Mice that win fights every day for weeks and are then kept from fighting do not calm down. When they meet an opponent again, they attack harder and longer than before. This is the "aggression withdrawal effect," and it is one of the clearest animal demonstrations that winning can become rewarding in itself. Online platforms run a version of this experiment millions of times. They suspend users for hostile behavior, usually for a fixed period, and then let them back in. The trouble is that we cannot simply watch what returning users do and read off whether their aggression rebounded. Someone who knows they are being watched behaves the way a survey respondent answers: they tell the observer what the observer wants to hear. Post-suspension behavior in the place that punished you is, in effect, a biased self-report.
This proposal treats that bias as the central measurement problem. It splits each sanctioned user's behavior by how closely it is watched and how much is at stake. It also uses the written appeals users submit to get unblocked as an explicit self-report of intent. The aim is to separate three explanations that current data cannot tell apart: a real rebound (supercompensation, meaning the suppressed behavior overshoots its old level once the restriction ends), strategic masking, and genuine calming. The design also tests the two predictions from the mouse work that matter most and that animal studies have barely probed. The first is that rebound should appear mainly in users whose aggression previously paid off. The second is that its size should depend on how long the deprivation lasted.
Core Claim
The ban-evasion literature appears to contradict the withdrawal effect. Characterizing, Detecting, and Predicting Online Ban Evasion studied 8,551 verified Wikipedia sockpuppet pairs. It found that users who go on to evade bans are "subdued" malicious actors: they use less overtly aggressive language and survive far longer before suspension (18 days versus 4 hours). Meanwhile, the mouse work in Neurotransmitter Genes in the Nucleus Accumbens That Are Involved in the Development of a Behavioral Pathology After Positive Fighting Experiences and Their Deprivation reports that 14 days of fighting deprivation after 20 days of winning produced higher aggressive motivation and longer attacks. In those mice, 1,111 of 2,629 winning-induced gene expression changes in the nucleus accumbens persisted through abstinence.
Our claim is that the contradiction is a measurement artifact of one specific, testable kind. Behavior observed under high observation and high stakes understates aggressive drive, just as socially desirable survey answers do. If that is right, then after a temporary block ends we should see three things:
- Aggression in the venue that imposed the block falls.
- Aggression in less-watched venues rises above the user's own pre-block level.
- Both effects are concentrated among users whose earlier aggression "won", for example by prevailing in edit disputes.
If instead aggression falls everywhere, regardless of prior winning, the subdued profile reflects genuine attenuation and the mouse analogy fails for humans online. Either outcome is informative. Neither can be reached by the obvious study, which compares a user's toxicity just before and just after a block.
Research Program
The deprivation setting. Wikipedia supplies the core data because its block log, edit histories, talk pages and unblock requests are public, and because its sanctions vary in exactly the ways we need. Full blocks of differing durations remove a user from the whole site. Page- or namespace-restricted blocks remove the user only from the place where the conflict happened while leaving the rest of the site open. Sockpuppet child accounts from the ban-evasion dataset represent a third channel, where the user's original identity is no longer watched at all. Together these give, for the same person, behavior in a high-observation venue, in lower-observation venues, and under an unmonitored identity.
The "winner" history. The analogue of the mouse A20 group is defined from behavior, not text. A user is a prior winner if, in the weeks before the block, their contested reverts persisted and their opponents' did not. The user who loses edit wars is the analogue of the chronically defeated mouse. Users with comparable activity who were not in conflicts serve as controls. This mirrors the three-group design (controls, winners, winners-then-deprived) of the chronic social conflict model described in Prospects for applying the chronic social conflict model in medical and biological research. It turns the model's central moderator into something observable in logs.
Measuring aggression without trusting one label. Text classifiers are not neutral instruments here. On the TRAC-1 benchmark, Benchmarking Aggression Identification in Social Media reports a best weighted F-score of only 0.64 on English Facebook comments and 0.60 on English tweets, with covert aggression the hardest category. Moderation in the Wild finds that annotator subjectivity explains 34% of the variance in rater disagreement on comment style.
