A simulated emergency department triage desk at night. A triage nurse in the headset sits at the desk while a man stands at the counter, agitated but not aggressive.

41 of 50 studies stop in the training room: what VR de-escalation research has and has not measured

Of 50 VR and simulation studies published 2015 to 2026 (Europe PMC, PubMed), 41 of 50 stopped at reaction or learning.

By Kim Loza, Head of Product at People Tech Revolution

Published 25 September 2026 Last updated 28 September 2026 Version 2.0

Connect with Kim on LinkedIn Cite this research

41 of 50

studies stopped at reaction or learning

PTR evidence table, 2015 to 2026

13 of 14

VR studies measured on the day alone

PTR evidence table, 2015 to 2026

6.4 times

growth in training papers, 2015 to 2025

PubMed, read 26 September 2026

RR 1.14

long-term aggression, trained against untrained

Cochrane review, 2020

A man waits at the triage bay counter. A first person VR training scenario for healthcare, seen through the headset. — A man waits at a triage counter.

A support worker talks with a young man in a plain room, seen through the headset, a floating panel of replies beside them. — A support worker talks with a young man.

A young man stands with hands covering his eyes at an after-hours crisis service desk; a night worker wearing the headset stands opposite, rehearsing a crisis conversation. — A night worker at a crisis desk.

The answer

What has VR de-escalation research measured?

Mostly the training room: of 50 studies we coded, 41 of 50 measured reaction or learning, and 8 of 50 reached incident or restraint data.

Highest outcome level measured, 50 studies, 2015 to 2026

Level 1, reaction6 of 50

Level 2, learning35 of 50

Level 3, behaviour at work1 of 50

Level 4, results8 of 50

See the data (studies)

Measure	Studies
Level 1, reaction	6 of 50
Level 2, learning	35 of 50
Level 3, behaviour at work	1 of 50
Level 4, results	8 of 50

Studies by the highest Kirkpatrick level measured, published 2015 to 2026, coded by PTR. Source: Europe PMC; PubMed; Kirkpatrick Partners levels.

Learning is where most studies stop

Level 2 is confidence, knowledge, attitudes and skill inside a simulation, so 82% of studies never looked past the session. The behaviour result was self-reported, and every Level 4 result came from an uncontrolled package or a feasibility trial on 3 wards.

So what for you: ask a provider which level its evidence reached before reading its percentages.

A learner faces you in a coaching room, seen through the headset. — A learner faces the camera in coaching.

VR studies stop earlier than simulation studies

13 of 14 VR studies stopped at reaction or learning, against 28 of 36 studies with actors, standardised patients or role play. Just 13 of 50 studies had a comparison group.

A ward round seen from inside the headset. — A ward round seen through the headset.

How the table was built

Coded from Europe PMC and PubMed: staff training with VR, 360 video, actors or role play.

The problem

Why does the evidence matter to health employers?

The violence the training is bought for has risen: US health care recorded 6.4 cases per 10,000 full-time workers (2011) and 11.9 (2023 and 2024).

US health care violence injuries with days away from work, 2011 to 2024

Annual rateAnnualised 2-year average

See the data (per 10,000 full-time workers)

	2011	2012	2013	2014	2015	2016	2017	2018	2019	2020	2021-22	2023-24
Annual rate	6.4	7.0	7.8	8.2	8.0	8.4	9.1	10.4	9.7	10.3
Annualised 2-year average											9.8	11.9

Per 10,000 full-time workers, 2011 to 2024; the last 2 points are annualised 2-year averages, with a new event code from 2023. Source: BLS fact sheet, 2018; BLS Table R8, 2019 to 2023-24.

Health care violence rates have climbed for over a decade

The annual rate for private health care and social assistance went from 6.4 (2011) to 10.3 (2020). The 2-year average for 2023 and 2024 was 11.9, against 2.6 across private industry, under a revised event code.

So what for you: training is bought against this rate, so its evidence has to reach incidents.

A community nurse greets an older woman. A VR training scenario for healthcare. — A community nurse greets an older woman.

Psychiatric hospitals sit far above the average

In 2023 and 2024, psychiatric and substance abuse hospitals recorded 138.8 such cases per 10,000 full-time workers (BLS Table R8).

