Effect of acyclovir on herpetic ocular
recurrence using a structural nested
model (and some context)
Stephen R. Cole and Haitao Chu
Department of Epidemiology
Johns Hopkins Bloomberg School of Public Health
[email protected]
http://statepi.jhsph.edu/~scole/index.htm
9 Jan 06
Symposium on Causal Inference
1
9 Jan 06
Symposium on Causal Inference
2
9 Jan 06
Symposium on Causal Inference
3
Herpetic Eye Disease Study (HEDS)
•
HEDS Study Group NEJM 1998; 339: 300
•
703 adults with ocular HSV inactive in prior month
•
Randomized between 1992.67 and 1997.0
•
365 days of 400 mg oral acyclovir or placebo
•
Study visits at 1, 3, 6, 9, 12 months, and as needed
•
Patients and physicians well masked
•
Endpoint was first recurrence of ocular HSV
9 Jan 06
Symposium on Causal Inference
4
Characteristics at randomization
Male
Caucasian
Years of age
Severe
9 Jan 06
Acyclovir
N = 357
Placebo
N = 346
55 %
80 %
50 ± 18
39 %
52 %
78 %
48 ± 18
38 %
Symposium on Causal Inference
5
Number of days followed and treated
9 Jan 06
Symposium on Causal Inference
6
Methods
•
Nested structural AFT model
U(i) = 0∫T(i) exp[α0X(i,t)]dt
- U(i) is event time for subject i under no exposure
- T(i) is observed event time
- X(i,t) is 1 if exposed to acyclovir on day t, else 0
- exp(-α0) is the time expansion factor
•
G-estimation
- line-search across {α} for α* = α0
- α* is the member of {α} yielding ƒ[U(i, α*)|R(i)] =
ƒ[U(i,α*)], where R(i) indicates randomized group
9 Jan 06
Symposium on Causal Inference
7
Alpha {α} versus test statistic
9 Jan 06
Symposium on Causal Inference
8
Results
Intent-to-treat
Structural model
As-treated
9 Jan 06
Hazard
Ratio
95% CI
0.55
0.41
0.62
0.41, 0.75
0.22, 0.76
0.45, 0.84
Symposium on Causal Inference
9
Survival curves
9 Jan 06
Symposium on Causal Inference
10
Assumptions, Limitations
•
Assume potential outcomes of one participant are
independent of potential outcomes of others
•
Assume compliance information is correct
•
Assume no heterogeneity in effect
•
Weibull distribution assumed to obtain hazard ratio
•
Dropout model assumed correct
9 Jan 06
Symposium on Causal Inference
11
Context: We do not account for
noncompliance in biomedical RCTs
NEJM volumes 352 and 353 in 2005
203
92
≥74
0
9 Jan 06
research articles
randomized clinical trials
intention-to-treat analyses
analyses accounting for noncompliance
Symposium on Causal Inference
12
Context: When we do, it will only make
a difference roughly half the time
9 Jan 06
Symposium on Causal Inference
13
Context: When we do, it will only make
a difference roughly half the time
Mark,
Robins
CCT
1993
Cole,
Chu
CCT
2005
Joffe
SIM
2001
9 Jan 06
Korhonen, et al
SIM 1999
Symposium on Causal Inference
14
Context: Some of the times it makes no
difference are particularly insidious
U
X
Y
Z
9 Jan 06
Symposium on Causal Inference
15
Context: Some of the times it makes no
difference are particularly insidious
U
X
Z
9 Jan 06
Y
W
Symposium on Causal Inference
16
Effect of acyclovir on herpetic ocular
recurrence using a structural nested
model (and some context)
Stephen R. Cole and Haitao Chu
Department of Epidemiology
Johns Hopkins Bloomberg School of Public Health
[email protected]
http://statepi.jhsph.edu/~scole/index.htm
Thank you for your time
9 Jan 06
Symposium on Causal Inference
17