Causal interpretation rules for encoding and decoding models in neuroimaging

Abstract

Causal terminology is often introduced in the interpretation of encoding and decoding models trained on neuroimaging data. In this article, we investigate which causal statements are warranted and which ones are not supported by empirical evidence. We argue that the distinction between encoding and decoding models is not sufficient for this purpose: relevant features in encoding and decoding models carry a different meaning in stimulus-and in response-based experimental paradigms. We show that only encoding models in the stimulus-based setting support unambiguous causal interpretations. By combining encoding and decoding models trained on the same data, however, we obtain insights into causal relations beyond those that are implied by each individual model type. We illustrate the empirical relevance of our theoretical findings on EEG data recorded during a visuo-motor learning task.

Assessing the Driver’s Current Level of Working Memory Load with High Density Functional Near-infrared Spectroscopy: A Realistic Driving Simulator Study

“…As the multivariate fNIRS regression model linearly combines multiple channels to predict current working memory load, one approach would be to interpret the weights of these decoding models. However, decoding model weights are hard to interpret for various reasons (Reichert et al, 2014; Weichwald et al, 2015). Therefore, we performed a univariate regression analyses separately for each channel.…”

“…Because weights of regularized multivariate decoding models can be hard to interpret (Reichert et al, 2014; Weichwald et al, 2015), we used univariate linear regression to determine for each channel separately if it can predict the current working memory load level. These analyses revealed multiple predictive brain areas.…”

“…Multivariate decoding models can be hard to interpret regarding the specific brain areas which are sensitive to working memory load-related changes (Reichert et al, 2014; Weichwald et al, 2015). In order to simplify the interpretation of brain activation measurements, we regressed the HbR fNIRS measurements on the current n-back working memory load levels for each channel separately and calculated univariate Pearson’s correlations ( r uvr ) from the regression for each participant.…”

Transcranial magnetic stimulation reveals complex cognitive control representations in the rostral frontal cortex

“…Two recent studies with frontal lesion patients demonstrated that complex cognitive control processes involve rostral frontal regions and less complex cognitive control processes engage caudal frontal regions Azuar et al, 2014). From a broader perspective, the level of inference of lesion (Rorden and Karnath, 2004) and TMS (Dayan et al, 2013) studies is generally higher than the level of inference of functional magnetic resonance imaging (fMRI) studies (Poldrack, 2011;Weichwald et al, 2015). Thus, previous frontal lesion studies together with present TMS findings suggest that complex cognitive control functions are represented in rostral lateral frontal cortex.…”

Interpretability of Multivariate Brain Maps in Linear Brain Decoding: Definition, and Heuristic Quantification in Multivariate Analysis of MEG Time-Locked Effects

“…Currently, brain decoding is the gold standard in multivariate analysis for functional magnetic resonance imaging (fMRI) (Haxby et al, 2001; Cox and Savoy, 2003; Mitchell et al, 2004; Norman et al, 2006) and magnetoencephalogram/electroencephalogram (MEEG) studies (Parra et al, 2003; Rieger et al, 2008; Carroll et al, 2009; Chan et al, 2011; Huttunen et al, 2013; Vidaurre et al, 2013; Abadi et al, 2015). It has been shown that brain decoding can be used in combination with brain encoding (Naselaris et al, 2011) to infer the causal relationship between stimuli and responses (Weichwald et al, 2015). …”