| name | latex-tables |
| description | Generates publication-ready LaTeX tables produced by code (Stata `esttab`/`outreg2`, R `modelsummary`/`fixest::etable`, Python `stargazer`/`pyfixest.etable`) and writes them directly to the paper's `tabs/` folder. Defaults to DIME's "full replicability" standard (no copy-paste from any console), `booktabs` styling, three-line tables, threeparttable notes, sensible significance stars, and `\Cref{}`-friendly labels. Aligned with DIME Analytics' [Exporting Analysis](https://dimewiki.worldbank.org/Exporting_Analysis) and [Submit Table checklist](https://dimewiki.worldbank.org/Checklist:_Submit_Table). Use when the user asks for regression tables, summary statistics, balance tables, descriptive statistics, mean-comparison tables, multi-spec tables, or any code-to-LaTeX-table workflow.
|
| workflow_stage | writing |
| compatibility | ["claude-code","cursor","codex","gemini-cli"] |
| author | JonasWeinert |
| version | 2.0.0 |
| tags | ["latex","tables","regression","booktabs","esttab","modelsummary","stargazer","dime","reproducibility"] |
LaTeX Tables
Generate publication-ready LaTeX tables that come out of code (esttab, modelsummary, pyfixest.etable, stargazer) and \input{} cleanly into the paper. The default style follows DIME's "full replicability" tier: no manual editing between code and PDF.
Operating Principles
- Tables are produced, never typed. DIME identifies four levels of replicability: full, good, basic, and none. "None" (copy-paste from a console window) is unacceptable for any output that leaves the analyst. Default to full: code writes a
.tex file that the paper \input{}s directly.
- One source of truth per table. Every numeric cell, star, sample size, and footnote comes from the same script that ran the regression. Never edit the
.tex file by hand — re-run the script and rebuild the paper.
booktabs styling. Three horizontal rules (\toprule, \midrule, \bottomrule), no vertical lines, consistent decimal alignment via siunitx S columns when precision matters.- Notes via
threeparttable. Significance levels, standard-error type, sample restrictions, and clustering belong in a single notes block at the bottom — not in the caption.
- Cross-reference with
\Cref{}. \label{tab:main} + \Cref{tab:main} reads "Table 2" in body text and behaves correctly at sentence start.
Decision Policy
This skill follows the repo-wide Agent Policy.
ASK before proceeding (blocking):
- Source language of the regression (Stata / R / Python) — determines the canonical table package.
- Table type (regression / summary stats / balance / multi-panel / DiD-estimator comparison).
- Standard-error type and cluster level (these belong in the table notes).
- Output target (LaTeX
.tex, Word .rtf, Excel .xlsx).
DEFAULT + flag (use this default; tell the user how to override):
esttab (Stata), modelsummary or fixest::etable (R), pyfixest.etable or stargazer (Python).booktabs styling: \toprule, \midrule, \bottomrule; no vertical lines.- Three-star convention
* 0.10 ** 0.05 *** 0.01; pin once per paper.
threeparttable for the notes block.- Output path
paper/tabs/<name>.tex; the paper \input{}s it.
- Within-R² (
r2_within) reported instead of overall R² for FE models.
DOCUMENT and proceed (write into the decisions log of the table-generating script):
- The estimating equation, sample, cluster level, and FE included for each column.
- Variable label mapping (
coef_map / coeflabels).
- Any subsample restriction shown in a column.
PROCEED items: tables produced by code (never copy-paste); generated .tex never hand-edited; rebuild path runs the table script before LaTeX.
Pre-flight Checklist
Before generating code, confirm with the user — and write the answers in the script header:
- Source language. Stata, R, or Python? (Each has its own canonical table package.)
- Table type. Regression (multi-spec), summary statistics, balance / difference-in-means, descriptive cross-tabs, or model-comparison panel?
- Sample. Single or stratified? If panel, which fixed effects are on/off across columns?
- Standard errors. Cluster-robust at which level? Wild-bootstrap p-values?
- Stars and notes. Conventional
*** ** * cutoffs (1/5/10), or journal-specific (some require none)?
- Output path. Almost always
paper/tabs/{name}.tex; the paper \input{}s that file.
- Compile target.
\begin{table} floating environment in body, or naked tabular snippet to be included from a parent macro?
Decision Tree (which table package to use)
Source = Stata
└── esttab (from estout) is canonical.
eststo store; esttab using "tabs/x.tex", booktabs label se ...
Source = R
├── fixest models (most common in 2024-26): use fixest::etable
└── Mixed model classes: use modelsummary (works across lm, glm,
PanelOLS, fixest, brms, ...). modelsummary writes .tex via
`output = "tabs/x.tex"`.
Source = Python
├── pyfixest models: use pyfixest.etable
├── linearmodels models: use compare(...) + manual to_latex(), or
pass models through stargazer
└── statsmodels OLS / GLM: use stargazer (Python port).
Output for Word
└── Same packages with .rtf or .docx output (esttab supports rtf;
modelsummary supports docx via flextable).
Output Path Convention
paper/
├── paper.tex
├── tabs/ # all .tex tables; never hand-edited
│ ├── table_main.tex
│ ├── table_balance.tex
│ ├── table_summary.tex
│ └── table_robustness.tex
├── figs/
└── code/
├── stata/
│ └── make_tables.do
└── r/
└── make_tables.R
Body text:
\Cref{tab:main} reports our preferred specification.
\input{tabs/table_main.tex}
The .tex file should produce the entire \begin{table} ... \end{table} environment so you only have one \input{} per table in the paper.
Common Pitfalls
- Hand-editing the
.tex file after generation. The next run will overwrite or diverge from the published numbers.
- Mixing significance-star conventions across tables (
* at 0.05 in one, at 0.10 in another).
- Using
\hline everywhere; readers expect three-line booktabs tables.
- Vertical lines in regression tables — never necessary, always ugly.
- Putting key information in the caption (e.g. "Standard errors clustered by firm") instead of in the notes block, where it can be longer and less awkward.
- Using
_ in variable labels without \_ escaping; LaTeX silently interprets them as subscripts.
- Including raw column names like
treat_post instead of human-readable labels.
- Numbers with inconsistent decimal precision across columns.
- Missing N row at the bottom of regression tables; reviewers will ask.
- Not exporting at all — running the regression in interactive mode and screenshotting the result. (This is "no replicability" in DIME's framework.)
Additional Resources
reference.md — extended patterns: balance tables, summary stats, panel tables, threeparttable notes, siunitx S-column alignment, multi-panel tables, journal-specific deviations.examples/ — runnable scripts that produce real .tex files:
examples/make_tables_esttab.do — Stata regression + summary + balance via esttabexamples/make_tables_modelsummary.R — R via modelsummary + fixest::etableexamples/make_tables_pyfixest.py — Python via pyfixest.etableexamples/make_tables_stargazer.py — Python statsmodels via stargazerexamples/table_main_template.tex — minimal copy-paste-able tabular for "good replicability" casesexamples/table_balance_template.tex — DIME-style balance table layout
Requirements
- LaTeX: TeX Live 2022+ or MacTeX 2022+; packages
booktabs, threeparttable, siunitx, caption, cleveref.
- Stata:
ssc install estout (provides esttab, eststo, estadd, estpost).
- R:
install.packages(c("modelsummary", "fixest", "kableExtra")).
- Python:
pip install stargazer pyfixest pandas.
References
DIME
Style
