more on fMRI
Developing Intelligence has a nice post today summarizing last week’s flurry of posts on the utility of functional neuroimaging. I should point out that while these posts offer a nice introduction to some of the issues involved, they by no means offer a comprehensive assessment of the pros and cons of fMRI. For one thing, with the exception of David Dobbs’ post, no one’s really said much about the benefits of fMRI. My own posts were intended mainly to deflect specific criticisms of fMRI; I haven’t really focused on the advantages of functional neuroimaging methods as compared to other methods so far. That’s something I hope to get to in future posts.
On the other side of the ball, I think there are also a number of reasonable criticisms of fMRI that haven’t been highlighted yet. Some of these highlight important problems with current practices (e.g., limitations on power, overly liberal statistical analyses, or rampant overinterpretation) or interpretative (drawing rather speculative conclusions from thin results). These are the types of criticism non-cognitive neuroscientists tend to focus on, and they’re certainly worth paying attention to. Still, I think the largest and most important class of criticisms of fMRI are essentially constructive ones that take the form of “well, why not do it this way?” There’s just so much you can do with fMRI that relatively few people are doing because of a lack of awareness or an absence of consensus procedures and corresponding tools.
The more quantitatively sophisticated groups (e.g., the folks at UCL, who’ve pioneered most of the statistical approaches currently in use) regularly publish all sorts of papers detailing clever approaches to data that everyone concludes sounds really neat in principle but that most people (myself included) don’t really understand. Collectively, neuroimaging researchers are only just starting to talk about things like functional connectivity and parametric analyses; there are all sorts of places people want to go, but the methods haven’t caught up to the ideas yet.
Part of the problem is that functional neuroimaging datasets are much more flexible than behavioral datasets. A single subject’s brain might contain 100,000 voxels (each an interesting measurement in its own right) per image, with several hundred images per scan. And these observations aren’t independent of one another; what one part of your brain is doing at a given point in time depends not just on the experimental manipulation, but on interactions with other regions (David Dobbs provided an excellent overview of he problem in his Scientific American article). Given that level of complexity, it’s not surprising that people have generally stuck with familiar methods imported from other areas of behavioral research. T-tests on subtractive contrasts aren’t necessarily the most natural way to explore brain activity, but they’re what’s closest to standard analyses of experimental treatments in other fields. As time goes on, we’ll probably see a shift toward more sophisticated multivariate modeling. Hopefully, such studies will be less susceptible to many of the common criticisms.
July 10th, 2006 at 4:34 am
“Speculative conclusions from thin results” describes much of the fMRI work related to marketing and advertising in the last year or two, although time will solve much of this problem. As we get more experience in relating observed brain activity to future behavior probabilities, and as the spatial and temporal resolution of scanning systems improves, the techniques will become a lot more useful.
July 10th, 2006 at 11:16 am
Excellent post - I think you’ve hit the nail on the head with “the methods haven’t caught up to the ideas yet.” Although the geniuses at UCL may have methods of analysis that *have* caught up with everything we’d like to do with fMRI, even some of the most famous journals are publishing imaging studies with embarassing analysis probelms (problems with baseline control tasks, or unjustified conclusions about hemispheric assymetries based on inappropriate comparisons).
It might be interesting to “walk the walk” that we’re talking here by analyzing - as a group - a specific fMRI study that appears to be well done. My guess is that we will be able to uncover methodological problems in just about any fMRI study that might be volunteered for this exercise … Of course, it’s not hard to find methodological issues in any paper, if you’re looking hard enough.)