Voodoo versus me–you correlations in relationship neuroscience

Correlational approaches to social interaction in social neuroscience

For decades relationships researchers have investigated physiological synchrony as a measure of various relationship processes (cf., Levenson & Gottman, 1983; Levenson & Ruef, 1997). Levenson and Gottman (1983) found that physiological linkage between married partners during a marital disagreement was predictive of marital distress and empathic accuracy for negative emotions (Carstensen, Gottman, & Levenson, 1995; Levenson, Carstensen, & Gottman, 1994; Levenson & Ruef, 1992). Alternatively, Feldman (e.g., 2006, 2007) found evidence that behavioral and physiological synchrony during parent–infant play covaries with infant self-regulation, cognitive and theory-of-mind abilities.

Emotional synchrony may also characterize the bonds experienced by adult relational partners (cf., Konvalinka et al., 2007). In a particularly clever and compelling study, Konvalinka and colleagues tested this idea by looking at physiological linkages between relational partners during a public fire-walking celebration. They found that the physiological arousal experienced by fire-walkers was tightly correlated to the arousal of spectators if those spectators were relationally significant, such as a relative or a friend. If the spectator was a stranger, no similar correlation occurred. From this observation, Konvalinka et al. suggested that physiological synchrony is mediated by social information rather than (or at least in addition to) synchronized behavior—that physiological systems among relational partners become linked in part via symbolic information exchanges, almost like routers communicating on a wireless network. If true, then it behooves fMRI researchers to consider how such phenomena can be measured at the neural level.

Functional MRI has rarely been used to investigate physiological synchrony and covariation. This is likely due to several issues, including the physical limitations of the technology, analytical challenges unique to neuroimaging and fears that correlations between self-report or observational measures and fMRI-based brain activations constitute non-independent “voodoo” (cf. Lieberman, Berkman, & Wager, 2009; Vul et al., 2009). A major hurdle in establishing correlations between the brains of two or more interacting subjects concerns the technology’s logistical limitations. Safety issues, the confining nature and loudness of the MRI scanner, sensitivity in fMRI measurement to movement artifacts that make speech virtually impossible—all of these things make it very difficult to measure two interacting brains simultaneously. Despite these limitations, researchers are developing methods that show considerable promise. Hyperscanning, for example, allows quasi-simultaneous imaging of two peoples’ brains as they interact through an internet connection (Montague et al., 2002) and, although they have not to our knowledge yet been used in a neuroimaging context, indirect measures of interaction may also hold promise. A good example of these can be found in the pioneering work of Levenson and Gottman (1983). Their basic design starts by asking marital partners to engage in a 15-minute conflict discussion—a discussion that is videotaped (cf., Roberts, Tsai, & Coan, 2007). Subsequently, participants are asked to view the videotape of their conversation (called a “video recall” session) twice. During one viewing, they are asked to rate their own emotional experience. During the other, they are asked to judge the emotional experiences of their spouse. During all of these experiences—the conflict discussion, the self-rating and the other rating—various channels of autonomic physiology are recorded. This pool of data allows for a multitude of covariation—across behavior and physiology, across persons within behavior and physiology and all of the above across time (Coan & Gottman, 2007; Coan, Gottman, Babcock, & Jacobson, 1997; Gottman, Coan, Carrere, & Swanson, 1998; Levenson & Ruef, 1992). Moreover, early on Levenson and Gottman (1983) made the valuable methodological observation that people were quite accurate in recalling and “reliving” their conflict discussions both physiologically and behaviorally—even up to five days later. These advantages have yet to be exploited in a functional neuroimaging environment, but their implementation there is eminently possible.

