{{title}}
{{textarea}}
{{task}}
{{select}}
{{team}}
{{input}}
{{solution}}
{{result}}
ICDS Working Paper Digital Reputation
Anger Gets Likes. Fear Gets Shares.
The emotional coloring of a message predicts how far it travels: an analysis of 34,783 negative mentions during corporate crises.
Working paper · five cases · technology vendors · 2022–2024
VM
Vadim Matyushkin Behavioral Scientist & Sociologist · ICDS
Social media analytics treats a message as a symptom by default: someone posted something angry, therefore someone is angry. Sum the symptoms and you get “brand sentiment,” read as a gauge of mood.
But a message has a second nature. It doesn’t only express an emotion in its author — it provokes one in the thousands who read it. And that raises a question sentiment never asks: different emotions push the reader toward different actions. Some make you nod and scroll past. Others make you pass it on.
We measured this directly. All else equal, a fearful message travels three times deeper than an angry one. Anger collects approval and burns out. Fear collects transmission and spreads.
About the data
34,783 negative mentions from five corporate crises at technology vendors (CrowdStrike, Unity, Atlassian, Progress/MOVEit, Datadog), in a window of ±3 weeks around each event. Each message was tagged by its dominant emotion using the NRC Emotion Lexicon (Mohammad & Turney) — an open, standard dictionary of 6,468 words mapped to eight basic emotions. Propagation metrics — reposts and likes — come from social-listening platform data.
A Message Has Two Natures
Emotion research has long distinguished feelings not by their sign (pleasant vs. unpleasant) but by the action they prompt. Fear mobilizes: a danger signal evolved to demand an immediate response — warn, flee, check. Anger discharges in the very act of expression: voicing outrage is itself the release. Sadness turns inward and prompts inaction.
If that holds, then on social media each emotion should leave a different behavioral trace. Anger — many messages, but each self-contained: the person vented and settled. Fear — fewer messages, but each drags a chain of transmission behind it: read, alarmed, forwarded.
Sentiment doesn’t see this distinction. To a negativity counter, an angry {chr(0x201c)}what a disgrace{chr(0x201d)} and a fearful {chr(0x201c)}our data is exposed{chr(0x201d)} are identical units of negative. Yet they behave in opposite ways.
Depth of Transmission: Fear Travels Farther
We tagged every negative message by its dominant emotion and looked at what happened next — whether it was reposted, and if so, how deep the chain ran.
| Emotion | Messages | Reposted | Depth (median) | Like / repost |
|---|---|---|---|---|
| Fear | 6,977 | 4.2% | 3 | 1.3 |
| Sadness | 2,892 | 4.1% | 2 | 3.1 |
| Anger | 3,961 | 3.5% | 1 | 4.6 |
The share of messages that get reposted at all is nearly identical across the three emotions (3.5–4.2%, difference not significant). The gap is not whether a message is picked up — it’s how far it goes once it is.
When a fearful message is reposted, it spreads three times deeper than an angry one: median depth of 3 versus 1, and this is robust (fear > anger: p = 0.0004; fear > sadness: p = 0.022, on the logarithm of reposts, outlier-resistant). Anger — the most numerous in the moment — travels the shortest.
A like and a repost are two different acts
The ratio of likes to reposts exposes the mechanism. For anger it is 4.6: an angry message is liked nearly five times more often than it is passed on. A like is agreement — “yes, outrageous” — and there the reader’s involvement ends. For fear the ratio is 1.3: a fearful message is transmitted almost as often as it is endorsed. Anger seeks agreement. Fear seeks distribution. Meeting anger, the reader nods and likes — the tension is released. Meeting fear, the reader hits “share” — because fear demands that others be warned.
Why This Explains How Long Crises Last
The finding has a sequel at the level of the crises themselves. In a separate analysis of the same corpora, we measured how long a crisis persists in the media field after its peak — and what that duration tracks.
Crises with a high share of fear at the peak smolder markedly longer (fear-to-duration correlation r = 0.57; fear and sadness combined against afterglow r = 0.60). On our five-case sample this is a signal, not proof — but the direction matches the transmission mechanism.
The loop closes. A fearful crisis lasts longer not because people stay afraid longer — but because a fearful message is more viral. It reproduces itself through the reader: everyone who was alarmed and forwarded it creates a fresh wave of reads. An angry crisis is louder on day one, but anger discharges into likes and never pulls a chain — it burns out.
A crisis’s loudness and its longevity are different things, and different emotions govern them. Anger delivers loudness. Fear delivers longevity.
What Follows for Practice
Measure the composition of negativity, not its share.
Two crises with the same share of negativity can have opposite fates. The one dominated by anger will blow over in a week. The one dominated by fear will smolder for months and return in waves. A dashboard that counts negativity as a single mass cannot tell them apart — and the difference determines the entire communication strategy.
Fear is a call to act; anger, a call to hold steady.
An angry crisis can often simply be waited out: the outrage discharges on its own. A fearful crisis cannot — it doesn’t discharge, it accumulates until the source of uncertainty is removed. Fear is answered with information that dispels the unknown; anger, with time.
The early emotional mix predicts the trajectory.
The emotional profile of the first days — not the anger, but specifically the share of fear — gives an early read on whether a crisis will have a sequel. It’s an indicator found in neither mention volume, nor sentiment, nor market data.
What Is Established — and What Isn’t
Firmly established
On 34,783 messages: a fearful message travels deeper than an angry or sad one (p < 0.001), and the three emotions provoke different reader actions — anger collects likes, fear collects reposts (ratio 4.6 vs. 1.3).
Established as a signal
The link between fear and crisis duration (r ≈ 0.6) rests on five cases and needs a larger sample to confirm.
Method limits
Emotion is tagged by dictionary, without context, negation, or irony. The dictionary is standard and open, which ensures reproducibility, but a contextual classifier would sharpen accuracy. A repost as a measure of transmission does not separate an endorsing repost from a condemning one.
Analytics measures what was written. Here we measured what the writing does to those who read it.
Methodology note
Corpus & method
Corpus: 34,783 messages auto-labeled negative in sentiment, from five crisis corpora, window −3 to +21 days around the event. Emotional tagging: NRC Word-Emotion Association Lexicon (Mohammad & Turney, 2013); a message’s dominant emotion was set by the majority of dictionary matches among fear / anger / sadness; messages with no emotional words or a tied profile were excluded from pairwise tests. Propagation metrics: a repost is treated as an act of transmission (a reader action), a like as an act of approval. Depth of transmission was measured by repost count among reposted messages; for outlier resistance, comparisons used the logarithm (log10(reposts+1)), one-sided Student’s and Mann–Whitney tests. The share reposted was compared with χ². Duration link: at the aggregate level of five cases, we measured the share of fear/anger/sadness at the peak and the length of elevated activity after the peak; Pearson and Spearman correlations. The case sample (n = 5–9 depending on data availability) is insufficient for statistical generalization — those results are reported as a directional signal. Institute for Communications and Data Science. Working draft for internal discussion.
VM
About the author
Vadim Matyushkin
Behavioral Scientist & Sociologist · ICDS
Vadim Matyushkin is a psychologist and sociologist with close to twenty years of research into digital behavior, trust, and information dynamics, and a researcher behind the Institute of Communication and Data Science (ICDS). His work has supported organizations including Coca-Cola, PepsiCo, Mars, Danone, Nestlé, and Bayer in moving from self-reported survey data toward direct behavioral evidence.
ICDS — Institute of Communication and Data Science is an independent research institute focused on understanding how trust, behavior, and reputation are formed in digital environments. ICDS operates as an intellectual institution — not an agency, not a platform, and not an educational provider.