Health Services Research & Outcome
Health services research tests study design, bias, outcome measures and quality improvement.
Core rules
- Randomisation reduces confounding; intention-to-treat preserves randomisation; per-protocol analysis can bias results.
- Bias types include selection, information, lead-time, recall, attrition and publication bias.
- Standardised mortality ratio (SMR) = observed deaths/expected deaths; case-mix adjustment is essential.
- Composite endpoints increase event numbers but can obscure component effects.
- Quality improvement uses PDSA cycles, run charts and balanced measures in a blame-free culture.
- NNT = 1/absolute risk reduction; relative risk reduction can exaggerate small absolute effects.
Worked example
An ICU has 120 observed deaths and 100 expected deaths by a risk model: SMR 1.20. This suggests 20% more deaths than predicted, but it does not prove poor care: the model may be miscalibrated and case-mix may differ. Investigate with adjusted analysis and review of avoidable factors.
Traps
- Treating SMR as a league table without confidence intervals.
- Replacing mortality with surrogate biomarkers.
- Lead-time bias in screening studies.
- Hawthorne effect in quality improvement projects.