A returning user who shifts from overt to covert hostility is therefore exactly the case a classifier will miss. Missing it would manufacture the "subdued" finding. We therefore pair every text-based score with behavioral outcomes that need no annotator: revert counts, participation in new edit wars, reports filed against the user, and re-sanction. We also test whether the classifier's output means the same thing before and after a block. Following the measurement-invariance logic that Measurement Invariance of the Bergen Social Media Addiction Scale Across Genders applied across genders, we check whether the mapping from text features to the behavioral outcomes is stable across the pre- and post-block periods. That paper itself flags longitudinal invariance as untested.
The explicit self-report. Unblock requests are self-reports written under the strongest possible demand characteristics, because the reader decides whether the writer gets back in. We score them for contrition, stated commitments and acknowledged fault. We then ask whether these statements predict post-block behavior, and whether administrators' grant decisions track them. The prediction from the masking account is specific. Among prior winners, appeal contrition will be unrelated or even inversely related to subsequent aggression in low-observation venues, yet it will still strongly predict being unblocked. This is the online counterpart of the pattern in Impact of Push Notifications on Physical Activity and Sodium Intake Among Patients with Hypertension, where self-reported engagement improved while objective outcomes did not move.
Identification. Blocks are imposed when a user is at a behavioral peak, so naive before-and-after comparisons will show calming from regression to the mean alone. We handle this in two ways. First, we compare blocked users with users who reached a similar escalation, were warned on their talk page, but were not blocked, matched on activity and conflict history. Second, we exploit the fact that different administrators assign systematically different durations for similar offenses. Administrator leniency, measured on that administrator's other cases, becomes an instrument for deprivation length. This gives a dose-response curve over durations ranging from a day to several weeks. The mouse literature, by contrast, has essentially one point on that curve: 14 days. To avoid selecting on persuasive appeals, the main analysis uses only blocks that ran their full term. Appeals are analyzed separately.
Baselines
The naive reference point is the within-user comparison of aggression in a fixed window before the block versus the same window after it, on the same venue and with the same classifier. This is what a platform dashboard would report. Because of regression to the mean, we expect it to show a decline for everyone. The decline it reports is the number our corrected estimates must explain away. Our method "beats" it if the matched and instrumented estimates diverge from it in sign or in significant magnitude for prior winners.
The strongest existing method for anticipating what sanctioned users do next is the feature-based predictor of Characterizing, Detecting, and Predicting Online Ban Evasion. It uses edit-history, psycholinguistic and embedding features, and reports an AUC of 0.78 for predicting evasion and 0.85 for early detection after three child-account edits. We will re-implement its feature set on our cohort, with the same data splits and the same hyperparameter-search budget as our model, to predict post-block re-sanction and low-observation-venue aggression. The reported 0.78 is for a different target, so it is context, not the bar. The bar is the re-implemented model's AUC on our target.
Our model adds two things: the prior-winning history and the observation-level split of outcomes. It beats this baseline if it improves held-out AUC with a confidence interval that excludes zero improvement. More importantly, it must also explain where the extra signal comes from.
For the macro view, the reference is the community-level comparison style of The Effect of Moderation on Online Mental Health Conversations. That study found toxicity fell from 5.2% to 1.0% of messages after moderation was introduced. We reproduce the analogous aggregate on our Wikipedia cohort to show how much individual-level rebound such aggregates can hide.
Two ablations isolate the key ingredients. Removing the winner/loser/control distinction tests whether rebound is specific to rewarded aggression, as the mouse work predicts. Collapsing all venues into one outcome tests whether the observation-level split is what reveals the rebound. A third ablation swaps the text classifier for behavior-only outcomes. If the conclusions reverse, the result was a classifier artifact.
Why not just…?
