http://www.pnas.org/content/early/2013/12/26/1321664111.full.pdf
Statistical Analysis.
Data were screened manually for anomalous painting
behavior (e.g., drawing symbols on bodies or scribbling randomly). Moreover,
participants leaving more than mean
+2.5 SDs of bodies untouched were
removed from the sample. Next, subjectwise activation and deactivation
maps for each emotion were combined into single BSMs representing both
activations and deactivations and responses outside the body area were
masked. In random effects analyses, mass univariate
tests were then used
on the subjectwise BSMs to compare pixelwise activations and deactivations
of the BSMs for each emotional state against zero. This resulted in statistical
t
-maps where pixel intensities reflect statistically significant experienced
bodily changes associated with each emotional state. Finally, false discovery
rate (FDR) correction with an alpha level of 0.05 was applied to the statistical
maps to control for false positives due to multiple comparisons.
To test whether different emotions are associated with statistically dif-
ferent bodily patterns, we used statistical pattern recognition with LDA after
first reducing the dimensionality of the dataset to 30 principal components
with principal component analysis. To estimate generalization accuracy, we
used stratified 50-fold cross-validation where we trained the classifier sep-
arately to recognize one emotion against all of the others (one-out classi-
fication), or all emotions against all of the other emotions (complete
classification).To estimate SDs ofclassifieraccuracy,thecross-validationscheme
was run iteratively 100 times.
To assess the similarity of the BSMs associated with different emotion
categories, we performed hierarchical clustering. First, for each subject we
created a similarity matrix: for each pair of emotion categories we computed
the Spearman correlation between the corresponding heatmaps. To avoid
inflated correlations, zero values in the heatmaps (i.e., regions without paint)
were filled with Gaussian noise. The Spearman correlation was chosen as the
optimal similarity metric due to the high dimensionality of the data within
each map: with high dimensionality, Euclidean metrics usually fail to assess
similarity, as they are mainly based on the magnitude of the data. Fur-
thermore, as a rank-based metric, independent of the actual data values, it is
also less sensitive to outliers compared with Pearson’s correlation. We also
evaluated cosine-based distance as a possible metric, but the normalization
involved in the computation lowered the sensitivity of our final results, as
cosine distance uses only the angle between the two vectors and not their
magnitude. We averaged individual similarity matrices to produce a group
similarity matrix that was then used as distance matrix between each pair of
emotion categories for the hierarchical clustering with complete linkage.
The similarity data were also used for assessing reliability of bodily top-
ographies across languages and experiments.
