Skip to content

USER-RESEARCH.md

Latest commit

 

History

History
424 lines (387 loc) · 28.1 KB

USER-RESEARCH.md

File metadata and controls

424 lines (387 loc) · 28.1 KB
Raw

User Research — Research-Backed Synthetic Persona Panel

Warning

These personas and interviews are SYNTHETIC. They are a structured brainstorming device — composite users modeling real audience segments for nearmiss — not interviews with real people. No real user said any of this. Treat every "quote" as a hypothesis to validate, not evidence of demand. This is consistent with how the project labels its synthetic datasets (see docs/research/ and the planted-hotspot fixtures): a synthetic instrument is useful for pressure-testing and useless as proof.

Last assembled: 2026-06-30.

This panel complements the earlier, deliberately exhaustive 2026-06-20 synthetic user interviews (24 personas; a R1–R70 / E1–E60 backlog). That panel was an internal expert walkthrough with no external evidence base. This one is narrower and research-backed: a smaller, fully-covered stakeholder cast whose frictions and wants are each anchored to the published literature on near-miss reporting, underreporting of vulnerable-road-user injury, exposure normalization, and spatial-hotspot statistics. Where a finding restates a prior backlog item it is tagged [corroborates R#/E#] (independent triangulation — and now with a citation behind it); where it is genuinely new it is tagged [NET-NEW]. The companion RESEARCH-ROADMAP.md turns these into a sequenced, cited backlog.

Method

  • Sampling frame. Every stakeholder who touches the dataset and the analysis (the product — not an app): the people who report and advocate, the people who plan and decide, the people who reuse and research, the people who assure and audit, and the person who operates it. Each persona maps to a real audience the README, data card, and threat model already name.
  • Protocol. For each persona: a goal; a walkthrough of the real surfaces they would touch today (the two-map exposure-vs-counts view, the authoritative sortable table, the Byar/Poisson confidence intervals, the Getis-Ord Gi* + Benjamini-Hochberg hotspots, bias.py, the exposure unknown honesty, the k-anonymity withholding, the bilingual brief, make reproduce, the BikeMaps/OSM fetchers, the Davis/Sacramento real configs, the source-only submit prototype + moderation queue); where they get stuck; what they want next; and the one thing that makes them adopt or walk.
  • Research basis. Each interview's friction is checked against the evidence below, so the panel cannot quietly invent a need the literature contradicts — and high-stakes statistical/epidemiological claims are cross-checked against ≥2 reputable sources. Full citations and the roadmap mapping are in RESEARCH-ROADMAP.md; the anchors most load-bearing for this panel (all accessed 2026-06-30):
  • Synthesis. Frictions → remediations (RR-#); wishes → expansions (RE-#), in RESEARCH-ROADMAP.md. The RR-/RE- namespace is deliberately distinct from the prior panel's R-/E- so the two backlogs never collide; bare R#/E# references point at the 2026-06-20 panel.
  • Effort scale. S ≈ an afternoon · M ≈ a day or two · L ≈ a week+ · XL ≈ multi-month / research.

Persona roster

# Persona Group Primary goal Top friction (research-anchored)
P1 Dana — daily bike commuter, would-be reporter Report & Advocate Log the truck that just buzzed her Near-misses leave no official record, yet the path in is a CLI/JSON to most users
P2 Marisol — Safe-Routes-to-School parent (pedestrian) Report & Advocate Prove the school crossing is dangerous Data is cyclist-centric; her kids walk; no time-of-day lens
P3 Theo — manual-wheelchair pedestrian Report & Advocate Find curb-ramp / blocked-crossing hazards Schema has wheelchair mode but BikeMaps coverage is near-zero for rolling
P4 Priya — safe-streets advocate building a council campaign Report & Advocate Win a specific redesign, survive cross-examination The national live site cannot yet answer a local exposure-normalized near-miss question; screenshots strip caveats
P5 Karim — city traffic engineer / active-transport planner Plan & Decide Defend a project list with defensible evidence Needs the method next to official data, not instead of it
P6 Dr. Okafor — Vision Zero coordinator (self-report skeptic) Plan & Decide Not stake a plan on biased self-report Crowdsourced near-misses ≠ KABCO collisions; wants validation
P7 Lena — data / investigative journalist Reuse & Research Publish a claim she can defend No download/permalink to cite; premise asserted, not sourced
P8 Prof. Halvorsen — transport epidemiologist (dataset reuser) Reuse & Research Cite the method; reuse the rates+CIs CI covers the count, not the denominator; no DOI
P9 Sam — data scientist evaluating the statistics Reuse & Research Decide if the numbers survive scrutiny Poisson assumed; overdispersion check not yet implemented
P10 Marcus — open-data / reproducibility reviewer Assure & Audit Re-run and get byte-identical output requirements.lock not committed; no tagged release
P11 Grace — blind screen-reader user (NVDA), + low-vision lens Assure & Audit Get every finding without the map Structural a11y is designed-for; manual SR pass still pending
P12 "the brigade" — bad-faith reporter / astroturf threat Assure & Audit Manufacture (or bury) a hotspot Publishing the form without edge defenses would invite flooding, doxxing-by-report, and poisoning
P13 Chelsea — owner / maintainer Operate Keep it honest, cheap, and unfundable-proof Honesty scales worse than features; every expansion grows HR3/HR4 surface

