Team Stability:
Theory and Evidence
Abstract
Teams emerge organically when individuals decide on their own their activities and collaborators. I develop and test a theory in which teams form with consent, interactions reflect team composition, and payoffs feature externalities and several forms of congestion. Stability constraints concern group interactions and require each team member to be willing to remain rather than withdraw or pursue a project alone.
Empirical analysis of research collaborations among economists at the Federal Reserve Board suggests that:
- these constraints help explain observed teams beyond the payoffs to collaboration;
- estimated violations point to large teams of entrants that appear to be assigned;
- a compensatory mechanism mitigates the effects of variation in policy workload, which appears to affect team composition more than total research output; and
- shifting workload toward experienced economists increases entrants' research output and collaboration, at the cost of lower output among experienced economists who compensate for this shift in policy workload with increased interactions with entrants.
Example: k-team stable dynamics (k ∈ {1, 2})
Individuals choose projects and teammates. A locally feasible change can destabilize another team. Its next adjustment may lower the potential—even as the process approaches equilibrium.
Players’ payoff functions
U1(A) = 2p1 − p12 + 0.5g12 + 3g13
U2(A) = 2p2 − p22 + 0.5g12 − 10g23
U3(A) = 2p3 − p32 + 3g13 − 10g23
Here pi = ∑r ai,r is individual i’s number of projects, and gij = ∑r ai,raj,r counts projects shared by i and j. Each project returns 2; total workload costs pi2. Collaboration values are symmetric and earned per shared project. All other payoff terms are zero in this example.
Individuals & projects
Potential progress
3.5Starting value
Individual Project Unstable team link
One possible meeting sequence, with k = 1. Choices are deterministic; meetings are random. Playback stops at the absorbing equilibrium.
The meeting sequence
- Period 0. Initial configuration. Project 1: {1, 2}; project 2: {3}. Both initial teams are feasible.
- Period 1. i = 1, R = {2}. Individual 1 considers only project 2. With k = 1, she can join it but cannot simultaneously drop project 1. Stability is checked only for project 2, whose constraints pass. Her membership in project 1 becomes unstable, but it is outside this meeting’s consideration set.
- Period 2. i = 2, R = {1}. Individual 2 must leave project 1 because retention fails its stability constraints. Her payoff and potential fall by 1.5.
- Period 3. i = 1, R = {1}. Individual 1 drops the now-solo project 1, gaining 1. Project 2 remains stable.
- Period 4. i = 2, R = {3}. Individual 2 starts a new solo project, gaining 1. The resulting configuration is an absorbing equilibrium.
Each consideration set has one project, so k = 1. If period 1 instead used R = {1, 2}, individual 1 would drop project 1 and join project 2 in one update. That alternative would reach equilibrium directly and would not illustrate a potential decrease.
The meeting sequence
- Period 0. Initial configuration. Project 1: {1, 2}; project 2: {3}. Q = 3.5.
- Period 1. i = 1, R = {1, 2}. Individual 1 drops project 1 and joins project 2. The stability constraints pass for the retained shared membership on project 2. Project 1 is now solo for individual 2. Q = 6, and the state is an absorbing equilibrium.
This particular two-project meeting reaches equilibrium directly. The general k = 2 dynamics need not increase potential at every update; meetings are random and singleton consideration sets remain possible.
The meeting sequence
- Period 0. Project 1: {1, 2}; project 2: {3}. Q = 3.5.
- Period 1. i = 1, R = {2, 3}. Individual 1 joins project 2 and leaves the fresh slot 3 empty. Project 1 cannot be dropped in this meeting and becomes unstable. Q = 5.5.
- Period 2. i = 2, R = {1, 3}. Stability forces individual 2 to leave project 1. She starts a solo project 3 instead. Her payoff falls from 1.5 to 1, so Q falls to 5.
- Period 3. i = 1, R = {1, 2}. Individual 1 drops her solo project 1 and retains project 2, gaining 1. Q = 6; the configuration is an absorbing equilibrium.
Every meeting considers exactly two projects. An unused fresh slot disappears after a meeting; the next fresh slot is again shown as project 3 until it is used. This is a verified extension of the paper’s example, with unchanged payoffs and initial configuration.
Why convergence is almost sure
Under the assumptions of Section 3.1, the number of active projects remains bounded throughout the k-team stable deterministic dynamics. With finitely many individuals, only finitely many configurations can therefore arise. At each period, a randomly selected individual chooses her best participation profile among the k-sized considerations set subject to the stability constraints, retaining her current choice whenever it is feasible and optimal. These updates define a finite Markov chain in which every allowed meeting has positive probability.
