4 min read

Economics and social science: regression tables that journals accept

booktabs, threeparttable, siunitx stars, notes under estimates, biblatex habits, and working-paper versioning for econ, poli-sci, and sociology.

Empirical social science lives and dies by tables. A misaligned decimal or a missing significance star can burn a revise-and-resubmit cycle that had nothing to do with your identification strategy. Journals care about readability almost as much as coefficients.

What follows is for students and RAs who export from Stata or R and then clean LaTeX by hand.

booktabs is not optional

Vertical rules and double horizontal lines look like 1990s Word. Use the three horizontal rules from booktabs:

\usepackage{booktabs}
\begin{tabular}{lccc}
  \toprule
  & (1) & (2) & (3) \\
  \midrule
  Treatment & 0.12 & 0.09 & 0.08 \\
  \bottomrule
\end{tabular}

See three-rule tables. Export from R (modelsummary, estout-style tools, older stargazer) or Stata to LaTeX fragments, then clean the header row yourself. Auto-generated headers are usually too wide or too cryptic for the final PDF.

Notes under the table, not random footnotes

threeparttable keeps the caption, tabular, and notes the same width:

\begin{table}
\begin{threeparttable}
  \caption{Main results}\label{tab:main}
  \begin{tabular}{...} ... \end{tabular}
  \begin{tablenotes}
    \small
    \item Notes: Robust standard errors in parentheses.
    Sample is ... Fixed effects: ...
  \end{tablenotes}
\end{threeparttable}
\end{table}

Put sample size, fixed effects, clustered SE notes, and the significance-star legend here. Do not hang a footnote off a single cell that only some readers will notice. Panel labels (A/B) belong in the caption or in midrules with clear text, not in a figure-style legend that never appears in the PDF.

Numeric columns and significance stars

siunitx column types (S) align on the decimal point. That alone makes a results table look intentional. Stars break pure numeric columns, so you usually need a custom format or a text-ish column. Common patterns:

Report coefficients and SEs on separate rows (standard in econ), or wrap stars so alignment still works (0.012^{***} with a dedicated column type).

Be consistent across tables: same star thresholds, same SE style, same number of decimals for the same metric. Reviewers compare Table 2 to Table 3 on purpose.

A visual first draft is fine. The browser table generator can sketch structure; the final table should still go through booktabs and your SE notes.

From regression software to camera-ready

Typical loop looks like this. Freeze the estimation script and seed. Export a LaTeX fragment or CSV of estimates. Drop it into a table shell with stable labels (tab:main, tab:robust). Write the caption so a skimmer knows the dependent variable, sample, and estimator without reading the notes twice. Compile, then fix overfull boxes by cutting columns; rotate only as a last resort.

If a table is longer than a page, switch to longtable (or split panels) before deadline week, not after the journal complains about 7pt fonts.

Citations: author-year is the dialect

Most econ and poli-sci venues expect author-year. biblatex with style=authoryear (or the journal’s style) is easier long-term than fighting .bst files. If the lab still uses natbib, that is fine; just do not mix both stacks in one preamble. Overview: BibTeX or biblatex.

Keys should be stable (author:year:keyword). Re-exporting from Zotero under new keys mid-project is how \cite breaks the week of submission. Validate the .bib for missing years and journals: BibTeX validator.

Working papers and SSRN

Working-paper versions change monthly. Keep the title-page date in a macro:

\newcommand{\paperdate}{June 2026}

Bump it when you re-export the PDF. Local Git history is more honest than final_final_v7.pdf on a shared drive. Commit when the main table set compiles; tag or message the submission version explicitly.

A research editor that treats the project as a real Git repo (checkpoints after compile, restore, optional private GitHub) matches how empirical work already versions code. The paper and the estimation scripts should not live under different backup habits.

Identification and table ethics (short version)

LaTeX cannot fix a weak design, but it can make a strong one readable. State the estimator and sample in the caption. Do not hide the fact that column (3) drops half the sample in a footnote nobody reads. If you show robustness tables, keep the same dependent-variable order as the main table so readers can scan.

Pre-submission checklist

Things that bite people late:

  • Three-rule tables; no vertical rules.
  • Decimal alignment; consistent stars and SE notes.
  • Caption states outcome, sample, estimator.
  • Labels stable; text cites \ref{tab:...}.
  • Bibliography style matches the venue; .bib validated.
  • Working-paper date and Git commit match what you uploaded.

Replication packages

More journals want code and data. Keep table-export scripts next to the paper repo (or linked with a commit hash). When a coefficient changes because the cleaning script changed, regenerate the LaTeX fragment and recommit both. A table that cannot be reproduced is a liability even if it looks clean.

Beamer for seminars

Reuse the same booktabs fragments in slides when you can, or export PDF crops of the tables. Rebuilding numbers by hand into a slide deck is how seminar slides disagree with the paper.

Get the tables right and the identification strategy gets a fair hearing. Get them wrong and you spend a round arguing about formatting.

All posts