Chuẩn bị Q&A — Bảo vệ nhóm A2

You are right, and that is exactly why we chose them carefully rather than conveniently. TIPU does not build roads and does not decide anything — it runs the statutory procedure on behalf of the Secretary of State and is explicitly neutral on the merits of a case. What a Transport and Works Act order can authorise, however, includes the closure or alteration of roads and footpaths as an ancillary power. That is the single hook, and we claim nothing beyond it.

So our analysis is not a funding bid. It is evidence for one of the six tests the Secretary of State weighs — "the need for the scheme", "impacts upon communities and mitigations", and "whether the scheme can reasonably be funded". When a promoter alters a rural single carriageway under a TWA order, the question "which condition on that road is doing the most harm" is directly a need-case question. Our answer is 27,792 accidents that carry 13.6 per cent of every fatal collision in the dataset while being 4.2 per cent of it.

Under Mitchell, Agle and Wood, TIPU holds power (the procedural gate) and legitimacy (statutory role), but the urgency is supplied by the evidence we bring, not by them. That makes them a dominant stakeholder, and only definitive once a case with a road component is live. We are not claiming they are definitive by default.

The report and the presentation describe the same decision from two levels, and the presentation is the more disciplined of the two. The report named the function — the engineering team that specifies physical countermeasures. Reviewing it against your checklist, we found that "a traffic safety engineering team" is a role, not an identifiable body, so it cannot be checked against any real mandate. We therefore named a real organisation with a documented statutory remit.

Nothing downstream changed: the same five variables, the same three research questions, the same target condition, the same investment order. That is the test we applied — if naming the stakeholder more precisely had changed a single number, the stakeholder would have been driving the analysis backwards. It did not, which tells us the narrowing is stakeholder-robust.

We used three measures because the three layers ask three different questions, and we should have said so explicitly in the report. The metric family is fixed — severity, casualties, accident count — and each layer uses the one that answers its own question, not whichever one looks best.

Layer 1 asks where does one accident cost more, so intensity is right: rural 2.09 casualties per fatal accident against urban 1.55, 1.42 against 1.28 for slight. Layer 2 asks where is the harm concentrated, which is a share question: single carriageways carry 70.47 per cent of rural casualties, 247,952 of 351,833. Layer 3 asks which condition kills, and only the fatal share separates the lighting states — the fatal share nearly doubles from 2.16 to 4.25 per cent while average casualties barely move, 1.46 to 1.54.

That last point is the honest version of the answer. Had we held average casualties fixed all the way down, Layer 3 would have shown almost nothing and we would have missed the effect entirely. The discriminating power of a measure is itself a finding: lighting changes who dies, not how many people are in the car.

RQ3 names both lighting and weather precisely because we did not know which one would win, and weather was the one we expected. The question does not ask "why does darkness kill" — it asks which of two environmental conditions separates outcomes. Weather lost. If the question had been written to lead, we would not have carried a losing variable through to a figure.

The sequence was exploratory first, in Tukey's sense. The overview established scale (660,679 accidents, Birmingham the largest single district at 13,491 and still under 3 per cent of the total), which told us no location could be treated on its own and forced a condition-based approach rather than a place-based one. The severity distribution — 85.3 / 13.4 / 1.3 per cent — told us fatal cells would be thin, and following King and Zeng on rare events we narrowed through splits that leave tens of thousands on both sides rather than hunting the extreme value.

The honest self-criticism: the questions in the written report were finalised after the exploration, so they read tidier than the process was. What we can defend is that each layer's entry condition was fixed before the next layer was opened, and the losing candidates are still in the document — weather in Figure 6, and three screened variables in Figure 3.

Correct, and we should raise this before you do. Rural single carriageways hold 168,010 accidents. The three lighting states we compare hold 165,755. The gap is 2,255 records — 1.34 per cent — split into "darkness, lighting unknown" (1,655) and "darkness, street lights present but unlit" (600).

We tested whether excluding them could overturn the conclusion. It cannot. Those 2,255 records carry a fatal share of 2.26 per cent and KSI of 18.4 per cent — closer to daylight than to unlit darkness. Reassigning all of them to the unlit group would move that group's fatal share from 4.25 to 4.10 per cent, still nearly double daylight. The direction, the ranking and the recommendation are unchanged.

The same test applies to the declared gaps: road type is missing for 4,520 records (0.68 per cent) and weather for 14,128 (2.14 per cent). We did not impute and we did not drop rows globally — every figure states its own population, so a reader can see which base each percentage sits on.

Because the dataset contains its own control group, and it is the comparison we lean on hardest. Compare unlit darkness with lit darkness. Both are night. Both are rural. Both are single carriageway. The fatal share is 2.59 per cent where street lighting exists and 4.25 per cent where it does not — a 39 per cent relative difference with night held constant. Speed, alcohol and fatigue do not disappear when a lamp is switched on, so they cannot be the whole of that gap.

A second signature points the same way. Accidents in the target condition involve 1.50 vehicles on average against 1.88 in daylight. Fewer vehicles per crash means single-vehicle roadway departure — a driver leaving the carriageway, not a driver colliding with someone else. That is a visual-information failure, exactly the mechanism lighting and edge delineation act on, and it is consistent with Jafari Anarkooli and Hadji Hosseinlou (2016) on two-lane rural roads.

