Citation Snowball Sampling & Literature Saturation

Screening Criteria

Relevance threshold τ 0.55

Reviewer noise σ 0.12

Snowball direction

Review Progress

Round

Screened

Included

Coverage

Marginal yield / round —

Seed papers Candidate (this round) Included (relevant) Excluded (screened out) Not yet reached

How it works

A systematic literature review rarely reads the whole literature — it grows a sample by snowball sampling (Wohlin, 2014): starting from a small seed set of known-relevant papers, each round follows citation edges outward — backward to the papers a seed references, forward to papers that cite it — and screens every newly reached candidate against a relevance criterion.

round r:  screen each candidate in frontier(r)
          observed = trueRelevance + noise(0, σ)
          include  if observed ≥ τ
frontier(r+1) = neighbors(included papers) − visited

Two numbers decide when to stop:

Coverage(r)      = ΣtruePositives(≤r) / totalRelevantInGraph
MarginalYield(r) = truePositives(r) / screened(r)

τ (threshold) — how strict the inclusion criterion is; too low admits noise, too high misses real work.
σ (reviewer noise) — models imperfect, inconsistent human screening (dual-reviewer disagreement in real PRISMA workflows).
Marginal yield — the bar chart's declining trend is the empirical saturation signal: once new rounds return almost no new relevant papers, the review has diminishing returns from continuing to snowball, exactly the heuristic reviewers use to justify stopping the search.
Coverage — the fraction of the (ground-truth, hidden) relevant literature actually found so far — unknowable in a real review, but instructive here to see how threshold and noise trade off precision against completeness.

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.