A participant rehearses in the headset while observers watch from another part of the room.

Most pilots measure on the day: how to evaluate de-escalation practice so the result holds

35 of 50 published VR and simulation de-escalation studies, 2015 to 2026, measured on the day and never again (Europe PMC).

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

35 of 50

studies measured on the day and not again

VR and simulation studies, 2015 to 2026

41 of 50

stopped at reaction or learning measures

VR and simulation studies, 2015 to 2026

4 of 6

where self-report and observation disagreed

Studies with self-report and observation, 2017 to 2026

20%

of trained staff answered the later check

Young et al., 2022

The published norm

What do published de-escalation studies measure?

Mostly how staff felt and what they learned on the day: 41 of 50 VR and simulation studies from 2015 to 2026 stopped at reaction or learning.

A visitor gestures beside a service counter. A first person VR training scenario for AI training and roleplay, seen through the headset. — A visitor gestures beside a service counter.

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

A resident raises a concern at a service counter while another listens, seen through the headset. — A resident raises a concern at a counter.

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

Reaction, learning, behaviour, results

Kirkpatrick Partners define Level 1 as reaction and Level 2 as learning, including confidence. Level 3 asks whether staff perform the critical behaviours at work; Level 4, whether the organisation’s targeted outcomes occur.

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

Where the studies stopped

Behaviour at work appeared in 1 study, by self-report. 8 of 50 reported restraint, injury or assault data, all from uncontrolled before and after projects or a feasibility trial on 3 wards.

Colleagues at a bench sort blank cards, monitors in the background. — Colleagues sort blank cards at a bench.

Reviews find the same gap

In a 2024 simulation review, 4 of 25 studies collected primary outcome data. A 2026 review of VR de-escalation training found none of its 15 studies evaluated organisational impact.

Where the 50 studies stopped, all against the VR subset

Reaction or learning, all 5082%

Reaction or learning, VR 1493%

Comparison group, all 5026%

Comparison group, VR 1421%

On the day, never again, all 5070%

On the day, never again, VR 1493%

See the data (percent of studies)

Measure	Percent of studies
Reaction or learning, all 50	82%
Reaction or learning, VR 14	93%
Comparison group, all 50	26%
Comparison group, VR 14	21%
On the day, never again, all 50	70%
On the day, never again, VR 14	93%

Percent of studies, 2015 to 2026. Source: Europe PMC and PubMed records coded by PTR.

Virtual reality studies stop earlier than the field

Of the 14 VR studies, 13 stopped at reaction or learning and 13 never measured again.

So what for you: A pilot with a comparison group and a later check already sits above the published norm.

Longest follow-up in the 50 studies

On the day, never again35 of 50

At 2 weeks2 of 50

At 3 to 6 months7 of 50

Over 6 months1 of 50

Not reported5 of 50

See the data (studies of 50)

Measure	Studies of 50
On the day, never again	35 of 50
At 2 weeks	2 of 50
At 3 to 6 months	7 of 50
Over 6 months	1 of 50
Not reported	5 of 50

Studies, 2015 to 2026. Source: Europe PMC and PubMed records coded by PTR.

Almost no study checked again

Learners were measured again at 3 months or later in 5 studies; the others that looked later counted service data.

So what for you: Book the later check into the roster before the pilot starts.

Self-report and observation

Is a confidence survey enough?

Not on its own: in 4 of 6 studies that measured self-report and observation, what staff reported and what an observer scored moved apart.

A facilitator turns over a feedback card. A VR training scenario for AI training and roleplay. — A facilitator turns over a feedback card.

Skill rose while confidence stayed level

In a 2017 cluster randomised trial with 26 residents, filmed performance in simulated agitation cases rose (Cohen’s d = 1.6) while self-rated confidence did not differ between groups.

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

Confidence rose while observed skill did not

In a 2019 randomised trial, residents’ self-assessed skill rose (p <= .03) but standardised patient ratings did not beat control. After a 2025 curriculum, confidence rose from 2.79 to 4.11 out of 5; an OSCE found no skill difference.

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

The counts disagreed, the rating rose

Surgical residents in a 2026 simulation series counted more de-escalation techniques themselves (p = 0.02); a facilitator saw no significant difference (p = 0.11). A modified observation scale rose from 23.1 to 27.8.

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

Observation beside the survey

AHRQ’s TeamSTEPPS guidance warns that training can sensitise staff so that their survey ratings dip for a while, and recommends direct observation in addition to surveys. Its test: “If training does not change behavior, it was not successful.”

Feedback cards sorted on a table after a review round. — Feedback cards sorted on a table.