Relative gestures beside a facilitator. A first person VR training scenario for healthcare, seen through the headset. — A relative gestures beside a facilitator.

Accreditation now asks for de-escalation training

Since 1 January 2022, Joint Commission hospital requirements expect de-escalation training at hire, yearly and on program changes; home care followed on 1 January 2025 (R3 Report Issue 45, the home care report).

Attention

Is research on this training growing?

Fast: PubMed records on de-escalation and workplace violence training rose 6.4 times from 2015 to 2025, while PubMed as a whole grew 1.5 times.

PubMed records per year, de-escalation and workplace violence training, 2015 to 2025

De-escalation or workplace violence trainingVR and aggression or de-escalation training

See the data (records per year)

	2015	2016	2017	2018	2019	2020	2021	2022	2023	2024	2025
De-escalation or workplace violence training	37	42	38	55	63	107	102	139	121	135	237
VR and aggression or de-escalation training	1	0	0	2	1	5	8	8	5	5	12

Records per year, 2015 to 2025; 2026 is partial and not shown. Source: PubMed counts via NCBI E-utilities, read 26 September 2026.

Attention has grown faster than the evidence base

The count reached 237 records (2025), up from 37 (2015), while all of PubMed moved from 1,258,176 to 1,881,905. VR aggression or de-escalation training papers were 12 of 812 VR health education papers in 2025.

So what for you: expect more pilots each year, and judge each by its design rather than its date.

A participant discusses an XR rehearsal with a facilitator. — A participant talks it through afterwards.

What the counts measure

Attention, not quality: title and abstract matches, with an indexing lag.

Across reviews

Do published reviews find the same gap?

They do: Tang 2026 found no VR study testing organisational impact, and Kyaw 2019 none reporting behaviour change.

Studies that measured incidents, patients or the organisation, by review

Tang 2026, VR de-escalation0 of 15

Kyaw 2019, VR in health education0 of 31

Dafny 2025, nurses on placement1 of 13

Mitchell 2024, simulation4 of 25

PTR table, 2015 to 20268 of 50

Geoffrion 2020, Cochrane, controlled designs5 of 9

See the data (studies with outcome data)

Measure	Studies with outcome data
Tang 2026, VR de-escalation	0 of 15
Kyaw 2019, VR in health education	0 of 31
Dafny 2025, nurses on placement	1 of 13
Mitchell 2024, simulation	4 of 25
PTR table, 2015 to 2026	8 of 50
Geoffrion 2020, Cochrane, controlled designs	5 of 9

Studies with outcome data over studies included, reviews published 2019 to 2026. Source: Tang 2026; Kyaw 2019; Dafny 2025; Mitchell 2024; Geoffrion 2020.

Outcome data are the exception in every review

VR reviews found none. The Cochrane review found 5 of 9 controlled studies counting aggression, and its long-term estimate crossed no effect.

So what for you: treat an incident claim from a VR pilot as new evidence and ask for its design.

A clinician wearing a headset stands at a hospital bedside facing a simulated patient propped up on pillows, practising the bedside conversation in a ward that looks like the ward. — A clinician at a hospital bedside.

The review’s behaviour ratings were therapy studies

Tang and colleagues rated 3 studies at Level 3. All 3 treated aggressive patients; every staff training study stopped at Level 1 or 2.

A charge nurse in green scrubs briefs nurses at a ward station, a headset visible among them. — A charge nurse briefs the ward.

Simulation reviews carry high risk of bias

Of 25 simulation studies with 2,790 participants, 24 of 25 were at high or serious risk of bias and 4 of 25 collected patient outcomes (Mitchell 2024).

A simulated ward bay with a low light over the bed. A woman lies awake as a nurse wearing the headset crouches to her level. — A nurse crouches beside a patient at night.

A facilitator adjusts a headset for a seated learner in a workshop room. — A facilitator adjusts a seated learner’s headset.

An interviewer faces you across a spare room, seen through the headset. — An interviewer seen through the headset.

Confidence

Does higher confidence mean better de-escalation?

Not reliably: confidence rises in almost every study, yet in 4 of 6 studies measuring self-report and observed skill together, the measures disagreed.

Standardised effects reported for confidence, knowledge and skill

Different outcomes and designs; read each bar alone.

