At a team meeting, nobody objects to a flawed plan. A manager concludes that everyone feels safe and agrees. Afterward, several members tell a researcher they expected a challenge to be punished. The silence was visible, but the belief behind it was not. Psychological safety concerns whether people expect interpersonal risk taking, such as admitting an error or asking a difficult question, to be tolerable in a particular group. Measuring it requires more than counting how often someone speaks.
Measuring psychological safety means specifying the referent, setting, time, and risk. A private survey can ask about a person’s perception of a team. An observation can document how a team responds when someone raises a concern. An interview can expose meanings a fixed item misses. Each method supplies a different piece of evidence. This guide explains how to combine them without treating a favorable score as proof that every member is safe in every situation.
Define the interpersonal risk
Edmondson’s foundational study defined team psychological safety as a shared belief that a team is safe for interpersonal risk taking. The study examined 51 work teams in a manufacturing company and linked team safety to learning behavior. The key measurement question is what a member expects would happen if they spoke candidly, made a mistake visible, sought help, or challenged an assumption. Feeling cheerful, liking colleagues, and being free from all pressure are not equivalent to that expectation.
A measure should therefore contain a clear social referent. “I feel safe” can point to a workplace, a supervisor, a meeting, or a current mood. “Members of this team can bring up problems” narrows the target. If a survey asks about the immediate supervisor while the analysis labels the result “team climate,” the level has shifted without evidence. Name whether the intended construct is a person’s felt safety, a dyadic relationship, a team belief, or an organization’s climate.
Psychological safety is also conditional. Someone may speak freely about a technical error but fear raising discrimination, workload, or misconduct. A scale based on generic risk items may miss those domain differences. Researchers can choose a broad measure for broad comparisons, but they should not interpret it as comprehensive protection in every high-stakes setting. The design should state which kinds of interpersonal risk are included and whether any are excluded by the wording.
Edmondson and Bransby’s review describes how the literature expanded across levels, contexts, and outcomes. That breadth makes construct clarity more important, not less. A score in a hospital ward, a software team, and a classroom may use the same label while respondents imagine different acts of speaking up. Preserve the item wording and the team boundary. If language is adapted, check that the revised item still asks about interpersonal consequences rather than merely satisfaction with leadership.
Use surveys with a defined unit of analysis
Edmondson’s original team scale is widely used because it asks about errors, help seeking, rejection, and speaking up in the group. A response is initially an individual’s report. Turning several reports into a team score requires evidence that members are describing a shared environment and that the team mean is estimated with useful precision. Within-team agreement and between-team variation answer related but different questions. A high average can coexist with one member’s markedly different experience.
In the original field study, analyses examined whether team-level aggregation was defensible. That precedent is not a permanent license to average any modern sample. Team sizes, item interpretation, sampling, and organizational structure change. Report how many members responded per team, the response rate, distribution of answers, within-team agreement, and the statistical basis for group comparison. A team mean computed from one enthusiastic respondent is a person’s report, not a reliable team climate estimate.
Survey administration itself can affect candor. When a manager sends the link, sees who completed it, or asks for raw responses, employees may protect themselves by endorsing safety. Confidential collection and transparent reporting rules are therefore part of validity. Explain who can see individual answers, whether quotations could identify someone, and whether results will be shown only at a level that protects respondents. A low score may be informative, but a high score under weak confidentiality deserves caution.
Missing responses deserve an analysis rather than a footnote. The members least willing to criticize a team may also be least willing to complete a survey whose results reach management. In that case a high response average over the remaining members can exaggerate the sense of safety. Researchers can report participation by role without revealing identities, compare available characteristics of respondents and nonrespondents, and test how plausible unobserved answers would change the team estimate. These checks cannot recover private beliefs that were never reported. They do make the uncertainty visible. Repeated invitations should also be voluntary and neutral; pressuring reluctant members to answer can itself change the interpersonal risk that the survey is meant to assess.
Time frame matters. A question about “this team” may invite a general impression; a question after a specific meeting may reflect that event. Neither is inherently superior. A general scale may be useful for comparing stable climates, whereas repeated event-linked measurement may capture change after a leadership action or conflict. A longitudinal design should not assume the same referent if the membership or task changes. Record who belonged to the team at each wave and what period the answer was meant to summarize.
Alternative instruments can broaden the measured domains. Plouffe and colleagues developed a Psychological Safety Inventory with content spanning interpersonal risk, respect, organizational support, clarity, and leadership. That development work offers a different operationalization, but a broader item set can also blend antecedents with the target belief. Before choosing a scale, inspect what its items ask and whether the proposed outcome is already embedded in them. Two scales bearing the same name need not estimate the same construct.
Observe behavior without equating it with belief
Observations can record invitations to speak, interruptions, questions, admissions of error, defensive reactions, and how a group handles dissent. O’Donovan and colleagues developed an observational approach for healthcare teams to complement survey methods. Observation can expose concrete interaction patterns that members may not describe well in a questionnaire. It cannot directly reveal a silent person’s private expectation. A quiet meeting could be safe but uneventful, unsafe, or simply tightly scheduled.
Code the opportunity as well as the behavior. The number of voiced concerns means little if the meeting had no decisions to challenge or if only one person had relevant information. A useful coding frame describes when a concern was possible, who had the floor, who initiated a question, how others responded, and whether the issue affected later action. Normalize counts for opportunity or meeting length where appropriate. Without that denominator, a busy team can appear safer merely because it had more occasions to speak.
Observer presence can alter conduct. Leaders may invite opinions more often when watched; members may suppress sensitive topics. A camera can change who participates. If feasible, sample several meetings and document how familiar the team was with the observer. The healthcare observational work was designed as a complementary measure rather than a substitute for surveys. That distinction should remain in reports: observed behaviors are evidence about interaction, and survey reports are evidence about expectations.
