Crowd dynamics and crush — evidence synthesis (draft v4)

Domain: `crowd-dynamics` Revised: 2026-08-16 — supersedes v3 (2026-08-12). What changed in v4: the 2026 Purple Guide tranche is folded in — Still's definitional 1 person/m² threshold with its citation; the 2–4 persons/m² risk-assessed planning range from the Venue & Site Design capacity flow chart; the egress arithmetic (66/82 per metre width per minute, evacuation time bands, the largest-single-exit discount); the campsite-density category distinction; parental counterflow; and the show-stop / keep-the-show-running doctrine with its detection layer. All new material is guidance-grade and cited to the chapter entries now in the library.

Sources read: Fruin 1993 (full text); Purple Guide 2026 chapters — Zone Ex, Crowd Management (10/30 sections, working group includes G. Keith Still), Venue & Site Design (15/22 sections including Figure 1 read visually), Campsites (complete), Contingency & Emergency Planning (13/27) — via authenticated subscriber session; Still's published density figures; MGMI's 51-jurisdiction regulatory atlas; and abstracts of Hsieh 2009, Moitinho de Almeida & von Schreeb 2019, Bain & Bartolo 2019, Chang 2024, Alaska 2017, Johansson 2012, Moussaïd 2011, Helbing 2000, Burkle & Hsu 2011, Ngai 2009.

The claim in one paragraph

Deaths in crowd disasters are asphyxial, not traumatic, and occur at densities that are known, measurable, and routinely reached at events that pass inspection. Between 1980 and 2007 alone, 215 recorded stampede events killed 7,069 people and injured at least 14,078. The field has largely abandoned "panic" as the explanation and replaced it with successive systemic failure. But the prevention measures that follow from that understanding — crowd management and venue design — have never been tested for effectiveness, and the medical response failures that recur are administrative rather than clinical: delayed decisions, poor triage, lost records. The lethal variable is density, not crowd size, and density is the one variable event medical plans almost never specify.

Scale

Hsieh and colleagues analysed 215 human stampede events reported between 1980 and 2007, assembled from LexisNexis and multiple internet news agencies because the peer-reviewed record was too sparse to support the analysis. The events produced 7,069 deaths across 213 events with fatality data, and at least 14,078 injuries across 179 events with injury data.

Their multivariate analysis identifies what raises mortality: events occurring in developing countries and outdoors. Sports, religious, music and political events all increased the relative number of deaths, as did a unidirectional mechanism — a crowd moving one way with no relief. In bivariate analysis the highest median death counts were in the Middle East, in developing countries, outdoors, and at religious events.

Two things are worth saying about that dataset. It is the best epidemiology the field has, and it is built substantially on news reports because nothing better exists. The authors say so themselves and recommend standardised data collection. Eighteen years later, in the review below, the same recommendation is still outstanding.

Mechanism

Fruin states it flatly: "Virtually all crowd deaths are due to compressive asphyxia and not the 'trampling' reported by the news media."

Chest excursion is prevented by sustained circumferential pressure. The victim commonly remains upright, remains conscious for a period, and is frequently invisible to responders metres away — in a crowd at lethal density there is nowhere to fall. Fruin describes pressures that "make it difficult to breathe," and notes that "the heat and thermal insulation of surrounding bodies cause some to be weakened and faint." Access to those who fall is impossible; removal "can only be accomplished by lifting them up and passing them overhead to the exterior of the crowd."

The sentence that should govern event medical planning: "Victims of compression asphyxia can be revived only if resuscitation begins quickly."

And we do not actually know the injury spectrum. Chang and colleagues, writing on Itaewon in 2024, make the uncomfortable point directly: while traumatic asphyxiation has been widely implicated as the primary cause of death, "the wider spectrum of injury patterns and their causative mechanisms remain poorly identified." They attribute this to limited and restricted access to official reports, incomplete documentation, and reliance on unofficial sources. For the most scrutinised crowd disaster of the last decade, in a high-income country with a functioning health system, the clinical picture is still incomplete.