13 personas · 5 groups · every stakeholder type in the brief covered.

Interviews

Compact transcripts. Each: Goal · Values today (real, shipped features) · Gets stuck · Wants next · Adopts / walks. Frictions in italics are the ones the literature directly supports.

Group 1 — Report & Advocate

P1 — Dana, daily bike commuter (would-be contributor)

  • Goal. Report the close pass that just happened, from the curb, in seconds.
  • Values today. That the project exists at all: near-misses leave no police report, so an official dataset literally cannot contain them — Nelson et al. call this exactly the gap BikeMaps was built for. The accessible source-only web/submit.html prototype and moderation queue show a privacy-conscious path (no name/email/account by construction), but they are not deployed.
  • Gets stuck. The public site has no intake path; she must use a local/CLI workflow or coordinate with a maintainer — more than the "20 seconds with adrenaline" she has. The local two-map view is framed for cyclists; fine for her, not for a friend on foot.
  • Wants next. A one-tap hazard-type + pin with a true POST endpoint; offline capture; the "your report helped flag B St" acknowledgement.
  • Adopts if reporting is genuinely sub-30s and her precise spot never goes public. Walks if it feels like a 311 queue that does nothing, or if she can't tell her exact location stays private. (corroborates R40–R43)

P2 — Marisol, Safe-Routes-to-School parent (pedestrian framing)

  • Goal. Show the city that the 3 p.m. crossing by the school is dangerous.
  • Values today. The exposure-normalized rate with a confidence interval is exactly the chart that beats "everyone knows that corner is bad" — and the bilingual (EN/ES) brief reaches her neighbors.
  • Gets stuck. The published dataset has no time dimension — per-report timestamps are withheld under HR4, so "dangerous at the school bell" is invisible. The dataset is BikeMaps-sourced and cyclist-centric; pedestrian coverage in a given city may be sparse even though the schema supports it.
  • Wants next. A privacy-safe aggregated time-of-day band; pedestrian framing as first-class; a table filtered to her school's blocks (the name filter shipped — extend to mode).
  • Adopts if she can hand a council member one honest sentence + a footnote. Walks if the tool quietly implies pedestrian coverage it doesn't have. (corroborates E3, R33)

P3 — Theo, manual-wheelchair pedestrian

  • Goal. Find the curb with no ramp, the blocked crossing, the heaved sidewalk.
  • Values today. The report schema already carries a wheelchair mode and surface_hazard / sightline types; the honesty rules mean the project won't pretend to cover him if it doesn't.
  • Gets stuck. His world is nearly absent from the data. BikeMaps is a cycling instrument; rolling and walking exposure aren't measured, so even if his hazards were reported there is no denominator to rate them against (HR1 → exposure unknown). Self-selection compounds it: the contributor pool skews to app-equipped, confident riders, not disabled pedestrians.
  • Wants next. A pedestrian/rolling intake source; walk/roll exposure; an explicit per-city mode-scope label so absence isn't read as safety.
  • Adopts if the dataset states plainly "this city is cycling-only; rolling coverage is sparse." Walks if it lets a city cite "no reports" on a sidewalk as evidence it's fine. (corroborates R33, E12; HR3)

