mssense-eval-benchmark / docs /statistical_power_analysis.md
Novylab's picture
Upload 13 files
b6a6d0d verified
|
Raw
History Blame Contribute Delete
6.21 kB

Statistical Power Analysis — mssense-eval-benchmark v1.1

V5.15.h.3 artefact — JOT submission readiness.

A priori power analysis for the four research questions declared in the manuscript (Section 7). Demonstrates that mssense-eval-benchmark v1.1 (N = 1865) carries enough samples to detect the effect sizes that the empirical follow-up will need to claim, at conventional α = 0.05 and power = 0.80.

Method

Power was computed analytically with classical formulas:

  • Two-proportion comparison (RQ1, RQ2, RQ4):

    n per arm ≈ ((zα/2 + zβ)² · (p₁(1−p₁) + p₂(1−p₂))) / (p₁ − p₂)²

    with zα/2 = 1.96 (two-sided α = 0.05) and zβ = 0.84 (power = 0.80).

  • Per-stratum (RQ3) comparisons follow the same formula, applied inside each (modality × channel) cell.

Equivalent computations can be reproduced with statsmodels.stats.power or G*Power.

Sample sizes available in mssense v1.1

Stratum N v1.1
Whole corpus 1865
Seeds only (filter augmentation_type == None) 1772
Web 569
Desktop 285
Other (aggregated) 1011
Text modality 1237
Demo modality 225
Capture modality 188
Audio modality 125
Easy 595
Medium 817
Hard 453
Per expected_decision label INVALID 857, VALID 450, UNDERSTOOD 230, QUESTION 218, THINKING 60, DONE 50

RQ1 — Dynamic skill retrieval reduces closed-vocabulary violations

Comparison. Two-proportion test of vocabulary_violation_rate between dynamic-retrieval and prompt-only baselines on the same held-out set.

Pilot estimate. Prompt-only RPA generation literature (FlowMind, SmartFlow) reports vocabulary-violation rates between 25 % and 45 %. Conservative baseline: p₁ = 0.30.

Δ to detect (absolute) Required n per arm mssense v1.1 has Verdict
15 percentage points (0.30 → 0.15) 138 1865 ✅ ≫
10 percentage points (0.30 → 0.20) 291 1865 ✅ ≫
5 percentage points (0.30 → 0.25) 1075 1865
3 percentage points (0.30 → 0.27) 2800 1865

Conclusion. Detection of ≥ 5 % absolute reduction is supported on the full corpus. A 3-point reduction would require a v2.0 expansion.

RQ2 — Full pipeline > each component alone

Comparison. Two-proportion test of executable_trace_rate for the full pipeline vs each of four ablations.

Pilot estimate. p_full ≈ 0.70, p_baseline ≈ 0.50.

Δ to detect Required n per arm mssense v1.1 has Verdict
20 percentage points (0.50 → 0.70) 97 1865 ✅ ≫
10 percentage points (0.50 → 0.60) 388 1865 ✅ ≫
5 percentage points (0.50 → 0.55) 1564 1865 ✅ marginal
3 percentage points (0.50 → 0.53) 4350 1865

Conclusion. Pairwise comparison supports Δ ≥ 10 % comfortably and Δ ≥ 5 % with a tight margin.

RQ3 — Which modalities × channels produce highest residual errors?

Comparison. Per-cell estimation in the modality × channel grid; pairwise comparisons across cells.

Per-cell N (approximate).

Modality \ Channel web desktop other (aggregated) Total
text high high high 1237
demo low medium medium 225
capture medium low low 188
audio low low medium 125
mixed low low low 90

Detailed sub-cell counts in reports/evaluation_balance_report_v1_1.md.

Effect-size targets. ±10 % half-width 95 % CI requires n ≈ 96 at p̂ ≈ 0.50; ±15 % requires n ≈ 43.

Conclusion. Cells with n ≥ 96 (text × web, text × desktop, text × other, demo × any, capture × any, audio × other) support ±10 % half-width CIs. Cells with n < 96 (audio × web, audio × desktop, mixed × any) support ±15 % half-width CIs only and should be reported with the larger CI explicitly.

RQ4 — Targeted clarification reduces silent workflow errors

Comparison. Two-proportion test of silent_workflow_error_rate on the subset where expected_decision ∈ {QUESTION, INVALID} (n = 1075).

Δ to detect Required n per arm Available n Verdict
15 percentage points 138 1075 ✅ ≫
10 percentage points 291 1075
5 percentage points 1075 1075 marginal
3 percentage points 2800 1075

Conclusion. Comfortably powered for Δ ≥ 10 %.

Per-label confidence intervals after V5.15.h augmentation

Wilson 95 % CI half-widths for proportion estimates per expected_decision:

Label n Half-width at p̂ = 0.50 Half-width at p̂ = 0.80
INVALID 857 ±3.3 % ±2.7 %
VALID 450 ±4.6 % ±3.7 %
UNDERSTOOD 230 ±6.4 % ±5.2 %
QUESTION 218 ±6.6 % ±5.4 %
THINKING 60 ±12.4 % ±10.5 %
DONE 50 ±13.5 % ±11.6 %

THINKING and DONE remain the weakest cells; v1.2 should target n ≥ 200 per label to reduce CI half-width below ±7 %.

Threats and limits

  • Methods may produce correlated outputs (in-context cache, session memory). The empirical version should report bootstrap CIs in addition to the analytical figures above.
  • Pilot estimates (p₁ ≈ 0.30, p_full ≈ 0.70) are working hypotheses; empirical follow-up should recompute power post-hoc.
  • Augmented THINKING/DONE samples share their template with their seeds. A benchmark consumer restricting to seed-only (augmentation_type == None) reverts to n = 12 (THINKING) and n = 5 (DONE) for those two labels and should report only descriptive statistics on them.

Bottom line

mssense-eval-benchmark v1.1 supports detection of:

  • 5–10 % absolute differences on the full corpus for RQ1 / RQ2 / RQ4
  • per-cell estimates with ±10 % to ±15 % 95 % CI for RQ3
  • non-degenerate per-label CIs for all six expected_decision labels after the V5.15.h augmentation

These figures exceed the manuscript's Section 8.1 target of 1000 samples and are consistent with the JOT-published precedent of dataset papers at smaller scale (XCorpus 76 programs, JOT 2017).