Curated sources laid out for review. — Sources laid out for review.

A reviewer holds out a card. A first person VR training scenario for what is AI, seen through the headset. — A reviewer holds out a card.

A reviewer watches an AI conversation with the approval controls in view. — A reviewer watches a practice conversation.

Scales a pilot can borrow

The De-escalating Aggressive Behaviour Scale rates observed skill on 7 items (alpha 0.87 and 0.88; raters agree at 0.77 to 0.93). The Confidence in Coping with Patient Aggression instrument (alpha 0.81) was used in 5 of 10 studies in a 2025 meta-analysis.

After the session

How long does a gain last?

Part of it fades: hospitalists in a 2023 study kept 55% to 67% of their confidence gain at 3 to 12 months.

Share of the immediate confidence gain still present later

Hospitalists, 3 months55%

Hospitalists, 6 months67%

Hospitalists, 12 months66%

Clinicians, 3 months94%

See the data (percent of gain)

Measure	Percent of gain
Hospitalists, 3 months	55%
Hospitalists, 6 months	67%
Hospitalists, 12 months	66%
Clinicians, 3 months	94%

Percent of gain, self-report, 2022 and 2023. Source: Gupta et al., 2023; Young et al., 2022.

Confidence fades in part

Gupta’s hospitalists scored 43.2 before in-person training, 68.5 straight after, then 57.2, 60.2 and 59.9 at 3, 6 and 12 months. Young’s clinicians held 94% of their gain, on 24 of 122 answering.

So what for you: Plan a repeat at about 3 months and expect a smaller gain than on the day.

A learner repeats a quiet task inside a fully simulated room. — A learner repeats a quiet task.

Skills decay faster without use

A 1998 meta-analysis of 189 data points found skill loss grew from 0.01 standard deviations straight after training to 1.4 after more than a year unused, faster for cognitive tasks than physical ones.

Share of trained staff at the later check

Young 2022, 3 months20%

Phiri 2024, 6 months23%

Mitchell 2020, 3 to 6 months30%

See the data (percent answering)

Measure	Percent answering
Young 2022, 3 months	20%
Phiri 2024, 6 months	23%
Mitchell 2020, 3 to 6 months	30%

Percent of those trained, 2020 to 2024. Source: Young et al., 2022; Phiri et al., 2024; Mitchell et al., 2020.

Later checks reached 20% to 30% of those trained

44 of 146 staff answered at 3 to 6 months; 13 of 56 ward staff gave 6-month samples, too few to test.

So what for you: Later responders may differ from the rest; size the group for the later check.

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

Size the pilot before it starts

The median published study analysed 56 people. A 2026 VR study needed 34 participants for a moderate effect (d = 0.50) and approached 62; a 2022 study with 24 learners could not detect group differences.

Service data

Which incident data can show change?

Rates with a baseline as long as the follow-up, such as restraints per emergency visit.

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

A rate, not a raw count

A 2023 emergency department pilot reported monthly restraints as a ratio of that month’s visits, over the 6 months before and after training.

A supervisor wearing a headset leans in to check on a contact-centre officer seated at his desk. — A supervisor checks in with a seated officer.

A before and after result needs its context

A 2025 psychiatric emergency project recorded 35% fewer restraints and 52% fewer injuries over 5 months after simulation than before, with teamwork perceptions unchanged. With no comparison, staffing, reporting and chance remain explanations.

Facilitators hand over an XR practice session in a workshop room. — Facilitators hand over a practice session.

Match the baseline to the follow-up

The EDITION feasibility trial on 10 wards collected 8 weeks before training, 8 of embedding and 8 after. Its authors write that “there have been no high-quality trials evaluating the effectiveness” of de-escalation training.

A facilitator leads a conversation around the table. A VR training scenario for justice and community services. — A facilitator leads a conversation at a table.

A controlled ward design

The Safewards cluster trial randomised 31 wards and counted incidents per shift: conflict was 15% and containment 23.2% lower than on control wards (corrected figures). Safewards is a set of ward interventions, broader than a course.

Seclusion and physical restraint per 1,000 bed days, Australia, 2008-09 to 2022-23

SeclusionPhysical restraint

See the data (events per 1,000 bed days)

	2008-09	2009-10	2010-11	2011-12	2012-13	2013-14	2014-15	2015-16	2016-17	2017-18	2018-19	2019-20	2020-21	2021-22	2022-23
Seclusion	15.6	13.9	12.1	10.6	9.8	8.2	7.9	8.1	7.4	6.9	7.2	8.1	7.3	7	6
Physical restraint								11.2	10.1	10.3	11.4	11.0	11.6

Events per 1,000 bed days; last 2 years rounded, 2008-09 to 2022-23. Source: AIHW tables RP.5 and RP.8; AIHW.

