| name | results-analysis |
| description | This skill should be used when the user asks to "analyze empirical results", "generate results section", "interpret regression output", "create regression tables", "visualize DID results", "plot event study", "check pre-trends", "first stage F-statistic", "run robustness checks", "interpret causal estimates", or mentions connecting empirical analysis to social science paper writing. Covers DID, IV, RDD, and Synthetic Control designs. |
| tags | ["Research","CausalInference","Statistics","Visualization","Econometrics"] |
| version | 1.0.0 |
Results Analysis for Social Science Empirical Research
A systematic empirical results analysis workflow for political science and economics papers — from raw regression output to publication-ready tables, figures, and Results section text.
Core Features
- Identification Assumption Checks — Pre-trend tests, first-stage F, McCrary density, MSPE ratios
- Main Results Presentation — Regression tables with proper SE clustering, LaTeX output
- Robustness Battery — Alternative specifications, placebo tests, subgroup analysis
- Visualization — Event study plots, coefficient plots, RD plots, binned scatter
- Results Section Writing — Social science journal structure (APSR/AJPS/AER/QJE)
When to Use
- Validate identification assumptions (pre-trends, exclusion restriction, continuity)
- Generate LaTeX regression tables ready for insertion
- Create event study / coefficient plots
- Conduct and report robustness checks
- Write or revise the Results / Empirical Findings section
- Interpret causal estimates in substantive terms
Workflow
Data Loading → Identification Checks → Main Results → Robustness → Visualization → Writing → Quality Check
Step 1: Data and Output Loading
Common Input Formats:
.dta — Stata dataset (load via pystata or pandas + pyreadstat).csv — Panel / cross-sectional data- Stata log (
.smcl / .log) — Regression output to parse
- Python objects — Results from
statsmodels, linearmodels, econml
Initial Data Checks:
- Sample dimensions: N units, T periods, total observations
- Panel structure: balanced vs. unbalanced, attrition pattern
- Missing values: which variables, which groups/periods
- Treatment/control proportions and overlap
Step 2: Identification Assumption Checks
Always validate before reporting main results.
DID / Event Study
- Pre-trend test: plot event-study coefficients for periods t = −k to t = −1; all should be statistically indistinguishable from zero
- Parallel trends visual: plot raw outcome trends for treated and control groups pre-treatment
- Anticipation effects: inspect t = −1 and t = −2 coefficients
- Callaway-Sant'Anna / Sun-Abraham for staggered treatment: check for heterogeneity-robust estimates
Instrumental Variables
- First stage F-statistic: report Kleibergen-Paap (clustered) F, not Cragg-Donald; Stock-Yogo critical values
- Weak instrument: F > 10 as rule of thumb; F > 104.7 for 5% size distortion (Stock-Yogo)
- Exclusion restriction: theoretical argument + placebo outcome test
- Endogeneity test: Hausman test (if OLS and IV differ significantly, report both)
RDD
- Density continuity: McCrary (2008) test; plot histogram around cutoff
- Covariate balance: run RDD on pre-determined covariates (should be ≈ 0)
- Optimal bandwidth: IK (2012) or CCT (2014); report results across multiple bandwidths
- Donut-hole robustness: exclude observations very close to cutoff
Synthetic Control
- Pre-period fit: MSPE ratio (treated vs. donor pool); target < 2
- Placebo in-space: run SC for each donor unit; treated effect should be an outlier
- Placebo in-time: use an artificial earlier treatment date; effect should be near zero
- Leave-one-out: drop each donor and re-run; check stability
See references/identification-checklist.md for detailed protocols and code snippets.