Density thresholds

Where crowd dynamics begins is now on the record with its author's citation. The 2026 Purple Guide Crowd Management chapter — whose original working group includes Keith Still — defines crowd dynamics as "the study of the how and where crowds form and move above the critical density of more than one person per square metre," attributed in the source to Still's doctoral thesis ^[1]. Below that line, individuals walk freely and the discipline's models do not apply; above it, density, flow and space have to be managed.

And the planning range now has a primary source. The Venue & Site Design chapter's capacity flow chart sets expected densities of 4, 3 or 2 persons per square metre, selected by risk assessment of the audience profile — profile, not headcount, chooses the figure ^[2]. The published figures across sources:

| Threshold | Value | Source | |---|---|---| | Crowd dynamics begins (definitional) | >1 person/m² | Still, cited in ^[1] | | Expected-density planning range, risk-assessed by audience profile | 2–4 persons/m² | Figure 1, ^[2] | | Green Guide, standing viewing areas | 4.7 persons/m² | SGSA Green Guide | | Still's stated upper limit for standing space | 5 persons/m² | Still | | "Crowd becomes almost a fluid mass" | ~7 persons/m² | Fruin 1993 |

Read as a single scale, UK practice plans up to the bottom edge of the danger band: the top of the planning range (4/m²) abuts the region where the crush literature places physiological compromise, and sits well below Fruin's 7/m² fluid-mass regime. Operational guidance is two to three times more conservative than the number the literature is famous for — a point v3 made from inference and v4 can make from the Guide's own figure.

The category boundary matters as much as the thresholds. A campsite built to the Guide's own density standard — 476–543 tents per hectare at 2–2.4 occupants — runs at roughly 0.1 persons/m², an order of magnitude below the 1/m² line at which crowd dynamics is defined. Campsite density is governed by fire-safety legislation (fire spread and appliance access), not by crush mechanics, and the Guide says so explicitly ^[3]. Conflating the two density literatures is a category error this synthesis exists to prevent: a campsite is not a low-risk crowd; in the technical sense it is not a crowd at all — until its population converges on a stage.

Still's criticism of the Green Guide figure carries: it "fails to highlight that the demographics (individual sizes) need to be factored into the density analysis." A limit in persons per square metre assumes a person of a given size. It is a different limit for adults in winter coats, and a different limit again where children are present — see the pediatric synthesis, where children showed 2.5 times the adult odds of presenting after a crush injury at the same event. The Guide operationalises the same logic at admission: children under five "may not be appropriate" at events by reason of crowd scale or density ^[1].

Above roughly seven per square metre, Fruin reports shock waves with force "sufficient to lift people off of their feet and propel them distances of 3 m (10 ft) or more." People "may be literally lifted out of their shoes, and have clothing torn off."

Helbing and Molnár's social-force model is the foundation the whole crowd-safety simulation literature rests on — pedestrians modelled as particles subject to attractive and repulsive forces. Its value to an event planner is indirect but real: it is what makes layout and flow review possible before anyone is on site. It does not yield density thresholds, and should not be cited as if it did. The practitioner-side counterpart is Still's DIM-ICE model — Design, Information and Management across Ingress, Circulation and Egress — in use over thirty years, and honest about its own epistemics: it "evolved from research into crowd incidents" and lets the user "see risks through the lens of past failures." A checklist distilled from disasters: weaker than a physical model, much stronger than intuition ^[1].

Moussaïd, Helbing and Theraulaz supply the mechanism from the modelling side: combining simple visual-cognitive walking heuristics with body collisions generates crowd turbulence at extreme densities — "a phenomenon that has been observed during recent crowd disasters." Turbulence is not a metaphor here; it is the emergent regime where individual control ends.

