Emotional Contagion Metrics

• Emotional contagion metrics are quantitative frameworks that assess the propagation of affective states through individuals, dyads, groups, and online systems using statistical, network, and diffusion analysis.
• They integrate direct annotation, time-series, and epidemic-based methodologies to offer precise benchmarks like CCC and Granger causality for measuring emotional synchronization.
• These metrics empower researchers to analyze affect dynamics across diverse settings, enabling real-time interventions and predictive modeling of emotional transmission.

Emotional contagion metrics provide quantitative frameworks for measuring how affective states propagate through individuals, dyads, groups, and broader online/offline systems. Such metrics encompass time-series concordance, network-based influence indices, epidemic-theory analogues, multivariate diffusion models, tie-weighted propagation, and cross-modal coherence estimators. This domain unifies approaches from affective computing, social network analysis, and statistical physics, yielding operational benchmarks for both experimental and observational studies. Below is a detailed review of the state-of-the-art metrics, protocols, and findings underpinning the quantification of emotional contagion across key paradigms.

1. Direct Annotation and Agreement Metrics

The most direct approach operationalizes emotional contagion as the alignment between a “sender’s” affective signal and a “receiver’s” affective state over time. In the OMG-Empathy framework (Barros et al., 2019), the key procedures are:

• Continuous Valence Annotation: Listeners annotate their own affective responses in real time (range –1 to +1; sampled at 25 Hz), yielding “impact traces.”
• Evaluation Protocols: Two protocols are defined:
  • Personalized empathy: target is each listener’s framewise trace on unseen stories.
  • Generalized empathy: target is per-frame average across listeners for a given story.
• Concordance Correlation Coefficient (CCC): The objective metric is the CCC between predicted and true valence time series:

$$\mathrm{CCC}(x,y) = \frac{2\,\rho_{xy}\,\sigma_x\,\sigma_y}{\sigma_x^2 + \sigma_y^2 + (\mu_x - \mu_y)^2}$$

where $\rho_{xy}$ is the Pearson correlation, $\mu_x, \mu_y$ are means, and $\sigma_x, \sigma_y$ are standard deviations, all computed across the series.

• Benchmarking: Baseline models use frame-level visual (VGG-16 + LSTM) and audio (SoundNet) features regressed onto valence via SVM. Effect sizes (CCC) reflect the model’s ability to explain recipient affect (e.g., CCC=0.23 for “listener only” in generalized protocol).

This approach operationalizes emotional contagion as real-valued, framewise synchronization of affect between affect generators and recipients (Barros et al., 2019).

2. Time-Series and Network-Based Metrics (Longitudinal and Granger-Based)

In extended workplace settings, as in the WELD study (Sun, 17 Oct 2025), a comprehensive suite of affective-dynamics and contagion metrics is derived from face-extracted emotional probabilities:

• Valence & Arousal: Scalar reductions of categorical probabilities (e.g., $$v(t) = p_{happy}(t) - p_{sad}(t) - 0.5\,p_{angry}(t) - 0.5\,p_{fear}(t)$$).
• Volatility ($\sigma_v$): Sample standard deviation over a sliding window.
• Inertia ($\rho_1$): First-lag autocorrelation of valence.
• Predictability ($R^2_{AR}$): $R^2$ from AR(p=5) fit to valence series.
• Granger Causality (Emotional Contagion Strength): For each pair $(i \to j)$, construct restricted (self-history only) and unrestricted (add $i$’s history) AR models for $j$, and compute

$F_{i \to j} = \ln \frac{RSS_0}{RSS_1}$

where $RSS_0, RSS_1$ are residual sums of squares. All $N(N-1)$ pairwise $F_{i\to j}$ log-ratio values form a weighted, directed contagion network.

• Windowing and Smoothing: Metrics can be calculated globally or in fixed-length windows (e.g., 30-day rolling windows for longitudinal trend analysis).

The significance of $F_{i\to j}$ is validated via psychological regularities (e.g. synchrony among role peers, AUC=1.0 for turnover prediction with these metrics included) (Sun, 17 Oct 2025). This paradigm generalizes to any setting where time synchronization and reliable emotion probability estimates are available.

