Modelling causal pathways in health services part 2 - Sam Watson
Download as PPTX, PDF1 like452 views
The document discusses the complexities of modeling causal pathways in health services, particularly the challenges posed by observational data and confounding variables. It highlights the Bayesian causal networks approach to identify causal effects, as well as studies examining the impact of weekend admissions on mortality. Additionally, it addresses the use of expert elicitation when data is unavailable, emphasizing the need for transparent assumptions in health policy modeling.
![Weekend mortality
• Examine data that measure day of admission, mortality, and errors
• SPI2 data
– Patients aged >65 with acute respiratory illness
• Crude mortality relative risk: 1.17 [0.79, 1.60]
• Adjusted (age, sex, number of comorbidities) RR: 1.19 [0.79, 1.75]
• Similar point estimates. Under powered (n=670)](https://image.slidesharecdn.com/modellingcausalpathwaysinhealthservicespart2-sw-150612114342-lva1-app6892/85/Modelling-causal-pathways-in-health-services-part-2-Sam-Watson-19-320.jpg)
![Weekend mortality
• Front-door estimator
• 𝑃 𝑌 𝑊 = 𝑤 = 𝐸 𝑃 𝐸 𝑊 = 𝑤 𝑊′ 𝑃 𝑌 𝑊′, 𝐸 𝑃 𝑊′
• RR: 1.03 [1.00, 1.06]](https://image.slidesharecdn.com/modellingcausalpathwaysinhealthservicespart2-sw-150612114342-lva1-app6892/85/Modelling-causal-pathways-in-health-services-part-2-Sam-Watson-20-320.jpg)
Modelling causal pathways in health services part 2 - Sam Watson
- 1. Modelling causal pathways in health services, part 2 12/06/2015
- 2. Modelling • Representations of the world – Models of data and models of phenomena • Make our assumptions clear and transparent
- 3. Why? • For policy we need a causal effect • Usually ATE or ATET – E.g. 𝐸 𝑌 𝐷1 − 𝐸 𝑌 𝐷0 • Barriers: – Observational data – Can’t measure endpoints • But data, even observational data, tell us something
- 5. Outline • Interested in the effect X->Y • Some information on 𝑝𝑞 • Lots of information on 𝑝 X Z Y p q
- 6. Outline • Interested in X->Y • But confounded by 𝑢 • Can still identify causal effect by making use of 𝑍 X Z Y u
- 7. Outline • Model describes relationships between variables • Can combine information on different data sources Intervention Upstream endpoint Patient outcomes
- 10. CPOE ME ADE 𝑅𝑅 = 𝑃(𝐴𝐷𝐸|𝐶𝑃𝑂𝐸 = 1) 𝑃(𝐴𝐷𝐸|𝐶𝑃𝑂𝐸 = 0) = 𝑃(𝐴𝐷𝐸|𝑀𝐸)𝑃(𝑀𝐸|𝐶𝑃𝑂𝐸 = 1) 𝑃(𝐴𝐷𝐸|𝑀𝐸)𝑃(𝑀𝐸|𝐶𝑃𝑂𝐸 = 0) = 𝑃(𝑀𝐸|𝐶𝑃𝑂𝐸 = 1) 𝑃(𝑀𝐸|𝐶𝑃𝑂𝐸 = 0) Using only studies with ADE endpoint Using studies with ADE and ME endpoint
- 13. Nuckols et al.
- 18. Weekend mortality • Many studies have examined the effect of weekend admission on risk of mortality (at least 105). • In the UK the estimated relative risk 1.1-1.2 (Meacock, Doran, and Sutton, 2015, Freemantle et al., 2012) • Confounded by patient health
- 19. Weekend mortality • Examine data that measure day of admission, mortality, and errors • SPI2 data – Patients aged >65 with acute respiratory illness • Crude mortality relative risk: 1.17 [0.79, 1.60] • Adjusted (age, sex, number of comorbidities) RR: 1.19 [0.79, 1.75] • Similar point estimates. Under powered (n=670)
- 20. Weekend mortality • Front-door estimator • 𝑃 𝑌 𝑊 = 𝑤 = 𝐸 𝑃 𝐸 𝑊 = 𝑤 𝑊′ 𝑃 𝑌 𝑊′, 𝐸 𝑃 𝑊′ • RR: 1.03 [1.00, 1.06]
- 21. Weekend mortality • Assumption of no relationship between errors and health may be too strong: – Sicker patients more exposed to risk of error – Sicker patients more likely to die, less exposed to risk of error Weekend admission Errors Mortality Health
- 22. Weekend mortality • Examine performance of estimators under different assumptions using simulated data – Two types of individual: sick v healthy. Sick 4x more likely to die. • Only when there is no unobserved confounding due to health is the ‘standard’ estimator preferred, even with fairly large relationship between errors and health. • No evidence of a difference in errors by weekend or by health in SPI2 data.
- 24. Expert Elicitation • What happens when there are no data? • Can use expert elicitation. Figure: Example group subjective prior, from Yao et al. (2012) BMJ Qual Saf. See also Lilford et al. (2014) BMC Health Serv Res.
- 25. Conclusions