P4 — Priya, safe-streets advocate building a council campaign

  • Goal. Win a redesign on 5th St and survive a hostile traffic engineer.
  • Values today. The two-map "busy ≠ dangerous" view is the argument she's made for years; the planted-fixture proof (the busy decoy seg-03 ranks low on exposure-normalized rate while the genuinely-hot seg-06 lights up) is the slide she wants. make reproduce means no one can wave it away.
  • Gets stuck. At the time of this research the live site was the synthetic Davis demo. The deployed site now uses reviewed nationwide FARS records, but it still cannot represent her city's exposure-normalized near-miss rates without a reviewed local release. And a screenshot of the surface, legend stripped, becomes "the most dangerous street" — the exact misread the threat model (T4) warns about and cannot prevent once republished.
  • Wants next. A council export (PNG/PDF of both maps + ranked table with the caveats baked in) and a per-segment permalink; the literature behind the premise cited so a skeptic can't call it activism with a map.
  • Adopts if the export survives cross-examination. Walks if the only shareable artifact is a caption-less heat map. (corroborates R23, E19; NET-NEW: cite-the-premise)

Group 2 — Plan & Decide

P5 — Karim, city traffic engineer / active-transport planner

  • Goal. Add defensible evidence to a project list without overclaiming.
  • Values today. He respects what most crowdsourced maps skip: exposure normalization, confidence intervals, Gi* with FDR, the documented network spatial-weights, and the exposure unknown honesty. The GeoJSON loads straight into QGIS with embedded metadata.
  • Gets stuck. He won't swap near-misses in for his KABCO/MMUCC collisions — and he's right not to: near-miss is a surrogate measure whose predictive validity for crashes is an open research question, not a settled one. He wants the snapping/dedup thresholds and a sensitivity note exposed.
  • Wants next. Crowdsourced near-miss shown next to official collisions (a tri-view), a documented crosswalk to MMUCC/KABCO, an exportable methodology appendix for a staff report.
  • Adopts if it agrees with his collision data where he has it. Walks if it asks him to treat self-report as ground truth. (corroborates E5, R29; surrogate- safety literature)

P6 — Dr. Okafor, Vision Zero coordinator (skeptical of self-report data)

  • Goal. Build a High Injury Network she can defend, without staking it on a biased volunteer signal.
  • Values today. That the project names its biases instead of hiding them (HR3, bias.py), refuses to publish a rate without a denominator (HR1), and withholds low-count segments (HR4). Vision Zero best practice is explicitly to supplement police data with health and community sources and to map under-representation — nearmiss is built in that spirit.
  • Gets stuck. Self-selection is structural: the contributor pool oversamples confident, app-equipped, often recreational riders, so streets used by under-represented groups are under-reported, and exposure normalization fixes the volume confound, not the who-reports one. She also worries a low per-segment rate on a busy corridor will be read as "safe" — but safety-in- numbers is weak and contested at the micro (junction/segment) level even where it holds city-wide.
  • Wants next. A validation against official collisions; an equity overlay that surfaces under-reporting (handled with consent, not stigma); the macro-vs- micro caveat written into the brief.
  • Adopts if the dataset is positioned as a complement that admits its bias. Walks if it's pitched as a replacement for collision records. (corroborates R48, E5, E7; NET-NEW: micro-SiN caveat)

Group 3 — Reuse & Research

P7 — Lena, data / investigative journalist

  • Goal. Publish a checkable claim about where it's dangerous to ride.
  • Values today. make reproduce is "a dream" — a number she can regenerate. The data card and limitations page pre-state the caveats so she won't get burned.
  • Gets stuck. The site's premise — "vulnerable users absorb the risk and produce almost none of the data" — is asserted but uncited; her editor will ask for a source. There's a download affordance and per-segment deep links (shipped). At research time the live data was still the demo; the national site now provides versioned real FARS evidence, but no reviewed local near-miss release or DOI yet answers her original reporting need.
  • Wants next. The underreporting and safety-in-numbers literature cited in the data card; an embeddable map; a machine-readable version feed.
  • Adopts if she can link the exact row, re-run the number, and footnote the premise. Walks if the central claim rests on the author's word. (corroborates R22/R32; NET-NEW: cite-the-premise)

P8 — Prof. Halvorsen, transport epidemiologist (would reuse the dataset)

  • Goal. Reuse the per-segment rates + CIs in a peer-reviewed health-equity analysis, and cite the method.
  • Values today. The rate-with-interval framing is her language; the choice of Byar's Poisson interval (well-behaved to count 0, never negative) over Wald is the right call; Benjamini-Hochberg FDR across segments is exactly what she'd demand; the planted-fixture coverage simulations are reassuring.
  • Gets stuck. The interval covers the numerator, not the denominator — the CI is Poisson-on-counts with exposure treated as fixed, and "your numerator has error bars; your denominator pretends it doesn't." Report counts also cluster (one viral post, one active group), so the Poisson assumption likely understates variance — and the methodology itself flags that the overdispersion check is not yet implemented.
  • Wants next. Exposure-uncertainty propagation (or a louder scope statement); the overdispersion/quasi-Poisson check landed; empirical-Bayes smoothing for small areas; a versioned DOI and a documented power analysis ("how many reports until a block is rankable").
  • Adopts if the uncertainty is honest end-to-end and citable. Walks if the CI looks rigorous but silently fixes the shakiest input. (corroborates R28/R34, E9/E11; NET-NEW: overdispersion)