Triangulation is particularly useful when measures disagree. A healthcare study combining surveys, observations, and interviews found that favorable survey scores could coexist with observed or described situations of lower safety. The discrepancy is not a statistical nuisance to smooth away. It asks which events the survey missed, whose view dominated the average, and whether people interpret “safe” differently in routine versus consequential situations. The methods become stronger by explaining disagreement.
Distinguish related outcomes and antecedents
Psychological safety is often studied with voice, learning behavior, engagement, and performance. These are not its synonyms. Someone may believe speaking up is safe yet have nothing to add. Another may voice a concern despite anticipating retaliation because the stakes are high. An observed suggestion therefore cannot by itself score psychological safety. Separate the belief measure from the behavior measure and test their relationship rather than building one into the definition of the other.
Trust and psychological safety overlap but also differ. Trust can concern another person’s reliability or benevolence; psychological safety concerns the expected interpersonal consequence of risk taking in a setting. A participant may trust a colleague with a confidential task yet fear embarrassment when challenging the group publicly. If a survey uses only generic trust questions, it may not measure the risk-taking construct. The comparison should examine discriminant validity rather than rely on labels.
Leadership style can be an antecedent, but placing leader support items inside the safety score can make an association between leadership and safety partly circular. A respondent who likes the manager may endorse all favorable items. Use distinct scales or behavioral measures when testing whether leadership predicts safety, and consider common-method bias when both are collected at one time from the same person. A cross-sectional association is not proof that a particular leader action caused the safety belief.
Frazier and colleagues’ meta-analysis aggregated many samples and found relationships between psychological safety and a range of outcomes. Such synthesis shows that the construct is consequential across studied settings, yet heterogeneous measures, designs, and levels limit a simple causal reading. A validation study should ask whether the local scale functions as intended, not just whether the broader literature reports positive correlations. Strong population evidence cannot repair a badly administered local survey.
Interpret difference and change carefully
Teams are not randomly assigned to every condition. A team facing a recent crisis, new membership, or a demanding project may report different safety for reasons unrelated to a proposed intervention. Measure baseline conditions and describe changes in membership, tasks, and leadership. If the goal is causal evaluation, compare with a suitable group and prespecify timing. A before-and-after increase following a workshop may reflect temporary optimism or a changed interpretation of the questions.
Repeated surveys can also sensitize members to the topic. After a training session explains psychological safety, participants may recognize risks they had previously normalized and score the team lower, even while conditions improve. Conversely, they may infer that a higher answer is expected. Pair repeated ratings with concrete observations and interviews, and examine item-level changes. A lower number is not automatically deterioration; it may reflect greater awareness of the construct being asked about.
Survey scores can hide unequal exposure. A junior member, contractor, or marginalized colleague may face higher interpersonal cost than a senior member. If the team average is high, inspect the distribution and protect confidentiality when subgroup counts are small. Do not expose a dissenting respondent through a chart that makes them identifiable. A measurement that causes retaliation undermines the very construct it tries to assess.
Cross-cultural or cross-language comparison needs more than translation. Norms about public dissent, hierarchy, and the meaning of a “mistake” can differ. Check whether items have comparable meaning and response patterns across groups before comparing means. A low endorsement of speaking up may reflect fear, deference, language nuance, or an accurately perceived role boundary. Context does not make measurement impossible; it makes local validation necessary.
What the evidence does not show
No single survey proves that a team is safe for every member and every topic. A high mean can conceal dissent, and a low mean does not by itself identify which interaction caused concern. An observational count cannot read private expectation; silence can have several causes. The original association between team safety and learning behavior is informative but does not establish that increasing a scale score alone will cause learning or performance to improve.
The cited literature does not authorize using a safety score as a label for individual worth, loyalty, or willingness to cooperate. It is a research estimate of a belief in a defined setting. The Institute’s synthesis is to make the inference chain explicit: item responses support perceived interpersonal consequences; agreement may support a team-level estimate; observation can test whether interaction is consistent with that estimate; longitudinal or experimental designs are needed for stronger claims about change.
Build a defensible report
A transparent study reports the item set, wording, response scale, language, referent, collection mode, confidentiality practice, timing, team size, response rate, and reason for aggregation. If observation is used, it provides the coding scheme, sampling period, opportunities to voice, and agreement between coders. If interviews are used, it explains selection and how quotations were protected. Results can be presented as a distribution rather than a single score.
The opening meeting can then be described accurately: members were quiet, and some privately expected negative consequences for challenging the plan. That evidence supports concern about psychological safety in that situation. It does not require a sweeping verdict about everyone in the organization. The point of careful measurement is to make an invisible interpersonal expectation discussable without exposing the people who most need protection.
Research references
- Edmondson, A. (1999). Psychological safety and learning behavior in work teams. Administrative Science Quarterly, 44, 350–383. View source
- Edmondson, A. C., & Bransby, D. P. (2023). Psychological safety comes of age: Observed themes in an established literature. Annual Review of Organizational Psychology and Organizational Behavior, 10, 55–78. View source
- Frazier, M. L., et al. (2017). Psychological safety: A meta-analytic review and extension. Personnel Psychology, 70, 113–165. View source
- O’Donovan, R., et al. (2020). Measuring psychological safety in healthcare teams: Developing an observational measure to complement survey methods. BMC Medical Research Methodology, 20, 203. View source
- O’Donovan, R., et al. (2020). Exploring psychological safety in healthcare teams to inform the development of interventions: Combining observational, survey and interview data. BMC Health Services Research, 20, 810. View source
- Plouffe, R. A., et al. (2023). Feeling safe at work: Development and validation of the Psychological Safety Inventory. International Journal of Selection and Assessment, 31, 443–455. View source