Egress arithmetic

New in v4, and the numbers every evacuation-time claim resolves to. UK escape rates follow the Green Guide 6th edition, which rests on BS EN 13200-1:2003: for a 1.2 m width, 79 people per minute on a stepped surface (66 per metre width per minute) and 100 on the level (82 per metre width per minute). Higher rates "should only be used if specifically justified," and the risk assessment must weigh alcohol, drugs and child buggies/prams as factors that slow escape below the standard rates — the same audience variables that drive presentation rates also degrade egress ^[2].

Three structural rules travel with the constants. Egress width is discounted by the largest single exit — worst-single-failure design. Evacuation time is banded (>5, 5–10, >10 minutes, set by risk assessment), and rate × width × time yields evacuation capacity. And capacity overall is the minimum of three independent capacities — ingress, holding, evacuation: "lowest figure = safest capacity." A venue can be egress-limited with holding space to spare ^[2].

Two asymmetries complete the picture. Ingress is "a steady flow over a longer time"; egress is "a condensed flow over a shorter time" — a system sized for arrival is undersized for departure, because departure is triggered by a single event, the last chord ^[1]. And the provenance chain itself is a finding: the rates the entire UK events industry plans on are three documents deep and terminate in a European seating standard, not empirical crowd research. Anyone wanting to refine them must go to the standard.

Forces

From Fruin, via post-incident deformation and guardrail experiment:

- Bent steel railings after fatal incidents indicate forces exceeding 4,500 N (1,000 lb) - Ibrox 1971, bodies piled 3 m high: chest pressures of 3,600–4,000 N (800–900 lb) at the bottom, assuming half the weight above concentrated on the upper body - Three people leaning on a guardrail generated 792 N; pushing, 609 N (US NBS) - Under simulated panic, five people generated 3,430 N (766 lb) (Australian Building Technology Centre)

Five people generate most of the force that bends structural steel. That is the number for a planning meeting, because it defeats the intuition that crush requires a very large crowd.

The panic myth is dead in the literature — and it matters clinically

This is the most important shift since v1, and it is now documented rather than asserted.

Moitinho de Almeida and von Schreeb reviewed 64 publications on human stampedes, 34 of them published between 2013 and 2016. Their finding: "The common belief that they result from an irrational and panicking crowd has progressively been replaced by studies suggesting that successive systemic failures are main underlying causes."

Fruin had already made the distinction in 1993 and it deserves restating because it is still absent from most event plans. Flight follows a real or perceived threat and is "frequently mislabelled a panic"; investigation "usually shows that flight was a reasonable group reaction," often showing "mutual cooperation and assistance." A craze is "a competitive rush to obtain some highly valued objective" — and general admission and festival seating "cause craze like competition."

Helbing, Farkas and Vicsek's Nature simulation is the paper most often cited for panic, and reading it undercuts that use: they find an optimal escape strategy consisting of a mixture of individualistic behaviour and collective herding — that is, neither pure panic nor pure order is optimal, and their simulations "suggest practical ways to prevent dangerous crowd pressures." The paper is about jamming by uncoordinated motion, not about irrationality.

Why this matters at the treatment tent: if the cause is systemic failure, the intervention is upstream and structural. If the cause is panic, the intervention is crowd control — restricting behaviour — which is what the Purple Guide correctly distinguishes from crowd management, and which is the response most likely to make a unidirectional crush worse.

Why the crowd keeps pushing while its front rank is dying

Fruin: crowds behave as "a series of intermeshing behavioural cells" with limited communication between them. "Most crowd incidents exhibit a lack of front to back communication. People in the rear of the crowd press forward while those in front experience severe distress." Lighter densities at the rear allow free movement while the front is immobile, and "the collapsing of front ranks gives a false perception of forward movement."

Bain and Bartolo measured this. Using tens of thousands of road-race participants in starting corrals — real dense human crowds, not simulations — they established that speed information propagates over system-spanning scales through polarized crowds, while orientational fluctuations are locally suppressed. They build a hydrodynamic theory from it and demonstrate predictive power, describing human groups as active continua and offering the result as "quantitative guidelines for crowd management."

