King's College London
Question
Asked 15th Aug, 2012
BOLD signal change data analysis. Problems with correlations analyses.
I am looking for advice about conducting correlations using fMRI BOLD effect (% signal change) data. I am trying to correlate these data with subjective appetite sensation scores (collected using visual analogue scales), appetite hormone concentrations and body temperature data. I have analysed the fMRI data using FSL Library software and conducted the correlations analyses using SPSS, however I cannot find any meaningful correlations. Previous research consistently identifies correlations between neural activation within areas of the brain which regulate the rewarding properties of food, and appetite sensations and gut hormone concentrations. Can correlations be conducted using FSL Library software?
Most recent answer
Dear Daniel,
in principle, the equations used for calculations of the Pearson correlation coefficient should be the same, regardless whether you calculate by hand, by pocket calculator or by using inference software ☺
However, it may make a difference if you are using fMRI inference software, for e.g. whole-brain analyses versus using masks for ROIs (as in SPM). Other procedures implemented in inference software can influence the results. Some inference packages, such as XBAM or also SPM (on demand) use permutation resampling procedures, which lower alpha levels for statistical significance. Therefore, such randomization procedures typically make fMRI software more “sensitive” to detect correlations.
It is advisable, however, to be able to demonstrate a covariation both in fMRI inference software and in behavioral statistics. This is more convincing to reviewers…
One more issue I had previously forgotten to speak about is: One should also look at whether one’s (behavioral and physiological) data conform to a normally distributed population. This can be done e.g. in STATA using Shapiro-Wilk tests and/or Q-Q plotting. Deviations from normality may bias your results. Often, physiological data are not really normally distributed for reasons e.g. of internal secretion. In psychophysiology it is then state-of-the-art to use transformations (log, log-10, arcsin etc. ) to remedy for that. Such procedures are implemented in SPSS and can be used conveniently.
So if you have some nil correlation, it is useful to look for possible reasons outside the fMRI inference software first. If the reasons can be identified, this will also improve your correlation images.
Good luck, Erwin
2 Recommendations
Popular answers (1)
Forschungszentrum Jülich
Dear Daniel
You can (and should) use the obtained behavioural scores as covarates (explanatory variables) in your FSL analysis. This could either be done on the single-subject level (for trial by trial variations) or at the 2nd (group) level when looking for correlations across subjects. This will allow you to identify brain regions where the BOLD signal covaries with your behavioural scores. Just do a quick web or pubmed search on "parametric modulation" for lots of material on this topic. Also, have a look at the FSL course notes for more detailed descriptions on the technical aspects of implementing these.
Extracting percent signal changes and correlating them offline in SPSS, however, is probably not the best idea. One of the main problems here: Where to extract these. If you don't use an independent localizer, you quickly end up in the "double-dipping" problem. When using a localizer, however, you are making the rather strong assumption, that the same area identified by the localizer is also modulated.
Simon
5 Recommendations
All Answers (11)
University of Nebraska at Lincoln
A helpful document: http://woldorfflab.ccn.duke.edu/files/uimages/fMRI_data_behavior_NeuroBio381Psych362_title.pdf
I think you should be able to use FSL software: flame
1 Recommendation
Forschungszentrum Jülich
Dear Daniel
You can (and should) use the obtained behavioural scores as covarates (explanatory variables) in your FSL analysis. This could either be done on the single-subject level (for trial by trial variations) or at the 2nd (group) level when looking for correlations across subjects. This will allow you to identify brain regions where the BOLD signal covaries with your behavioural scores. Just do a quick web or pubmed search on "parametric modulation" for lots of material on this topic. Also, have a look at the FSL course notes for more detailed descriptions on the technical aspects of implementing these.
Extracting percent signal changes and correlating them offline in SPSS, however, is probably not the best idea. One of the main problems here: Where to extract these. If you don't use an independent localizer, you quickly end up in the "double-dipping" problem. When using a localizer, however, you are making the rather strong assumption, that the same area identified by the localizer is also modulated.
Simon
5 Recommendations
Universitäre Psychiatrische Kliniken Basel
Dear Daniel
Just one advice: Read the Yarkoni et al. (2009) paper on "Big correlations in little studies: inflated fMRI correlations reflect low statistical power" regarding a) statistical power or number of subjects needed for correlative analyses in fMRI and b) the danger of applying a too high statistical threshold.
Best
Annette
3 Recommendations
University Medical Center Utrecht
I think that correlations can be calculated through FSL. However, why would those results be any different from the SPSS results? They use the same statistics...
