| """Table 16: how raters used the rating scale, by condition. |
| |
| python analysis/08_cross_study/05_scale_usage.py |
| |
| The table answers one question: is B1's low ICC caused by raters |
| DISAGREEING with each other, or by raters not DIFFERENTIATING the stimuli at |
| all? Those look identical in an ICC -- both shrink the stimulus share of |
| variance -- but they mean opposite things, so the columns here separate them. |
| |
| Mean rating leniency: where on the 0-100 scale the condition sits |
| Per-rater SD spread: how much of the scale a typical rater used |
| Raters with SD < 10 how many used almost none of it |
| Tie rate how often no winner was declared |
| Variance components the ICC's own numerator and denominator, split out |
| |
| THE BASIS PROBLEM, and why two of these are computed differently. |
| |
| The four conditions do not have the same data available. B1 and B2 raters each |
| did 15 trials on the 30 audio-context prompts -- 30 ratings, their whole session. |
| The original text and audio raters each did 20 trials drawn from 100 prompts, |
| so restricting them to the same 30 prompts leaves roughly 6 prompts, i.e. 12 |
| ratings each. |
| |
| A standard deviation estimated from 12 numbers is far noisier than one from 30. |
| Ordinary raters cross an "SD < 10" line by chance. Computed on the ctx-30 |
| subset the counts are 6/46 text and 7/46 audio; computed on each rater's full |
| session they are 2/47 and 1/46. The subset version is measurement noise, not |
| behaviour, so the RATER-LEVEL columns use full sessions. |
| |
| The VARIANCE COMPONENTS cannot do that. They are the decomposition behind the |
| ICCs in Table 13, and an ICC is a property of a specific stimulus set, so all |
| four conditions must be fit on the shared 30 audio-context prompts or the numbers |
| are not comparable to each other or to Table 13. |
| |
| So the table mixes bases deliberately, and says so in its caption. Note this is |
| a real trade-off rather than a bug being papered over: matching on stimuli costs |
| precision, matching on precision costs stimulus comparability. Mean rating and |
| tie rate also shift between bases, but for a different reason -- the 30 |
| The audio-context prompts were selected for conversation history and capped duration, |
| so they genuinely tie more often than the corpus average. The conclusion is |
| unchanged either way: B1's share of non-differentiating raters is 45% against |
| 9-15% (subset) or 2-9% (full session). |
| """ |
|
|
| from __future__ import annotations |
|
|
| import sys |
| from pathlib import Path |
|
|
| import numpy as np |
| import pandas as pd |
|
|
| ROOT = Path(__file__).resolve().parent.parent.parent |
| sys.path.insert(0, str(Path(__file__).resolve().parents[1] / "common")) |
| from _lme4 import lmer_varcomp |
|
|
| SD_DIR = ROOT / "data/main_study" |
| VOICE_DATA = ROOT / "data/voice_study" |
| CONTEXT_DATA = ROOT / "data/audio_context_study" |
| DERIVED = ROOT / "results/08_cross_study" |
|
|
| SOURCES = [ |
| ("Text (reference)", SD_DIR / "long_text_table_normalized.csv"), |
| ("Audio (baseline)", SD_DIR / "long_audio_table_normalized.csv"), |
| ("Ablation B1", CONTEXT_DATA / "long_audiocontext_onevoice_normalized.csv"), |
| ("Ablation B2", CONTEXT_DATA / "long_audiocontext_normalized.csv"), |
| ] |
|
|
|
|
| def load(path: Path) -> pd.DataFrame: |
| d = pd.read_csv(path) |
| if "IsAttentionCheck" in d.columns: |
| d = d[~d["IsAttentionCheck"].astype(bool)] |
| return d.dropna(subset=["Rating"]) |
|
|
|
|
| def rater_columns(d: pd.DataFrame) -> dict: |
| """Leniency and spread, one rater at a time then averaged. |
| |
| Per-rater SD is the MEAN of each rater's own SD, not the SD of all ratings |
| pooled. Pooling would fold between-rater differences in leniency into the |
| number, which is the opposite of what the column is for -- it is meant to |
| say how much scale an individual used. |
| """ |
| per_rater_sd = d.groupby("ParticipantID")["Rating"].std() |