3. Population and Network Diffusion Models

Many studies model emotional contagion as a function of direct or indirect exposures within complex networks, yielding a range of operational metrics:

• Exposure-Based Overrepresentation ($\Delta P$): In social media, emotional contagion is measured by the difference in pre-exposure to a sentiment class prior to producing matching content versus shuffle-based null exposure (e.g., $\Delta P^{-} = 21.63\% - 17.29\% = 4.34\%$ for negative posts in Twitter (Ferrara et al., 2015)).
• Valence Linear Regression: Regression of response valence on stimulus window valence shows tight linearity ($R^2=0.975$), establishing highly predictable direct contagion (Ferrara et al., 2015).
• Susceptibility Classification: Fraction of posts/tweets following the over-exposed emotion template (using Euclidean distance to class means for each exposure window), allowing identification of "highly" or "scarcely" susceptible individuals (Ferrara et al., 2015).
• Agent-Based Simulations: Incorporate empirically derived emotion assortativity ($c_i$), tie-strength distributions, and finite-attention decision rules. The effective branching number ($\sigma$), retweet tendency ($t = s\,e^{c_i-1}$), and thresholding ($\tau$) calibrate when and how emotion dominates (Fan et al., 2017).
• Critical gap analysis: Shows that when the raw posting-rate gap $\Delta P = P_{joy} - P_{anger}$ falls below 0.10, anger takes over both retweet and user dominance—an early warning for viral outrage (Fan et al., 2017).

Table: Key Exposure-Based and Agent-Based Metrics

Metric	Formula/Definition	Context
Overexposure $\Delta P$	$P_{obs,s}(e) - P_{baseline}(e)$	Twitter contagion (Ferrara et al., 2015)
Valence Regression	$y = a\,x + b$, $R^2$	Exposure→response (Ferrara et al., 2015)
Retweet tendency $t$	$s_{uv}\,e^{c_i-1}$, thresholded by $\tau$	Simulation (Fan et al., 2017)
Branching number $\sigma$	$\langle k\rangle\,p_e\,p$	Epidemic model (Jarynowski et al., 2013)

4. Tie-Weighted and Multimodal Propagation

Tie structure and propagation path specificity are central in discriminating types and rapidity of contagion:

• Tie-Strength Metrics:
  • Common-friends ($S^{CF}_{ij}$), Reciprocity ($S^{Rec}_{ij}$), and Retweet Frequency ($S^{RT}_{ij}$) reflect the embeddedness and bidirectionality of the communication channel (Fan et al., 2016, Fan et al., 2020).
  • Weaker ties are empirically found to foster faster spread of emotions with higher arousal/valence (notably anger) (Fan et al., 2016, Fan et al., 2020, Fan et al., 2017).
• Dynamic SI Models: Propagation velocity ($v$) and coverage ($C$) are tracked between outbreak "awakening" and "peak" via time-series landmarks (Fan et al., 2016).
• Signed Simplicial Models and Higher-Order Interactions: Emotional group influences are captured via higher-order mean field parameters ($\lambda_2$), with critical thresholds for bistable epidemic transitions, jump sizes ($\Delta\rho$), and hysteresis areas ($A$) (Ma et al., 2024).
• Diffusion Process Scores: Net “contagion score” per user ($\xi_c = \sum_{i}(\mathcal{I}_i - \mathcal{O}_i)$) integrates inflow (exposure to others) and outflow (user engagement back to others), modulated by content and profile homophily, and temporal causality (Mittal et al., 2022).

5. Biobehavioral and Cross-Modal Contagion Metrics

In controlled dyads and embodied contexts, metrics extend to neural, postural, and speech signal domains:

• Granger-Causal Analysis of Affective States: Time-resolved cross-influence via Granger-causality network reconstruction, with reproduction rate ($R_0$), clustering ($C$), and amplification factors (system/individual variance ratio) mapping the epidemic “strength” of emotional instability (Sun, 16 Oct 2025).
• Hyperscanning Neurograph Metrics: In EEG-based settings, emotional contagion is captured as rising cross-subject cosine similarity in trialwise embedding spaces (derived from functional Graph Contrastive Learning), DGC classifier performance drifts, and behavioral regression slopes over time—signalizing neural synchrony and performance “spillover” (Huang et al., 2023).
• Kinematic Synchrony: In social robotics, contagion is measured via convergent shifts in torso inclination, increased frequency of micro-movements, and high-frequency power-spectral density in observer posture—each parametrized as response to robot motion rate (Casso et al., 2022).

6. Event-Driven and Branching Process Estimators

For bursty, viral, or protest-driven contagion, classical epidemic and Hawkes-process formulations are employed:

• SEI/SE Branching Estimators: The probability of exposure per contact ($p_e$), infection per exposure ($p$), and per-invitation resend rate ($p_s$), plus derived branching number ($\sigma$), track amplification due to emotion in viral message cascades (Jarynowski et al., 2013).