P9 — Sam, data scientist evaluating the statistics

  • Goal. Decide, adversarially, whether the numbers hold up.
  • Values today. Gi* run on the rate, not the raw count (the crucial choice that stops it re-telling the heat-map lie with a p-value attached); network-based spatial weights, not straight-line; raw and FDR-adjusted significance reported; the banned-Wald discipline; the null-fixture test that a method finding hotspots in pure noise fails.
  • Gets stuck. Three things the project already half-concedes: (1) MAUP — the block is an arbitrary unit and a hotspot at one segmentation can dissolve at another, and there's no rank-stability check yet; (2) Gi* significance rests on the normal approximation, with conditional-permutation inference only noted as future work; (3) BH-FDR assumes a structure that spatial dependence strains — Caldas de Castro & Singer's spatially-aware FDR is the relevant refinement.
  • Wants next. A re-segmentation sensitivity report; a permutation Gi* option; a note on (or move to) spatial FDR; overdispersion handling.
  • Adopts if the sensitivity analyses are published, not promised. Walks if significance is asserted on assumptions the data violates. (corroborates R28; NET-NEW: MAUP sensitivity, permutation Gi*, spatial FDR)

Group 4 — Assure & Audit

P10 — Marcus, open-data / reproducibility reviewer

  • Goal. Independently re-run the pipeline and get byte-identical output.
  • Values today. make reproduce asserting a clean git diff on data/published/; content-hashed artifacts + metadata sidecar; the committed planted-hotspot fixtures with known answers; ADRs and CITATION.cff; the read-only server that refuses any path under data/raw/.
  • Gets stuck. The README admits requirements.lock is generated but not committed yet, so a from-scratch install isn't pinned/hashed for him; there's no tagged release or DOI to pin a citation to; "reproducible" is true on the maintainer's machine but not yet push-button for an outsider.
  • Wants next. The committed hashed lock; a tagged release + Zenodo DOI; a documented "paste this to reproduce" path.
  • Adopts if a clean clone reproduces the published bytes. Walks if reproduction needs the maintainer in the room. (corroborates R35; NET-NEW: commit the lock)

P11 — Grace, blind screen-reader user (NVDA) — also the low-vision lens

  • Goal. Reach every finding the map shows, without the map.
  • Values today. The architecture is right: the sortable data table is authoritative, not a second-class caption; sort buttons announce through an aria-live region; the segment-name column is sticky for 200% zoom; significance is conveyed in text and pattern, never color alone; a generated prose hotspot-geography summary shipped (R8).
  • Gets stuck. It's designed-for but not yet measured: the README and ACR are honest that the manual NVDA/VoiceOver pass is still pending and some VPAT rows are "Partially Supports." At 200% zoom the two side-by-side maps get tiny (the single-map toggle helps). She can't yet confirm streamed/announced behavior by lived use.
  • Wants next. The real NVDA + VoiceOver + zoom pass run and its results committed; "Partially Supports" predictions converted to measured PASS/FAIL.
  • Adopts if the conformance claim is backed by a real assistive-tech session. Walks if she's handed a colored blob with a promise. (corroborates R15)

P12 — "the brigade", bad-faith reporter / astroturf threat (red-team lens)

  • Goal. Manufacture a hotspot to push a project (or bury a real one to defeat one), or craft a report to expose where a specific person rides.
  • Values today (against them). The published artifact is genuinely hard to abuse: exposure normalization blunts volume floods (you must beat the denominator, not just add counts); intervals + Gi* show an injected burst as uncertain, not a confident top rank; dedupe collapses near-identical submissions; k-anonymity withholds low-count segments and no per-report coordinate/timestamp is ever published; the moderation queue means no public submission reaches the dataset until a human approves it, with identifier-leak and near-duplicate flagging.
  • Gets stuck (the gaps the project itself names). The network-edge defenses — rate limiting, proof-of-work, per-source influence caps, burst/outlier detection, trust tiers — are designed but not yet built, so an open public endpoint is not yet safe to expose. A patient, distributed, low-and-slow campaign of plausible unique reports still passes (the threat model concedes this).
  • Wants next (what must exist before "open"). The full B2–B7 abuse stack shipped and fixture-tested before the form is opened beyond a closed/invite pilot.
  • Defeated if intake stays gated until the toolkit lands. Wins if the form opens publicly before the defenses do. (corroborates R44, E15; HR4)

