Lindquist et al. seek to distinguish between locationist and constructionist models of emotion by performing a meta-analysis of brain activations during various types of emotional stimuli. Methodologically speaking, the discovery of activated brain areas using the general linear model is primarily geared towards the locationist framework because it is a univariate method, which assumes that each voxel time series is independent of others. However, interaction between voxel time series from different brain regions is imperative for the constructionist model to work. Consequently, any evidence for the constructionist model from an analysis of activations is tenuous and indirect. Therefore, in our opinion, although this target article provides strong evidence against the locationist view, the evidence for the constructionist view is not conclusive. The authors do offer the future possibility of pattern classification of the meta-analytic database showing associations between emotional category and a set of co-activated brain regions. This might be a step forward, but it still does not directly model the interactions between brain regions.

Here we present a rigorous approach for testing the constructionist view. If a set of brain regions co-activate in response to an external stimulus of emotional value, then there must be significant information transfer between the co-activated regions. For example, a fearful stimulus may first activate the sensory regions, which transmit that information to the amygdala, which determines the stimulus to be motivationally salient. The insula may encode the visceral reaction to it and then transmit it to the orbitofrontal cortex. The visceral information from the insula, the salience information from the amygdala, and executive attention from the dorsolateral prefrontal cortex, may all converge in the orbitofrontal cortex (or some other prefrontal area) and be integrated to create the experience of fear (or any other emotion). The constructionist hypothesis predicts that different regions encode some basic psychological operations, which, when integrated, “feels” like different emotions. In the aforementioned example, the co-activation of a set of brain regions is only a necessary condition for the constructionist view, but is not sufficient. Only a model which can capture and quantify the task-associated connectivity between regions can provide sufficient evidence for the constructionist hypothesis. We will now describe how such models could be used for directly testing the constructionist hypothesis.

Granger causality analysis (GCA) is based on the principle that if the past activity of region A can predict the present and future activity of region B, then A must have a causal influence on B. Traditionally, GC is obtained from the vector autoregressive model wherein GC from time series j to time series i is given by the ijth element of coefficient matrix (Deshpande et al. Reference Deshpande, LaConte, James, Peltier and Hu2009). GCA is completely data-driven, accommodating a large number of time series in the model, and is primarily an exploratory technique. Many refinements to GCA have been proposed, such as correlation-purged/dynamic/nonlinear GC (Deshpande et al. Reference Deshpande, Sathian and Hu2010; Marinazzo et al. Reference Marinazzo, Pellicoro and Stramaglia2008; Sato et al. Reference Sato, Junior, Takahashi, Felix, Brammer and Morettin2006) and hemodynamic deconvolution (Havlicek et al. Reference Havlicek, Friston, Jan, Brazdil and Calhoun2011; Ryali et al. Reference Ryali, Supekar, Chen and Menon2011), which have increased the applicability of GCA to fMRI. On the other hand, dynamic causal modeling relies on modeling underlying neuronal causality using state-space equations and Bayesian inference for comparing model evidence. Although dynamic causal modeling is restricted by the number of time series (up to 8), it is a robust confirmatory technique. We propose combined use of GCA and dynamic causal modeling for testing the constructionist hypothesis. First, GCA could be employed on time series from a large number of nodes, similar to the number used in the target article, provided there are enough subjects for such an analysis to be adequately powered. Using an iterative network reduction procedure we have previously proposed (Deshpande et al. Reference Deshpande, Hu, Stilla and Sathian2008), the large exploratory network could be reduced to a smaller network by removing network redundancies, which could then be confirmed by either dynamic causal modeling (if the reduced network has fewer than 8 nodes) or other methods such as Patel's tau (τ) (Sathian et al. Reference Sathian, Lacey, Stilla, Gibson, Deshpande, Hu, LaConte and Glielmi2011), which are not based on autoregressive principles but rather on Bayesian inference. In this way, converging evidence can be obtained from multiple and distinct methodologies for quantifying the magnitude and direction of connectivity between co-activated brain regions. Such an analysis could provide sufficient evidence for the constructionist hypothesis.

In section 6.1 of the target article, paragraph 2, the authors suggest that studies of resting states will be useful in determining if there are “intrinsic” functional networks corresponding to emotion categories or basic psychological operations underlying them. Here we address some pertinent methodological advances in fMRI research in this direction. While most neuroimaging studies of resting state networks have concentrated on functional connectivity based on instantaneous correlation (which is a non-directional influence) in a single network, we investigated both functional connectivity and effective connectivity (which is a directional influence) of four different resting state networks using a single multivariate model (Deshpande et al. Reference Deshpande, Santhanam and Hu2011). This enabled an explanation of the basic psychological operations during the resting state, such as episodic memory, self-referential processing, and cognitive integration in terms of significant pathways in the network. We believe that this study will provide significant insights for testing the constructionist hypothesis using resting-state data.