Confidence, controlled studies (Chung 2025)0.85

Confidence, uncontrolled studies (Chung 2025)0.71

Confidence after a 20-minute VR session (Johnson 2026)0.70

Knowledge (Geoffrion 2020)0.86

Skills (Geoffrion 2020)0.21

See the data (standardised effect)

Measure	Standardised effect
Confidence, controlled studies (Chung 2025)	0.85
Confidence, uncontrolled studies (Chung 2025)	0.71
Confidence after a 20-minute VR session (Johnson 2026)	0.70
Knowledge (Geoffrion 2020)	0.86
Skills (Geoffrion 2020)	0.21

Standardised mean difference or Cohen’s d, published 2020 to 2026. Source: Chung 2025; Johnson 2026; Geoffrion 2020.

Confidence and knowledge move; skill barely does

Pooled confidence sits at 0.85 in controlled studies and knowledge at 0.86. The Cochrane skill estimate was 0.21, with an interval that includes no effect.

So what for you: put an observed skill measure beside every confidence survey.

An allied-health trainee wearing a headset stands at eye level with a simulated client seated on a treatment bed, practising coming to eye level before it matters. — A trainee at eye level with a client.

Where confidence and skill parted

In a cluster randomised trial of 26 psychiatry residents, simulated performance rose (d = 1.6) while confidence matched control. In a paediatric trial and a medical school OSCE, observed skill did not differ while self-ratings rose.

A clinician wearing a headset sits knee to knee with a simulated client who is talking and gesturing, in a plain consulting room. — A clinician sits knee to knee with a client.

The largest VR cohort measured confidence

Johnson and colleagues ran a 20-minute VR session with 223 participants in Australian nursing programs. Confidence rose (d = 0.70), 93% called VR suitable and 89% wanted more, with no comparison group or later check.

A man waits on the footpath. A first person VR training scenario for healthcare, seen through the headset. — A man waits on the footpath.

Newer VR studies share the design

In Taiwan, 54 nurses rated confidence at 60.04 before 360 video VR and 71.93 after, measured on the day (Lin 2026).

Incidents

Do assaults fall after de-escalation training?

Controlled studies have not shown it: the Cochrane review (1,688 participants) put the long-term risk ratio at 1.14, an interval spanning no effect.

A community nurse sits alone in a carpark. A VR training scenario for healthcare. — A community nurse alone in a carpark.

Earlier and later reviews agree

Heckemann 2015 found no change in patient aggression and Brenig 2023 no relevant impact on forensic ward incidents; in Okubo 2022, 4 of 11 controlled studies reported less violence and the pooled estimate was null.

A simulated community mental health room. A man in his mid-twenties sits forward, elbows on his knees; a clinician in the headset sits back, giving him room, holding the silence. — A clinician gives a young man room.

Service gains came with wider packages

A psychiatric emergency program recorded 35% fewer restraints and 52% fewer injuries over 5 months after simulation than before, with no control group, alongside teamwork and crisis prevention programs (Bowllan 2025).

A supervisor wearing a headset sits beside a simulated driver in a depot crib room, the driver quiet and holding a mug, practising a return-to-work conversation. — A supervisor beside a quiet driver.

Assaults rose again once the program stopped

In an emergency medicine residency, 28% of residents reported assault before its program, 11.3% a year after, and 30.6% 5 years later, once in-person training had stopped (Roppolo 2025).

Longest follow-up in each study, 50 studies

Immediate, end of session35 of 50

Under 3 months2 of 50

3 to 6 months7 of 50

Over 6 months1 of 50

Not reported5 of 50

See the data (studies)

Measure	Studies
Immediate, end of session	35 of 50
Under 3 months	2 of 50
3 to 6 months	7 of 50
Over 6 months	1 of 50
Not reported	5 of 50

Studies by their longest follow-up, published 2015 to 2026, coded by PTR. Source: Europe PMC; PubMed.

Almost nothing is measured after the session

35 of 50 studies measured on the day and not again, and 1 of 50 looked beyond 6 months. A reading taken at the session overstates what lasts.

So what for you: build a later check into the pilot plan before the pilot starts.

Head to head

Can VR stand in for live role play?