Group 5 — Operate

P13 — Chelsea, owner / maintainer

  • Goal. Keep nearmiss statistically honest, accessible, cheap, and survivable without a grant — and resist the gravity toward shiny features that erode the hard rules.
  • Values today. Config-over-code (config/*.toml), make verify as one merge gate (lint, types, tests, a11y, security), the planted-hotspot fixtures, the scale-to-zero / static-friendly footprint, the audit-as-artifact discipline.
  • Gets stuck. Honesty scales worse than features. Every expansion the cast wants — public intake, equity overlays, an API, before/after evaluation — multiplies the HR3 (bias) and HR4 (privacy) surface, and the anti-features list (no per-reporter dashboards, no plate-reporting, no "safest route" product, no selling risk to insurers) is part of the roadmap, not separate from it. Single- maintainer means no second reviewer on every change.
  • Wants next. To ship the cheap, research-grounded correctness and honesty items first (they protect the franchise) and to keep the expensive, privacy- expanding ones gated behind proven defenses.
  • Adopts the discipline of "cite the premise, fix the stats gap, then build." Walks from any feature that asks the project to overclaim. (corroborates the prior panel's cross-cutting theme 6)

Cross-cutting themes

  1. The premise is true but uncited — fix that first, cheaply. Every "reuse" and "decide" persona (P5, P6, P7, P8) independently wants the project's founding claim — that VRU near-misses and even injuries are massively under-recorded in official data — sourced. The literature is unambiguous and convergent (police reporting of cyclist injury ≈ 7–46%; an international average near 10%), yet the data card asserts it without a single citation. This is the highest-leverage, lowest-cost change: it converts the thesis from advocacy to evidence. [NET-NEW]
  2. Exposure normalization is the moat — and the literature warns it doesn't fix the bias people will assume it fixes. Normalizing by exposure removes the volume confound (safety-in-numbers, Jacobsen/Elvik). It does not remove self-selection (who reports) and is weak/contested at the micro level, so a low per-segment rate on a busy street is not proof of safety. P6 and P9 both hit this; the brief should say it. [NET-NEW caveat on a [corroborated] feature]
  3. One real statistical-correctness gap, named by the project itself. The methodology concedes the overdispersion check (quasi-Poisson / negative- binomial) is not yet implemented, so intervals on clustered report counts may be too narrow. The epidemiologist (P8) and data scientist (P9) both find it. The prior panel has no item for it. [NET-NEW]
  4. "Designed-for" vs "measured" recurs in two places: accessibility and reproducibility. Grace (P11) and Marcus (P10) tell the same story from opposite ends — the architecture is excellent (table-authoritative a11y; make reproduce) but the proof is pending (the manual NVDA/VoiceOver pass; the committed hashed lock + DOI). Closing both is cheap and converts assertions to evidence. [corroborates R15, R35]
  5. Honesty about coverage is a feature, not a disclaimer. Theo (P3), Marisol (P2), and Dr. Okafor (P6) all need the dataset to refuse to imply coverage it lacks (rolling/pedestrian modes; time-of-day; who's missing). A per-city mode-scope label and a literature-grounded representativeness panel are what keep "no reports here" from being misread as "safe." [corroborates R33, R48]
  6. The write path triples the threat model — keep it gated. Three reporters (P1, P2, P3) want a door in; the brigade (P12) and the maintainer (P13) both say that door must stay invite-only until the network-edge abuse stack ships. This is a privacy project, not a form. [corroborates R44, E15]

Honest limits of this exercise

This panel is synthetic. Role-playing a research-grounded cast surfaces gaps and pressure-tests the framing, but it cannot tell you which needs are real, how many of each stakeholder exist, or what they would actually do. It over-represents the author's mental model and the literature's emphases, and it will miss what only real users surprise you with — and the access findings (P11) are structural predictions, not a substitute for the real NVDA/VoiceOver session. Do not prioritize off this alone. Its job is to (a) attach external evidence to the intuitions the 2026-06-20 panel already captured, and (b) design cheaper, sharper questions for real interviews with ≥1 of each role — especially a real Vision Zero coordinator (P6), a disabled pedestrian (P3), and a transport epidemiologist (P8), whose needs most shape the statistics.

→ The triaged, cited, sequenced backlog is in RESEARCH-ROADMAP.md.