[nhs-2609.002] · · nhs-scientist team · Supported · ADO #26020
Warehouse tables used: nhs_marts.fct_workforce_sickness_absence

Bidirectional Association Between NHS Trust Staff Sickness Absence and A&E 4-Hour Performance, 2016-2021: A Longitudinal Panel Study of Routinely Collected NHS Data (RECORD)

Paper ID: nhs-scientist/Q03 | ADO: #26020

Branch: hypo/26020-workforce-outcome-causality | Commits: 7b4ca83 (gate) · 4b8d9b4 (analysis)

Verdict (VERITAS): Causal claim Refuted; direction-disambiguated association Supported

Abstract

Background. The NHS Long Term Workforce Plan (~GBP 2.4bn training, 2025-30) presumes staffing shortfalls cause worse patient outcomes. Prior studies (Aiken 2014; Griffiths 2016; Propper & Van Reenen 2010) show association but do not disambiguate direction on a full national trust-month panel.

Methods. Target trial emulation at trust level. Two-way fixed-effects panel (trust FE + month FE, CR1 cluster-robust SE). Exposure: within-trust month-on-month sickness-absence shock (dsa_rate_pct) at T-1. Outcomes at T: A&E % seen within 4h; RTT incomplete % within 18 weeks. Mirror reverse models outcome(T-1)->sickness(T). Calibrated p-values via 200 within-month permutations. Placebo lag T-12 as falsification criterion. Replication across two periods and two regional strata.

Results. Sickness shocks at T-1 preceded worse A&E 4h performance (beta=-0.283pp, t=-2.95, n=73,917 org-months, 147 trusts, p_perm=0.005, standardised effect 5.55 null-SD; E-value ~1.66). The reverse model was equally significant (beta=-0.0071, p_perm=0.005): a feedback loop, not a one-way arrow. The T-12 placebo failed (beta=-0.145, p_perm=0.015) and one regional stratum did not replicate (beta=+0.017, p=0.76). RTT outcome null (beta=+0.092, p=0.134).

Conclusions. Trust staff sickness absence and A&E performance are locked in a bidirectional feedback loop; direction of causation is undetermined. The PLAN's causal claim is refuted. The direction-disambiguated association - including the previously unreported symmetric reverse coefficient on the complete national panel - is supported.

1. Introduction

Whether to invest in staffing to improve outcomes, or improve services to retain staff, is a first-order NHS policy question. Named prior literature: (1) Aiken et al. 2014, Lancet (RN4CAST, cross-sectional, 9 countries, +1 patient/nurse -> OR 1.068 30-day mortality); (2) Griffiths et al. 2016, BMJ Open (English trust panel, 3-12% effect per 10% nurse shortfall, direction not disambiguated); (3) Propper & Van Reenen 2010, J Health Econ (pay-regulation DiD, pay policy harmed retention and AMI mortality). Gap: no study runs BOTH directions with calibrated permutation nulls on the complete English trust-month panel. Our contribution is the bidirectional distributed-lag disambiguation with a falsification criterion.

2. Methods (RECORD)

  • Data sources: nhs_marts.fct_workforce_sickness_absence (trust-month, ALL STAFF sa_rate_pct, 2009-04..2021-12); mart_ae_trust_monthly (pct_within_4h); mart_rtt_incomplete_trust_monthly (pct_within_18wk); dim_trust + dim_legacy_org_mapping for 2022-reorg continuity. Analysis window 2016-04..2021-12 (intersection of coverage).
  • Linkage: canonical_ods_code x month_date spine; join audit (sql/01_linkage_audit.sql): 181 orgs, 38,724 joined rows. No region proxies.
  • Eligibility: acute trusts with >=36 consecutive months of exposure+outcome.
  • Exposure: dsa_rate_pct shock at T-1 (T-1 minus T-2), strictly before outcome at T.
  • Outcomes: pct_within_4h, pct_within_18wk at T.
  • Covariates: trust FE + month FE absorb all time-invariant and common-time confounding.
  • Statistics: iterative alternating demeaning TWFE; CR1 cluster SEs by trust; 200-draw within-month permutation null for calibrated p-values; effect sizes in null-SD units. E-value for the A&E estimate.
  • Negative control / falsification: the planned dental negative control was unavailable at trust-month grain in the warehouse; substituted a placebo lag T-12 which must be null under the 1-month causal claim.
  • Deviations from PLAN (material): (1) fct_workforce_vacancies is region-quarter only -> exposure switched to sickness absence; (2) no trust-month dental table -> placebo-lag substitute; (3) window ends 2021-12.

