Short communicationThe face-responsive M170 is modulated by sensor selection: An example of circularity in the analysis of MEG-data
Introduction
Given the large quantity of multivariate data provided by methods such as functional magnetic resonance imaging (fMRI), multi-array electroencephalography (EEG) or magnetoencephalography (MEG), selective data analysis is often necessary to study signals of interest. For instance, in MEG studies on face processing the analysis of event-related fields (ERFs) within sensor-space such as the face-selective M170 is typically restricted to regionally arranged clusters of sensors. These sensors of interest can be identified making use of previous research using different methodologies suggesting a consistent topography and chronology of activations by face stimuli occurring between 140–200 ms in posterior and lateral sensors over occipital and temporal areas (for review see Eimer, 2011). However, if sensor-selection is not only based on physiological or anatomical assumptions a phenomenon termed ‘circular analysis’ (Kriegeskorte et al., 2009) might lead to distorted and eventually flawed results. This problem would for example arise when the comparison of M170 amplitudes elicited by different types of faces were based on the averaged signal strength for only one of the face-categories. This can be verified by comparing the contrasts used for selection and analysis as vectors, which should be orthogonal to each other (i.e., their inner product is zero) to yield an independent analysis (given that presentation times, experimental design, etc. are matched between conditions). An example can be found in the article by Kriegeskorte et al. (2009, p. 536; see also the supplementary discussion in the online version of the article): when the responses to two stimulus-categories are equally used for sensor-selection (contrast A + B) before statistical comparison between both stimuli (contrast A − B); cselection [1 1]T × ctest [1 −1]T = 0.
Although the issue of circularity is not confined to a particular methodology such as fMRI but is instead relevant whenever data analysis is selective (Vul and Pashler, 2012), most of the articles on this topic concentrate almost exclusively on neuroimaging techniques (e.g., Kriegeskorte et al., 2010, Vul et al., 2009). Specifically, the extent of distortions caused by a non-independent analysis of fMRI data has been demonstrated by means of comparing artificial and realistic data as well as conducting parallel or repeated analyses (Baker et al., 2007, Kriegeskorte et al., 2009, Poldrack and Mumford, 2009, Vul et al., 2009). On the contrary, only one article concerned with the topic of selection bias in EEG and MEG data analysis has demonstrated that selection based on signal strength massively increased the type-1 error in artificial, pure noise data (Kilner, 2013). Whereas there is no doubt about the relevance of the topic for electrophysiological methods, there is an apparent lack of concrete examples and, possibly awareness of the problem. Therefore, we conducted a simplified re-analysis of MEG-data from a previously published experiment on processing of personally familiar and loved faces (Tiedt et al., 2014, Tiedt et al., 2013), adopting two different (independent compared to circular) sensor-selection strategies. Aim of the current analysis was to illustrate how sensor selection ‘favoring’ each one of three face categories for the analysis of the M170 component might not only distort the results but lead to statistically significant results.
Section snippets
Participants and experimental procedure
We analyzed data from 28 healthy subjects (mean age 25.5 years, range 20–35 years; 14 female, two were left-handed). All participants viewed photographs of three opposite-sex faces of their romantic partner as well as another two close friends for six seconds. Pictures (size 100 × 100 pixels; viewing angle 11.5°) were in color and showed only the faces of the depicted persons. Stimuli from each category were presented 30 times in randomized order with variable inter-stimulus intervals.
Results
In each of the sensor-selection conditions, an event related field (ERF) peaking within the time-window of the M170 could be identified as an outward-going magnetic flux (positive polarity) over the right hemisphere. Graphs showing the M170 obtained from the different sensor-selection strategies as well as topographic maps are depicted in Fig. 1.
In the non-circular analysis, there was no discernible effect of face-category with almost identical peak amplitudes across face categories. In
Discussion
Our re-analysis of the M170 component obtained from sensors selected by signal strength averaged across all face-categories replicated our earlier results and no effects of face-category on peak-amplitude of the right-hemispheric M170 were observed. On the contrary, selecting sensors by signal strength elicited by each one of and testing for differences between the three face categories yielded a significantly increased M170 component for the respective experimental condition. In the
Conflict of interest
None of the authors has potential conflicts of interest to be disclosed.
Acknowledgements
This research was supported by Lutz Trahms and Tilman Sander of the Biomagnetism group at the Physikalisch-Technische Bundesanstalt (PTB) in Berlin.
References (15)
- et al.
Good practice for conducting and reporting MEG research
NeuroImage
(2013) - et al.
Affective processing of loved familiar faces: integrating central and peripheral electrophysiological measures
Int. J. Psychophysiol.
(2012) Bias in a common EEG and MEG statistical analysis and how to avoid it
Clin. Neurophysiol.
(2013)Understanding face perception by means of human electrophysiology
Trends Cogn. Sci.
(2014)- et al.
A different pattern of lateralised brain activity during processing of loved faces in men and women: a MEG study
Biol. Psychol.
(2014) - et al.
Voodoo and circularity errors
NeuroImage
(2012) - et al.
Does the fusiform face area contain subregions highly selective for nonfaces?
Nat. Neurosci.
(2007)
Cited by (3)
Narrowband multivariate source separation for semi-blind discovery of experiment contrasts
2021, Journal of Neuroscience MethodsCitation Excerpt :Next, we examined the top sources that csGED found at these frequencies and quantified how accurately the sources captured the ground-truth signals. This source reconstruction accuracy was compared to the accuracy at the best single sensor for each signal (which is commonly taken to be representative for phenomena detectable in M/EEG; e.g., Asanowicz et al., 2019; Tiedt et al., 2016). For all but the highest and lowest signal-to-noise ratios, the csGED source reconstruction was markedly more accurate than the single-sensor representation of the source.