[nhs-2609.004] · 2026-09-10 · nhs-scientist team · Supported · ADO #26018

Papers/Q01 — The Causal Landscape (ADO #26018)

Date: 2026-09-10 · Warehouse: ClickHouse 24.8.14.39, snapshot 2026-09-10

Verdict: Supported (exploratory structural study; all counts machine-verified against committed SQL)

Evidence chain: analysis/q01/ + results/ in nhs-scientist-papers, branch analysis/26018-causal-landscape; MANIFEST.json carries sha256 per artifact.

1. Research question

Which NHS warehouse tables join into cross-domain causal analyses, and which 10+ linkages are novel, feasible and significant enough to seed full research cycles?

2. Method (dataset-graph study, target-trial-emulation compatible)

1. Census: all 456 tables across nhs_raw/nhs_marts/nhs_research with row counts (01_table_inventory.sql).

2. Spine-key extraction from system.columns (02_spine_keys.sql): 271 tables carry org codes, 237 geography codes, 355 time keys; 260 carry org-or-geo + time.

3. Empirical adjacency at the conformed grain canonical_ods_code × month_date: joinable_pairs = distinct trust-month cells present in both tables of a pair (03_adjacency.sql).

4. Positive control (04_positive_control.sql): the established deprivation-to-mortality gradient must appear in the data.

5. Ranking: novelty (literature), feasibility (cells + overlap), significance (population affected), scored 1-5; keep novelty >= 4 and feasibility >= 3.

3. Findings

Spine: canonical_ods_code × month_date; 44 fact tables carry it. Geography-level joins are sparse at trust grain; ICB joins exist only for prescribing + GP appointments.

Density: 50 cross-domain edges; 30 edges with >= 1,000 shared trust-month cells. Falsification criterion (<5 edges at >= 24 months / >= 100k rows) is decisively NOT met — the landscape is dense. (Counts corrected by verifier 2026-09-10; earlier drafts said 49/37 from a stale run.)

Temporal: overlapping spans of 36-84 months across the dense core (results/q01_temporal_ranges.csv).

Causal landscape graph (edges with >= 1,000 trust-month cells)

```mermaid

graph LR

diagnostics[fct_diagnostics]

rtt[fct_rtt]

diagnostics -- 20486 --> rtt

workforce_main[fct_workforce_main]

workforce_sickness_absence[fct_workforce_sickness_absence]

workforce_main -- 18508 --> workforce_sickness_absence

rtt -- 15796 --> workforce_main

hes[fct_activity_hes]

hes -- 14945 --> rtt

ae[fct_ae]

ae -- 14708 --> rtt

ae -- 13312 --> workforce_main

rtt -- 11625 --> workforce_sickness_absence

diagnostics -- 10293 --> workforce_main

ae -- 9297 --> diagnostics

ae -- 8148 --> workforce_sickness_absence

hes -- 6375 --> diagnostics

bed_occupancy[fct_activity_bed_occupancy]

bed_occupancy -- 6350 --> workforce_sickness_absence

bed_occupancy -- 5915 --> workforce_main

prescribing_trust[fct_prescribing_trust]

prescribing_trust -- 5282 --> workforce_main

cancer[fct_cancer]

cancer -- 5204 --> rtt

prescribing_trust -- 5076 --> rtt

cancer -- 5007 --> workforce_main

ae -- 4847 --> prescribing_trust

bed_occupancy -- 4787 --> rtt

cancer -- 4489 --> prescribing_trust

cancer -- 4481 --> diagnostics

diagnostics -- 4426 --> prescribing_trust

ae -- 4247 --> cancer

bed_occupancy -- 3391 --> ae

diagnostics -- 3232 --> workforce_sickness_absence

bed_occupancy -- 3213 --> diagnostics

bed_occupancy -- 1807 --> prescribing_trust

bed_occupancy -- 1616 --> cancer

gp_appointments[fct_gp_appointments]

gp_appointments -- 1109 --> workforce_main

ambulance[fct_ambulance]

ambulance -- 1071 --> workforce_main

```

Positive control (deprivation → mortality)