Step 3: Main Results
Regression Table Standards (social science journals):
| Element | Requirement |
|---|
| Standard errors | Clustered at appropriate level (state, county, individual); heteroskedasticity-robust as minimum |
| Significance markers | †p<0.10, *p<0.05, **p<0.01, ***p<0.001 (or journal-specific) |
| Fixed effects rows | Include "Unit FE ✓", "Time FE ✓" rows at bottom |
| Sample info | N (observations), N_clusters if clustered |
| Model fit | R² (within R² for FE models), first-stage F for IV |
Recommended Column Progression:
(1) OLS / Baseline
(2) Main specification (DID / IV / RDD)
(3) + Controls
(4) Alternative clustering
(5) Alternative fixed effects / specification
Effect Size Interpretation:
- Raw coefficient + percentage of outcome mean (e.g., "0.23 SD increase, or 14% of the sample mean")
- For log outcomes: elasticity ("a 10% increase in X → Y% change in outcome")
- For binary outcomes: marginal effects at the mean alongside LPM/logit
Step 4: Robustness Checks
Standard Robustness Battery:
| Check | Description |
|---|
| Alternative window/bandwidth | ±50%, ±200% of main specification |
| Different control sets | No controls → parsimonious → saturated |
| Alternative outcome definitions | Different measurement, normalization |
| Alternative clustering level | One level up / down |
| Wild bootstrap | For small N_cluster (< 50) |
| Placebo treatment | Assign treatment to untreated units |
| Placebo outcome | Test on predetermined variable |
| Subgroup analysis | Key heterogeneity dimensions |
| Balanced panel | Restrict to always-observed units |
Present robustness as a coefficient plot across specifications (point estimate + 95% CI per column), or a robustness table in the appendix.
Step 5: Visualization
Event Study / Coefficient Plot:
- Point estimates + 95% CI for each event-time period
- Reference line at zero; vertical dashed line at treatment date
- Pre-period serves as visual identification test
- Use
matplotlib + pandas or plotnine (ggplot2 equivalent)
Coefficient Plot (across specifications):
- One row per specification, all on same scale
- Ordered by model complexity or specification dimension
RD Plot:
- Binned scatter with local polynomial fit on each side of cutoff
- Confidence band around polynomial
- Raw scatter as background (lighter color)
Binned Scatter:
- For non-parametric relationship visualization
- Residualize on controls / FE before binning
Publication Standards:
- Vector format (PDF preferred for journals)
- Colorblind-safe palette (Okabe-Ito or similar)
- Font size ≥ 10pt
- Axis labels in plain English (not variable names like
ln_gdppc)
- Caption must be self-contained
See references/visualization-guide.md for Python/Stata code templates.
Step 6: Writing the Empirical Results Section
Social Science Results Structure:
## [N]. Empirical Results
### [N.1] Main Results
[Lead with headline number and substantive significance]
[Table X presents the main results. Column (1) shows... Column (2) adds...]
[Interpret magnitude: "a one-standard-deviation increase in X is associated with a Y-unit / Z% change in outcome"]
### [N.2] Identification Tests
[Pre-trend test: "Figure X plots event-study coefficients... none of the pre-treatment coefficients are statistically distinguishable from zero"]
[First stage: "The Kleibergen-Paap F-statistic is XX, exceeding the Stock-Yogo critical value"]
### [N.3] Robustness
[Brief summary: "These results are robust to..."; point to Appendix Table X]
### [N.4] Heterogeneity / Mechanisms
[Which subgroups show larger/smaller effects and why — connect to theory]
Writing Principles:
- Lead with substantive significance ("equivalent to 0.3 years of average wage growth"), not just stars
- Always interpret magnitudes in context
- Explicitly link identification evidence back to the causal claim
- Acknowledge limitations honestly; reviewers will raise them anyway
See references/results-writing-guide.md for journal-specific style guidance.
Step 7: Quality Checklist
Integration with Paper Writing
Causal analysis complete (Stata / Python)
↓
results-analysis: validate assumptions, generate tables/figures
↓
ml-paper-writing: integrate into full paper structure
↓
Complete Empirical Results section
This skill handles:
- Assumption validation and reporting
- Regression tables (LaTeX)
- Visualization (event study, coefficient plots, RD plots)
- Results section drafting
ml-paper-writing handles:
- Full paper structure (Introduction, Theory, Data, Conclusion)
- Citation formatting
- Journal submission requirements
Reference Resources
references/identification-checklist.md — DID/IV/RDD/SC assumption testing with codereferences/visualization-guide.md — Python + Stata code templates for coefficient plots, event studiesreferences/results-writing-guide.md — Results section writing standards for APSR/AJPS/AER/QJEreferences/statistical-methods.md — Statistical tests reference (clustered SE, wild bootstrap, etc.)
