NHS Bed Occupancy Analysis

Replication code and results for The 85% Myth, published in The Bevan Briefing.

Summary

Analysis of approximately 5,400 trust-quarters covering ~140 NHS acute trusts in England from 2018-19 to 2024-25. Key findings:

• No threshold effect at 85% — or at any level from 80% to 92% — for A&E performance, 12-hour waits, or cancelled operations
• Occupancy risk is mediated by bed base size and geographic position, not by crossing any universal percentage
• Trusts within 10km competitor rings are substantially buffered against occupancy pressure (p < 10^-10 for 12-hour waits)
• For isolated trusts (e.g. Royal Cornwall), the 85% target is structurally unachievable and represents surveillance without agency

Structure

analysis/
  cleaning/           Data cleaning pipeline
    00_clean_eric.py          ERIC estate data (backlog, age profiles)
    load_master.py            Load panel into DuckDB
    01a-clean-kh03.py         KH03 bed occupancy
    02a-clean-ae.py           A&E SitRep
    03a-clean-canc.py         Cancelled operations
  04-merge-quarterly-datasets.py   Panel assembly
  05a-regressions_*.py        Threshold and spline models
  05b-regressions_DEBUGGED.py Full interaction model (geographic terms)
  01_within_between.py        Mundlak within-between decomposition
  02_covid_subperiods.py      Structural period analysis
  03_iv_2sls.py               IV/2SLS estimation (estate instruments)

tools/
  occupancy_target_tool.py          Trust-specific ceiling calculator
  01-build_proximity_features.py    Compute proximity features from ODS
  streamlit_app.py                  Interactive web version

outputs/
  models_threshold.csv              Threshold models (80-92%)
  models_spline.csv                 Piecewise linear spline models
  models_capacity_interactions.csv  Occupancy x bed base interactions
  models_scaled_interactions.csv    Full geographic interaction model
  models_scaled_threshold.csv       Threshold x geography interactions
  models_size_quintiles.csv         Effects by trust size quintile
  regression_results.csv            Primary results table

data/
  master-quarterly-trust.csv        Trust-quarter panel (all sources)

Quickstart

pip install pandas numpy statsmodels linearmodels duckdb matplotlib openpyxl

Run in order:

python analysis/cleaning/00_clean_eric.py
python analysis/cleaning/load_master.py
python analysis/05b-regressions_DEBUGGED.py
python analysis/01_within_between.py
python analysis/02_covid_subperiods.py
python analysis/03_iv_2sls.py

Local occupancy ceiling calculator

python tools/occupancy_target_tool.py REF
python tools/occupancy_target_tool.py RRV --edr 0.38
python tools/occupancy_target_tool.py        # interactive prompt

Data sources

All source data is publicly available from NHS England:

• KH03 bed availability and occupancy
• A&E Attendances and Emergency Admissions SitRep
• Urgent and Emergency Care daily SitRep
• Cancelled Operations quarterly return
• ERIC NHS estates and facilities return

Trust geographic proximity features computed from ODS postcode data.

Citation

Bevan Briefing (2025). The 85% Myth: What the Data Actually Says
About NHS Bed Occupancy. https://thebevanbriefing.com

Licence

Code: MIT. Data: Open Government Licence v3.0.