# Advanced methods QA

Generated: 2026-06-11T00:04:45+12:00

## Analysis population
- Three-period road-month panel: 65,116 observations; 662 roads.
- Controlled corridor reversal panel: 78 roads; 26 raised and 52 comparison.
- March 2025 is excluded from controlled reversal models because changes occurred on different days.
- All-severity outcomes stop at 30 April 2026. May 2026 injury counts are complete but non-injury records still lag in CAS, so May is excluded to avoid understating recent crashes.

## Primary all-road before/during/after phase contrasts

| Class | Outcome | Term | IRR (95% CI) | p |
|---|---|---|---|---|
| corridor | total | reduction_level | 0.927 (0.793-1.083) | 0.327 |
| corridor | total | reversal_level | 1.054 (0.907-1.224) | 0.478 |
| corridor | injury | reduction_level | 0.963 (0.757-1.225) | 0.750 |
| corridor | injury | reversal_level | 0.930 (0.701-1.233) | 0.600 |
| corridor | dsi | reduction_level | 0.934 (0.487-1.791) | 0.829 |
| corridor | dsi | reversal_level | 0.842 (0.370-1.913) | 0.669 |
| corridor | dsi_people | reduction_level | 0.856 (0.423-1.732) | 0.653 |
| corridor | dsi_people | reversal_level | 0.709 (0.322-1.563) | 0.379 |
| local_street | total | reduction_level | 0.831 (0.700-0.986) | 0.034 |
| local_street | total | reversal_level | 1.070 (0.918-1.248) | 0.387 |
| local_street | injury | reduction_level | 1.136 (0.811-1.590) | 0.457 |
| local_street | injury | reversal_level | 0.870 (0.649-1.167) | 0.351 |
| local_street | dsi | reduction_level | 0.651 (0.325-1.304) | 0.222 |
| local_street | dsi | reversal_level | 0.698 (0.359-1.357) | 0.286 |
| local_street | dsi_people | reduction_level | 0.691 (0.340-1.406) | 0.304 |
| local_street | dsi_people | reversal_level | 0.642 (0.330-1.249) | 0.189 |

## Controlled corridor reversal difference-in-differences

| Outcome | IRR (95% CI) | p |
|---|---|---|
| total | 1.177 (0.886-1.563) | 0.260 |
| injury | 0.885 (0.582-1.346) | 0.569 |
| dsi | 1.200 (0.308-4.682) | 0.793 |
| dsi_people | 0.626 (0.142-2.767) | 0.537 |

## Interpretation limits
- Conditional Poisson road fixed effects control stable road differences without selecting roads on post-change crashes.
- The controlled reversal model uses raised-segment geometry and pre-period matched comparison roads whose lower limits were not raised.
- The comparison-group before/after analysis is not a true Empirical Bayes estimate; a calibrated Auckland safety-performance function and untreated segment inventory are still required for that.
- Negative-binomial GEE checks overdispersion and within-road correlation as an alternative to the primary Poisson fixed-effects specification.
- Matched-set total-crash randomization-inference p-value: 0.191.
- Total-crash synthetic-comparison ratio: 2.024; placebo p-value: 0.421.
- Suspected speed is investigator-estimated and available only for a subset of DSI crashes. It is not operating-speed monitoring.
- Event-study rows written: 46. Comparison-group rows written: 4.