Measuring Psychological Safety
Edmondson's seven items and the audit of what was built on them, self-report against behaviour, measures that become targets, the demand for evidence, and how to measure people without mistaking them for data: the thinking behind Measure, our survey tool.
The landscape
Our way in is the premise our own instrument is built on: a survey is an intervention rather than a thermometer, because observing changes what is observed, people cannot always explain their own silence, and a score locates where to look without saying what is happening or what to do about it (Measure). Ridgway (1956) said it first and cleanly, at the dawn of management science: once a quantitative measure of performance is installed, people reorganise their behaviour around the measure rather than the purpose it was meant to track, so measurement is not a neutral readout but an intervention that changes the thing it measures. Jade's Principles of Ethical Measurement adds the politics: someone always chooses what to count, what it means and how to present it, and it is rarely the people being counted, so a badly run survey can teach that speaking up is pointless just as easily as a good one can help, a trap Jade calls survey theatre.
The instrument and its audit. Edmondson (1999) turned psychological safety into a team-level construct with a measure: seven items, several descended from quotations gathered in interviews, with an alpha of .82. The limitations are Edmondson's own and travel with it: the data could not conclusively separate the construct from trust, the design is cross-sectional and single-company, and 51 teams is small for multivariate work. Edmondson & Lei (2014), Frazier et al (2017), synthesising 136 samples, and Newman, Donohue & Eva (2017), across 62 studies, summarise what was built on it, and Newman and colleagues flag the gaps in cross-cultural and longitudinal research. Liu et al (2024) is the audit: 217 studies and 504,136 people, internal consistency adequate and very little evidence of anything else, with test-retest reliability, factorial validity and criterion validity all major gaps. Edmondson's scale supplied 130 of the studies, 54.6 per cent, and was rated the least rigorous of the five named measures; Detert and Burris's measure scored highest, despite appearing in only five studies. It is a preprint, not peer reviewed, and its search was run in October 2021. Edmondson & McManus (2007) present the original as a worked example of methodological fit, with team-design variables rated by blind manager interviews to reduce common-method bias. Kerrissey, Satterstrom & Edmondson (2020) ask what the instrument does when the team itself is fluid and respondents may not agree on who is on it.
What self-report can and cannot see. Ericsson & Simon (1980) set the terms: verbal reports are reliable when they capture what the participant was already attending to, and reconstructive rather than recollective when they require inference about processes never heeded. Tetlock (1985) supplies the mechanism for accounts given to a known audience: people optimise them for that audience rather than for accuracy. Boothby et al (2018) find people reliably underestimating how much a conversation partner liked them. Tom's article on the SANE effect explains why neuroscience framing makes weak or false claims seem more credible; Väth et al (2024) replicate the original effect and find it much weaker than advertised, and the lesson is even-handed scepticism, of the sciencey gloss and of the neat result alike.
When the measure becomes the target. Ridgway's (1956) three cases, the single measure that produces tunnel vision, the multiple measures that conflict, and the composite that buries judgements inside a figure that looks objective, predate the vocabulary that later attached to them. Kohn (2011) finds the mechanism in grades, which turn attention from the material to the mark and push students to the easiest possible task, and Campbell (1979) explains why asking whether an intervention worked is almost always harder than it looks. Tom's articles on Goodhart's Law, applied to surveys and measurement programmes, the Streetlight Effect, where what gets measured is what is easy to measure and measurement choices are power choices about whose experience counts, and the safety to dissent, on targets, KPIs, benchmarks and diagnostics that produce the appearance of safety while suppressing its substance, carry the argument into practice. The critique of the Psychological Safety Index tests one instrument against it, granting real diagnostic value alongside the Goodhart risk, and Jade's Strange Confidence of 360 Degree Feedback examines a practice whose authoritative meta-analysis, Smither, London & Reilly (2005), found improvements generally small.
Outcomes, and the demand for evidence. Higgins et al (2022), Montgomery et al. (2025), Grailey et al (2021) and Vogus & Sutcliffe (2007) carry the outcome evidence, and the tensions below take them in turn. Tom's Parachutes and Evidence answers the demand that psychological safety prove itself by randomised trial, Tom's critique of Eldor, Hodor and Cappelli's nonlinear limits argues that paper's conclusions are not supported by the evidence it presents, and Bea Poyton's account of Project Aristotle covers what Google's research found, what it did not, and how its popularisation has misrepresented it.