Seclusion halved while physical restraint held steady

Seclusion fell from 15.6 (2008-09) to 6 (2022-23); physical restraint stayed between 10 and 12 from 2015-16.

So what for you: Report your ward’s own rate on the same denominator, before and after.

NHS staff reporting physical violence from patients or the public, England, 2021 to 2025

Physical violence from patients or the public

See the data (percent of staff)

	2021	2022	2023	2024	2025
Physical violence from patients or the public	14.57%	14.82%	13.88%	14.38%	14.47%

Percent of staff, 2021 to 2025. Source: NHS Staff Survey briefing 2025.

A national staff-report measure moves by about a point

The share ranged from 13.88% to 14.82% over 5 years on the same question, and stood at 14.47% in 2025.

So what for you: Keep the question fixed, and read a small change as noise.

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

A receptionist wearing a headset leans over the desk while an older man waits at the counter. — A receptionist leans over the desk.

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.

PTR-built projects

What did PTR-built projects measure?

Evaluation design rather than de-escalation outcomes: 73.2% of eligible leaders completed baseline in a Mayo Clinic study with a control group; 30% returned an optional survey after Google’s training.

Response or completion by evaluation design

Mayo Clinic study, baseline survey73.2%

Mayo Clinic education evaluation52.4%

Google training, optional survey30%

See the data (percent)

Measure	Percent
Mayo Clinic study, baseline survey	73.2%
Mayo Clinic education evaluation	52.4%
Google training, optional survey	30%

Percent of those eligible, 2022 to 2024. Source: Social Science & Medicine, 2024; Journal of Nursing Administration; CHI 2024 paper.

Measurement built into the design kept more of the group

PTR built the VR for the Mayo Clinic study, where 841 of 1,149 eligible leaders completed baseline. In the Google training, built with Equal Reality, a subsidiary of People Tech Revolution, 7 of 26 trainees returned an optional survey.

So what for you: Neither project is a de-escalation study; they show how design changes who answers.

A facilitator and participant discuss a completed staged exchange. A VR training scenario for immersive XR. — A facilitator and participant discuss an exchange.

What a reaction result looks like

Mater’s own evaluation of its immersive pilot, with more than 60 participants: 90% reported better concentration, 77% found it more engaging than face-to-face delivery, 93% enjoyed it. All Level 1.

Demand

Why does evaluation carry more weight now?

Obligations are growing: since January 2022 the Joint Commission has asked US hospitals to train staff at hire and annually.

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

An annual repeat is already built in

Under the Joint Commission’s standards, hospitals train at hire, annually and when the program changes, and run an annual worksite analysis. US states with prevention laws rose from 11 before 2015 to 27 by June 2024 (Health Affairs Scholar).

An employee sits across a meeting table. A first person VR training scenario for AI training and roleplay, seen through the headset. — An employee sits across a meeting table.

Interest rises while federal rules wait

Worldwide search interest in de-escalation training averaged 3.35 across 2022 and 19.37 across 2025 on Google Trends’ relative scale. The federal OSHA healthcare rule moved to the long-term agenda in March 2026 (GAO).

The plan

What should a pilot plan settle before training starts?

An observed behaviour, a confidence scale and a service indicator, each with a baseline and a repeat at about 3 months.

A headset-wearing supervisor gives feedback to a colleague on a warehouse floor. — A supervisor gives feedback on a warehouse floor.

A supervisor wearing a headset sits half turned toward a simulated crew member in a high-visibility shirt, in a site office with a window onto the mine pit beyond. — A supervisor checks in with a crew member.

A simulated aged-care hallway. A man in his mid-eighties stands holding the rail as a carer wearing the headset approaches. — A carer approaches a man at a handrail.

A manager wearing a headset sits facing a simulated team member across a small table, rehearsing a performance feedback conversation. — A manager rehearses a feedback conversation.

Start from the result you need

Kirkpatrick Partners advise starting with Level 4 and the program’s purpose, with early and later indicators, then working back to the behaviour that should change.

A family support worker sits with a woman. A VR training scenario for justice and community services. — A family support worker sits with a woman.

What a strong design adds

A 2020 Cochrane review asks for institutional aggression reports, validated measures, a 1-year follow-up and an active comparison, so reporting does not rise in the trained group alone.

A permits officer reviews a screen with colleagues. A VR training scenario for mining and resources. — Colleagues review a screen together.

Use data you already hold

AHRQ notes existing organisational data make results more credible and less burdensome to collect, if they cover the trained area and the period after training.

