Using Network Analysis for Examining Interpersonal Emotion Dynamics

Abstract

Several contemporary models conceptualize emotion as inherently interpersonal. We demonstrate how network analysis, a class of statistical methods often used to examine intrapersonal dynamic processes, provides a potential avenue for parameterizing interpersonal emotion dynamics (and interpersonal dynamics in general). We claim that this method allows (a) observing interpersonal dynamics at various temporal levels; (b) examining interpersonal dynamics occurring through various emotional pathways; and (c) capturing variations in interpersonal networks, which can subsequently be used to predict changes in outcomes. To demonstrate the potential of this method, we used dyadic daily diary data on emotion dynamics from two samples; Sample 1 involved couples in their routine daily lives, whereas Sample 2 involved couples in their transition to parenthood. Graphical Multilevel-Vector-Autoregressive modeling was used to estimate partners’ emotional networks, whereas in a second step, LASSO was used to test the predictive value of couple-level differences of the obtained networks. The analysis revealed several patterns. For example, the between-couple network of Sample 1 was more interpersonally dense, but couple-level differences in the networks’ interpersonal associations were predictive of partners’ relationship satisfaction over time only in Sample 2. We also include commented code implementing a new dyadmlvar R package developed for conducting this analysis.

Keywords:

• Network analysis
• graphical multi-level VAR
• interpersonal emotion dynamics
• emotion
• relationships

Notes

1 In graph theory density refers to the sparseness of a graph, i.e. how many edges exist out of the maximum of possible edges. As in our model, all of the edges are estimated, but with different weights, we refer to density as the average of absolute weights of edges in a network (See Pe et al., 2015 for a similar approach).

2 Available at the Github link: https://github.com/haranse/dyadmlvar/

3 Multivariate multilevel estimation is conceptually possible (see Baldwin, Imel, Braithwaite, & Atkins, 2014), but in network analyses, the estimated parameters grow exponentially with the inclusion of each new network node (variable). Therefore, such network models are computationally expensive and most often not feasible (do not converge). Thus, the mlVAR package applies sequential estimation using univariate models. In addition, even with univariate models, estimating correlated random effects with many predictors becomes unfeasible. As such, in the mlVAR package, networks with more than eight nodes are estimated using uncorrelated random effects (i.e., orthogonal estimation). In such cases the level-2 covariance structure is simplified into: $\left[\begin{array}{cc}{\omega }_{{\mu }_i}& 0\\ 0& {\mathit{\omega }}^{\left({\mathit{\beta }}_{\mathit{i}}\right)}\end{array}\right]$

4 In Epskamp et al. (2018) this network is referred to as the between-subject network; however, when applying network to dyadic data the highest unit of analysis is actually the couples and thus we opted to refer to this network as the between-couple network.

5 The use of residuals (rather than the raw scores) in the estimation of the contemporaneous network ensures that this network represents unique information which does not overlap with that of the temporal or between-couple networks. The advantage of running a second set of multi-level models on the obtained residuals (rather than using, for example, simple-order correlations) is that MLM leads to shrinkage of person-specific parameters (i.e. random effects), and thus parameters that are far from the sample’s parameters (i.e. fixed effects) are often estimated closer to those of the sample. This leads to conservative and more reliable estimates (aka empirical Bayes estimates or best linear unbiased predictors), which take into account the information derived from the entire sample. Another advantage of using MLM is that it accommodates for (the quite common phenomenon of) non-balanced data (e.g. when subjects contribute different amounts of data), thus often leading to a more precise estimation of the fixed and random effects (for a detailed discussion please see chapter 3 in Hoffman (2014)).

6 As noted above, the between-couple network estimates associations at the couple-level (e.g. whether women with higher average reported sadness across the diary period are coupled with men with higher average reported sadness), and thus does not include parameters tapping couple-level differences in the associations’ strength (i.e. random effects). Therefore, this network cannot be used to predict couple-level differences in couples’ outcomes.

7 To facilitate this practice, the dyadmlvar package implements the functionality to conduct such an analysis.

8 We wish to thank an anonymous reviewer of a previous version for suggesting within couple partitioning.