Agent Skills
› brycewang-stanford/Awesome-Journal-Skills
› expecon-tables-figures
expecon-tables-figures
GitHub专为实验经济学论文设计，用于构建展示处理对比及不确定性的图表。强调直接呈现分布、正确误差线及动态趋势，遵循原始数据优先、报告效应量等规范，确保读者无需阅读正文即可理解核心结果。
触发场景
需要直观展示处理效应及其不确定性
表格仅显示回归系数而缺乏原始处理均值和离散度
时间/轮次动态（如学习、收敛）被聚合数据掩盖
图表中的误差棒、单位或推断水平模糊或错误
安装
npx skills add brycewang-stanford/Awesome-Journal-Skills --skill expecon-tables-figures -g -y
SKILL.md
Frontmatter
{
"name": "expecon-tables-figures",
"description": "Use when an Experimental Economics (ExpEcon) manuscript's exhibits must make the treatment contrast and its uncertainty visible at a glance. Builds design-carrying tables and figures; it does not run analysis or draft prose."
}
Tables and Figures (expecon-tables-figures)
When to trigger
- A reader cannot see the treatment effect from your main figure without reading the text
- Tables report regression coefficients but never show the raw treatment means and their spread
- Time/round dynamics (learning, convergence, decay) are hidden inside aggregate numbers
- Error bars, units, or the inference level on a plot are ambiguous or wrong
The treatment-comparison figure is the paper's spine
At a method-defined journal, the central exhibit is almost always a comparison across treatments, and a referee should grasp the result from it alone. Design it to do three jobs.
- Show the contrast directly. Plot the primary outcome by treatment — bars or dot-and-CI for means, but prefer showing the distribution (box/violin, or a CDF when distributions differ in shape, not just level). A CDF often reveals that a "no mean effect" masks a shift in the tails.
- Show uncertainty at the right level. Error bars / CIs must reflect the independent unit (session/matching-group), not per-decision noise. State the level in the note. Mismatched error bars are an instant credibility hit at this journal.
- Show dynamics when the game is repeated. A round-by-round line plot per treatment reveals learning, convergence to equilibrium, and end-game effects that an average conceals. Mark the period averaging window you use for tests.
House conventions experimentalists expect
- Raw behavior before models. Lead with a figure of actual choices (e.g., contribution distributions), then a table of treatment means with group-level SEs, then regressions for adjustment/heterogeneity. Do not open with a regression table.
- Report effect sizes, not just stars. Give the treatment difference in outcome units and as a fraction of the control mean / endowment (e.g., "+3.2 tokens, 16% of endowment"), with a CI. Significance asterisks alone are weak evidence here; many ExpEcon papers report exact p-values and CIs.
- A summary/design table. Include a treatments table (conditions × what is manipulated × n sessions/groups/subjects) so a referee can audit the design and balance at a glance.
- Self-contained notes. Each exhibit names the outcome, the unit of observation, the inference level, the test used, n at each level, and the period window. A figure a referee must decode from the body is a figure that loses trust.
- Color/encoding discipline. Keep treatment colors consistent across every figure; ensure grayscale legibility; label axes in the subjects' units (ECU) with the money conversion noted once.
Exhibit-by-exhibit plan for a typical paper
A clean ExpEcon results section usually runs a small, fixed set of exhibits — build them in this order:
| # | Exhibit | What it must show |
|---|---|---|
| 1 | Treatments/design table | conditions × manipulated dimension × n (sessions / groups / subjects) |
| 2 | Primary outcome by treatment | distribution (box/violin or CDF) + mean with group-level CI |
| 3 | Round-by-round dynamics (if repeated) | per-treatment trajectory; averaging window for tests marked |
| 4 | Primary test table | effect size in units + share of endowment + CI + exact p, at the right unit |
| 5 | Heterogeneity / mechanism | regression with covariates, clustered at session/group |
| 6 | Robustness panel | exclusions, learning halves, order, permutation test |
Resist the urge to add a seventh table of exploratory contrasts to the main text; push it to the appendix and label it exploratory.
Execution bridge (StatsPAI / Stata MCP)
Generate exhibits from the fitted result, not by retyping numbers. Full map:
execution-with-mcp. Experimental Economics is lab/field experiments; randomization inference, romano_wolf for many treatments/outcomes, and power are decisive — observational tools secondary.
- Tables:
etable(multi-model) ordid_summary_to_latexstraight from theresult_id. - Figures:
plot_from_result/enhanced_event_study_plot/event_study_table— axis units and the SE/clustering note baked in. - Every note names the estimator + clustering and states the magnitude in interpretable units.
See a full fitted-result → exhibit chain in the JF execution walkthrough.
Checklist
- One main figure shows the primary treatment contrast and is interpretable on its own
- Distributions (not just means) shown where shape matters; CDF used when relevant
- Error bars / CIs reflect the session/matching-group unit; the level is stated in the note
- Repeated games include a round-by-round dynamics plot with the averaging window marked
- Effect sizes reported in outcome units and as a share of endowment/control, with CIs
- A treatments/design table lists conditions, manipulation, and n at each level
- Every exhibit note is self-contained (outcome, unit, level, test, n, window)
- Exploratory contrasts are confined to the appendix and labeled
Common scale and encoding traps
- Truncated y-axes that magnify a trivial treatment gap — start contribution/effort axes at the meaningful floor (often 0 or the endowment min), not at a level chosen to dramatize.
- Pooling rounds silently — if you average over rounds for a bar, the reader cannot see decay; pair it with the dynamics plot and state the window.
- Inconsistent treatment colors across figures, forcing the reader to re-learn the legend each time.
- Over-plotted scatter of individual decisions — bin, jitter, or show the density instead.
- CIs that imply more precision than the design has — at the matching-group unit, intervals are usually wider than per-decision intervals; show the honest ones.
Each of these is a small thing a referee notices instantly and reads as carelessness about the inference unit, which is the trait this journal least forgives.
Anti-patterns
- A bar of means with no distribution, hiding that treatment shifts tails not the center
- Error bars computed on per-decision data when the unit is the group — overstated precision
- Opening with a dense regression table before the reader has seen the raw treatment behavior
- Stars-only reporting with no effect size, CI, or share-of-endowment context
- A repeated-game result with no dynamics plot, so learning/end-game effects are invisible
- Exhibit notes that omit the inference level or the rounds used for the test
Worked vignette (illustrative)
A risk-elicitation paper reports only mean "number of safe choices" by treatment, with t-test stars — and finds nothing. Rebuilding as a CDF of safe choices reveals the treatment shifts the upper tail (more extreme risk-aversion) while the mean is flat. The new main figure shows the two CDFs diverging above the 7th choice; the test moves to a Kolmogorov–Smirnov / distributional comparison at the session level, and the effect is reported as "the 80th percentile rises by 1.5 safe choices (illustrative)." The story was in the distribution the mean hid.
Output format
【Journal】Experimental Economics (ESA method flagship)
【Skill】expecon-tables-figures
【Verdict】clear / rebuild
【Main contrast figure】outcome by treatment; distribution shown? CDF?
【Uncertainty】CI at session/group unit, stated in note? [Y/N]
【Dynamics】round-by-round plot for repeated game? window marked?
【Effect size】in units + share of endowment + CI
【Design table】conditions × manipulation × n present? [Y/N]
【Exploratory】confined to appendix and labeled? [Y/N]
【Next skill】expecon-writing-style
版本历史
- 1839142 当前 2026-07-05 13:14