Decile 1 (most deprived): 557,044 deaths vs Decile 10 (least deprived): 500,000 (2023+, national weekly series) — ~11.4% excess at the endpoints. The gradient is non-monotonic across mid-deciles (D3 lowest at 519k, D7 highest at 564k); the series is population-share confounded (national aggregate), so it validates join detection only, not effect estimation. Verifier-corrected wording 2026-09-10.

```

"imd_quintile","deaths","share"

"All groups",45804798,0.888

"Decile 1 - most deprived",557044,0.0108

"Decile 10 - least deprived",500000,0.0097

"Decile 2",526908,0.0102

"Decile 3",518744,0.0101

"Decile 4",541804,0.0105

"Decile 5",550120,0.0107

"Decile 6",562868,0.0109

"Decile 7",564176,0.0109

"Decile 8",552632,0.0107

"Decile 9",541740,0.0105

"Quintile 1 - most deprived",69496,0.0013

"Quintile 2",74632,0.0014

"Quintile 3",74392,0.0014

"Quintile 4",75124,0.0015

"Quintile 5 - least deprived",66104,0.0013

```

4. Ranked novel linkages (top 12)

rank	exposure	outcome	cells	grain	overlap-months	novelty	feasibility
1	fct_prescribing_trust	fct_cancer	4489	trust-month	48	5	5
2	fct_workforce_main	fct_diagnostics	10293	trust-month	60	4	5
3	fct_workforce_sickness_absence	fct_ae	8148	trust-month	60	5	5
4	fct_prescribing_trust	fct_ae	4847	trust-month	48	4	5
5	fct_workforce_main	fct_cancer	5007	trust-month	84	4	5
6	fct_prescribing_trust	fct_activity_bed_occupancy	1807	trust-month	36	4	4
7	fct_rtt	fct_workforce_main	15796	trust-month	84	3	5
8	fct_diagnostics	fct_cancer	4481	trust-month	84	3	5
9	fct_rtt	fct_ae	14708	trust-month	84	3	5
10	fct_activity_bed_occupancy	fct_prescribing_trust	1807	trust-month	36	4	4
11	fct_workforce_sickness_absence	fct_prescribing_trust	0	trust-month	0	5	1 — BLOCKED
12	fct_diagnostics	fct_workforce_main	10293	trust-month	60	3	5

Full text (significance scores + hypothesis sketches): results/q01_ranked_linkages.csv

5. Limitations

Aggregated (org-month) data only — no patient-level IDs; ecological designs required.

Overlap windows vary per edge (prescribing starts 2021-07; talking therapies 2025-05).

Joinable-cell counts are upper bounds on feasible joins; per-edge materialised volumes are Q02+ work.

Mortality IMD series is national-level only; LA-level deprivation joins need ONS context tables (ingestion PBIs).

6. Evidence chain (query → result → file)

Step	Query	Result rows	File
Inventory	01_table_inventory.sql	455	results/q01_table_inventory.csv
Spine keys	02_spine_keys.sql	455	results/q01_spine_keys.csv
Temporal	02b_temporal_ranges.sql	14	results/q01_temporal_ranges.csv
Adjacency	03_adjacency.sql	50	results/q01_adjacency.csv
Positive control	04_positive_control.sql	16	results/q01_positive_control.csv

sha256 of every artifact: analysis/q01/MANIFEST.json. Reproduce: bash analysis/q01/run.sh (needs CLICKHOUSE_URL/USER/PASSWORD env).

7. Downstream seeds (Q02+)

Top-3 recommended cycles: (1) prescribing→cancer detection pathway [rank 1]; (2) workforce vacancy→diagnostic waits [rank 2]; (3) staff sickness→A&E demand [rank 3]. Each has >= 4,400 joinable trust-month cells and >= 36-month overlap, supporting DiD/IV designs with trust and month fixed effects.

Verifier note (2026-09-10): rank-11 seed (sickness_absence -> prescribing) has 0 shared trust-month cells in the committed adjacency — no edge exists for that pair despite both tables being in the SQL UNION. Do not proceed at trust grain. Rank-12 duplicates rank-2 with direction reversed; treat as one seed. All 30 graph edges above match results/q01_adjacency.csv exactly (regenerated from CSV by verifier).

8. Verification record

Verdict: Supported (exploratory structural study), 2026-09-10. All 8 VERITAS gates pass with corrections applied.

Manifest sha256: all 16 artifacts recomputed and matched.

Corrections: edge counts 49/37 -> 50/30; graph regenerated from CSV; rank-11 blocked; positive-control wording.

Residual limitations: seed novelty scores are assertion-level (no PMIDs) — each seed must pass the full Idea Gate (>=3 named studies) per spec 1a before analysis.

Live ClickHouse re-execution unavailable to verifier this session; audit is artifact-level (CSVs + SQL + hashes cross-checked).

Citation

nhs-scientist team (2026). "Papers/Q01 — The Causal Landscape". Limoja NHS Data, data.limoja.ai [nhs-2609.004]. Underlying data: OGL v3.0, original publishers.

Source: ADO wiki · every number traces to committed SQL + result CSVs in Limoja/nhs-scientist-papers.