Loading 3D assets...
...
Drag to orbit - wheel / pinch to zoom - the pile in the left tray is the review queue

Peer Review Bottleneck: Why Reviewer Shortage Makes Decision Times Explode

Random submissions, a limited pool of reviewers, one in-tray. Small changes in load near 100% cause very large delays.
Load (utilisation)-
Waiting now / reviewers busy-
Wait for a reviewer: theory (Erlang C)-
Wait for a reviewer: this run so far-
Decisions finished-
Chance a paper must wait-
Mean wait before review starts versus submissions per week. Orange: your number of reviewers; blue: one more; grey: one fewer. The wait climbs steeply as the load approaches 100%. White dot: your setting.
Independent numerical checks
Each line compares two separate calculations; a FAIL would show here.
The science

The model. Manuscripts arrive at random (a Poisson process, rate λ per week) and are handled first-come first-served by c reviewers who each hold a paper for an exponentially distributed time with mean T (service rate μ = 1/T). This is the classic M/M/c queue.

Load and the cliff. Utilisation is ρ = λ/(cμ). For ρ < 1 the queue settles; the probability that an arriving paper must wait is the Erlang C formula C(c,a) with a = λ/μ, and the mean wait before review starts is Wq = C(c,a)/(cμ − λ). As ρ approaches 1 the denominator goes to zero, so a system that is "only a little busier" can have a much longer queue. For ρ ≥ 1 the queue grows without bound at about λ − cμ papers per week.

Simplifications and honesty. Real review times are not exponential (they cluster around deadlines and have a long tail), reviewers are not interchangeable, editors can skip or reassign, and the workload is seasonal. This model counts only the wait for a reviewer to be free, not reminder emails, revisions or second rounds. Numbers shown are for the abstract queue, not for any real journal.

Checks. (1) A separate long recursion over 300,000 simulated arrivals gives the mean wait; it is compared with the Erlang C result (computed by the Erlang B recursion). (2) The simulated fraction of papers that wait is compared with C(c,a) written out with factorial sums. (3) For an overloaded setting the simulated queue growth is compared with λ − cμ.

The desks, lamps, trays, paper and figures were modelled and baked in Blender; the queue and rendering run in Three.js r160.

Made with 🟠 Blender · Three.js