empirical-analysis-skill-python

name	empirical-analysis-skill-python
description	Parameterized Python empirical-analysis and machine-learning workflow for applied economics, public health epidemiology, supervised ML, and ML causal inference. Use when the user asks for data cleaning, feature engineering, train/test/validation splits, feature matrix X and target y, Table 1, diagnostic tests, OLS/panel/IV-style formulas, LinearRegression, Ridge, Lasso, ElasticNet, decision trees, random forests, GBDT, regression/classification metrics, DID/event-study formulas, DML/double machine learning with LinearDML, causal forests, robustness checks, mechanism or heterogeneity analysis, mediation, publication-ready tables, figures, or an end-to-end empirical paper pipeline. The skill must route execution through fixed step scripts under scripts/ instead of writing ad hoc Python code in markdown.
triggers	["full empirical analysis in Python","classical econometrics pipeline","applied economics paper replication","end-to-end empirical workflow","data cleaning empirical","winsorize and standardize","variable construction","machine learning basics","supervised learning","feature matrix X","target vector y","train validation test split","validation set","regression metrics","classification metrics","MAE","MSE","RMSE","R2","accuracy","precision","recall","F1","LinearRegression","Ridge regression","Lasso regression","ElasticNet","Elastic Net","decision tree","random forest","GBDT","Table 1 summary statistics","balance table","correlation matrix","normality test","heteroskedasticity test","autocorrelation test","multicollinearity VIF","baseline regression","panel fixed effects","PanelOLS","random effects","first difference panel","DID workflow Python","difference in differences","event study","pretrend test","instrumental variables regression","2SLS","LIML","GMM IV","double machine learning","DML","LinearDML","causal forest","meta learner","regression discontinuity","propensity score matching","synthetic control python","robustness checks","placebo test","specification curve","alternative clustering","heterogeneity analysis","mechanism analysis","mediation analysis","moderation analysis","publication-ready regression table","coefplot","binscatter","event study plot","target trial emulation","IPTW","TMLE","Mendelian randomization","survival analysis","Cox model","AFT model","E-value sensitivity","policy tree","conformal causal","fairness audit"]

Empirical Analysis Skill - Python

This skill runs a full empirical workflow through fixed, decoupled Python scripts. The agent may choose variables, formulas, modes, and arguments, but must not invent Python snippets inside markdown responses or reference files.

Hard Rules

• Run the fixed step scripts under scripts/ for analysis work.
• Use CLI parameters, formula strings, and input/output paths to customize behavior.
• Do not paste new Python code into markdown as the primary solution.
• If the fixed scripts do not support a requested estimator, either express it through --formula or --formulas, or state the limitation and ask whether to extend the relevant script.
• Keep generated outputs on disk under the project output directory, usually output/empirical.
• Use reference files only to decide which subcommand and parameters to call.

Script Layout

Use the per-step scripts directly:

Step	Script	Purpose	Reference
1	scripts/clean_data.py	Inspect, coerce dtypes, trim strings, handle key missingness, deduplicate, and log sample construction.	references/01-data-cleaning.md
2	scripts/transform_data.py	Create logs, IHS variables, winsorized columns, z-scores, dummies, lags, leads, differences, and treatment timing.	references/02-data-transformation.md
2M	scripts/prepare_ml_data.py	Build supervised ML feature matrix X, target vector y, train/valid/test splits, imputation, and standardization.	references/09-machine-learning.md
3	scripts/describe_data.py	Generate Table 1, balance table, categorical frequencies, correlation matrix, heatmap, and trend plot.	references/03-descriptive-stats.md
4	scripts/run_diagnostics.py	Run residual normality, heteroskedasticity, autocorrelation, condition-number, and VIF diagnostics.	references/04-statistical-tests.md
5	scripts/run_model.py	Fit OLS, logit, probit, poisson, or IV formulas; export regression tables and coefficient plots.	references/05-modeling.md
5A	scripts/run_panel.py	Run panel FE/RE/between/first-difference models with clustered covariance support.	references/05-modeling.md
5A	scripts/run_did.py	Run 2x2/TWFE DID and event-study specifications with pretrend tests.	references/05-modeling.md
5A	scripts/run_iv.py	Run 2SLS/LIML/GMM IV estimates with first-stage and overidentification diagnostics where available.	references/05-modeling.md
5A	scripts/run_rd.py	Run sharp/fuzzy regression-discontinuity estimates, bandwidth checks, and RD plots.	references/05-modeling.md
5A	scripts/run_matching.py	Run propensity-score IPW, nearest-neighbor matching, and covariate balance diagnostics.	references/05-modeling.md
5A	scripts/run_synth.py	Run synthetic-control weights, trajectory, gap, and fit diagnostics.	references/05-modeling.md
5M	scripts/run_supervised_ml.py	Run LinearRegression, Ridge, Lasso, ElasticNet, tree models, random forest, GBDT, and classification models with fixed metrics.	references/09-machine-learning.md
5B	scripts/run_dml.py	Run fixed econml.dml.LinearDML specifications for double machine learning and ML causal inference.	references/05-modeling.md
5B	scripts/run_cate.py	Run T-learner, S-learner, or econml causal forest CATE estimation.	references/07-further-analysis.md
6	scripts/run_robustness.py	Run fixed robustness variants: alternative SEs, clusters, filters, placebo variables, and control sets.	references/06-robustness.md
6	scripts/run_sensitivity.py	Run Oster delta, E-value, and randomization-inference sensitivity analyses.	references/06-robustness.md
7	scripts/run_further_analysis.py	Run heterogeneity, mechanism/outcome-ladder, and simple Baron-Kenny mediation tables.	references/07-further-analysis.md
7E	scripts/run_survival.py	Run Kaplan-Meier, Cox, and Weibull AFT survival-analysis workflows.	references/07-further-analysis.md
8	scripts/table_factory.py	Format tidy model results into publication-oriented wide tables with significance stars.	references/08-tables-plots.md
8	scripts/plot_factory.py	Create coefficient, event-study, binscatter, and love plots from fixed input files.	references/08-tables-plots.md
8	scripts/render_manifest.py	Collect tables and figures into an artifact manifest.	references/08-tables-plots.md

