Measuring Manager Effects on Attrition: Fixed Effects, Sorting, and What an Identification Claim Requires
“People leave managers, not companies” is the most quoted sentence in retention practice, and most of the analyses offered in support of it are correlations. A dashboard showing that some managers’ teams lose people at three times the rate of others establishes that attrition varies by manager. It does not establish that the manager caused it. The claim “bad managers cause attrition” is an identification claim, and identification requires a design. This piece covers manager fixed effects, the sorting problem, leave-in and leave-out estimation, and what a defensible manager-impact programme can actually estimate.
Manager fixed effects and what they absorb
The standard approach fits a model of individual exit with a separate intercept for each manager: exit_12m ~ X + manager_fe, where X holds employee-level controls such as tenure_days, compa_ratio, job family, level, and location. The estimated manager_fe is each manager’s attrition rate conditional on the composition of their team.
What that intercept captures is everything shared by employees under that manager that is not in X. That includes the manager’s behaviour. It also includes the team’s work content, its location’s local labour market, its project cycle, the business unit’s restructuring history, and any characteristics of the people assigned to that manager that the controls do not capture. The fixed effect is a manager-position effect until proven otherwise.
Small-team estimates add noise. A manager with six reports and two exits has an extreme estimated effect driven largely by chance. Shrinkage toward the mean (empirical Bayes or a random-effects specification) is the minimum treatment before any ranking is shown.
The sorting problem
The central threat is non-random assignment of employees to managers. Sorting runs in several directions:
- Difficult teams get experienced managers, or the reverse, depending on how the organisation deploys talent.
- Flight-risk employees seek transfers away from managers they dislike, and toward ones with reputations, so team composition reflects prior attrition intent.
- New managers inherit unstable teams. A team that just lost its manager is already in a high-attrition state when the new manager arrives.
Each of these produces a correlation between manager identity and attrition without any manager behaviour causing exits.
Leave-in/leave-out and mover designs
The designs that make identification plausible exploit movement.
Manager movers. When a manager moves between teams, compare attrition in the destination team before and after arrival, against comparable teams that did not change manager. If a manager’s effect is real, it should travel with them.
Employee movers. When employees change managers through reorganisation rather than choice, their before-and-after attrition hazard, conditional on their own characteristics, identifies differences between managers. Reorganisation-driven moves are especially valuable because they weaken the sorting story.
Leave-out estimation. To avoid an employee’s own outcome contaminating the estimate of their manager’s effect, compute each manager’s effect leaving that employee out — or, more strictly, estimate the manager effect from one period and evaluate it against attrition in a later period with a different team.
Mobility constraints
Mover designs need movement, and many organisations have little of it. If managers rarely change teams and employees rarely change managers, the network of connections between managers is sparse, and individual manager effects are not separately identified from team and unit effects. Before estimating, map the mobility network: how many managers are connected through movers, and how many movers each estimate rests on.
To illustrate — numbers constructed for demonstration, not measured: in a population of 400 managers where only 60 have observed movers in or out over three years, at most those 60 have effects estimated from movement. The remaining 340 have only cross-sectional estimates, which carry the full sorting problem.
What a defensible programme estimates
A defensible manager-impact programme claims less than the slogan does. It can estimate:
- The variance of manager effects in the population, which bounds how much attrition could plausibly be addressed through management quality.
- Effects for well-connected managers with confidence intervals and stated mover counts.
- Effects of specific manager events — a change of manager, a manager departure — which are more cleanly identified than individual manager quality.
It should not publish league tables of individual managers ranked by raw team attrition, and it should not use unidentified estimates in performance evaluation.
What we cannot claim
Even mover designs assume that moves are not themselves timed to anticipated attrition, and that is rarely fully true. A well-identified estimate is stronger evidence than a correlation; it is not proof. We can say managers plausibly matter; we cannot, from observational data alone, say which managers caused which exits. The data always wins over the narrative — and here the narrative arrived long before the identification.