Vera Data Binary Generating

Binary Outcome --- Full Analysis & Manuscript Generation

Open-source skill. Read reference/specs/output-variation-protocol.md before every generation --- apply all variation layers.

Workflow

Continues from where vera-data-binary-reviewing stopped (PART 0-2 done).

Additional Inputs

Collect if not already provided:

• Target discipline (for reporting conventions)
• Target journal or style (APA 7th, STROBE, etc.)
• Research question / hypothesis
• Subgroup variable (if subgroup analysis desired)

Output Structure

output/
├── methods.md
├── results.md
├── tables/             ← Markdown + CSV per table
├── figures/            ← PNGs, 300 DPI
├── references.bib
├── code.R              ← Style-varied
└── code.py             ← Style-varied

Key References (read before generation)

Reporting Standards

Same as vera-data-binary-reviewing, plus:

• Logistic coefficients: report OR (not raw B) in results text; raw B with SE in supplementary table
• Pseudo-R2: report McFadden and Nagelkerke; say "the model accounted for" --- never "explained"
• AUC: report with 95% CI; note "in-sample" if no cross-validation performed
• Classification: report sensitivity, specificity, and threshold used
• Tree-based with small N: frame as "exploratory"; never claim predictive validity
• Hosmer-Lemeshow: report chi-sq, df, p; non-significant = adequate fit. HL is known to have low power with large N — when N > 1000, supplement with a calibration plot (predicted vs observed, loess-smoothed) and report Brier score.
• Rare events or quasi-separation (minority < 10% OR any model coefficient |β| > 5 with very wide SE): switch to Firth's penalized likelihood (logistf in R; statsmodels with method='bfgs' + bias-reduction, or the firthlogist package in Python). Report Firth-penalized OR with profile-likelihood 95% CI. Flag in the manuscript that standard ML estimates were unstable and Firth was used.

Classification Threshold

Default threshold is 0.5 (balanced misclassification cost). Adjust only when justified:

• Youden's J (maximizes sensitivity + specificity − 1): use when sens and spec are equally important.
• Cost-optimal threshold (minimizes expected loss given cost ratio): use when FP and FN have documented asymmetric costs (e.g., medical screening, fraud detection).
• Always report the threshold used and, when non-default, state the objective (Youden, cost ratio C_FP:C_FN).

Cross-Skill Interface

Method Unit Contract:
├── code_r           → .R script (style-varied)
├── code_python      → .py script (style-varied)
├── methods_md       → methods.md (varied structure)
├── results_md       → results.md (varied phrasing)
├── tables/          → Markdown + CSV
├── figures/         → PNGs 300 DPI (varied layout)
├── references_bib   → .bib with cited references
└── comparison       → cross-method narrative (in results.md)

Invoked directly after vera-data-binary-reviewing or orchestrated by vera-data-application-pipelining.