In the short term, perhaps: 22 police officers in VR and 20 in live action each beat 21 controls by similar margins (Lavoie 2023).

A facilitator and a colleague rehearse a feedback conversation. — A facilitator and colleague rehearse feedback.

The comparison is thin

1 of 14 VR studies compared VR with live practice. Officers were allocated by scheduling availability, outliers were removed, and follow-up was short.

A caseworker wearing a headset sits at a desk across from a simulated distressed client holding his head. — A caseworker across from a distressed client.

The ward trial lost its sample

An NHS feasibility trial allocated wards to VR training (34 staff) or usual care (22 staff). Restrictive practice fell with fluctuation on wards that started higher, and 8 and 5 remained at the 6-month follow-up.

A headset-wearing practitioner raises a hand to begin. An AI roleplay scenario for AI roleplay and practice. — A practitioner raises a hand to begin.

Usability shapes the result

Western Sydney clinicians (28 participants, Moore 2022) found talking with a virtual agent feasible but flagged motion sickness and privacy; a US pilot with 13 providers scored usability 63.30 out of 100.

Demand

Are buyers looking for this training?

Increasingly: worldwide searches for de-escalation training rose 5.8 times from 2022 to 2025, and for workplace violence training 5.1 times (Google Trends).

A facilitator briefs participants in a workshop room. — A facilitator briefs participants in a workshop.

State obligations are widening

US states with laws to prevent health care workplace violence went from 11 before 2015 to 27 by June 2024 (Lombardi 2026).

A sheriff's officer directs an older man inside. A VR training scenario for justice and community services. — An officer directs an older man inside.

The federal rule is not scheduled

The OSHA rule on workplace violence in health care sits among long-term actions, with no proposal date.

A clinician sits with a young woman outside. A VR training scenario for healthcare. — A clinician sits with a young woman outside.

A practitioner discusses feedback with a coach. An AI roleplay scenario for AI roleplay and practice. — A practitioner reviews feedback with a coach.

A team leader discusses escalation with a colleague. A VR training scenario for cross-sector. — A team leader talks through an escalation.

Evidence table

Which studies does this rest on?

Design, sample, measure and result for each source.

Evidence table: reviews and primary studies
Study	Design	Sample	What was measured	Result
Tang et al. (2026), Frontiers in Medicine	Review, VR	15 studies	Kirkpatrick level	None tested organisational impact
Geoffrion et al. (2020), Cochrane Database of Systematic Reviews	Review, controlled designs	9 studies, 1,688 participants	Aggression, knowledge, skills	RR 1.14; knowledge 0.86; skills 0.21
Mitchell et al. (2024), International Journal of Nursing Studies	Review, simulation	25 studies, 2,790 participants	Learning, performance	24 of 25 high or serious bias risk
Chung et al. (2025), International Nursing Review	Meta-analysis	10 studies	Confidence	SMD 0.85; d 0.71
Kyaw et al. (2019), Journal of Medical Internet Research	Meta-analysis, VR education	31 trials	Knowledge, skills	Skills SMD 1.12; no behaviour outcomes
Jameyfield et al. (2026), Academic Medicine	Review, verbal de-escalation	56 studies	Evidence strength	12 of 56 strong, all blended
Okubo et al. (2022), Revista Latino-Americana de Enfermagem	Meta-analysis	11 studies	Violence	4 of 11 positive; pooled null
Johnson et al. (2026), Teaching and Learning in Nursing	Uncontrolled, VR	223 enrolled, 221 analysed	Confidence	d = 0.70
Lavoie et al. (2023), Policing	Non-randomised, 3 arms	63 officers	Scored de-escalation	VR and live comparable
Phiri et al. (2024), World Journal of Psychiatry	Feasibility, wards allocated	34 staff, 22 staff	Restrictive practice	Fell, with fluctuation
Vestal et al. (2017), Academic Psychiatry	Cluster randomised trial	26 residents	Performance, confidence	d = 1.6; confidence same
Bowllan et al. (2025), J Am Psychiatr Nurses Assoc	Before and after service data	Emergency program staff	Restraints, injuries	35% and 52% fewer; no control
Roppolo et al. (2025), AEM Education and Training	Repeated survey	Residents	Reported assault	28%, 11.3%, 30.6%

For buyers

What should a buyer do with this?

