9a0fb5f2d5
Implements a comprehensive suite of AI SDK v6 tools across multiple categories: - Research (5): page-brief, compare-pages, source-credibility, claim-checklist, timeline-from-text - Web (10): fetch-text, links-catalog, extract-meta, extract-json-ld, redirect-trace, sitemap-read, rss-read, table-extract, robots-policy, url-normalize - Data (15): csv-parse, csv-stringify, json-repair, json-schema-validate, yaml-parse, yaml-stringify, text-chunk, normalize-whitespace, dedupe-by-key, pivot, rows-filter, rows-sort, rows-group-aggregate, rows-join, schema-infer - Doc (12): toc-generate, glossary-build, faq-from-text, executive-brief, decision-record-adr, prd-outline, acceptance-criteria, style-rewrite - Eng (12): diff-text-unified, env-var-docs-generate, dependency-audit-lite, conventional-commit-suggest, markdown-lint-basic, test-case-generate, stacktrace-parse, release-notes, changelog-entry, release-checklist - Security (7): redact-secrets, secret-scan-text, url-risk-heuristic, csp-compose, hardening-checklist-web, access-control-matrix, data-classification-heuristic - Stats (9): effect-size-suite, bootstrap-ci, permutation-test, multiple-testing-adjust, linear-regression-ols, logistic-regression, time-series-decompose-lite, anomaly-detect-mad - Ops (7): slo-draft, runbook-draft, postmortem-draft, postmortem-action-extractor, error-log-triage, coverage-tracker, monitoring-gap-analysis - Agent (15): prompt-to-workflow-skeleton, workflow-validate-io, workflow-explain, workflow-cost-estimate, tool-call-accuracy-score, eval-fixture-build, guardrail-policy-draft, workflow-auto-repair, tool-selection-plan, novelty-score-workflow, workflow-variant-generate, config-normalize, recipe-* - Utility (8): base64-encode, base64-decode, hash-text, regex-extract, template-render, date-parse, json-path-query, url-parse - HTML (3): html-sanitize, html-to-markdown, markdown-to-html All tools follow AI SDK v6 pattern with tool() and jsonSchema<T>(). 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
130 lines
3.8 KiB
Markdown
130 lines
3.8 KiB
Markdown
# Difference-in-Differences (DiD)
|
|
|
|
Causal inference estimator for measuring treatment effects using before/after comparison with a control group.
|
|
|
|
## Installation
|
|
|
|
```bash
|
|
npm install @tpmjs/tools-diff-in-diff
|
|
```
|
|
|
|
## Usage
|
|
|
|
```typescript
|
|
import { diffInDiffTool } from '@tpmjs/tools-diff-in-diff';
|
|
|
|
// Example: Evaluate impact of a policy intervention
|
|
// Treatment group: Cities that implemented the policy
|
|
// Control group: Cities that did not implement the policy
|
|
const result = await diffInDiffTool.execute({
|
|
treatmentBefore: [100, 105, 98, 102], // Before policy
|
|
treatmentAfter: [120, 125, 118, 122], // After policy
|
|
controlBefore: [95, 100, 92, 98], // Before (no policy)
|
|
controlAfter: [98, 103, 95, 101], // After (no policy)
|
|
confidenceLevel: 0.95,
|
|
});
|
|
|
|
console.log(result);
|
|
// {
|
|
// effect: 17.5, // Treatment caused 17.5 unit increase
|
|
// standardError: 2.1,
|
|
// tStatistic: 8.33,
|
|
// pValue: 0.0001,
|
|
// significant: true,
|
|
// confidenceInterval: {
|
|
// lower: 13.2,
|
|
// upper: 21.8,
|
|
// level: 0.95
|
|
// },
|
|
// interpretation: "The treatment effect is 17.5 (increased by 17.5 units)...",
|
|
// groupMeans: {
|
|
// treatmentBefore: 101.25,
|
|
// treatmentAfter: 121.25,
|
|
// controlBefore: 96.25,
|
|
// controlAfter: 99.25
|
|
// },
|
|
// differences: {
|
|
// treatmentDiff: 20.0, // Treatment group change
|
|
// controlDiff: 3.0 // Control group change
|
|
// }
|
|
// }
|
|
```
|
|
|
|
## API
|
|
|
|
### Input
|
|
|
|
- **treatmentBefore** (required): Treatment group values before intervention
|
|
- **treatmentAfter** (required): Treatment group values after intervention
|
|
- **controlBefore** (required): Control group values before intervention
|
|
- **controlAfter** (required): Control group values after intervention
|
|
- **confidenceLevel** (optional): Confidence level (default: 0.95)
|
|
|
|
### Output
|
|
|
|
- **effect**: Estimated causal treatment effect (DiD estimator)
|
|
- **standardError**: Standard error of the estimate
|
|
- **tStatistic**: Test statistic for significance testing
|
|
- **pValue**: Two-tailed p-value
|
|
- **significant**: Whether effect is statistically significant
|
|
- **confidenceInterval**: Confidence interval for the effect
|
|
- **interpretation**: Plain English interpretation
|
|
- **groupMeans**: Mean values for all four groups
|
|
- **differences**: Within-group changes over time
|
|
|
|
## Algorithm
|
|
|
|
The DiD estimator removes time-invariant confounders by differencing:
|
|
|
|
**Formula**: `DiD = (T_after - T_before) - (C_after - C_before)`
|
|
|
|
Where:
|
|
- T = Treatment group
|
|
- C = Control group
|
|
|
|
This double-differencing removes:
|
|
1. Time trends (via control group)
|
|
2. Group differences (via before/after comparison)
|
|
|
|
**Key Assumption**: Parallel trends - Without treatment, both groups would have changed similarly.
|
|
|
|
## Use Cases
|
|
|
|
- **Policy evaluation**: Measure impact of new regulations
|
|
- **Marketing**: Test effectiveness of campaigns
|
|
- **Medicine**: Clinical trials with before/after measurements
|
|
- **Economics**: Evaluate economic interventions
|
|
- **Education**: Assess program effectiveness
|
|
|
|
## Example: Minimum Wage Study
|
|
|
|
```typescript
|
|
// States that raised minimum wage (treatment)
|
|
// vs states that didn't (control)
|
|
const result = await diffInDiffTool.execute({
|
|
treatmentBefore: [5.2, 5.5, 5.1, 5.4], // Employment before
|
|
treatmentAfter: [5.1, 5.3, 5.0, 5.2], // Employment after
|
|
controlBefore: [5.3, 5.4, 5.2, 5.5],
|
|
controlAfter: [5.4, 5.5, 5.3, 5.6],
|
|
});
|
|
|
|
// Effect tells us the causal impact on employment
|
|
```
|
|
|
|
## Interpretation
|
|
|
|
A **positive effect** means treatment increased the outcome.
|
|
A **negative effect** means treatment decreased the outcome.
|
|
|
|
**Statistical significance** (p < 0.05) suggests the effect is real, not due to chance.
|
|
|
|
## Limitations
|
|
|
|
- Requires parallel trends assumption
|
|
- Can't control for time-varying confounders
|
|
- Sensitive to outliers
|
|
- Needs sufficient sample size
|
|
|
|
## License
|
|
|
|
MIT
|