[Disclaimer: I have just typed this after returning home from a long day, please excuse the unpolished form]
In their recent paper, Litwin and Miłkowski showcase our fMRI study on the Rubber hand illusion (RHI) – thanks for this. Unfortunately, they use it as a “bad example” for post-hoc interpretations in the light of Predictive Processing (PP) theories. While I understand Litwin and Miłkowski’s rage, I feel they have – in parts – misread our paper (in parts, but important ones), and misunderstood parts of the DCM framework – much of this could have been avoided by asking me (or others) about these questions before writing this paper. Since our DCM results are the only figure in their paper – which sort of puts me on the spot – I’m writing the following clarification.
First off, however, let me emphasize that I agree with the message of this paper, and I believe it is an important one: don’t go overboard with data interpretation!
In our paper, we set out to investigate a multisensory illusion, the RHI – in a new way: we conducted 2 fMRI experiments using a novel, robotic setup; i.e., focussing on BOLD signal activation changes, correlations of these with our participants’ illusion ratings, and then using Dynamic causal modelling (DCM) of these results to test mechanistic hypotheses about their generation.
While we never (!) claimed to test PP directly, we could indeed have been more careful when interpreting our results in the light of our ‘favourite’ framework. That’s something I have tried to do better since starting my PhD (when this paper was written).
Some of the stuff that I found problematic:
- You say that our paper is contradictory based on this: “their [i.e. ours, JL] winning model predicted the attenuation of intrinsic connectivities in both the LOC and SII”. This is false. In fact, we did not even test for modulations of instrinsic connections; our winning model identified extrinsic (forward) connections as being modulated. If you mean that the values of each node’s intrinsic (recurrent) connections were negative at baseline (‘default’ endogenous connectivity) – this is a standard procedure in DCM, where self-connections are always negative; this keeps a (brain) network ‘stable’.
- What is also false is your claim that in our model “effective connectivity from IPS to SII was altogether absent” – these values were just lower than for the IPS to LOC connectivity.
- You conclude that ours and Zeller’s results and interpretation are “irreconcilable”. This is wrong, as I will explain:
First off, a DCM analysis is always motivated by significant effects found in a cognitive contrast – which obviously varies with brain scanning (sorry EEG folks) method. Our fMRI design identified secondary rather than primary visual and somatosensory regions, which is an important difference to Zeller et al’s EEG results. So there are empirically – not theoretically! – motivated differences in how we set up the DCMs to begin with.
However, even given those differences, our winning model – just as Zeller’s – identified enhanced ascending connectivity from visual to multisensory (fronto-parietal) regions. This is, if one wants, interpretable as enhanced visual information processing (and you can, if you want, call “information” the unpredicted error part… you see where I’m getting at). We find this very plausible given that in the RHI ‘one now has a new visual hand position’. The notable difference is between Zeller’s attenuation of recurrent connections within S1, and our enhanced extrinsic connections from S2 to parietal cortex. This was a bit puzzling, indeed, but note that in a follow-up study, we found reduced S1 activity during the RHI using fMRI (Limanowski & Blankenburg 2016 SCAN), which very well matched Zeller’s EEG results. However, our DCM was focussed on ‘secondary’ visual and somatosensory regions, meaning S1 was not included in the 2015 paper’s DCM. We had speculated that the S2 was therefore – in contrast to the S1 – subject to cross-modal information (“prediction error”) from the visual system (see our paper for references to empirical evidence for exactly this sort of cross-modal interactions via higher-level areas).
- You ask about our interpretation: “why does error‐related bottom‐up effective connectivity from LOC to IPS arise? None of these problems is even discussed in the article” – In fact, we do discuss this at lengths, e.g. on p. 2300, which can be summarized as: the seen location of the rubber hand does not fully match where one expects one’s arm to be (based on an internal body model). So there is ‘unpredicted’ visual information passed upwards. I believe this is the same interpretation that Zeller and colleagues arrive at; and one that seems easily in line with other groups’ interpretation of how the RHI comes about (I’m not naming anyone since this is a private rant, but you’ll know them).
- You present our study as a “post‐hoc PP interpretation”. I find this very unfair, since if you read the paper we make it very clear that we never set out to ‘test PP’ in this study – and we never claimed to do so. Specifically, you write that “Both the PP‐based account and the traditional multisensory integration approach seem quite likely in the face of the data (Fig. 1A). However, the authors waive off this concern with a short remark that the alternative hypothesis is not inconsistent with PP.” Please note that our “PP” interpretation is very carefully – but apparently not carefully enough – being introduced only on the third (!) page of the Discussion (with “There is one intriguing interpretation of these results…”, p. 2300). We also avoid using definitive terms, and instead use “may/might/could” etc. On p. 2301, we discuss alternative interpretations after saying that “Although in conjunction, our results comply with the idea that the information flow within the brain network underlying body ownership follows a predictive coding scheme (i.e., the forward-propagation of multisensory prediction errors), predictive coding is only one candidate explanation for the mechanisms underlying the brain’s hierarchical inference about the world and the body.” (p. 2301). These interpretations are indeed not mutually exclusive, and we focussed on the one we liked – as one is allowed to do in a Discussion section of an empirical paper.