  Table: Branching Process Metrics in Natural and Artificial Virals

  Campaign	$p_e$	$p_s$	$p$	$\sigma$
  Non-incentivized	0.31	0.07	0.24	0.23
  Incentivized	0.14	0.07	0.48	0.26
  Stop-ACTA (emotional)	0.87	0.33	0.38	0.28

  The fourfold increase in effective branching under emotional stimuli is a robust operational marker of emotional amplification (Jarynowski et al., 2013).

• Hawkes Process Self/Cross-Excitation: In multivariate event models for chatroom emotion, self-excitation coefficients ($\alpha^{e,e}$), cross-excitation ($\alpha^{e,f}$), memory decay ($\gamma^e$), and the endogenous-to-exogenous ratio ($R^e_{endo}/R^e_{exo}$) statistically parse the relative roles of peer-driven and stimulus-driven emotional outbreaks (Luo et al., 2024). Positive emotions propagate with higher branching rate ($3\times$ vs. negatives), while negative emotions linger longer ($2\times$ mean memory).

7. Contagion Metrics in Online Discourse and Multimodal Interactions

Specialized metrics operationalize how emotions propagate in microtext and collaborative content:

• Emoji Cascade Metrics: In developer communities (GitHub), sentiment is encoded as normalized, weighted aggregates of reaction counts, with positivity “cascades” defined by early, consistent valence. Signal strength is quantified by correlation ($r=0.679$), odds ratio (OR=3.2), effect size (Cohen’s $d=2.393$), and positivity-to-negativity ratio (23:1) (Kraishan, 4 Nov 2025).
• Empathy (Mood Coherence $\chi^2$): On Facebook, emotional alignment between post and comments is formalized as the Pearson $\chi^2$ statistic on a 2×2 post/comment contingency table (high values indicate strong emotional tuning, i.e., “empathy” scores above 4 are statistically significant) (Guazzini et al., 2016).

Summary Table of Core Emotional Contagion Metrics

Metric Type	Mathematical Expression / Definition	Paradigm / Setting	Primary Reference
CCC	Concordance correlation of predicted vs. true valence series	Dyadic annotation	(Barros et al., 2019)
Granger Score	$F_{i\rightarrow j} = \ln(RSS_0/RSS_1)$	Longitudinal network	(Sun, 17 Oct 2025)
Branching $\sigma$	$\langle k \rangle p_e p$	Epidemic/viral spread	(Jarynowski et al., 2013)
Overexposure $\Delta P$	Difference vs. null in pre-post sentiment class frequencies	Microblog exposure	(Ferrara et al., 2015)
Tie Preference	Tie-strength selection per emotion ($S_{ij}^{CF}, S_{ij}^{RT}$)	Diffusion networks	(Fan et al., 2016, Fan et al., 2017)
Hawkes $\alpha^{e,f}$, $\gamma^e$	Peer-to-peer (self/cross) excitation, memory	Event-based chat systems	(Luo et al., 2024)
Contagion Score $\xi_c$	Inflow–outflow balance adjusted by homophily/correlation	Video sharing	(Mittal et al., 2022)
Positivity Ratio	$C^{+}/C^{-}$ in emoji cascade detection	Collaborative coding	(Kraishan, 4 Nov 2025)
Empathy $\chi^2$	Pearson goodness-of-fit for comment-post emotional match	Wall posting platforms	(Guazzini et al., 2016)

Conclusion

Emotional contagion metrics encompass a broad toolkit: continuous time-series concordance, Granger-based influence indices, branching factors from epidemic theory, Hawkes-process excitation/memory, tie-weighted diffusion selectivity, event-driven overexposure, susceptibility classifications, and cross-modal coherence statistics. Their design and interpretation are context-specific, but all serve the fundamental goal of quantifying the dynamics, intensity, and selectivity of affective transmission in human and human–machine systems. Benchmark values and thresholds—such as the CCC in dyadic affect modeling, critical posting gaps for competitive outrage, branching numbers for viral spread, or peak correlation and effect size for emoji-driven contagion—provide replicable ground truths for both experimental replication and real-time intervention design (Barros et al., 2019, Sun, 17 Oct 2025, Ferrara et al., 2015, Fan et al., 2017, Luo et al., 2024, Kraishan, 4 Nov 2025).