons of the part of the body in the scanner. It then send a particular frequency of electro magnetic energy into the brain which knocks the protons out of alignment and ionises them. As the protons fall back to their previous alignment they emits the frequency of energy that they were given to knock them out of alignment. This is the ‘magnetic resonance’ in functional magnetic resonance imaging (fMRI). The differentiation of a functioning and non-functioning area relates to signal strength. Theory suggests that oxygenated blood gives a stronger signal. So, if we assume that de-oxygenated blood is blood that has been ‘used’ by a particular brain area and we can see where this blood is because of a weaker blood-oxygen level signal (or BOLD signal)- we can therefore see the parts of the brain that are being used while the scan was taking place and it is done in 3 dimensions by taking slices through the brain.
However fMRI technology does not have the spatial resolution to identify individual neuron activation – rather it generalizes and averages activation over a small area called a voxel. One voxel is the equivalent to one pixel of colour on the final fMRI image and even with the highest possible voxel resolution one voxel interprets the activation of millions of neurons. It does this by statistically analysing the signal in one particular area to find significant differences in signal strength and highlight said differences for us to see.
It is this part of the process that is especially subject to error and the reason why the (very dead) salmon looks as though it might be using its (very dead) brain. The sheer volume of statistical inference taking place at this stage means that statistically it is 5% likely that some error will occur and resulting in a false positive.
But it should be said that this is ‘uncorrected’ analysis, as with comparative statistics of this type it is working on the basis of a p-value <0.001 to just statistical probability. Unfortunately a large amount of fMRI studies use an uncorrected analysis and are unaware of this problem. However, things can be done to the data to take into account the large amount of multiple comparisons taking place and generate a more conservative analysis. This is just what the study found – when controlling for family-wise error and false discovery rate no active voxels were found.
This study gives us a number of cautionary lessons. Firstly, just because an impressive technology has been used, it is important to understand its application and whether it has been used correctly, being aware of the fallacies and biases of a method as well as its strengths is necessary for good scientific enquiry. Secondly, even fMRI, a particularly hard social science is subject to statistical inference which relies on p values, probabilities and can suffer from causal misconnections. It is (very very very) unlikely that the salmon in the study was responding to the stimulus, but if this study was on a more believable topic it is not hard to imagine people taking the methodology for granted and jumping to surprising, attention grabbing conclusions.