Another potential roadblock to studying synchrony in fMRI involves the statistical and methodological challenges of neuroimaging. Neuroscientists are constantly broadening their analytical toolkit, but issues related to the problem of multiple comparisons and the establishment of inferential criteria for localization of brain function create perennial problems. We suggest that meeting these challenges will require the incorporation of measurement theory into otherwise traditional neuroscience training programs, particularly in the social and affective neurosciences. Far from holding correlation in suspicion, measurement theory frequently places correlational approaches at the center of determining the status and meaning of psychological constructs (Coan, 2010) and—for identifying mediation—potential causal mechanisms linking those constructs to other behavioral and physiological phenomena (Coan & Allen, 2004). A prominent example is the classic work of Campbell and Fiske (1959), who gave us the multitrait-multimethod matrix (MTMM). However, recent decades have also seen vital advances in structural equation modeling (using correlational data) to powerfully isolate sources of variability associated with our psychological constructs of interest and our measurement methods (e.g., Kline, 2011). Moreover, innovative approaches, such as Social Relations Modeling, rely heavily on correlational dynamics across interacting dyads (see the actor partner interdependence model, Kenny, Kashy, & Cook, 2006). Increased training in these domains would promote ever more creative approaches to measurement, design and analysis in the social and affective neurosciences, any or all of which would be a boon to the study of relationships.

In considering correlational methods in fMRI research, it has also become necessary to discuss the unfortunately catchy and otherwise misleadingly named “voodoo correlation” problem (Vul et al., 2009). At first glance, and to an alarming degree among lay consumers of fMRI research, Vul et al. appear to have argued against the reporting of correlation coefficients in fMRI research—or at least to arouse suspicion that such correlations represent a form of statistical “voodoo.” In reality, the issue they have raised is much more specific. FMRI approaches require correction for multiple comparisons in order to identify significant regions of activation. Because there are thousands of voxels (three-dimensional units analogous to pixels in a digital image) upon which statistical tests are applied, the typical approach requires a certain number of voxels that are physically contiguous at a given level of statistical significance. Such methods of correction for multiple comparisons are used to avoid reporting the large number of false positives one might otherwise expect. Importantly, in multiple comparisons including a covariate, only highly correlated voxels are selected. So if 25 physically contiguous voxels significant at p < .01 are needed for a region to be considered “activated,” and a .01 p-value for a voxel requires a minimum correlation of .52, then all significant regions of interest (ROIs) associated with the covariate will be correlated with the variable of interest at r ≥ .52. From this distribution of correlation coefficients, some researchers have chosen to report (and graph) the peak voxel in the covariate-identified cluster. The peak voxel is that which has the highest correlation with the covariate, and it may indeed overestimate the size of the covariate effect. There is considerable debate about the degree to which this is a problem (Lieberman et al., 2009), but our concern regards the degree to which the debate itself creates an unnecessary reluctance to implement, report and interpret correlations in fMRI analyses. There are many ways to avoid this problem (or indeed to assess its gravity), and we refer the reader to Lieberman et al. (2009) for a fuller discussion, but a simple method we have used is to define ROIs in a way that is statistically independent of correlational tests. ROIs can be specified theoretically, via previous findings, or empirically using contrasts that are identified independently of the covariate(s) of interest. In this way, correlational approaches can be employed and concepts like physiological synchrony studied without fear of correlational “voodoo.”

Correlational methods applied: empathy and identification

The psychological construct empathy has enjoyed tremendous attention from neuroscientists over the last decade. Recent findings suggest that neural circuits which represent threats directed at the self (e.g., shock or pain) are similar to those that represent threats directed at others (e.g., Lamm, Batson, & Decety, 2007; Lamm, Meltzoff, & Decety, 2010; Singer et al., 2004). Many of these studies feature an analytical technique called conjunction analysis (Nichols, Brett, Andersson, Wager, & Poline, 2005; Price & Friston, 1997). As with most approaches to functional neuroimaging data, conjunction analysis relies on the subtractive method—contrasting blood oxygen level dependent (BOLD) activity in one context from that of another—to locate areas of the brain that are putatively more active in the one condition than the other. Applied to threat-responsive circuits, these would be areas of the brain that are more active during threat cues than safety cues. Conjunction analysis begins with separate, parallel contrasts—for example, a self-directed threat minus safe contrast, and an other-directed threat minus safe contrast. By this approach, areas in which self- and other-directed threats are both active (adjusted for statistical significance) are considered “conjoined,” and statistical procedures have been introduced to provide evidence that identified conjunctions are valid (cf., Nichols et al., 2005). As applied to empathy research, conjunctions representing self- and other-directed threats are viewed as evidence for a putative empathy network. Theorists have argued that activity in these regions during empathy tasks represents a “breach of individual separateness” (Singer et al., 2004, p. 1158) in which “…similar neural networks mediate the simulation of pain for self and other” (Decety, 2011, p. 104). There may be problems with these interpretations, however, as none of these studies has ever actually assessed the correlation between self- and other- directed threat contrasts.