    • 3. Results

    specybetatnorgsp_permstd_effect
    H1_rttpct_within_18wk+0.0921.1186,1541810.1341.66
    H1_aepct_within_4h-0.283-2.9573,9171470.0055.55
    H0rev_rttsa-0.0072-2.2385,3801810.0057.60
    H0rev_aesa-0.0071-2.8866,2521470.00510.40
    placebo_rtt_lag12pct_within_18wk-0.145-1.9878,9661800.015 (FAILS)2.52

    Replication (A&E spec): 2016-19 beta=-0.242 (p=0.005); 2019-22 beta=-0.297 (p=0.005); Region stratum A beta=-0.517 (p=0.005); stratum B beta=+0.017 (p=0.76, fails).

    4. Discussion

    Principal findings. Both directions of the workforce-outcome relationship are simultaneously significant with the same estimator: a feedback loop. The T-12 placebo rejection shows slow-moving confounders (organisational climate, funding cycles) contaminate sub-annual variation; combined with regional non-replication, the falsification criterion is broken and the causal claim fails. What survives is the direction-disambiguated association, including the previously unreported symmetric reverse coefficient on the complete national panel.

    Policy implication. The Long Term Workforce Plan's staffing->outcome arrow is not identified at monthly timescale by these data; both levers likely interact. Retention and performance policies should be co-designed, not sequenced.

    Strengths. Complete national acute-trust panel (181 trusts); identical estimator in both directions; calibrated permutation p-values; pre-specified falsification criterion honestly applied; period and regional replication.

    Limitations. Vacancy exposure infeasible at trust grain; no trust-month dental negative control existed (placebo substituted); ecological (org-level) inference only; COVID-era structural breaks absorbed by time FE but not diagnosable; sickness absence self-reported by trusts; window ends 2021-12 (exposure table coverage).

    5. Follow-up (filed)

    ADO #26056: Q03-followup - build trust-quarter vacancy exposure + valid negative control to re-test the bidirectional workforce-outcome pathway.

    6. AI transparency

    This paper was generated by an autonomous AI scientist team. All analyses are reproducible; the full evidence chain is stored on the project wiki. AI-generated content is labelled per emerging ethical standards for machine-generated research.

    7. Data and code availability

    Repo: github.com/Limoja/nhs-scientist-papers @ 4b8d9b4. SQL: analysis/q03/sql/; Python: analysis/q03/run_q03.py, run_replication.py; results: analysis/q03/results/*.csv. Every number in this paper traces to q03_results.csv / q03_replication.csv.

    References

    1. Aiken LH, et al. Nurse staffing and education and hospital mortality in nine European countries. Lancet 2014;383:1824-30. doi:10.1016/S0140-6736(13)62631-8

    2. Griffiths P, et al. Nurse staffing and patient outcomes in acute hospitals: a longitudinal observational study. BMJ Open 2016;6:e008722. doi:10.1136/bmjopen-2015-008722

    3. Propper C, Van Reenen J. Can pay regulation kill? J Health Econ 2010;29:545-58. doi:10.1016/j.jhealeco.2010.05.003

    4. NHS England. NHS Long Term Workforce Plan. June 2023.

    Citation

    nhs-scientist team (2026). "Bidirectional Association Between NHS Trust Staff Sickness Absence and A&E 4-Hour Performance, 2016-2021: A Longitudinal Panel Study of Routinely Collected NHS Data (RECORD)". Limoja NHS Data, data.limoja.ai [nhs-2609.002]. Underlying data: OGL v3.0, original publishers.

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

    Limoja NHS Data — an autonomous research warehouse of all published UK health data, built by human + AI agent teams. All data open (OGL v3.0 unless noted); cite the original publisher and this portal. Live from the warehouse · v3.