Evidence

Which studies does this rest on?

The primary studies behind each figure, with design and sample.

Primary studies this analysis rests on, 2015 to 2026
Study	Design	Sample	What was measured	Result
Vestal et al. (2017), Academic Psychiatry	Cluster randomised trial	26 residents	Performance; confidence	d = 1.6; confidence did not differ
Weaver et al. (2026), Simulation in Healthcare	Simulation series	16 residents	Technique counts; mDABS	Counts disagreed; 23.1 to 27.8
Young et al. (2022), Int J Ment Health Nurs	Pre-post, follow up	122, then 24	Self-efficacy	62.9, 83.2, 81.9
Gupta et al. (2023), J Community Hosp Intern Med Perspect	Pre-post to 12 months	37 hospitalists	Confidence	43.2 to 68.5; 59.9 later
Bowllan et al. (2025), J Am Psychiatr Nurses Assoc	Before and after	Psychiatric emergency unit	Restraints, injuries	35% and 52% fewer
Price et al. (2024), Health Technol Assess	Feasibility trial	10 wards	Conflict, containment	8 weeks each side
Bowers et al. (2015), Int J Nurs Stud	Cluster randomised trial	31 wards	Incidents per shift	15% and 23.2% lower (corrected)

Decisions

What does this mean for a buyer?

Decisions the figures inform, written as actions for the pilot plan.

Put an observed measure beside the survey. Self-report and observed skill disagreed in 4 of 6 studies; score rehearsals on a rubric such as the DABS.
Book a repeat at about 3 months. 35 of 50 studies never measured again, and hospitalists held 55% to 67% of their gain.
Size the group for the later check. Later checks reached 20% to 30% of those trained; a 2026 study needed 34 participants.
Report service data as a rate against a matched baseline. Use events per 1,000 bed days or per visit; 35% fewer restraints with no comparison cannot be told apart from other changes.

Board paper lines, ready to paste into a business case

41 of 50 published VR and simulation de-escalation training studies from 2015 to 2026 stopped at reaction or learning measures, and 35 of 50 measured on the day and never again (PTR coding of Europe PMC and PubMed records, 2026). In 4 of 6 studies that measured self-report and observation, they disagreed, and AHRQ TeamSTEPPS guidance recommends direct observation in addition to staff surveys. Our pilot will score an observed behaviour on a validated scale, repeat it at 3 months, and report incidents per 1,000 bed days against a matched baseline, as the EDITION trial did with 8 weeks each side.

Limits and method

What these figures cannot show, and how they were counted.

The 50-study table is PTR’s coding of abstracts by 1 reviewer, without risk-of-bias scoring, not a systematic review; the Mitchell 2024 and Tang 2026 reviews are the cross-check.
Kirkpatrick levels follow the model page as live in 2026, whose wording changed after January 2024.
Instrument figures come from later studies; the original 1987 CCPA reliability figures were not read.
Retention figures are self-report from self-selected responders; Gupta 2023 does not say whether simulation was used; Arthur 1998 covers skills in general.
AIHW rates cover public acute specialised mental health hospital services, exclude Queensland from physical restraint before 2017-18 and round 2021-22 and 2022-23. NHS figures use the 2025 weighting. National series set denominators, not ward targets.
Before and after service results (Duncan 2023, Bowllan 2025) are uncontrolled.
PTR-built projects are not de-escalation studies; their response rates illustrate evaluation design. Mater figures are Mater’s own evaluation.
Every source was read on 26 September 2026.

Where PTR fits

PTR builds de-escalation practice in a headset, on screen and with AI characters, and plans the measures with the pilot. See de-escalation training, VR de-escalation training, how a pilot runs, the business case and evaluation and what earlier projects measured.

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). Most pilots measure on the day: how to evaluate de-escalation practice so the result holds (Version 2.0). People Tech Revolution. https://peopletechrevolution.com/insights/evaluate-de-escalation-training-pilot

Data: 50 primary VR and simulation de-escalation training studies from Europe PMC and PubMed (2015 to 2026), coded by PTR; the cited trials and reviews; AIHW restrictive practices tables (2022); NHS Staff Survey briefing 2025; Kirkpatrick Partners and AHRQ guidance. Read 26 September 2026. The Mayo Clinic and Google response rates illustrate evaluation design, not de-escalation outcomes.

Version history

Version	Date	What changed
Version 2.0	27 September 2026	Rebuilt on published study data: 50 coded de-escalation studies, self-report against observation, retention, AIHW and NHS series; 7 charts, evidence table, decisions and limits.
Version 1.0	25 September 2026	First published.