edwinhu/ds-tables

Contents

• Which Tool
• Table Enforcement
• Quick Start
• Additional Resources

Publication Tables in Python

Two libraries, one stack: pyfixest.etable() builds regression tables and renders them through great_tables, so anything etable() returns can be further styled with GT methods. Both render natively in marimo and Jupyter.

Which Tool

| Building | Use | Reference | |---|---|---| | Regression results (coefficients, SEs, stars, FE panel) | pf.etable([fit1, fit2]) | references/pyfixest-tables.md | | Summary stats, descriptives, any DataFrame | GT(df) method chain | references/great-tables.md | | Coefficient / event-study plot | pf.coefplot(), pf.iplot() | references/pyfixest-tables.md | | LaTeX or HTML for a manuscript | etable(..., type="tex" \| "html", file_name=...) | references/pyfixest-tables.md |

Table Enforcement

IRON LAW: NO TABLE CLAIM WITHOUT RENDERING IT

A table that "looks right" in code is not a table. Before claiming one is done:

1. RENDER it — display the GT object, or write the tex/html file
2. READ the rendered output (in a notebook: look at it; for tex/html: read the file)
3. VERIFY the numbers against the source fit or DataFrame
4. VERIFY the labels, stars, and notes say what the caller thinks they say
5. CLAIM done only after the render is inspected

Handing over an unrendered table is NOT HELPFUL — the user pastes it into a paper and discovers the stars are missing at submission time.

Table Facts

• The default etable() prints stars (1.873***) — but the moment you pass coef_fmt yourself, stars come only from a * token in it. coef_fmt="b \n (se)" silently drops the stars you had for free; "b* \n (se)" keeps them. Verified on pyfixest 0.60.0: default → 1.873***, explicit "b \n (se)" → 1.873. signif_code only sets the thresholds, never whether stars render.
• keep / drop take regex, not names. keep="x1" also matches x10 and log_x1; use r"^x1$" or exact_match=True.
• pyfixest drops singleton fixed-effect groups by default and prints a warning, not an error — the observation count silently differs from the same spec elsewhere unless you read the warning.
• etable() returns a great_tables GT object when type="gt" (the default). Reach for the GT method chain for anything etable doesn't parameterize instead of post-processing strings.
• fmt_* methods take columns= and rows=; omitting both applies the format to the whole table, which is rarely what a mixed-type table wants.

Red Flags — STOP If About To:

• Report a regression table as final without having rendered and read it → STOP. Unrendered is unverified.
• Write a custom coef_fmt without a * token → STOP. You just turned off the stars the default gave you; signif_code will not bring them back.
• Hand-edit LaTeX that etable() produced → STOP. Round-trip the change through etable()/GT parameters, or the next re-run silently reverts it.
• Retype numbers from a fit into a table by hand → STOP. Every hand-typed number is an unverified claim; pass the fit objects.

Quick Start

import pyfixest as pf

fit1 = pf.feols("y ~ x1 + x2 | fe1", data=df, vcov={"CRV1": "cluster_var"})
fit2 = pf.feols("y ~ x1 + x2 | fe1 + fe2", data=df, vcov={"CRV1": "cluster_var"})

pf.etable(
    [fit1, fit2],
    coef_fmt="b* \n (se)",              # stars REQUIRE the * token
    labels={"x1": "Treatment"},
    notes="Standard errors clustered by firm.",
)

from great_tables import GT, md

(GT(summary_df)
 .tab_header(title="Summary Statistics")
 .fmt_number(columns=["mean", "sd"], decimals=2)
 .cols_label(mean=md("**Mean**"))
 .tab_source_note("Sample: 1996-2024."))

Additional Resources

Reference Files

• references/pyfixest-tables.md — full etable() parameter reference: coef_fmt tokens, significance codes, keep/drop, labels, model headers, FE display, custom stats, coefplot/iplot, gotchas
• references/great-tables.md — GT constructor, method-chaining pattern, every fmt_* method, column management, headers/spanners, styling, export

Related

• /marimo, /jupytext — where these tables render
• ds-implement — the DS phase that produces exhibits