The second of a two-part talk from Richard Lilford and Sam Watson on modelling causal pathways in health services for the CLAHRC West Midlands Scientific Advisory Group meeting, 9th June 2015, Birmingham, UK
2. Modelling
β’ Representations of the world
β Models of data and models of phenomena
β’ Make our assumptions clear and transparent
3. Why?
β’ For policy we need a causal effect
β’ Usually ATE or ATET
β E.g. πΈ π π·1 β πΈ π π·0
β’ Barriers:
β Observational data
β Canβt measure endpoints
β’ But data, even observational data, tell us something
5. Outline
β’ Interested in the effect X->Y
β’ Some information on ππ
β’ Lots of information on π
X Z Y
p q
6. Outline
β’ Interested in X->Y
β’ But confounded by π’
β’ Can still identify causal effect by making use of π
X Z Y
u
7. Outline
β’ Model describes relationships between variables
β’ Can combine information on different data sources
Intervention
Upstream
endpoint
Patient
outcomes
10. CPOE ME ADE
π π =
π(π΄π·πΈ|πΆπππΈ = 1)
π(π΄π·πΈ|πΆπππΈ = 0)
=
π(π΄π·πΈ|ππΈ)π(ππΈ|πΆπππΈ = 1)
π(π΄π·πΈ|ππΈ)π(ππΈ|πΆπππΈ = 0)
=
π(ππΈ|πΆπππΈ = 1)
π(ππΈ|πΆπππΈ = 0)
Using only studies with ADE endpoint Using studies with ADE and ME endpoint
18. Weekend mortality
β’ Many studies have examined the effect of weekend admission on
risk of mortality (at least 105).
β’ In the UK the estimated relative risk 1.1-1.2 (Meacock, Doran, and
Sutton, 2015, Freemantle et al., 2012)
β’ Confounded by patient health
19. Weekend mortality
β’ Examine data that measure day of admission, mortality, and errors
β’ SPI2 data
β Patients aged >65 with acute respiratory illness
β’ Crude mortality relative risk: 1.17 [0.79, 1.60]
β’ Adjusted (age, sex, number of comorbidities) RR: 1.19 [0.79, 1.75]
β’ Similar point estimates. Under powered (n=670)
21. Weekend mortality
β’ Assumption of no relationship between errors and health may be too
strong:
β Sicker patients more exposed to risk of error
β Sicker patients more likely to die, less exposed to risk of error
Weekend
admission
Errors Mortality
Health
22. Weekend mortality
β’ Examine performance of estimators under different assumptions
using simulated data
β Two types of individual: sick v healthy. Sick 4x more likely to die.
β’ Only when there is no unobserved confounding due to health is the
βstandardβ estimator preferred, even with fairly large relationship
between errors and health.
β’ No evidence of a difference in errors by weekend or by health in
SPI2 data.
24. Expert Elicitation
β’ What happens when there are no data?
β’ 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.