Shared support files:

• scripts/common.py: shared file I/O, export, plotting, and argument helpers.
• scripts/model_utils.py: shared formula construction and model-fitting helpers.
• scripts/empirical_cli.py: thin compatibility dispatcher only. Prefer direct step scripts.

Use references/method-index.md to map a user request to the correct fixed script. Discover supported parameters with inline commands such as python scripts/run_model.py --help.

Default Workflow

1. Start with scripts/clean_data.py on the raw data. Always pass key variables and dedupe columns when known.
2. Use scripts/transform_data.py to create analysis variables rather than doing inline dataframe mutations.
3. Use scripts/describe_data.py to generate Table 1 and the main exploratory figures.
4. Use scripts/run_diagnostics.py on the baseline formula before reporting results.
5. Use scripts/run_model.py for the main results. Prefer progressive model columns with --progressive or explicit semicolon-separated --formulas.
6. Use scripts/run_robustness.py for alternative controls, clusters, subsamples, and placebo checks.
7. Use scripts/run_further_analysis.py for mechanism, heterogeneity, and mediation.
8. Use scripts/render_manifest.py to produce the final manifest of tables and figures.

Domain Modes

Default applied economics mode uses run_model.py, run_robustness.py, and run_further_analysis.py with OLS or IV formula strings. Fixed effects are expressed with --fixed-effects or explicit C(variable) terms in formula strings.

Epidemiology and public-health mode still uses the same cleaning, transformation, descriptives, and diagnostics steps. Use scripts/run_survival.py for Kaplan-Meier, Cox, and Weibull AFT work. For TMLE, target-trial, or Mendelian-randomization work that is not covered by fixed step scripts, add or extend a dedicated script first rather than placing code into markdown.

Supervised ML mode uses scripts/prepare_ml_data.py for X/y construction and scripts/run_supervised_ml.py for LinearRegression, Ridge, Lasso, ElasticNet, decision tree, random forest, GBDT, logistic classification, and regression/classification metrics.

ML causal inference mode uses the same file layout and reporting contract. For double machine learning, use scripts/run_dml.py directly. For CATE work, use scripts/run_cate.py with T-learner, S-learner, or econml causal forest. For policy-learning estimators not yet covered by fixed scripts, add a separate step script before use.

Output Contract

Default output root: output/empirical.

Tables should be exported as CSV, XLSX, and LaTeX unless the user specifies otherwise. Use --formats csv,xlsx,tex or add docx if python-docx is available.

Figures should be exported as PNG at 300 dpi and PDF when figure creation is enabled.

Core economics-style deliverables:

Artifact	File stem
Table 1 summary statistics	table1_summary
Table 1 balance table	table1_balance
Main regression table	table2_main and table2_main_tidy
Supervised ML metrics	ml_metrics
Supervised ML coefficients/importances	ml_feature_effects
Mechanism table	table3_mechanism
Heterogeneity table	table4_heterogeneity
Robustness table	table5_robustness
Sensitivity table	sensitivity_results
Trend figure	fig1_trend
RD plot	fig_rd_plot
Synthetic-control trajectory	fig_synth_trajectory
KM survival curve	fig_km_curve
CATE distribution	fig_cate_distribution
ML feature effect plot	fig_ml_feature_effects
Coefficient plot	fig3_coefplot
Sensitivity/specification figure	fig4_sensitivity
Final artifact index	artifact_manifest.json and artifact_manifest.md

Parameter Conventions

• Column lists are comma-separated, for example age,edu,tenure.
• Multiple formulas or filters are semicolon-separated.
• Use exact formula strings when the design is special. The script accepts formulas through --formula or --formulas.
• Use --cluster firm_id for clustered standard errors in supported model subcommands.
• Use --fixed-effects worker_id,year for categorical fixed effects in statsmodels formulas.
• Use --output-dir for tables and figures; use --output for transformed datasets.

When To Read References

Read only the relevant reference file before running a subcommand. The references do not contain Python implementation code. They explain parameter choices, outputs, and checks.

Maintenance Rule

When adding new empirical behavior, modify the relevant step script or shared helper and document only the invocation pattern in references/. Do not add executable Python code blocks to SKILL.md, README.md, or any references/*.md file.