Continuous-response vignette: scouting handoff

Produced by vignettes/continuous_scout_fetch.R + vignettes/continuous_scout_classify.R (scouting only, non-authoritative, no model fitting done). Full data: continuous_scout_candidates.csv (repo root), raw per-item cache in continuous_scout_data/continuous_scout_raw.rds.

55 IRW tables have n_categories >= 12 (the standard site summaries’ cutoff for “cleanly ordinal”). Each was fetched and classified bounded / unbounded / unclear from (a) whether the table’s observed min/max sit at round, interpretable endpoints and are consistently reached across items, and (b) irw_tags()’s item_format field (Slider/continuous = bounded evidence, Constructed Response = unbounded evidence). Where the two signals disagreed, the table was downgraded to unclear rather than resolved automatically — those are marked “CONFLICT” below.

Bounded (24) — beta-IRT (Noel & Dauvier) candidates

  • tears — 11 items, clean 0–100 VAS; 2 items top out at 92/98 (small-n, not a format issue).
  • much_tte_2025_currentmotivation — clean 0–100, “Current Motivation” survey.
  • klippel_irw — 0–100 but only n=32/item and per-item ranges vary a lot; Slider/continuous tag is what resolved this to bounded — verify per-item sample size before trusting.
  • nas_rogoza_2024_study5_nas — 0–100, Narcissistic Antagonism Scale, daily-diary study.
  • nas_rogoza_2024_study5_ngs — companion subscale (Grandiosity), same diary study.
  • nas_rogoza_2024_study5_nvs — companion subscale (Vulnerability), same diary study. These 3 could triangulate as a multi-subscale demo.
  • emoji_scheffler_2024 — 0–100, tagged Slider/continuous; valence/arousal/familiarity/clarity/complexity ratings.
  • opentsstvr_linnig_2025_vas — 0–100, name literally says VAS; VR-environment perception.
  • test_taking_much_2025_cm — 0–100, “Current Motivation” during test-taking.
  • ai_fear_dong_2026_ai — 0–100, tagged Slider/continuous; 1 of 4 related sub-measures.
  • ai_fear_dong_2026_requirement — companion sub-measure, same study.
  • ai_fear_dong_2026_own_fear — companion sub-measure, same study.
  • ai_fear_dong_2026_other_fear — companion sub-measure, same study. Good 4-subscale set from one study for a multi-subscale demo.
  • westhoff2023_stopd — 0–100 clean; no irw_tags() entry (table not found by tags lookup).
  • westhoff2023_pbat — same, no tags available.
  • ehealth_rioux_2025_triplep — 0–100 but some items miss the top (93 unique values); Slider/continuous tag resolved it; 1 of a 3-table eHealth family.
  • ehealth_rioux_2025_beam — companion, same family.
  • ehealth_rioux_2025_abiliti — companion, same family.
  • lsbq_maleki_2025_non_persian_proficiency — 0–10, 41% non-integer values (genuinely fractional), tagged Slider/continuous — strong candidate.
  • eammi_grahe_2018_marriage_identity_allocation — 0–100 clean.
  • gilbert_meta_95 — 0–10, Eysenck Personality Inventory Impulsivity subscale — check whether this is item-level or an already-aggregated subscale score.
  • thomeczek2025_les — 1–20, per-item ranges inconsistent (small sample?); Slider/continuous tag resolved the call — check before trusting.
  • double_marking_steele_2022 — 0–100 clean, observational rating (dissertation grading).
  • mclaughlin_samuel_2025_auditory_session_1 — 1.5–10, 39% non-integer, item ranges inconsistent, Slider/continuous tag resolved it — small table, check before trusting.

Unbounded (18) — Samejima-CRM candidates

  • DEMOS — huge range (0–1121.9), 98% non-integer, no metadata tags — likely a derived/aggregate score; confirm what “item” means here before use.
  • det_naismith_2023 — Constructed Response (writing test); range −3.4 to 6, negative values unusual — check codebook.
  • figure_skating — ISU judges’ scores, 2002 Olympics pairs figure skating, tagged Constructed Response, 95% non-integer — good genuine continuous-score candidate.
  • vollbracht_et_al_2026_ambulatory_assessment — −50 to 100, unusual bipolar-looking range, ambulatory/EMA study — check instrument.
  • chile_2023_social-welfare-survey_yy — construct_name “CASEN Income” — an income variable, not an item-response scale; likely not a real IRT candidate.
  • chile_2023_social-welfare-survey_u — “CASEN Use of Time”, 0–1440 = minutes/day — same caveat, a covariate not a scale.
  • cognitive_load_klimova_2023_know — 1–1000, “Task Knowledge” — unusual scale, check what’s measured.
  • simsalRbim_Mice_LargeValence — animal (mice) behavioral data, no metadata tags — check relevance to a human-response vignette.
  • simsalRbim_Monkey_LargeValence — animal (monkey) data, same caveat.
  • mclaughlin_samuel_2025_auditory_session_2 — 0–100, resp signal alone was unclear; Constructed Response tag (auditory speech transcription accuracy) is what pushed it to unbounded.
  • chile_2023_children-adolescents-survey_cp_c — “Cuidador Principal” (primary caregiver) — part of the CASEN covariate family, same caveat as the two above.
  • uti_newlands_2023_wpai — Work Productivity & Activity Impairment (WPAI), a known health-economics instrument — plausibly a real candidate.
  • florida_twins_par — “My parents Scale”, tagged Likert Scale/selected response despite n_categories=39 — possible item_format tag/data mismatch, worth a look.
  • chile_2023_social-welfare-survey_rr — “CASEN Social Relationships” — CASEN covariate caveat.
  • contreras_valdez_2022_edeq — EDEQ items are normally a 0–6 Likert; this table’s range suggests a subscale/summed score rather than raw items — check codebook.
  • opladen2025_edeq — same EDEQ caveat (range 0–112 looks like a summed score).
  • chile_2023_social-welfare-survey_ss — “CASEN Health” — CASEN covariate caveat.
  • piterova-slovak-science-related-populism — “Party Positions”, 0–98, feels VAS/thermometer-like even though untagged — worth a second look, might actually belong in bounded.