That is the physical substrate of Fruin's observation. Speed signals travel the length of the crowd; direction signals do not. A crowd can therefore transmit "we are moving" end-to-end while the information that would let the back understand why the front is moving — or that it is collapsing rather than advancing — stays local. Front-of-crowd appeals not to push are, in Fruin's words, "ineffective during a serious crowd incident in progress," and now there is a measured reason why.

Any plan whose crush mitigation is a public-address announcement to the front of the crowd is relying on the one intervention its own source literature says does not work.

Counterflow has a predictable, plannable generator. The Contingency chapter notes that where children are separated from parents — crèches, play areas — "parents will normally try to reach their children, even if this means going against the normal direction of escape," and requires that parents be told the evacuation arrangements at drop-off, when they are calm ^[4]. A subpopulation moving against egress is a recognised crush contributor; here is one that activates simultaneously, at the worst moment, and can be defused for free before the event starts. The engineering counterpart: separate exit lanes "where there is a possibility of conflicting crowd flow from people moving in opposite directions" ^[1].

Helbing and Mukerji's analysis of the Love Parade reaches the same conclusion by a different route, and supplies the vocabulary. They frame the disaster as a systemic failure rather than a crowd behaving badly, and describe the physical regime that killed people as crowd turbulence — a "crowd quake" — explicitly distinct from panic. They propose a criticality scale for how close a crowd is to that regime.

That framing is the one to carry, because it relocates responsibility. A crowd quake is a predictable consequence of density and geometry, not an eruption of irrationality, and the decisions that produce it are made months earlier on a site plan. Cite the peer-reviewed, open-access version (EPJ Data Science 1:7, 2012), not the SSRN working paper the library originally held.

What actually fails in the medical response

Moitinho and von Schreeb name the recurring response failures: "Delay in decisions, poor triage, or loss of medical records are common problems in the response, which may worsen the outcome."

None of those three is a deficit of clinical skill. They are decision latency, triage discipline under a casualty profile nobody rehearsed, and documentation collapse. All three are trainable and none appears in any US mass-gathering regulation we have surveyed. Connecticut's 1971 rule is the sole exception on the third point — it mandates patient records by name, address and tentative diagnosis, written into the staffing regulation itself.

Decision latency now has an addressable structure. The Crowd Management chapter's show-stop procedure names crowd collapse and excessive crowd density among its internal triggers, requires identifiable persons with authority to initiate, and requires the detection layer: trained crowd spotters in elevated positions at front of stage, operating to a pre-agreed protocol with event control ^[1]. What the listed decision group does not include is the event medical lead — although a cluster of syncope or crush presentations from one sector is density data the spotters cannot see, because it arrives at the medical tent rather than the barrier. The medical team is an independent density sensor with no defined route into the show-stop decision; a presentation threshold routed to control would close the loop cheaply. And the counter-intuitive doctrine deserves deliberate teaching: where an incident can be contained, it is "almost always best to keep the core event going," because the show is what holds a distributed crowd static — stopping it triggers the condensed-egress regime described above ^[4].

Where it happens: the space nobody owns

The Purple Guide's "Zone Ex" chapter names a concept US frameworks lack: the area "within the public realm, between an event perimeter and ports of entry/exit, that is activated as event attendees, transiting as pedestrians arriving or departing the event, form a crowd, which may require crowd management resources to manage safety risks." Its framing of the problem is the whole issue: this space is "outside the control of those organising the event and where there is a lack of clarity over who is responsible for their safety and behaviour." The Crowd Management chapter adds the legal edge: organisers are expected to liaise and plan for these Publicly Accessible Locations while holding no legal powers and no liability there ^[1].

This is where Itaewon happened. It is where Astroworld's ingress crush happened. It is outside the fence, so it is outside the medical plan, outside the insurance, and frequently outside anyone's assigned responsibility. The Guide also names background demand — people using normal transport and road functions "who are not associated with the event and where their presence may have an impact on crowd management operations." No US event medical plan I have examined accounts for it.

