Monday, June 13, 2011

Discriminate equally when recruiting subjects

Inclusion and exclusion criteria possibly don't get the scrutiny in fMRI studies that they should. After all, who reports the complete criteria in the methods sections of articles? At best we get headline exclusions. Without the full questionnaires we, as readers (and reviewers), must trust that no biases were introduced accidentally. Yet, like subtle parameter (mis)settings on the scanner, precisely how experimental and control groups are established is going to have profound effects on your results. They'd better!

A study included in Neuroskeptic's extremely useful weekly roundup of neuroscience makes the point of group bias quite clearly with a hypothetical example, and highlights the possibility that in selecting healthy controls we might accidentally set the bar higher than for the target group, introducing a potential confound to the experiment: Beware the "super well" - why the controls in psychology research are often too healthy.

Obvious? It should be, to a careful experimentalist. But there are insidious ways this selection bias can creep into your studies, making the point worth repeating ad nauseam in my opinion, especially to the waves of newcomers to our field. (Preach this lesson to all incoming students!) I'm going to make the unsubstantiated statement that subject selection (and its alliteration!) ranks above both acquisition and post-processing methods when it comes to biases and the ability to get an incorrect result with fMRI. It's critically important to balance physiological as well as psychological profiles as closely as possible between experimental and control groups.

The insidious biases? Whenever one or other group is difficult to recruit, for demographic reasons or whatever, there is a tendency to let certain things slide in order to net the requisite total. Don't cut this corner! Match as many factors as you possibly can, then note any factors that you can't match and include them in your experiment as covariates of no interest. In this way you might avoid the embarrassment of interpreting a neural difference for something that is better explained by physiology; hematocrit levels, say.

Remember, your fMRI experiment starts when you start recruiting subjects. And the less rigorously you do this fundamental, crucial step the more likely you are to get big error bars, or worse. Even I can't help your data at that point!