| trials = d.drop_duplicates(["ParticipantID", "PromptID"]) |
| return dict( |
| raters=len(per_rater_sd), |
| ratings_per_rater=d.groupby("ParticipantID").size().median(), |
| mean_rating=d["Rating"].mean(), |
| per_rater_sd=per_rater_sd.mean(), |
| n_low=(per_rater_sd < 10).sum(), |
| pct_low=(per_rater_sd < 10).mean() * 100, |
| tie_rate=(trials["overallPreference"] == "Tie").mean(), |
| ) |
|
|
|
|
| def variance_components(d: pd.DataFrame) -> dict: |
| """The ICC decomposition from Table 13's model, on raw ratings. |
| |
| Rating ~ presented_first + trial_z + (1|AudioClip) + (1|ParticipantID) |
| |
| Same specification as the paper's ICC (02_main_study/02_reliability_icc_alpha.py calc_icc), so the pieces |
| add up to the ICCs reported in Table 13. |
| """ |
| g = d.dropna(subset=["Rating", "TrialNumber", "AudioClip"]).copy() |
| g["pf"] = g["PresentedFirst"].astype(int) |
| g["trial_z"] = (g["TrialNumber"] - g["TrialNumber"].mean()) / g["TrialNumber"].std() |
| rv = lmer_varcomp(g, "Rating ~ pf + trial_z + (1|AudioClip) + (1|ParticipantID)") |
| |
| |
| |
| v = {k: float(rv.get(k, np.nan)) |
| for k in ("AudioClip", "ParticipantID", "Residual")} |
| k = g.groupby("AudioClip").size().mean() |
| tot = v["AudioClip"] + v["ParticipantID"] + v["Residual"] |
| return dict(v_stim=v["AudioClip"], v_rater=v["ParticipantID"], |
| v_resid=v["Residual"], k=k, |
| icc1=v["AudioClip"] / tot, |
| icck=v["AudioClip"] / (v["AudioClip"] + (tot - v["AudioClip"]) / k)) |
|
|
|
|
| def main() -> None: |
| ctx = set(pd.read_csv(ROOT / "stimuli/audio_context_study/prompts.csv")["PromptID"]) |
|
|
| print("RATER-LEVEL COLUMNS — each rater's FULL session") |
| print("(so every rater's SD rests on a comparable number of ratings)\n") |
| hdr = f"{'condition':20}{'raters':>7}{'ratings ea':>11}{'mean':>7}{'perRaterSD':>12}{'SD<10':>10}{'tie':>7}" |
| print(hdr + "\n" + "-" * len(hdr)) |
| full = {} |
| usage_rows = [] |
| for lbl, path in SOURCES: |
| r = rater_columns(load(path)) |
| full[lbl] = r |
| usage_rows.append(dict(condition=lbl, scope="full session", **r)) |
| print(f"{lbl:20}{r['raters']:7}{r['ratings_per_rater']:11.0f}{r['mean_rating']:7.1f}" |
| f"{r['per_rater_sd']:12.1f}{r['n_low']:5}/{r['raters']:<4}{r['pct_low']:3.0f}%{r['tie_rate']:7.2f}") |
|
|
| print("\n\nSAME COLUMNS on the ctx-30 subset — shown only to document the artifact") |
| print("(reference raters drop to ~12 ratings each, so low-SD counts inflate)\n") |
| print(hdr + "\n" + "-" * len(hdr)) |
| for lbl, path in SOURCES: |
| d = load(path) |
| d = d[d["PromptID"].isin(ctx)] |
| r = rater_columns(d) |
| usage_rows.append(dict(condition=lbl, scope="ctx-30 subset", **r)) |
| print(f"{lbl:20}{r['raters']:7}{r['ratings_per_rater']:11.0f}{r['mean_rating']:7.1f}" |
| f"{r['per_rater_sd']:12.1f}{r['n_low']:5}/{r['raters']:<4}{r['pct_low']:3.0f}%{r['tie_rate']:7.2f}") |
|
|
| print("\n\nVARIANCE COMPONENTS — all four fit on the shared 30 audio-context prompts") |
| print("(must match Table 13's ICCs, which are a property of that stimulus set)\n") |
| hdr2 = f"{'condition':20}{'v_stim':>9}{'v_rater':>10}{'v_resid':>10}{'k':>7}{'ICC(2,1)':>10}{'ICC(2,k)':>10}" |
| print(hdr2 + "\n" + "-" * len(hdr2)) |
| vc_rows = [] |
| for lbl, path in SOURCES: |
| d = load(path) |
| d = d[d["PromptID"].isin(ctx)] |
| v = variance_components(d) |
| print(f"{lbl:20}{v['v_stim']:9.1f}{v['v_rater']:10.1f}{v['v_resid']:10.1f}" |
| f"{v['k']:7.1f}{v['icc1']:10.3f}{v['icck']:10.3f}") |
| vc_rows.append(dict(condition=lbl, **v)) |
|
|
| print("\nICC(2,1) = v_stim / (v_stim + v_rater + v_resid)") |
| print("ICC(2,k) = v_stim / (v_stim + (v_rater + v_resid) / k), k = mean raters per clip") |
| pd.DataFrame(usage_rows).to_csv(DERIVED / "scale_usage.csv", index=False) |
| pd.DataFrame(vc_rows).to_csv(DERIVED / "variance_components.csv", index=False) |
| print(" -> results/08_cross_study/scale_usage.csv, variance_components.csv") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