The Guide separates two terms our own library had been using interchangeably: crowd management is "the systematic planning for, and supervision of, the orderly movement and assembly of people"; crowd control is "the restriction or limitation of group behaviour." Management is a system; control is an intervention. Both definitions originate with Fruin, thirty-three years earlier.

What this means for medical provision

Fruin's own section on emergency medical facilities is rarely quoted. He warns that "lives have been unnecessarily lost in large crowd incidents by the lack of simple equipment such as stretchers and oxygen," notes that many venues "accommodate the population equivalent of a medium sized city," and states that EMS response times and mass-casualty capability "should be established as part of the crowd management plan" — medical capability as a parameter of the crowd plan, not a separate document filed beside it.

Three consequences follow, each cutting against standard practice:

Access time is the survival variable. If resuscitation must begin quickly and casualties can only be extracted overhead, the distance from a clinician to the densest point of the crowd is a clinical parameter. It is not measured, not specified in any US regulation we have found, and not an input to any published staffing model. The Guide comes closest of any source we hold: all-standing-area ingress planning "will require an assessment of how stewards reach people in need of assistance" — access as a design problem solved before doors open ^[1].

Casualty count is the wrong trigger. A crowd at seven per square metre generates casualties who cannot be seen and cannot self-present. By the time presentations rise, the mechanism has run. Density is the leading indicator; presentation rate is lagging — which is exactly why the medical tent's sector-clustered presentations belong in the show-stop information flow rather than outside it.

Staffing ratios do not address it. There is no controlled evidence that any attendance-to-clinician ratio changes mortality once lethal density is reached. The regulatory literature routinely collapses the distinction between a staffing standard and a crush standard, and our atlas shows the collapse in statute.

The recommendation nobody adopted

Fruin, 1993: "It is recommended that every venue accommodating more than 500 persons be required by law to have a certified crowd manager on staff," with formal testing for certification.

MGMI's 51-jurisdiction survey found no US state that requires this. Thirty-three years, and the most concrete legislative recommendation in the field's most-cited paper has been enacted nowhere in the United States. That is a finding about American event regulation and it belongs in the manuscript.

Prevention strategies — and the hole underneath them

From Fruin, specific enough to put in a plan: metering and throttling arrival rates; early opening and delayed closing to spread ingress and egress; dispersed and balanced entry and exit points rather than one centralised location; access-tree diagrams to map pathway capacity and identify pressure points; arrival-time ticketing; standby power for lighting and communications; a centralised crowd management and communications centre with video coverage of blind spots. And a legal one: a duty to warn — litigation has established that managers know festival seating causes fainting from heat exhaustion, and that those affected are "virtually inaccessible within the crowd."

The 2026 Guide adds two site-plan controls worth naming. Capacity is frozen at ticket sale: site plans must not change after capacity is determined and tickets sold, because alterations to structures and sightlines silently convert sold capacity into overcrowding — with a hold-back of tickets until structures are up as the mitigation. And plans are gridded, version-controlled and distributed so that every agency, including the medical team, navigates by the same current map ^[2].

Saudi Arabia's Hajj programme is the largest sustained natural experiment in this: Alaska and colleagues describe crowd simulation models, grouping and scheduling of pilgrims, crowd management and control engineering, luggage management, video monitoring, and reconstruction of the transport system — and still conclude that the Hajj "holds an increasing risk for future disasters."

Johansson and colleagues review the computer models developed over the preceding decade for crowd movement and show how they can identify health and safety issues, extending them toward microscopic epidemic modelling.

And here is the hole. Moitinho and von Schreeb, having reviewed the whole literature, state that prevention measures "are mainly related to crowd management and venue design, but their effectiveness has not been studied."