2 Recommendations
King's College London
Dear Daniel,
it is always a good idea to examine your behavioral and physiological variables for possible confoundation: sociodemographic (age, sex, edu, ses), and/or measurement-related (daytime, seasonal fluctuations, etc.). As in behavioral or psychophysiological studies, it can be helpful to look at the degree of covariation by inspection of scatterplots in stats packages such as SPSS or STATA. This may reveal mediatory or modulatory effects of one of your measurements, whereby a true existing correlation is being suppressed by another variable (suppressor variable). A classical instruction is the paper by Baron & Kenny (1986): "The mediator moderator distinction..." Journal of Personality and Social Psychology Vol. 51, No. 6, 1173-1182. If you can identify one or more confounders, these should be entered as nuissance regressors in your FSL models. Have you tried whole brain correlations with your self-report and hormone data and experimental activation group maps before extracting percentage signal changes from certain blobs?
1 Recommendation
University of Groningen
Dear Daniel,
as a follow-up of Simon Eickhoff's suggestion (to use the behavioural scores as covariates), remember to demean them (google; Jeannette Mumford demean covariate).
Another approach you may want to try is the PPI (Psyco Phisyological Interactions), see http://www.fmrib.ox.ac.uk/Members/joreilly/what-is-ppi for the FSL implementation, where the BOLD signal is correlated with the behavioural data in an extremely refined way (google: friston gitelman ppi). Everything can be performed using SPM, of course.
Kind regards,
luca
2 Recommendations
University of Aberdeen
Thank you to everyone for your answers. I will, along with my collaborator at Imperial College London, begin to investigate the answers that you have all provided. I hope that I can contact you if I have any follow-up questions. One further question, could anyone answer Wouter Schellekens question regarding the differences between FSL analysis and SPSS analysis? Thank you all once again
Oregon Health and Science University
Perhaps I missed it but what is your sample size? To reinforce sine's and Annette's comments, unless you have a reasonable sample (N >30) you should be extremely cautious about any correlations you find, particularly if they are post hoc ROIs. With a smaller sample size you won't have adequate power to detect correlates and if you find significant effects, you should be suspicious of them.
1 Recommendation
University of Groningen
About the difference between FSL and SPSS: the key could be maybe sought in the amount of simultaneous tests. Functional brain = ~ 150k voxels = 150k correlations = dramatic high chances of false positives due to chance. If I am not mistaken, the default approach of FSL is to use the Gaussian Random Field theory to address the multiple comparison problem: basically, how likely it would be that one cluster ('blob of active voxels') of a certain size would happen only due to chance? I am not familiar enough with SPSS to rule out the possibility that such a correction exists; but if it is not applied, then the results would be different.
kind regards,
luca
King's College London
Dear Daniel,
in principle, the equations used for calculations of the Pearson correlation coefficient should be the same, regardless whether you calculate by hand, by pocket calculator or by using inference software ☺
However, it may make a difference if you are using fMRI inference software, for e.g. whole-brain analyses versus using masks for ROIs (as in SPM). Other procedures implemented in inference software can influence the results. Some inference packages, such as XBAM or also SPM (on demand) use permutation resampling procedures, which lower alpha levels for statistical significance. Therefore, such randomization procedures typically make fMRI software more “sensitive” to detect correlations.
It is advisable, however, to be able to demonstrate a covariation both in fMRI inference software and in behavioral statistics. This is more convincing to reviewers…
One more issue I had previously forgotten to speak about is: One should also look at whether one’s (behavioral and physiological) data conform to a normally distributed population. This can be done e.g. in STATA using Shapiro-Wilk tests and/or Q-Q plotting. Deviations from normality may bias your results. Often, physiological data are not really normally distributed for reasons e.g. of internal secretion. In psychophysiology it is then state-of-the-art to use transformations (log, log-10, arcsin etc. ) to remedy for that. Such procedures are implemented in SPSS and can be used conveniently.
So if you have some nil correlation, it is useful to look for possible reasons outside the fMRI inference software first. If the reasons can be identified, this will also improve your correlation images.
Good luck, Erwin
2 Recommendations
Similar questions and discussions
Recommendations
The purpose of this article is to present some main concepts about neurological-science, specifically the ones established by Francisco Varela, Antonio Damasio, Shaun Gallagher and Alva Noé. The aim is to analyze and to discover encounter points between the neurological-phenomenology of Francisco Varela and learning, based on the assumption that th...