Empathy has been expertly described as a process whereby:

But if empathy is an isomorphic affective state, then we should expect empathy to manifest as a correlation between self and other, and although conjunction analysis provides evidence of a common substrate, it does not provide evidence of correlation. Instead, conjunction analysis simply maps average activation at the group level across two different sets of contrasts. Thus, conjunction analysis cannot identify whether individuals process threats to themselves in these regions in a way that correlates with the way they process threats to others.

In this way, correlational methods are not something one simply can do, but indeed must do if one wants to know whether self-related threat processing and other-related threat processing are, indeed, correlated. Other distinctions between conjunction and correlational approaches are likely to be important as well. For example, I may use the same network to understand another person’s experience that I typically use to understand my own, but there may still be little or no correlation between activity in that circuit as applied to my self-understanding and similar activity applied to my other-understanding, any more than successfully predicting the experience of another person necessarily implies that the other person experiences things precisely the same way I do (Jacob, 2008). But what if activity in that circuit, as applied to both self- and other-directed responding, is in fact correlated? Psychological speaking, such a pattern may suggest that I am treating the other person as if they were me, or understanding them through some sort of self–other equivalence. ¹ In our work, we have observed both phenomena (Beckes et al., 2012). On the one hand, neural responses to threats directed at the self are not correlated with those directed at strangers, despite nearly equivalent mean levels of activation in those same threat-responsive regions across both self and stranger threat conditions. On the other hand, neural responses to threats directed at the self are highly correlated with those directed at close friends, also in a context of equivalent mean levels of activation in those same regions. We have suggested that these observations provide evidence that most people are more dependent upon theory-of-mind strategies for empathizing with strangers, but that the strong self–friend correlation reflects a subtly different psychological construct, that of interpersonal identification—the degree to which one includes another person into one’s representation of self (cf., Beckes, Coan, & Allen, under review; 2012; Decety, & Chaminade, 2003).

Evidence for the identification versus empathy hypothesis rests on correlational data

Specifically, we collected functional neuroimages from 22 participants while (1) they were threatened with shock; (2) their friend was threatened with shock; and (3) a stranger was threatened with shock. Using these data we superimposed a covariance approach onto a set of traditional conjunction analyses. By looking not only at group-level averages, but also at individual differences in threat-related activations across conditions, we observed evidence that high self–other correlations reflect a form of interpersonal identification associated with familiarity that can be distinguished from more generalized empathetic processes. We calculated a between-subjects correlation coefficient reflecting covariation in threat-responsive circuits during self- and other-directed threat conditions, for all brain regions initially activated during the self-threat condition, and tested for significant differences in correlation coefficients as a function of familiarity (e.g., friend versus stranger).

During identification, one should expect correlations between self- and other-directed threat activations to be high. Because people normatively identify with friends more than with strangers, we predicted that self and friend correlations would be greater than those between self and stranger. In numerous threat-responsive regions this is the pattern that we observed (see Beckes et al., 2012). Critically, the distinction between friend and stranger in these data was not detectable using conjunction analysis. That is, when attending only to group averages, many of the regions noted above were activated by threat cues to more or less the same degree whether threats were directed at the self, a friend or a stranger. To return to our earlier point about measurement training in the social and affective neurosciences, we have found that this overall pattern of results is in combination often difficult for people to comprehend without adequate training in understanding the difference between group and individual levels of statistical analysis. We also believe that the difference is critical for how we utilize fMRI designs and analyses to understand psychological constructs such as empathy and identification.

We hope our example provides persuasive evidence that correlation-based analyses can provide important information regarding interpersonal processes. Further, we wish to encourage the use of correlation-based statistical methods in fMRI data analytic traditions. We call as well for increased training in psychological measurement and correlational statistics, which can enhance our field’s ability to extract psychologically meaningful information from neuroimaging data and contribute enormously to the study of interpersonal relationships. Finally, we hope our readers take up the challenge of using correlational approaches generally in their own relationship research.