Where we must be honest: lit and unlit rural roads are not randomly assigned. Roads get lit because they are near settlements, which means lower speeds and shorter sight distances. So the lit group is probably safer for reasons beyond the lamp, and our 39 per cent overstates the causal effect of lighting alone. That is why we present the gradient as a prioritisation signal, and borrow the causal magnitude from the controlled evidence rather than from our own cross-section.

We dropped weather as an intervention, not as an explanation — and the reason is stated before the result, not after it. No engineer can specify the weather. Under the Safe System framing, weather sits permanently in the "cannot change" column, so it was never a candidate treatment regardless of what the numbers had shown.

What the numbers did show is worth interpreting rather than explaining away. KSI is highest in fine weather at 21.1 per cent, falling to 16.7 in rain and 12.5 in snow. That inversion is what Wilde (1982) calls risk homeostasis and Edwards (1998) documented directly in recorded weather data: drivers compensate for a hazard they can see. Rain is visible, so speed falls. Darkness removes information without announcing itself, so speed does not fall — and that is precisely why the two variables belong in different columns of the decision.

There is also a measurement reason we report openly. The two measures disagree in direction. On KSI, fine weather is worse. On average casualties per accident, rain is worse — 1.51 against 1.48. A variable whose ranking flips with the ruler is not a stable basis for spending money. Lighting shows no such flip: it is worse on severity, worse on KSI, and worse on fatal share, in every weather column.

We anchor to a road condition, not a postcode, and we made that trade deliberately. The overview showed why: the largest single district holds 13,491 accidents out of 660,679 — under three per cent. No district is large enough to treat on its own, so a district ranking would have produced a list of places rather than a diagnosis of a defect.

What we deliver instead is a filter any highway authority can run against its own asset register in an afternoon: rural, single carriageway, no street lighting. That is 27,792 accidents, 42,797 casualties and 1,182 fatal collisions — 13.6 per cent of every fatal accident in Great Britain in the period, sitting on 4.2 per cent of the records. And within it we can go one level finer: the fine-weather cell alone holds 19,006 accidents at 26.6 per cent KSI, so the target is not a thin outlier — it is the second largest cell in the matrix.

The investment order follows the same logic: unlit darkness first, lit darkness second, daylight last. Daylight is by far the largest group at 123,807 accidents and it ranks last, because it has no lighting defect for a lighting budget to fix. That inversion is the recommendation — spending follows the treatable defect, not the volume.

They should not be assumed to transfer at full strength, and we do not assume it. Beyer and Ker's Cochrane review reports a fatal rate ratio of 0.34 with a confidence interval of 0.17 to 0.68 — the interval is wide, the included studies are predominantly before-and-after designs, and before-and-after lighting studies are vulnerable to regression to the mean, because lighting tends to be installed where a cluster of crashes has just occurred. Elvik's 65 per cent comes from a meta-analysis with the same structural exposure, and Elvik himself notes publication bias in this literature.

So we treat those figures as an upper bound, and our own data supplies a more conservative lower one. Within darkness, where the night is held constant, the observed gap is 4.25 against 2.59 per cent fatal — around 39 per cent, not 65. The honest planning range is therefore roughly 30 to 60 per cent, not a point estimate, and we would rather present a range we can defend than a number we cannot.

On affordability, the stakeholder's own test is "whether the scheme can reasonably be funded", so the valuation must use the department's own basis: DfT table RAS4001, value of prevention of collisions. Applying even the conservative end of the range to 1,182 fatal collisions gives a prevented-casualty value large enough to carry a lighting or delineation programme — and where full lighting is not viable on an unlit rural road, edge delineation and retroreflective treatment address the same roadway-departure mechanism at a fraction of the capital and with no energy or light-pollution liability. That is why the recommendation is deliberately two-tiered.

The workbook is not a folder of charts — it is one story of four points, and each point inherits the filter set by the point before it. Point one is the overview on the full 660,679. Layer 1 splits urban against rural on the 660,653 classified records. Layer 2 adds Urban_or_Rural_Area = Rural. Layer 3 adds Road_Type = Single carriageway, and its three worksheets — lighting, weather, and the lighting-by-weather matrix — all sit on the same 168,010 population. The narrowing is not described in the workbook, it is enforced by it: a reader can see exactly what was carried forward at each step.

Structurally the data is a single fact table of accidents with severity, casualties and vehicle count as measures, and area, road type, lighting and weather as dimensions — a star schema with one fact and four dimensions. That is why hierarchy and filter actions carry a population downward cleanly, and it is the reason Tableau suits this problem specifically rather than generically.

Two limitations shaped how we used it, and we state them rather than working around them quietly. Tableau displays an aggregate without the base behind it, so a 4.25 per cent looks identical whether it rests on forty accidents or forty thousand — which is why every figure in this deck carries its n. And Tableau has no significance testing, so each layer was rebuilt in Python directly from the source file, not through the extract. That is the point of Appendix D: no conclusion in this report depends on a setting inside the workbook.