The Ubiquity of Sharp Thresholds

Dan Luu's 2020 post "Suspicious Discontinuities" catalogs how sharp cutoffs in systems—from U.S. tax policy to network queues to election results—create perverse incentives and artifacts. For developers, the lesson is clear: thresholds you design can have unintended consequences.

Tax Policy: $7,200 Cliff

In the U.S., the ACA health insurance subsidy cutoff at $48,560 for individuals creates a stark incentive: earning $55,000 can leave you worse off than earning $48,560 after insurance costs. Luu notes that losing $6,440 to drop below the threshold increases disposable income by roughly $720/year. People buy put options to deliberately lose money. This is rational individual behavior but suboptimal system design.

Network Queues: Random Early Drop

A naive queue has a binary behavior: drop when full, accept otherwise. This penalizes bursty workloads unfairly. Random Early Detection (RED) probabilistically drops packets as the queue fills, smoothing the discontinuity. The same principle applies to link aggregators where front-page placement creates a traffic cliff.

College Admissions: Pell Grant Proxy

Universities using Pell Grant eligibility as a proxy for helping low-income students introduced a discontinuity. Among non-Pell students, the lowest-income were most hurt; among Pell students, the highest-income benefited most. Savvy parents can manipulate income to cross the threshold, similar to tax gaming. Graphs from 2008 vs 2016 show a sharp drop in admissions just above the threshold.

Election Fraud: Round Number Spikes

Russian election results from 2004 onward show spikes at round turnouts (e.g., 95%) and vote shares (e.g., 90%). This indicates fabricated results where fraudsters pick nice numbers. Benford's law would detect such anomalies, but the discontinuity itself is a red flag.

Used Car Auctions: $10k Boundaries

Auction prices show discontinuities at every $10,000 increment—more cars priced at $9,999 than $10,000. This suggests psychological or administrative thresholds affect pricing strategies.

P-Values: The .05 Spike

Psychology papers show a suspicious spike of p-values just below 0.05. This is consistent with p-hacking, selective reporting, or publication bias. Gelman and others argue for abandoning significance thresholds entirely.

Drug Sentencing: Mandatory Minimum Thresholds

After the Fair Sentencing Act raised the 10-year mandatory minimum from 50g to 280g of cocaine, prosecutions spiked at 280g. This shows how thresholds influence charging decisions. A later change in evidentiary standards reduced the spike.

High School Exit Exams: Teachers Fudge for Passing

Polish language exam scores show a spike at 30 (passing) and a deficit just below. Teachers recheck borderline failing tests and award extra points, especially in subjective subjects. Math exams, being objective, show no such discontinuity.

Sports: Relative Age Effect

UEFA Youth League players are disproportionately born in the first months of the year. This is due to age cutoffs for youth teams. Younger players who make it are actually more valuable (more playing time), indicating they are better to overcome the bias.

What Developers Should Do

  • Use smooth transitions: RED for queues, gradual phase-outs for subsidies.
  • Monitor for spikes: histograms of your system's outputs can reveal unintended thresholds.
  • Avoid binary cutoffs: they invite gaming. If unavoidable, audit for discontinuities.
  • Design for robustness: consider how users will exploit sharp edges.

Luu's post is a masterclass in recognizing patterns across domains. The next time you set a threshold—rate limit, timeout, retry count—ask: what behavior will this incentivize?