Unclear (13) — needs a closer look before use

  • realpic_souza2021 — “RealPic” picture-norming ratings, 96% non-integer, item ranges inconsistent — real continuous data, worth checking despite the ambiguous label.
  • yu2025 — “Perceptual response variability in size estimation and reproduction” — a genuinely continuous perceptual-estimation task (0.1–200, 38% non-integer); strong Samejima-CRM candidate despite the “unclear” label.
  • estcrm_selfeff — CONFLICT: resp signal says unbounded, item_format says Slider/continuous. Table name references the EstCRM R package (Samejima CRM estimation) — likely a canonical CRM example dataset; prioritize.
  • estcrm_epia — same EstCRM-package naming pattern, consistent per-item range (1–111) — check together with estcrm_selfeff.
  • climatechange_geiger_2025 — 0–100 but item-inconsistent, tagged Mixed.
  • wine_luckett2021 — confirmed genuinely mixed format on manual check: 2 items on a 1–9 scale, 1 item on 0–100 (wine aroma intensity vs. familiarity/pleasantness) — good example of “unclear” being the right call; not usable as a single continuous outcome without splitting by item.
  • deception_professors — 0–100, tagged Mixed.
  • deception_game — 0–100, tagged Likert Scale/selected response despite the wide range — possible tag mismatch.
  • lsbq_maleki_2025_persian_comprehension — CONFLICT: resp signal (unbounded) vs. item_format (Slider/continuous). Companion to lsbq_maleki_2025_non_persian_proficiency, which landed in bounded — check both together.
  • simsalRbim_Mice_LowValence — 0–35, no metadata tags, part of the simsalRbim animal-data family.
  • immer12_immer — 29–58 range, non-zero floor, tagged Likert despite n_categories=29 — odd, check codebook.
  • identity_fusion_gomez_2025 — 0–6 but 27 unique values (genuinely fractional within a narrow range). The round-endpoints heuristic doesn’t recognize 6 as a “nice” bound — likely a false negative; probably belongs in bounded on manual inspection.
  • eammi_grahe_2018_marriage_timing — 10–26 range, non-zero floor, tagged Likert — check codebook (possibly age-related).

Recommendations

  • Bounded half is vignette-ready. The ai_fear_dong_2026_* (4 subscales) and nas_rogoza_2024_study5_* (3 subscales) families are good for demonstrating the model across related subscales from one dataset.
  • Unbounded half is thinner and messier. Prioritize estcrm_selfeff, estcrm_epia, and yu2025 first — most conceptually on-target for Samejima’s CRM (one pair literally named after the R package implementing it, the other a genuinely continuous perceptual-estimation task). figure_skating is a solid secondary real-data example.
  • Treat all chile_2023_social-welfare-survey_* and _cp_c tables as likely non-candidates despite their unbounded classification — construct names (CASEN Income, Use of Time, Health, Social Relationships) indicate these are demographic/covariate columns bundled as “items,” not a coherent single-construct instrument.
  • Check the EDEQ tables’ codebook (contreras_valdez_2022_edeq, opladen2025_edeq) — EDEQ items are normally bounded 0–6 Likert, so these continuous-looking ranges suggest subscale/summed scores rather than raw items; may need re-fetching at the true item level if so.
  • Don’t auto-resolve the 2 signal-conflict tables (estcrm_selfeff, lsbq_maleki_2025_persian_comprehension) — a human should look at the actual resp distributions before deciding.
  • Overall recommendation for the vignette: build the bounded/beta-IRT half directly from real IRW data (strong candidate set above). For the unbounded/Samejima-CRM half, lead with rt (response time) as the primary, explicitly-labeled continuous outcome, and use estcrm_* / yu2025 / figure_skating as secondary, more caveated real-data validation rather than the main demonstration — the honest read is that IRW doesn’t have a deep bench of clean, single-construct unbounded continuous item responses.