Contribution
It names team psychological safety as a group-level construct, “a shared belief that the team is safe for interpersonal risk taking”, gives it a seven-item scale, and shows that it varies between teams of one company and that members of a team agree about it. It places the construct in a model in which team structures (context support, leader coaching) work on learning behavior through it, and learning behavior works on performance. It is the paper the later literature on psychological safety starts from.
Key insight
Learning in a team is made of acts that put the actor’s image at risk: asking for help, admitting an error, seeking feedback, questioning the goal. Whether people do them depends less on whether they think the team can succeed than on whether they think the team will not embarrass, reject or punish them for trying. That belief is held tacitly, is shared within a team and differs between teams in the same organization, so it is a property of the team and not only of its members.
Assumptions
- Group level. A construct is treated as the team’s when it is meaningful at that level and members’ answers converge (intraclass correlation above zero, Kenny and LaVoie). Team means then make a data set of N = 51.
- Mediation by Baron and Kenny’s three steps: the mediator predicts the outcome, the predictor predicts the mediator, and the predictor’s effect drops or vanishes when both are entered. No test of the indirect effect itself is reported.
- Performance is meeting the needs of the team’s internal customers, rated by two or three managers who receive its work.
- The setting. One company (“Office Design Incorporated”, about 5,000 employees), teams listed by managers at the request of an internal organization-development contact; not a representative sample of its teams, by the author’s own statement.
Key results
- Sample (Methods). 53 teams listed: 34 functional, 9 self-managed, 5 product development, 3 project. 427 of 496 members in 51 teams returned the survey (86%); 135 of 150 observers returned theirs (91%). Members’ and observers’ ratings of the same team agree for learning (r = .60) and for performance (r = .36).
- Group-level agreement (Table 3). ICCs .29 to .39 for the group-level scales (psychological safety .39), against .03 and .04 for internal motivation and job involvement, which are individual constructs.
- H1, learning behavior → performance (Table 4). Self-reported learning predicts observer-rated performance, and of seven models it accounts for the most variance; context support and coaching add nothing to it.
- H2, safety → learning (Tables 5, 6). B = .76, t = 9.16, R² = .63 for self-reported learning; B = .46, t = 5.26, R² = .33 for observer-rated learning. Entered with efficacy, context support and coaching, safety is the only significant predictor of either measure.
- H4 and H5, efficacy. Efficacy alone predicts learning (B = .38, R² = .26 self-reported; B = .45, R² = .23 observed). With safety controlled it adds nothing to self-reported learning and remains significant for observed learning: “mixed support” for H5.
- H3, learning mediates safety → performance. Entered together, safety drops to B = .25, p = .42 and learning stays B = .60, p < .05.
- H6 and H7, structures → safety → learning (Table 6). Context support (B = .56, t = 3.82) and coaching (B = .33, t = 1.89, p = .06) predict safety (R² = .52). With safety entered, both lose significance for observed learning (.09 and .12) while safety stays (.29, p = .02).
- H8, efficacy as mediator (Table 7): not supported. Coaching does not predict efficacy (B = .01, p = .95); context support does (B = .77).
- Individual-level GLM (Table 8, N = 427). Team type does not predict safety (p = .12) or learning (p = .10) once context support, coaching and team membership are in; team tenure does not either.
- Follow-up cases. In the seven high- and low-learning teams safety goes with learning in all seven; coaching does in five. Context support is not tied to the split: two of four high-learning teams had inadequate support, and all three low-learning teams adequate support.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Team psychological safety is a shared, team-level belief that varies between teams in one organization | moderate | ICC .39 and F(50,427) = 6.98; Phase 1 interviews |
| C2 | Psychological safety is associated with team learning behavior, by members’ and observers’ ratings | moderate | Tables 5, 6; one company, cross-sectional |
| C3 | Learning behavior carries the association between safety and observer-rated performance | weak | Baron–Kenny steps at N = 51; no test of the indirect effect |
| C4 | Safety carries the association of context support and coaching with learning | weak | Table 6; predictors and safety come from the same survey |
| C5 | Team efficacy matters for learning less than safety does | moderate | Tables 5, 7; H5 mixed, H8 unsupported |
| C6 | High-learning teams can work round poor design; low-learning teams get stuck in a self-sealing pattern | weak | seven cases, interpretive |
Method
A multimethod field study in three phases: observation of eight teams and 17 interviews to build the survey items (Table 1); a member survey, an observer survey and structured interviews of 31 other managers by a researcher blind to the survey, to check the design measures; and follow-up observation and interviews of seven teams chosen from the survey’s six highest and six lowest on learning behavior. Analysis is OLS on team means plus a GLM on individuals with team membership as a random effect.
Concepts
- Team psychological safety: “a shared belief that the team is safe for interpersonal risk taking”; not cohesiveness, which can suppress disagreement, and more than interpersonal trust: a climate of trust and mutual respect “in which people are comfortable being themselves”.
- Learning behavior: “an ongoing process of reflection and action, characterized by asking questions, seeking feedback, experimenting, reflecting on results, and discussing errors or unexpected outcomes of actions”; a process, kept apart from learning outcomes.
- Team efficacy: a shared belief that the team can do its work.
Connections
The learning-as-process tradition it joins is Argyris and Schön’s (1978) detection and correction of error and Dewey’s inquiry; Argyris’s later statement of double-loop learning is LIT-745 (and LIT-739). The structural side is Hackman’s model of team effectiveness. Schein and Bennis (1965) are the source of “psychological safety” for individuals. Janis’s groupthink is the foil for cohesiveness.
Bearing on the record
- CLAIM-359 cites it as complementary: robustness to compromise depends on processes that protect dissent and audit. The paper supports a narrower point: in one company’s work teams, members’ belief that speaking up will not be punished goes with the team’s asking, admitting and questioning, and through that with its performance as its customers judge it. It says nothing about adversarial or compromised members; its teams are cooperative. Its case of NPD 2, whose members did not question a goal “for fear of sanction by management”, is the nearest it comes to the claim’s restriction of disagreement.
- CASE-108’s “honestly dissenting” specialists are the role this paper’s construct protects; the paper gives no evidence on what protecting them costs.
No instruction for machine-learning practice; nothing for the anthology.
Limitations
- Cross-sectional, by the author’s statement: no causal order is shown, and the dynamic in which unpunished risk taking builds safety is left out of the model.
- One company, 51 teams, a sample the author calls small for multivariate analysis; teams were not selected to be representative.
- Construct validity is not settled: the qualitative data “did not map onto it precisely”, and safety could not be cleanly separated from trust.
- Common source. Safety, context support and coaching come from the same survey, so the antecedent findings are, in the author’s word, tentative. Observers’ learning and performance ratings lacked discriminant validity and were not related to each other.
- Low alphas for context support and team efficacy (Table 2, stated in the text).
- Errors and slips. The text gives the ICC of team learning behavior as .33; Table 3 gives .27. Table 6 prints p = .21 for coaching at t = .78 (observer-rated learning, step 3); a two-sided test gives about .44, so the figure is a misprint or one-sided, and the conclusion (not significant) stands either way. The text says all ICCs are significant at p < .0001; Table 3 says < .001. The text writes “regressing team learning … on team performance” for the regression of performance on learning. “Six of the eight hypotheses” counts H6 as supported, though coaching’s own coefficient is p = .06.
Open questions
- Does safety precede learning behavior, or do unpunished acts of learning make the belief? A panel of teams would separate them.
- Does the association hold for highly constrained tasks, where the author expects learning behavior to matter little to performance?