BGPT: Paper Review: Factors associated with hospital deaths in the oldest old: a cross-sectional study

Explore by Goal

Quick Explanation Copied

Key result (England, age ≥85)

In a linked ONS–HES cross-sectional cohort (n=671,178), 62% died in hospital, and multivariable odds of hospital death were higher with greater comorbidity, more emergency admissions, higher deprivation, and non-malignant causes; odds were lower for care-home residence and for dementia mentioned on the death certificate.

Long Explanation

Paper review (skeptical, evidence-based): oldest-old hospital deaths

“Factors associated with hospital deaths in the oldest old: a cross-sectional study” — Age and Ageing, March 5, 2016.

1) Study design & data lineage (what is directly supported)

Design: cross-sectional, using linked ONS death registrations and Hospital Episode Statistics (HES) in England (2008–2012).
Analytic population: 671,178 adults aged ≥85 at death who had hospital admissions in the 12 months before death.
Exclusions described: from a broader death cohort of 855,492, ~22% were excluded for having no HES activity in the prior 12 months; 69 excluded for incomplete HES data.
Outcome: likelihood of dying in hospital vs dying elsewhere.
Covariates (as coded in the paper): gender, age group (85–89, 90–94, 95+), deprivation quintile, care-home residence at death (inferred from postcode on death certificate), cause-of-death categories, modified Charlson comorbidity group (from diagnoses recorded on hospital admissions in the last year), and mentions of dementia or hip fracture on death certificates.
Modeling approach: chi-squared associations followed by multivariable binary logistic regression reporting odds ratios (ORs) and 95% CIs.

2) Data-derived visuals (from the paper’s reported results)

2.1 Hospital vs elsewhere (overall proportion)

Source: hospital death proportion reported as 62% in the paper abstract/results.

2.2 Odds ratios: cause-of-death categories (multivariable)

Source: multivariable regression table values for heart, renal, respiratory, cerebrovascular, and other vs cancer reference.

2.3 Key sociodemographic/certificate covariates (care-home, dementia, hip fracture, sex, age)

Source: multivariable ORs reported for care-home residence, dementia mention, hip fracture mention, male sex, and age groups.

2.4 Comorbidity gradient (Charlson groups) & emergency admissions dose-response

Comorbidity source:
Emergency admissions source:

3) What the paper claims—and what we can and cannot infer

3.1 Directly supported associations (correlational, not causal)

Comorbidity: higher Charlson group associated with higher odds of dying in hospital.
Emergency admissions: odds of hospital death increase with more emergency admissions in the last year (with 1 emergency admission as reference in the model).
Age within the oldest-old: compared with 85–89, odds are ~unchanged for 90–94 (OR 0.99, CI 0.98–1.00) and reduced for ≥95 (OR 0.91, CI 0.89–0.92).
Care-home residence: care-home residence at death is strongly associated with lower odds of hospital death (OR 0.34).
Cognitive label on death certificate: “dementia mentioned” on the death certificate is associated with lower odds of hospital death (OR 0.32).
Hip fracture mention: hip fracture mentioned on the death certificate is associated with higher odds of hospital death (OR 2.47).

3.2 Skeptical interpretation: plausible mechanisms vs measurement/selection artifacts

Selection bias is explicit in the paper: excluding ~22% of deaths with no HES activity may change who appears in the analysis (final sample differs sharply in hospital-death proportion: ~9.4% vs 62% for excluded subset).
Outcome timing is constrained: the cohort uses hospital admissions within 12 months before death; the paper notes you cannot know admission timing relative to death beyond what is inferable from windowing.
Administrative-data measurement error: HES/ONS coding inaccuracy and incomplete records can distort comorbidity measurement and cause-of-death categories; dementia on certificates may be under-diagnosed/under-reported.
Charlson index depends on coding depth: Charlson score relies on hospital coding and can be influenced by length of admission (fewer comorbidities may be identified in short admissions).
Residual confounding remains possible: even with multivariable regression, unmeasured factors affecting both care trajectory and place of death (e.g., underlying severity patterns not captured by the coding schema) could remain. (The paper frames these as limitations and notes causality cannot be inferred.)

4) Directed “what would disprove this?” falsification checklist (from the study’s own structure)

If the selection mechanism is driving effects: replicating the analysis in a way that includes deaths with no prior HES activity (or harmonizes inclusion rules) should test whether the strong association pattern (especially care-home and dementia-certificate effects) persists.
If dementia/hip fracture labels reflect coding/reporting rather than biology/trajectory: comparing outcomes using alternative dementia ascertainment (e.g., multiple sources) would test robustness of the OR for “dementia mentioned on the death certificate” and would address known under-reporting.
If Charlson and emergency-admission counts are proxies for coding intensity: validating comorbidity scoring against independent clinical assessment or examining whether results track with objective illness severity measures would falsify the “more comorbidity/emergency admissions → more hospital death” interpretation.
If age-gradient is an artifact of survivorship within the ≥85 window: re-deriving age effects using alternative age cutpoints or including additional age strata around 80–84 could test whether “older age → lower odds after 85” is stable.

5) Author-review links

Feedback:

Updated: April 12, 2026