Reading the system rather than the score. Ross (1977) named the fundamental attribution error, the structural default of overweighting character and underweighting situation, which Tom's article on the fundamental attribution error treats as the cognitive root of blametropism; Understanding Variation carries Deming's distinction between common cause and special cause; Kletz (2011) lists the errors of accident investigation, confirmation bias the worst; and Tom's article on counterfactuals treats if-only thinking as hindsight bias that tidies messy reality into false clarity. Cilliers (1998) says no complete description of a complex system is possible, and Greenhalgh & Papoutsi (2018) that studying one needs a different epistemology, not just different statistics. On team process, Marks, Mathieu & Zaccaro (2001) supply the vocabulary of performance episodes and Zijlstra, Waller & Phillips (2012) code interaction as it happens rather than asking about it afterwards.
The key connections
The least safe person, and the distribution. Tom's article on the least safe person argues that measuring psychological safety as an average actively hides the outlier who is suffering, and the team performance quadrant that the least safe person is the team's actual measure. Measure builds that into its reports: a team answering 3, 3, 3, 3 and a team answering 1, 2, 4, 5 share a mean and need opposite conversations, so every report shows the spread, the lowest answer and the disagreement between colleagues, never a mean as the headline, and the practical guide says to find the statements with the lowest scores and the widest range, and to treat a perfect score as a warning rather than a triumph. The research side shows what licenses an aggregate at all: Edmondson (1999) treats the scale as a property of the team on the strength of an intraclass correlation of .39, the highest of any group-level variable in the study, and Vogus & Sutcliffe (2007) justify unit-level aggregation by a median within-unit agreement of 0.98; Measure reports the disagreement itself. Jade's 360 critique makes the same point about ratings: averaging many biased ratings does not cancel bias but blends it into something smoother and harder to attribute, with gender and racial bias accumulating behind a falsely objective aggregate.
The survey as intervention, and the loop. The practical guide gives the governing rule: only measure if you have a plan for what to do afterwards, because measuring and then doing nothing actively harms psychological safety, and it asks the reader to imagine making themselves vulnerable in a survey and then watching the results be discarded. Measure, the platform is built on the same premise, a survey that comes back with a plan rather than a score, since asking people how safe they feel and then doing nothing is one of the fastest ways to teach them that speaking up is pointless, and the Principles of Ethical Measurement end on close the loop, or do not open it. Tetlock (1985) says why the audience matters: accounts given to a known audience with predictable preferences are optimised for survival rather than truth, and post-incident investigations, performance reviews and compliance reporting are structured exactly that way.
Reports and behaviour. Liu et al (2024) found every one of 217 studies using self-report and none observing behaviour. Noort, Reader & Gillespie (2019b) put a number on the gap, a consistent mismatch between self-reported and observed safety voice across three studies of people asked to walk a plank that looked too weak to hold them: what the paper establishes is that the gap exists at all, and that the field's dominant instrument cannot see it. Their systematic review says the literature treats post hoc self-report as reliable data about a behaviour, and reports of behaviour are not the behaviour. Measure accepts the limit, measuring speaking up directly as the outcome and treating not knowing why one holds back as a real answer rather than a missing one. Other instruments ask about behaviour rather than climate: Vogus & Sutcliffe (2007) built a self-report measure of the concrete behaviours of safety organising rather than the context around them, and Zijlstra, Waller & Phillips (2012) observe and code the behaviour itself, finding the reciprocity of a crew's talk, exchanges running back and forth rather than one pilot persisting alone, visible within the first quarter of an hour and predicting effectiveness. Looking At Things Properly names the signal that cannot be faked, a junior challenging seniors and being thanked for it, dissent surfacing in meetings rather than corridors. Beaubien & Baker (2002), with 10,166 pilots, is the honest counter-case: a large, well-conducted reaction survey that measures satisfaction and perceived usefulness rather than behaviour or outcomes, and says so.
Mechanism, not score. Measure follows the calculus of voice: low scores trigger questions that distinguish ambiguity, not being able to predict how speaking would land, from consequence, predicting it accurately and finding the cost too high, from futility, expecting nothing to change, on the basis that each calls for a different response, and the power gradient is measured directly as the degree to which what is sayable depends on who is in the room. Kahneman & Tversky (1979) supply loss aversion against a reference point, which is why speaking up feels riskier than staying silent even when the expected value favours voice; we read Boothby et al (2018), who find people reliably underestimating how much they were liked, as showing that the self-assessment the cost is measured against is already running low; and Kassing (1998) arrives at a matching pair of judgements, whether the dissenter expects to be read as constructive and the likelihood of retaliation, from a communication tradition rather than a management one.