A closed recurrent class is a set of states in which every state can be reached from every other through transitions with positive probability, and which the process cannot leave once it enters. It need not be a simple cycle: it may contain many interconnected paths, or just one absorbing state.
Suppose the process could remain trapped in a closed recurrent class that does not contain a k-team stable equilibrium. Choose a state in that class with the highest potential. If all stability constraints hold but the state is not an equilibrium, some meeting produces a profitable feasible change and raises potential—a contradiction. If a stability constraint fails, the dissatisfied member prefers either leaving that project or replacing it with a solo project. Meetings that allow her to leave and, if needed, subsequently start a solo project lead to a new state. This new state has higher potential than the starting state, even if the first step lowers it. These meetings have positive probability, so the process cannot remain trapped in a closed recurrent class that does not contain a k-team stable equilibrium.
Every closed recurrent class must therefore consist of a single absorbing k-team stable equilibrium. The k-team stable deterministic dynamics reaches a k-team stable equilibrium after finitely many periods with probability one and stays there. Individuals make deterministic choices, but meetings are random: every allowed meeting has positive probability. Potential may fall along the way. The convergence result does not mean that every fixed sequence of meetings must lead to equilibrium.
What changes for k = 2? The same argument applies because consideration sets contain at most two projects, so singleton meetings remain available. In addition, a dissatisfied member can consider her current project and a fresh solo project together, R = {r, rnew}. She can then leave the team and start the solo project in one update. If her participation slack is negative, either exit or this replacement strictly improves her payoff; both are feasible because she retains no shared membership in the consideration set. Her constrained best choice therefore raises potential immediately, rather than requiring two successive meetings. This does not imply that every update raises potential: another member’s failed constraint can still force a payoff-reducing adjustment. In the illustrated example, however, the meeting i = 1, R = {1, 2} at the initial configuration lets individual 1 drop project 1 and join project 2 simultaneously. Her payoff rises from 1.5 to 4, potential rises from 3.5 to 6, and the resulting configuration is already an equilibrium.
About the paper
Which projects do individuals undertake, and with whom? Team stability connects these two choices. Individuals balance project returns, the cost of carrying several projects, team composition, and the benefits and costs of repeated collaboration.
A k-team stable equilibrium allows individuals to reconsider up to k project memberships at a time, subject to the willingness of the considered teams’ members to remain. Deterministic consensual updates converge under the paper’s assumptions and full-support random meetings. Stochastic updates provide a separate framework for studying the distribution of collaboration profiles.
The empirical application studies research teams at the Federal Reserve Board, using working papers, staff rosters, and FOMC meeting transcripts to examine project activity and collaboration.
Why can the potential fall?
The potential tracks the selected individual’s payoff change exactly:
In the k = 1 sequence, after individual 1 joins project r = 2, her workload makes membership in project r = 1 unattractive. That membership lies outside the current consideration set, so it does not block entry into project r = 2. When individual 2 subsequently considers project r = 1, staying is infeasible: individual 1 would prefer to leave. Individual 2 must exit, losing 1.5 units of utility. Potential falls by the same amount.
The convergence argument in Section 3.1 does not require improvement at every step. On the finite reachable state space, full-support meetings provide routes that repair unstable memberships and rule out non-equilibrium recurrent classes. This is an almost-sure convergence result, not a claim about every fixed meeting schedule.
Example parameters and verification
| Update | Meeting (i, R) | Unconstrained best | Constrained best / selected |
|---|
The first two updates reproduce the appendix’s “Why potential can decrease” example. The last two updates are a derived continuation, checked against the same payoff and feasibility rules.
Both k = 2 paths are derived comparisons using the same initial configuration and payoffs as the appendix example. All four choices at each meeting have been checked, as has optimality at each endpoint.
βi = 2; CP(p) = p²; V = CG = ψD = 0.
Symmetric pair values: w12 = 0.5, w13 = 3, w23 = −10.
Every retained or added shared membership in the selected project must satisfy σj,r ≥ 0 for every teammate, where σ compares staying with both leaving and replacing the membership with a fresh solo project. All meetings shown are singletons, so every change is permitted for k = 1.
| State | r = 1 | r = 2 | r = 3 | U₁ | U₂ | U₃ | Q |
|---|
At the endpoint, no individual can improve even by revising all existing memberships and adding a fresh solo project together. All participation constraints hold. It is therefore absorbing for every k ≥ 1, under status-quo inertia on ties.
Project indices remain fixed in the illustration for continuity. Empty projects have no economic role; Section 3.1 deletes them and identifies profiles up to relabeling.
Verification script