Most leadership advice stays fairly vague: be a good communicator, build trust, listen to your people. Useful in spirit, hard to act on directly. One of the more influential pieces of organizational research in the last few decades gets considerably more specific about what that vague advice actually looks like in practice, and why it matters enough to affect whether people stay or go.
The study is a 1999 paper by Amy Edmondson, published in Administrative Science Quarterly while Edmondson was at Harvard. She studied 51 work teams inside a single manufacturing company, measuring a chain of variables from antecedent conditions through to outcomes. The construct at the center of her model, team psychological safety, she defined as a shared belief among a team that it’s safe to take interpersonal risks: to ask a question that might sound obvious, admit an error, or flag a problem without expecting to be punished or embarrassed for it. Her results showed that team psychological safety was associated with learning behavior, the actual, observable habit of a team discussing mistakes, seeking feedback, and experimenting. Team efficacy, a team’s general confidence in its own ability, was not associated with learning behavior once psychological safety was accounted for. And learning behavior itself mediated the relationship between psychological safety and team performance, meaning safety mattered largely because it enabled the kind of open, error-surfacing behavior that made a team actually get better over time. Crucially for anyone asking what a leader specifically does to create this, Edmondson’s results pointed to team leader coaching, alongside structural context support, as antecedents that helped shape whether a team developed that shared sense of safety in the first place. Directing a team and coaching a team, in other words, are not the same act, and only one of them showed up as a foundation for the deeper condition her data linked to real performance.
A separate, much later body of research offers a real-world echo of the same finding at considerably larger scale. Between 2012 and 2015, Google ran an internal research effort known as Project Aristotle, studying more than 180 of its own teams to figure out what separated its most effective teams from the rest. The company’s own People Analytics researchers found that who was on a team, seniority, background, personality mix, mattered far less than how the team worked together. Of the team dynamics they examined, psychological safety came out as the most important factor distinguishing high-performing teams from the rest. It isn’t a peer-reviewed academic study in the way Edmondson’s paper is, but it’s a large, methodologically documented internal study from a company with obvious incentive to actually get the answer right, and its central finding lines up with Edmondson’s academic research from over a decade earlier almost exactly.
Put together, this research offers a fairly direct answer to why some leaders keep their best people motivated and others quietly bleed talent. A team where people are afraid to admit they’re stuck, disagree with a plan, or point out a problem early doesn’t just perform worse on Edmondson’s numbers. It’s also a team where the first signs of someone disengaging, growing frustrated, or actively job hunting are far less likely to reach a leader in time to do anything about it, since the same fear that suppresses a question about a mistake tends to suppress an honest conversation about wanting to leave. The specific leader behavior tied to preventing that, in Edmondson’s data, wasn’t charisma or vision. It was coaching: engaging with a team’s actual work and development rather than simply issuing direction and expecting compliance.
It’s worth being honest about the limits of this picture. Edmondson’s study was conducted in one manufacturing company in the 1990s, and while the underlying construct has since been replicated across many other industries and settings, the original 51-team sample is modest and specific to that context. Her model is also correlational at its core, drawing on survey and performance data collected together rather than a controlled experiment, which makes the causal direction harder to pin down with certainty than a true experiment would. Google’s Project Aristotle, valuable as its scale is, was not conducted or published with the same academic rigor and outside peer review as Edmondson’s paper, and it remains a single company’s internal research rather than an independently replicated study. Neither piece of research claims that psychological safety is the only thing that matters. Edmondson’s own model explicitly includes structural factors like team task design and resource support alongside the shared belief in safety, and treating either study as a complete leadership formula would overstate what the data shows.
Within those limits, what the research does support is a specific, learnable distinction rather than a vague trait. The leaders whose teams keep learning, performing, and staying appear to be the ones who coach rather than merely direct, creating the kind of shared safety that lets a problem, a mistake, or a quiet frustration surface early enough to actually do something about it.