Transcript Unsafety: Making no mockery of honest ad

Unsafety:
Making no mockery
of honest ad-hockery
Janet Wittes
Statistics Collaborative
ASA/FDA 2005
Topic du jour
Sleight of tongue

Game: Remove safety and add
virtue in two easy steps
DSMC
_ _ _
_ _ _ _
Sleight of tongue

Game: Remove safety and add
virtue in two easy steps
DSMC
DMC
_ _ _ _
Sleight of tongue

Game: Remove safety and add
virtue in two easy steps
DSMC
DMC
IDMC
Efficacy




Stop for “overwhelming” efficacy
=0.05; power = 90%; four looks
Probability of stopping early>70%
Early stop: estimate pulls toward
null
Our current approaches




Prespecify
Classify precisely
Give lots of data
Rely on mechanism
Divide and (un)conquer
(e.g. Neuropathy)
Event
Neuropathic pain
Neuropathy
Neuropathy NOS
Neuropathy peripheral
…
…
…
T
1
1
5
2
C
0
0
2
0
e.g. Neuropathy
Event
Anosmia
…..
Autonomic neuropathy
…
Cranial neuropathy
…
…
T
C
1
0
2
1
e.g. Neuropathy
Event
…
Parathesia
Parathesia NOS
Parathesia other
…
Peripheral motor neuropathy
Peripheral sensory neuropathy
T
C
3
4
0
2
0
1
6
3
0
2
Other examples
 Heart failure
• Separate near synonyms
• Allocate to heart and lung
Other examples
 Heart failure
• Separate near synonyms
• Allocate to heart and lung
 Bleeding: distribute over body
systems
Consider mechanism
 If you don’t get the drug, you can’t react
to it
• Eschew ITT: safety population
• Modified Daley’s Rule: censor early and often
• Don’t collect extraneous information
 Appeal to statistical conservatism
Populations





ITT
mITT
Safety: one dose of study med
ATP
Etc.
e.g., Vioxx- short follow-up
Through 36 months
With denominators
 Bresalier et al., NEJM, Feb 2005
The denominators
Mo Rx
0
1287
12 1129
18 1057
24
938
30
896
36
727
Po
1299
1195
1156
1042
1001
835
No data dredging
 We test hypotheses
 Too many type 1 errors if we dredge
What we typically present
 Current behavior
• Same tables for interim & final
analyses
• Long complete listings
 What we should do
• Interim data are different from final
• Presenting too much dulls the mind
Sentinel events
 Single event (e.g., death in a vaccine
trial)
 Several events (e.g., 3 retinal
thromboses)
 Sentinel event rate (e.g., WHI)
Lachenbruch and Wittes, SIM, to appear
How to handle
1. Identify sentinel event
2. Establish stat’l method for future
events


Need reasonable power
Type I error rate > 1-sided 0.025
Methods
 Individual
• Number of non-events until the k’th
event
Negative binomial
SPRT
• Time to the next (or k’th) event (dist’n)
 Rate: event rate in future patients
Normal
Poisson
Pitfalls
 Time is subtle
 Power is low
 Censoring is tricky
Pull-up
 Safety hypothesis:
• E.g., Cox-2 leads to 2 fold increase in
MI etc.
• Design: stop early if you reject
• Estimated relative risk must be pulled
toward 2
 Insight from Joe Heyse
The option
 Respect PI: Adjudicate adverse
events
 Precision: Reclassify, reorganize
 Mechanism:
Be an empiricist
 Dredging: Use sentinel events