Confounding in epidemiological studies, is SCCS the answer?

advertisement
08/08/2013
Confounding in epidemiological studies, is
SCCS the answer?
SCCS for censored, perturbed, or curtailed postevent exposures
Linda Wijlaars
PhD student
Department of Primary Care and Population Health
UCL
Why should I use a complicated little-used
method when I could also just do a
cohort/case control study?
Main limitation of SCCS method:
exposure distribution + observation period ⊥ event times
can’t use SCCS:
→ when event alters subsequent exposure risk
→ for terminal events
→ if observation of exposure is censored/disrupted by event
2
1
08/08/2013
Solution using counterfactuals
Counterfactuals:
What would have
happened if patient 1
would not have
received treatment?
Patient
Treatment Death
group
1
No
Yes
2
Yes
No
3
Yes
No
4
Yes
Yes
5
No
No
3
Patient 1
X=1, Y=1
C
Y
X
C
Patient 1
X*=0, Y*=?
X*
Y*
4
2
08/08/2013
Star Trek II: counterfactuals
Wrath of Khan
Khan=marooned
Khan
Into Darkness
Khan≠marooned
Khan
5
Counterfactuals & the SCCS
What if there were no subsequent exposures?
→ counterfactual used to estimate effects as if
there were no subsequent exposures
6
3
08/08/2013
Counterfactuals & the SCCS
Even if observation is censored, potential end of
observation period often known
 bi (factual or counterfactual) is known
 Event=non-recurrent so no events in unobserved period
Potential exposure = unknown
Event
Start of
observation period
ai
Start of risk
period ci1
Start of risk
period ci2
End of
observation period
7
bi
Counterfactuals & the SCCS
No assumptions about event-free exposure process
Analyse data for each exposure as if there could be
no subsequent exposures → counterfactual
Recursive principle: start at last pre-event exposure,
work back through previous exposures
8
4
08/08/2013
Counterfactuals & the SCCS
9
Assumptions SCCS for censored post-event
exposures
Exposure is binary
Risk period is (relatively) short
Event of interest is uncommon & non-recurrent
Risk returns to baseline level after each risk period
10
5
08/08/2013
Example: OPV and intussusception
Farrington CP et al. Biostatistics (2009)
Oral Polio Vaccine (OPV) – given in 3 doses
Intussusception = side effect & contraindication
HES data from 1991-1997
207 children aged 28-365 days
Up to 3 doses, risk period: 14-41 days after
vaccination
11
OPV and intussusception
4 analyses: standard model = standard SCCS
Analysis 1: gold standard
Analysis 2-4: censoring simulated
Dose
Analysis 1:
original data
Analysis 2:
observation
ends at event
Analysis 3:
censored data
Analysis 4:
censored data
Standard model
Standard model
Standard model
Censoring model
RI (95% CI)
RI (95% CI)
RI (95% CI)
RI (95% CI)
1
0.71 (0.33-1.41)
0.58 (0.05-3.80)
0.85 (0.37-1.65)
0.58 (0.26-1.17)
2
0.92 (0.50-1.64)
0.48 (0.17-1.17)
1.43 (0.73-2.66)
0.88 (0.44-1.63)
3
1.63 (1.04-2.59)
1.35 (0.49-5.10)
2.91 (1.69-5.02)
1.57 (1.00-2.52)
12
6
08/08/2013
Example 2: RV1 and intussusception
Patel MM et al. NEJM (2011)
Rotavirus cause of severe gastroenteritis in infants
and young children
– Worldwide 95% of children will be infected before age 5
Old rotavirus vaccine association with
intussusception: 1 : 10,000 recipients
New vaccine (RV1): added to national immunization
program in Mexico (2007)
13
RV1 and intussusception
Cases characteristics
N = 285
Age in months (range)
5 (1.5 – 8.0)
Boys
61%
Death
1%
Surgical treatment
87%
RV1 vaccination
Dose 1
95%
Dose 2
13%
14
7
08/08/2013
RV1 and intussusception
15
RV1 and intussusception
Dose and risk
Period
Cases
285
Controls
739
SCCS
(IRR)
CC
(OR)
91%
91%
-
1.0 (0.6-1.7)
1-7 days
9%
2%
5.3 (3.0-9.3)
5.8 (2.6-13.0)
8-14 days
2%
2%
1.1 (0.5-2.7)
1.0 (0.4-2.9)
15-21 days
2%
3%
0.9 (0.3-2.2)
0.8 (0.3-2.1)
5%
5%
1.8 (0.9-3.8)
1.1 (0.6-2.2)
Either dose
First dose
Second dose
1-7 days
8-14 days
7%
4%
2.2 (1.1-4.2)
2.3 (1.2-4.4)
15-21 days
7%
4%
2.2 (1.2-4.0)
2.0 (1.0-3.8)
16
8
08/08/2013
RV1 and intussusception
Very similar results between CC and SCCS
Slight increase in intussusception, but:
– Many more rotavirus deaths/hospitalizations averted
than intussusception d/h caused
17
Example 3: Antidepressants and suiciderelated behaviour
2003/04: MHRA warning against SSRIs
Possible association antidepressants ↔ ‘suicidality’
Trials: small and short
 Suicide-study needs large and long
Problem: confounding by (severity of) indication
18
9
08/08/2013
Cohort characteristics
Total
Girls
# taking
ADs
Completed
suicide
Attempted
suicide
Suicidal
ideation
Self-harm
General
population
81
1,496
1,178
2,361
952,892
29.6%
72.8%
60.1%
74.2%
48.4%
23.1%
36.6%
52.8%
5.7%
2.9%
19
10
08/08/2013
Day zero effect
Prescription day effect: recording artefact?
 Patient discloses earlier suicide-related event at GP
visit
 GP decides to prescribe antidepressant
 Both are recorded on same day
Prescription might be started by psychiatrist
21
Why should I use a complicated little-used
method when I could also just do a
cohort/case control study?
• (unmeasured) confounding + selection bias
• Case-only eliminates time-invariant confounders
• Efficient compared to cohort studies
• Code (also for extensions) available online
22
11
08/08/2013
Presentations available on UCL THIN page
bit.ly/UCLTHINteam
23
References
Farrington CP, Whithaker HJ, Hocine MN. Case series for analysis for censored, perturbed,
or curtailed post-event exposures. Biostatistics (2009), 10(1)3-16.
Nordmann S, Biard L, Ravaud P, Esposito-Farèse M. Case-only designs in
pharmacoepidemiology: a systematic review. PLoS one (2012).
Patel MM et al. Intussusception risk and health benefits of rotavirus vaccination in Mexico
and Brazil. New England Journal of Medicine (2011), 364(24)2283-92.
Pearl J. Causality: Models, Reasoning and Inference. Cambridge University Press (2000).
Whitaker H. (2010) The Self-Controlled Case Series Method. URL:
http://statistics.open.ac.uk/sccs
The Health Improvement Network (THIN) Research Team. URL:
http://www.ucl.ac.uk/pcph/research-groups-themes/thin-pub
24
12
Download