Stata Coef To DOCX

/stata-coef-to-docx — Stata → Coefficients → Word Table (no manual copy-paste)

Runs Stata models, extracts every coefficient automatically from the log, and generates a compact Word regression table. The key insight: manually copying numbers from Stata log → python dict is the #1 source of errors (wrong values, stale numbers). This skill automates the full chain.

Why this exists

In the BG Divestment project, TABLE A3 (first-stage IV) was submitted with HARDCODED WRONG VALUES (β=0.164 when actual was β=0.076) because the numbers were typed by hand from a different version. This skill prevents that by never having humans in the copy-paste loop.

When to use

• Building or updating any regression table (main results, robustness, IV)
• After running new Stata models and needing the Word table updated
• When you change model specification and need to regenerate the table

Workflow (4 steps, fully automated)

Step 1: Write the Stata do-file with printcoef extraction

* models_extract.do
log using "LOGPATH.log", replace text
use "CHECKPOINT.dta", clear

capture program drop printcoef
program printcoef
    args vname label
    capture local b = _b[`vname']
    if _rc != 0 { di "COEF|`label'|.|.|."; exit }
    local se = _se[`vname']
    local z  = `b' / `se'
    local p  = 2 * (1 - normal(abs(`z')))
    di "COEF|`label'|`b'|`se'|`p'"
end

* Run each model, print header, extract all coefficients
di "===MODEL1==="
YOUR_REGRESSION_COMMAND, vce(robust)
di "N=" e(N) " LL=" e(ll)
printcoef VAR1 "Variable 1 label"
printcoef VAR2 "Variable 2 label"
* [etc]

di "===DONE==="
log close

Step 2: Run Stata and parse the log

/Applications/StataNow/StataMP.app/Contents/MacOS/stata-mp -b do models_extract.do
grep "COEF\|===\|N=" LOGPATH.log

Step 3: Auto-generate the DATA dict (Python)

Parse log lines of the form COEF|label|b|se|p into a python dict:

import re, subprocess

def parse_coef_log(logfile):
    """Parse printcoef output → {label: (b, se, p)} per model"""
    models = {}
    current = None
    with open(logfile) as f:
        for line in f:
            m = re.match(r'===(\w+)===', line)
            if m:
                current = m.group(1)
                models[current] = {}
            c = re.match(r'COEF\|(.+?)\|([.\d-]+)\|([.\d-]+)\|([.\de-]+)', line)
            if c and current:
                label, b, se, p = c.groups()
                try:
                    models[current][label] = tuple(None if value == "." else float(value) for value in (b, se, p))
                except ValueError as error:
                    raise ValueError(f"Invalid coefficient record for {label}; retain the source log") from error
    return models

Step 4: Build Word table with python-docx

Use academic-docx-table and its format_3 helper with the full-precision extracted data. Apply its report presentation rules to titles, headings, body, tables, notes, headers/footers, and page numbers. Reformatting existing estimates does not authorize rerunning models or changing the project's data or significance-star thresholds.

The table style rules:

• Use Times New Roman 12 pt consistently throughout the report. Each variable has one row; each available coefficient cell has 3 paragraphs: β plus existing stars, [p], (SE), all upright 12 pt. Missing parts have empty paragraphs; a wholly absent entry is one empty paragraph. Never shrink p-values, SEs, notes, or page numbers.
• Use single line spacing and zero space before/after within each cell, with enough row height for 12 pt. Fit wide/long tables by wrapping, landscape, split panels, or continuation pages with repeated headers, without shrinking text.
• Display continuous statistics with three decimals from the unrounded source using Decimal(str(x)).quantize(Decimal("0.001"), rounding=ROUND_HALF_UP); normalize negative zero. For example: 0.6993 → 0.699, 0.2119 → 0.212, 0.1525 → 0.153. Keep observation counts, years, and identifiers as integers; preserve full precision in data/audit exports.
• Report p-values numerically with a leading zero: 0.0006 → [0.001], 0.0004 → [0.000]; no inequality signs in reported values. Displayed 0.000 is rounded, never an exact zero. Assign stars from unrounded p using the existing project convention, not from the rounded display; write legend thresholds using “p below …” while preserving their values. Do not invent a precise p-value from a threshold-only source.
• Leave missing, unavailable, not-applicable, and not-in-model cells genuinely empty. Map source missing values to None/NaN for display through format_3; do not print dash fillers, N/A, NA, dots, invented zeros, whitespace padding, or empty brackets/parentheses. Preserve actual zeros as 0.000 and real negative signs. Keep source values and missingness reasons in the internal audit; retain diagnostics for failed estimates or exports rather than presenting blanks as successful results.
• Keep every table cell and table note upright (roman), including p-values, coefficients, SEs, headers, and labels; bold is allowed. Non-table inline-statistic italics remain unchanged.
• Do not add process notes to outward-facing reports or tables: audit/review success, rerun status, version locks/source packages, generation steps, and rounding procedures belong in internal records. Retain only concise notes needed to interpret the statistics, such as estimator, outcome, SE treatment, star convention, and sample definition.
• Borders: thick top + thin under header + thick bottom only (no internal lines)
• Left-aligned table + left-aligned titles (for appendix)
• Merged rows for moderator variables: in the model that tests Hk, show the moderator coefficient; in other models, show the corresponding control variable coefficient in the same row
• Append the verified current manuscript hypothesis identifier once after the corresponding focal term label, such as Focal predictor (H1) when that mapping is verified. A moderation hypothesis belongs on the interaction row, not on the lower-order moderator/control row. Use the manuscript and project instructions (such as AGENTS.md) to establish the mapping; never infer H numbers from model-column order, significance, or generic examples, and never invent, reorder, or silently remap hypotheses.

Important: Merged moderator rows

When BG size is both a moderator (Model 2) and a control (Models 1,3,4), use a SINGLE row:

# List-type key = merged row: entry[j] is the DATA key for column j
ROWS = [
    ("BG size", ['log_bg_aff', 'c_size', 'log_bg_aff', 'log_bg_aff'], False),
    # M1 shows log_bg_aff, M2 shows c_size, M3 shows log_bg_aff, M4 shows log_bg_aff
]

Never show the same construct twice (once as moderator, once as control) — it confuses reviewers.

Stata executable

/Applications/StataNow/StataMP.app/Contents/MacOS/stata-mp

Checkpoint rule

Before running: check if analysis_manuscript_models.dta checkpoint exists. If it does, use checkpoint, clear and skip data construction (saves 30-60 min). Only rebuild from scratch if variable definitions changed or new data was merged.