Difference-in-differences estimator for causal inference
npm install @tpmjs/tools-diff-in-diffCausal inference estimator for measuring treatment effects using before/after comparison with a control group.
``bash`
npm install @tpmjs/tools-diff-in-diff
`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
// }
// }
`
- 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)
- 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
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.
- 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
`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
``
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.
- Requires parallel trends assumption
- Can't control for time-varying confounders
- Sensitive to outliers
- Needs sufficient sample size
MIT