That sentence should be read slowly. The field's entire prevention repertoire — the strategies above, the ones in every guidance document including the Purple Guide, the ones MGMI recommends — rests on mechanism, engineering judgement and post-hoc reconstruction. Not one has been shown to reduce mortality in a study designed to test it. That is not a reason to abandon them; they are well-reasoned and there is no ethical trial design available. It is a reason to state their evidentiary status honestly wherever we publish them, which is what this site exists to do.

Still not incorporated

| Source | Identifier | Why still open | |---|---|---| | Illiyas 2013, religious festivals | `10.1016/j.ijdrr.2013.09.003` | IJDRR not PubMed-indexed and Crossref carries no abstract. Matters because Hsieh identifies religious events as the highest-mortality category. Metadata verified 2026-08-12. | | Sime 1995, crowd psychology and engineering | `10.1016/0925-7535(96)81011-3` | Safety Science not PubMed-indexed, no Crossref abstract. The disciplinary fault line — likely the earliest statement of the anti-panic position later confirmed by Moitinho de Almeida. Metadata verified 2026-08-12. | | Sieben & Seyfried 2023, Love Parade eyewitnesses | `10.1016/j.ssci.2023.106229` | Priority. This domain holds no first-person account of a crush; this paper addresses what people inside one actually perceive — the human side of the communication failure Fruin described and Bain & Bartolo measured. Unread; full text not yet accessed. |

Also unread but recorded: Ngai 2009 (`10.1097/dmp.0b013e3181c5b494`) and Burkle & Hsu 2011 (`10.1016/s0140-6736(10)60442-4`) are indexed in PubMed with no abstract — Burkle is Lancet correspondence. Both need full text before their contents are summarised.

Greaves 2003 crush injury consensus is the clinical-management arm of this domain and is tracked separately; the correct identifier is the JRAMC version, `10.1136/jramc-149-04-02` — note that near-identical 2002 (Trauma) and 2004 (Accid Emerg Nurs) republications exist and must not be conflated with it.

Purple Guide chapters are chapter-level extractions with unread remainders listed in their entry notes; Crowd Management in particular is 10 of 30 sections, and its unread majority includes Greenfield Sites and Internal Circulation, both relevant here.

Two sources are unresolvable rather than unread: Fruin 1971/1987 (no digital edition; citable edition is the 2nd, Elevator World, 1987) and Still's 2014 textbook (in print, CRC Press). Still's 2000 Warwick doctoral thesis is free and is the stronger primary source.

Grading

Proposed evidence_grade: B, with the reservation stated in the body.

There is consistent observational support for the density thresholds and for the asphyxial mechanism, drawn from post-incident reconstruction across many decades and countries, and now from at least one quantitative measurement in real dense crowds (Bain & Bartolo). There is no prospective study measuring crowd density and clinical outcome at the same event. Every threshold in circulation descends from post-hoc reconstruction or engineering experiment. No prevention measure has been tested for effectiveness. The best epidemiology is built on news reports by its authors' own account.

A reader should treat "7 persons per square metre" as a well-supported engineering estimate of where control is lost, not a clinically validated cutoff — and should notice that the UK's own 2026 planning range sits at 2–4, selected by audience profile: operational guidance plans up to the bottom edge of the danger band, two to three times more conservative than the number the literature is famous for.

Citations added in v4

- ^[1] Events Industry Forum. The Purple Guide — Crowd Management, 2026 (published 2026-01-26; 10/30 sections extracted; working group includes G. Keith Still). - ^[2] Events Industry Forum. The Purple Guide — Venue & Site Design, 2026 (published 2026-01-26; 15/22 sections plus Figure 1 extracted). - ^[3] Events Industry Forum. The Purple Guide — Campsites, 2026 (published 2026-01-26; complete chapter extracted). - ^[4] Events Industry Forum. The Purple Guide — Contingency & Emergency Planning, 2026 (published 2026-01-26; 13/27 sections extracted).

Draft. Not reviewed. `reviewed_by` is not set by this document.