Why not just survey users before and after their suspension? Surveying is the textbook move, and it fails here for the reason the proposal exists. Respondents underreport stigmatized and addictive behavior, as noted in Measurement Invariance of the Bergen Social Media Addiction Scale Across Genders. Answers bend toward perceived researcher expectations, as argued in Tag2Risk. Instruments such as the aggressive-defending subscale in Development and Validation of the Defending Behavior Scale of Cyberbullying for Adolescents have never been checked against behavioral logs. Recruiting recently blocked users would also be heavily self-selected. Unblock appeals give us a self-report that already exists for every case, with known stakes. A small optional survey arm could be added later, but only after the trace-based ground truth exists to calibrate it.
Why not just run a toxicity classifier over each user's timeline? That is the naive baseline, and we run it. Its failure modes are concrete. It cannot separate a real decline from regression after a peak. Its accuracy is weakest on covert aggression, which TRAC-1 shows is hard even for the best systems. And a user who switches register after a block changes what the classifier's scores mean. The behavior-only ablation and the invariance check decide whether this is a real threat on our data. If text and behavioral outcomes agree, the classifier was adequate and we will say so.
Why not just do a matched difference-in-differences and skip the venue split? This is the most serious alternative. A good matched comparison alone may recover the average effect of blocks. What it cannot do is tell masking from genuine attenuation, because both predict a decline on the watched venue. The venue split and the sockpuppet channel are the only parts of the design that make these two accounts diverge. We will report the plain difference-in-differences estimate alongside the full design. If the venue-specific estimates do not differ, we will conclude that the simpler design was sufficient.
Why not just use the ban-evasion detector as-is? The detector predicts whether someone will evade. It does not estimate whether deprivation intensifies aggression. Its "subdued" finding comes from caught sockpuppets only, so it may reflect which evaders get detected rather than how evaders behave. Restricted blocks, where the user stays visible under their own name, avoid that detection filter. Comparing them with the sockpuppet channel tells us how much of the subdued profile is a selection effect.
Decisive Checks
The first decisive result is the sign of the instrumented post-block effect on low-observation-venue aggression among prior winners, with each user compared to their own pre-block baseline. A positive effect that is absent or reversed in prior losers and controls supports the withdrawal account. A negative effect across all groups refutes it for this setting.
The second is the dose-response shape over block duration. The baseline-inhibition-supercompensation idea, also invoked for children's activity after lockdown in The "supercompensation" effect of children's lockdown during COVID-19, predicts an overshoot that grows with the length of restriction, at least up to some point. A flat curve would count against it.
The third is the appeal check: whether contrition language predicts administrators' decisions while failing to predict behavior in low-observation venues among winners.
We are explicit about what the logs cannot show. They say nothing about reward circuitry, the dopamine and opioid gene changes reported in mice, or the theory-of-mind belief tracking modeled in A normative theory of social conflict. A positive result establishes a behavioral parallel with the same moderator and a measured duration dependence. It does not establish a shared mechanism. That parallel would still give animal researchers a reason to test deprivation lengths other than 14 days. It would also give platforms a reason to stop reading calm behavior after a block in the watched venue as evidence that the block worked.
References
- Characterizing, Detecting, and Predicting Online Ban Evasion
- Neurotransmitter Genes in the Nucleus Accumbens That Are Involved in the Development of a Behavioral Pathology After Positive Fighting Experiences and Their Deprivation: A Conceptual Paradigm for Data Analysis
- Prospects for applying the chronic social conflict model in medical and biological research
- Benchmarking Aggression Identification in Social Media
- Moderation in the Wild: Investigating User-Driven Moderation in Online Discussions
- Measurement Invariance of the Bergen Social Media Addiction Scale Across Genders
- Impact of Push Notifications on Physical Activity and Sodium Intake Among Patients with Hypertension: Microrandomized Trial of a Just-in-Time Adaptive Intervention
- The Effect of Moderation on Online Mental Health Conversations
- Tag2Risk: Harnessing Social Music Tags for Characterizing Depression Risk
- Development and Validation of the Defending Behavior Scale of Cyberbullying for Adolescents
- The “supercompensation” effect of children’s lockdown during COVID-19: based on the analysis of changes in physical activity, sleep, and psychology
- A normative theory of social conflict