File size: 5,011 Bytes
849f535 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 | """Post-hoc sensitivity split of B1 by trial duration (Appendix H).
python analysis/07_audio_context_study/03_duration_split.py
The claim being reproduced: splitting B1's raters at the cell's median trial
duration (a criterion independent of the ratings), the faster half yields a
singular variance-components fit with ~zero stimulus variance, while the
slower half's decision threshold matches the audio baseline.
Duration is the trial's totalTaskDurationMs from the Firestore export
(data/audio_context_study/trial_durations_audiocontext_onevoice.csv), one value per
trial; a rater's speed is the median of their trial durations. The split is at
the median of those per-rater medians. Variance components use the same model
as scale_usage.py / the paper's ICC:
Rating ~ presented_first + trial_z + (1|AudioClip) + (1|ParticipantID)
and the threshold is the trial-level mean gap on non-tie trials, as in
threshold_summary.py. Reported for both halves, alongside the audio-baseline
threshold on the same 30 prompts for reference. This is a post-hoc sensitivity
check, not a pre-specified exclusion rule.
"""
from __future__ import annotations
import sys
from pathlib import Path
import numpy as np
import pandas as pd
sys.path.insert(0, str(Path(__file__).resolve().parents[1] / "common"))
from _condition_contrasts import trial_gaps # noqa: E402
from _lme4 import lmer_varcomp # noqa: E402
ROOT = Path(__file__).resolve().parent.parent.parent
VOICE_DATA = ROOT / "data/voice_study"
CONTEXT_DATA = ROOT / "data/audio_context_study"
DERIVED = ROOT / "results/07_audio_context_study"
def durations() -> pd.Series:
"""(ParticipantID, PromptID) -> trial duration in seconds."""
t = pd.read_csv(CONTEXT_DATA / "trial_durations_audiocontext_onevoice.csv")
return (t.set_index(["ParticipantID", "PromptID"])["TaskDurationMs"]
.div(1000.0).rename("seconds"))
def varcomp(g: pd.DataFrame) -> dict:
g = g.dropna(subset=["Rating", "TrialNumber", "AudioClip"]).copy()
g["pf"] = g["PresentedFirst"].astype(int)
sd = g["TrialNumber"].std()
g["trial_z"] = (g["TrialNumber"] - g["TrialNumber"].mean()) / (sd if sd else 1)
return lmer_varcomp(
g, "Rating ~ pf + trial_z + (1|AudioClip) + (1|ParticipantID)")
def threshold(g: pd.DataFrame) -> tuple[float, int]:
x = trial_gaps(g)["gap"].dropna()
return float(x.mean()), len(x)
def main() -> None:
b1 = pd.read_csv(CONTEXT_DATA / "long_audiocontext_onevoice_raw.csv")
b1 = b1[~b1["IsAttentionCheck"].astype(bool)].dropna(subset=["Rating"])
# The raw table keeps attention-check failures (build_long_tables.py keeps
# them for the pass-rate stat); the analyzed cell is the normalized one.
analyzed = set(pd.read_csv(CONTEXT_DATA / "long_audiocontext_onevoice_normalized.csv")
["ParticipantID"])
b1 = b1[b1["ParticipantID"].isin(analyzed)]
b1["Arm"] = "B1"
dur = durations()
per_rater = (b1.drop_duplicates(["ParticipantID", "PromptID"])
.join(dur, on=["ParticipantID", "PromptID"]))
speed = per_rater.groupby("ParticipantID")["seconds"].median().dropna()
b1 = b1[b1["ParticipantID"].isin(speed.index)]
cut = speed.median()
fast = speed[speed <= cut].index
slow = speed[speed > cut].index
print(f"{len(speed)} raters with durations; median of per-rater medians "
f"{cut:.1f}s -> fast {len(fast)}, slow {len(slow)}")
rows = []
for name, ids in (("fast half", fast), ("slow half", slow)):
g = b1[b1["ParticipantID"].isin(ids)]
thr, n = threshold(g)
vc = varcomp(g)
print(f"\n{name}: {g['ParticipantID'].nunique()} raters, "
f"median trial {speed[ids].median():.0f}s")
print(f" threshold {thr:.1f} ({n} non-tie trials)")
print(" variance components: " +
", ".join(f"{k}={v:.2f}" for k, v in vc.items()))
if vc.get("AudioClip", 1) < 1e-6:
print(" -> singular fit: stimulus variance is zero")
rows.append(dict(group=name, raters=g["ParticipantID"].nunique(),
median_trial_s=speed[ids].median(), threshold=thr,
non_tie_trials=n, v_stim=vc.get("AudioClip"),
v_rater=vc.get("ParticipantID"), v_resid=vc.get("Residual")))
# Reference: audio baseline threshold on the same 30 prompts.
import _audio_context_replication as C
prompts = set(pd.read_csv(ROOT / "stimuli/audio_context_study/prompts.csv")["PromptID"])
base = C.load_arms(prompts)
base = base[base["Arm"] == "ORIGINAL af_heart"]
thr, n = threshold(base)
print(f"\naudio baseline (same 30 prompts): threshold {thr:.1f} "
f"({n} non-tie trials)")
rows.append(dict(group="audio baseline", threshold=thr, non_tie_trials=n))
pd.DataFrame(rows).to_csv(DERIVED / "duration_split.csv", index=False)
print("-> results/07_audio_context_study/duration_split.csv")
if __name__ == "__main__":
main()
|