Numbers that improve while the system decays. Hopkins (2009) gives the pathology a body count: BP's Texas City refinery had a recordable injury rate one third the industry average in the year before the explosion that killed fifteen people, and in 2004 the rate had improved so much that BP paid the whole workforce a bonus, in a year that included three fatalities. Bonus schemes tied to personal-injury rates drive the reported rate down, both by encouraging attention to minor hazards and by discouraging reporting, while the loss-of-containment events that predict disaster go uncounted; as we read it, each unreported small thing then makes the next one harder to report, because it contradicts an official story of excellence. Patterson & Wears (2015): managers assume they would know if something important were wrong, so the absence of an alarm is read as the absence of a problem. Montgomery et al. (2025): reporting patient safety problems can indicate both high and low psychological safety, and there is no way to tell which without knowing the culture and history of the organisation. Noort, Reader & Gillespie (2019) turn it round: if raising a concern carries a small cost and a prevented incident a much larger benefit, over-reporting is the predicted signature of a system that is working. Scheffer et al. (2009) find the warning in rising variance rather than the mean, which for us is why a dashboard showing a steady average can be accurate and uninformative right up to the collapse, and Pauly (1995) names the shifting baseline, which we read as a caution that a survey whose first round becomes the baseline may not see how far things had already moved.
The tensions and disagreements
Does the scale measure what it says? Frazier et al (2017) and Newman, Donohue & Eva (2017) summarise a field that has established a great deal on Edmondson (1999); Liu et al (2024) rate the scale that supplied more than half the samples as the least rigorous of five, and note that Edmondson and Lei have themselves since argued the construct is multidimensional and interpersonal where the scale was built as unidimensional; because the scale drives the overall averages, the reassuring headline and the unflattering detail come from the same place. Montgomery et al. (2025) add that nobody knows which team a respondent has in mind when answering about this team; Kerrissey, Satterstrom & Edmondson (2020) that fluid membership makes the question real; and Higgins et al (2022) measured organisation-level psychological safety with a school as the referent, which we read as a different construct from the team-level original. The most useful thing in the 1999 paper may be how carefully Edmondson declined to overclaim.
Fit to context or comparability. The practical guide offers ten questions adapted from The Fearless Organization and explicitly meant to be modified, on the argument that relevance to your own team matters more than replicability across organisations, and Choosing a Psychometric Tool says bespoke design is often preferable to proprietary tools, which can import their designers' assumptions. Measure takes the same side and names the cost: adaptation is precisely what destroys validation, so anyone needing research-grade scores comparable across teams and studies is directed to Edmondson's original seven items unmodified, with adaptation reserved for intervention in a single team. Edmondson & McManus (2007) warn from the research side that quantitative measures of a phenomenon not yet understood risk fishing expeditions and an uncertain relationship to the thing measured, and Grailey et al (2021) show what a field looks like when everyone measures differently: 62 studies from 19 countries, a patchwork of tools, scales and definitions so wide the data cannot be pooled.
Is more always better? Higgins et al (2022) is the paper most likely to be quoted at anyone arguing for psychological safety: across 545 New York City public schools, organisation-level psychological safety on its own was negatively associated with meeting the federal performance target, and the best-performing combination over time was relatively low psychological safety with relatively high felt accountability. The limits change what that can be used for: psychological safety never went low in absolute terms, the low condition scoring 2.6 on a four-point scale; the outcome is a contested test-based attainment metric the construct was never theorised to predict; and teachers work alone, which removes the interdependence team-level psychological safety is supposed to operate under. Read as a boundary-condition study, it says psychological safety works alongside accountability rather than instead of it; read as proof the construct is overrated, it carries an argument its authors do not make. Tom's critique of Eldor, Hodor and Cappelli argues that paper's nonlinear limits rest on confusing psychological safety with low standards and conflict avoidance, and Edmondson (1999) is precise that the construct is neither permissiveness nor relentless positive affect.