Use the evidence to set the questions, the measures and the promise.

Ask for the level, comparator and follow-up. 41 of 50 studies stopped at reaction or learning; 13 of 14 VR studies measured on the day.
Pair every confidence survey with an observed measure. In 4 of 6 studies that measured self-report and observed skill, the measures disagreed.
Frame fewer incidents as a system goal. The Cochrane risk ratio was 1.14; the 35% fall in restraints came with a wider package.
Test VR against live practice before replacing it. 1 of 14 VR studies made that comparison, in policing, over the short term.
Blend teaching with practice. 12 of 56 verbal de-escalation studies rated strong, and all combined teaching with role play or simulation.

A participant tests hand movement while a coach observes. A VR training scenario for immersive XR. — A coach watches a hand movement test.

A practitioner rehearses a conversation beside a facilitator. — A practitioner rehearses beside a facilitator.

A colleague waits for a reply. A first person VR training scenario for cross-sector, seen through the headset. — A colleague waits for a reply.

Board paper lines, ready to paste into a business case

Of 50 VR and simulation de-escalation studies from 2015 to 2026, 41 of 50 measured reaction or learning, and every study reporting incidents (8 of 50) used an uncontrolled or feasibility design (PTR evidence table, 2026). A Cochrane review of education to prevent aggression toward health workers reported a long-term risk ratio of 1.14, an interval that includes no effect (Geoffrion and colleagues, 2020). Confidence gains are large, a pooled standardised mean difference of 0.85 in controlled studies (Chung and colleagues, 2025), so evaluations need observed skill and incident baselines too.

A coach and learner arrange markers. A VR training scenario for AI training and roleplay. — A coach and learner arrange markers.

A team leader gives a colleague feedback. A VR training scenario for cross-sector. — A team leader gives a colleague feedback.

A palliative care nurse talks with a visitor. A VR training scenario for healthcare. — A nurse talks with a visitor.

Limits and method

What the data cannot show, and how each figure was read.

The table is a structured scan rather than a systematic review: 1 reviewer, 2 databases, English language, no risk of bias scoring. Review counts are the cross-check.
Levels are PTR’s coding against the Kirkpatrick Partners definitions, except the Tang ratings.
The pooled figures differ in outcome, instrument and design, and the Chung pools are heterogeneous.
Before and after service data cannot separate training from staffing, policy, reporting habits or chance.
Populations are mixed, and the head to head test is in policing.
PubMed counts match title and abstract words (violence, aggression or de-escalation terms with training terms; VR adds virtual reality); 2026 is partial.
BLS rates are private industry cases with days away from work (Table R8); 2021-22 and 2023-24 are annualised, with a new event code from 2023.
Google Trends values are relative interest, not volume; Australia sat below the reporting threshold until 2026.
Null pilot results are less often published, so the table can overstate positive results.
Every source was read on 26 September 2026.

Where PTR fits

PTR builds VR and AI practice for these moments, and designs pilots that reach past the training room with an observed skill measure, a comparator and a later check. See VR de-escalation training, de-escalation training programs, AI roleplay practice and training for safety and wellbeing leads. PTR built the VR program in a Mayo Clinic randomised study that followed nursing leaders for up to 8 months; it measured empathy, not de-escalation.

Cite this research

Free to quote and cite with a link to this page. Name the version you read, so readers can see what has changed since.

APA 7

Loza, K. (2026, September 27). 41 of 50 studies stop in the training room: what VR de-escalation research has and has not measured (Version 2.0). People Tech Revolution. https://peopletechrevolution.com/insights/vr-de-escalation-training-evidence

Data: PTR evidence table of 50 studies coded from Europe PMC and PubMed, 2015 to 2026; reviews by Tang, Geoffrion, Mitchell, Chung and Kyaw; PubMed E-utilities counts; BLS Table R8; Google Trends; Joint Commission R3 Reports. Read 26 September 2026.

Version history

Version	Date	What changed
Version 2.0	27 September 2026	Rebuilt on a 50-study evidence table, 30 reviews, PubMed counts and BLS incidence data: key numbers, 6 charts, an evidence table, decisions and limits.
Version 1.0	25 September 2026	First published.