Does it improve outcomes, and could we tell? Montgomery et al. (2025) searched for objective patient safety outcomes, found nine qualifying studies, five significant and pointing in opposite directions, and make the contradiction the finding: patient safety in that literature means the absence of harm, psychological safety a climate for interpersonal risk-taking, so a measure of harm avoided cannot register the value of a risk taken, and Goodhart is cited directly. Vogus & Sutcliffe (2007) do show a prospective link: a nine-item safety organising scale, validated with 1,685 nurses across 125 units, on which higher-scoring units had significantly fewer reported medication errors and patient falls over the following six months, with the objection that such scores capture willingness to report rather than safety answered directly. Campbell (1979) says the complexity of historical change makes attribution deeply difficult and argues for quasi-experimental designs as the honest alternative to either randomised trials or unevaluated implementation. Tom's Parachutes and Evidence rests the case on the catastrophe record, concerns that existed at the sharp end and did not travel, while granting that successes leave no trace, with Smith & Pell (2003) against the demand for a trial and Yeh et al. (2018), whose trial ran from a mean altitude of 0.6 metres, on who enters one; Glasziou et al. (2007) supports the two halves unequally, since falling without a parachute is the stable, near-certain outcome the dramatic-effect criterion was built for, while organisational psychological safety sits at the probabilistic end where the approach is weakest, so the catastrophe record is a forensic inference rather than a rate ratio.
Benchmark, or refuse to. Measure refuses benchmarking outright, no percentiles and no comparison with other teams, because comparison converts a conversation-starter into a target and invites competition, complacency, anxiety and gaming in turn, and the platform says of two organisations averaging 3.4 and 4.7 that we cannot say which is safer when the work, the context and what the questions mean all differ. Against that, the critique of the Psychological Safety Index grants the Fearless Organization Scan real diagnostic value alongside its Goodhart risk, and our own State of Psychological Safety survey sets out to produce a global picture of how psychological safety is experienced across sectors and cultures, a different scale and a different purpose. Hopkins (2009) warns that even adding process indicators to incentive pay must be handled carefully lest it produce attempts to manage the measure rather than to manage safety.
Anonymity: architecture, promise or obstacle. Measure treats anonymity as architecture rather than promise: no name, email, account or IP address, answers stored with a date but never a time, results sealed until at least four people have answered, comments returned shuffled. Jade's 360 critique says anonymity cannot be shortcut to safety, since what people need is consistent, lived experience that speaking honestly will not cost them, which an anonymous form cannot manufacture. Silverman, Kerrin & Carter (2005) complicate both: in a local authority that chose not to anonymise, five colleagues agreed to write the same criticism of their manager so that none could be blamed, yet most raters were content to be identifiable, because receivers feared anonymity would let personal agendas hijack the process; and they conclude that adoption of the practice reflects faith rather than proven validity. The Fist of Five critique sits underneath: the Fist of Five, a real-time measure, is subject to consensus bias and can only be answered honestly in conditions of existing safety.
Words and numbers. Jade's Qualitative Measurement critiques the assumption that numbers are rigour and words merely impressions: a heat map can tell you where psychological safety is low but not why, and a short conversation can reveal a tyrannical former manager or a culture of private channels. Edmondson & McManus (2007) are candid from the research side about the uneven status of evidence, recounting a mixed-methods paper that was strengthened considerably by removing its qualitative material, while holding that the 1999 study needed its final qualitative phase to explain what the numbers meant. Pauly (1995) says a grandfather's report of bluefin tuna in a mackerel net is as factual as a temperature record; Tom's masters coursework on social science and epidemiology argues qualitative evidence can be actionable, rapid and legitimate; and Beyond Metrics, with John Snow's cholera map, says over-reliance on metrics produces measurement-shaped rather than reality-shaped knowledge. Greenhalgh & Papoutsi (2018) give the general case, and its limit: a manifesto for a paradigm shift more than a manual for executing one.
What this means for practice
- Measure only with a plan, and close the loop. The practical guide, the Principles of Ethical Measurement, whose seven principles run from measure to benefit, not to surveil, to close the loop, or do not open it, and Measure, the platform, which returns suggestions offered as experiments, not prescriptions.
- Report the spread, not the mean, and read a perfect score as a warning. Measure and the practical guide, because a team is only as safe as its least safe member.
- Ask why, not only how much. Measure follows a low score with the ambiguity, consequence and futility questions; Jade's Qualitative Measurement offers open text fields, observation in meetings, interviews and focus groups; Measuring PS: Questions First asks about purpose and design before any instrument, since poor measurement design produces worse data than none.
- Watch the behaviour, not the dashboard. Looking At Things Properly: watching the dashboard substitutes for attending to the organisation, and once the indicator becomes the target it stops being an indicator. Things You Might Hear lists what to listen for, and 15/5 Reports build the channel.
- Choose the instrument for the job, and know when not to survey. Measure: Edmondson's original seven items unmodified for research-grade comparability, adaptation for intervention in a single team, and no survey at all for short-lived teams, teams of fewer than four, where anonymity becomes arithmetic rather than architecture, or teams in acute conflict expecting a survey to resolve it. Choosing a Psychometric Tool lists purpose, validity, cultural fit and the risk of Goodhart.
What the map does not yet hold
The measure that Liu et al (2024) rated most rigorous, Detert and Burris's, is on the map as a voice paper, Detert & Burris (2007), whose desc says nothing about its instrument. Kahn (1990), whose conditions May, Gilson & Harter (2004) operationalised, sits outside this region, and the May, Gilson and Harter entry is built from the abstract and secondary sources, as is Kassing (1998). Edmondson & Bransby (2023), the successor to Edmondson and Lei's review, is a foundations node and not in the region, and Bransby, Kerrissey & Edmondson (2025), the longitudinal study of how psychological safety changes with tenure, sits under team learning. O'Donovan and McAuliffe (2020) is named only in passing, inside one healthcare article, and nothing on the map describes it. The cross-cultural gap that Newman, Donohue & Eva (2017) flag has no node of its own on measurement; Tom's article on Hofstede's cultural dimensions covers how psychological safety manifests across national cultures, not whether the scale travels. Nothing covers the engagement-survey industry. The State of Psychological Safety survey node describes and promotes the 2025 survey, and the map holds nothing yet on what it found. The ten-statement quiz from 2020, a quick diagnostic and conversation starter, was retired on 25 October 2026 and its page now redirects to Measure; the node stays as a record of where the thinking started.
Explore the density
Start here
- Psychological Safety and Learning Behavior in Work Teams
The paper that turned psychological safety from a concept in the organisational change literature into a team-level construct with a measure, and the source of the definition almost every later study…
- Meta-Analysis and Systematic Review of the Measures of Psychological Safety
The measurement audit the field needed, and its most uncomfortable finding lands on the scale nearly everyone uses.
- Dysfunctional Consequences of Performance Measurements
Ridgway's two-page paper is the origin of the measurement-dysfunction argument.
- Thinking About Process Safety Indicators
Hopkins explains how an organisation can watch its safety numbers improve while a catastrophe assembles itself underneath, and the mechanism he describes is the most consequential measurement failure…
- Walking the Plank: An Experimental Paradigm to Investigate Safety Voice
The paper that put a number on the gap between what people say they would do about a hazard and what they actually do.
- Verbal Reports as Data
Established the theoretical basis for treating verbal reports as legitimate scientific data, within the human information processing framework.
- Accountability: The Neglected Social Context of Judgment and Choice
An empirical and theoretical paper arguing that accountability — being answerable to an audience for one's judgments — is one of the most powerful and neglected determinants of human thought and beha…
- Methodological Fit in Management Field Research
Introduces 'methodological fit', internal consistency among four elements of a field research project (research question, prior work, research design, contribution to the literature), as a criterion…
- Psychological Safety and Patient Safety: A Systematic and Narrative Review
A systematic attempt to find quantitative evidence that psychological safety improves objective patient safety outcomes, and the paper to have to hand when that link is asserted as settled.
- When Is Psychological Safety Helpful in Organizations? A Longitudinal Study
The first longitudinal test of psychological safety at organisation rather than team level, and the paper most likely to be quoted at anyone arguing for psychological safety, so worth reading precise…
- The Safety Organizing Scale: Development and Validation of a Behavioral Measure of Safety Culture in Hospital Nursing Units
Develops and validates the Safety Organizing Scale (SOS), a 9-item self-report measure of the concrete behaviours that constitute a unit-level safety culture, addressing a real gap in the prior liter…
- Measure: PS Survey Tool
A free measurement instrument built on the argument that a survey is an intervention rather than a thermometer: observing changes what is observed, people cannot always explain their own silence, and…
- Principles of Ethical Measurement
Seven principles for measuring people without mistaking them for data, inspired by UKRI's principles of ethical research, and the newsletter that gathers the practice's measurement writing into one p…
- Least Safe Person
Argues a team is only as safe as the least safe person — and that measuring PS as an average actively hides the outlier who is suffering.
- Qualitative Measurement of PS
Argues that qualitative measurement — conversations, observation, interviews, focus groups — tells you things that quantitative surveys can't: the context, nuance, and reasons behind the numbers.
Trails through this region
Where it borders other regions
- Critique & Boundary
- Interpersonal Practice
- Models & Critique
- Safety & Error
- Complexity & Systems
- Voice & Silence