“let’s get down to whatis reallytypes of missingness • missing at random (mar) –given the...
TRANSCRIPT
-
“Let’s get down to whatis really
wrong”
– what makes clinical sense when a
neuroscience trial has missing data?
-
Summary of this presentation
• Missing at random; missing not at random.
• Deciding what you want to estimate.
• Efficacy attributable to the experimental
treatment – basing assumptions on
evidence.
-
Types of missingness
• Missing at random (MAR)
– Given the model & observed data, missing-
ness is independent of the missing data.
– Corollary: one can model missing outcomes
based on the observed outcomes.
• Missing not at random (MNAR)
– Given the model & observed data, missing-
ness is not independent of the missing data.
– Corollary: we cannot estimate missing
outcomes from observed.Rubin (1978)
-
Types of missingness
• Missing at random (MAR)
– Given the model & observed data, missing-
ness is independent of the missing data.
– Corollary: one can model missing outcomes
based on the observed outcomes.
• Missing not at random (MNAR)
– Given the model & observed data, missing-
ness is not independent of the missing data.
– Corollary: we cannot estimate missing
outcomes from observed.
But withdrawals are probably
different than subjects who
remain in study!
-
Bias from missing data?
• Missing data could bias an estimate of
treatment effect…
• but bias depends on what you want to
estimate.
-
What do you want to estimate?
• The “estimand”
• Estimand should serve the needs of
patients, of clinicians, of regulators.
– Estimand may imply assumptions about
missings.
-
What do you want to estimate?
• The “estimand”
• Estimand should serve the needs of
patients, of clinicians, of regulators.
– Estimand may imply assumptions about
missings.
...perhaps
counterfactual
assumptions.
-
What do you want to estimate?
• What estimand would serve the needs of
patients? Treatment effect…
– if I adhere to regimen?
– observed in study population?
– taking into account that not all subjects
completed the course of treatment?
-
What do you want to estimate?
• What estimand would serve the needs of
patients? Treatment effect…
– if I adhere to regimen?
– observed in study population?
Per protocol
estimate
Completers only
– taking into account that not all subjectscompleted the course of treatment?
Withdrawals
are different:
“treatment
policy” or
MNAR
estimate
-
What do you want to estimate?
– if I adhere to regimen?
– observed in study population?
– if none had stopped treatment early
• What estimand would serve the needs of
patients/clinicians? Treatment efPfeer cprto…tocolestimate
Completers only
Model withdrawals
based on
observed: MAR
– taking into account that not all subjectscompleted the course of treatment?
Withdrawals
are different:
“treatment
policy” or
MNAR
estimate
-
What do you want to estimate?
• What estimand would serve the needs of
regulators? Treatment effect…
– if I adhere to regimen?
– observed in study population?
– taking into account that not all subjects
completed the course of treatment?
– if none had stopped treatment early
-
What do you want to estimate?
• What estimand would serve the needs of
regulators? Treatment effect…
– if I adhere to regimen?
– observed in study population?
– taking into account that not all subjects
completed the course of treatment?
– if none had stopped treatment early
-
What do you want to estimate?
• What estimand would serve the needs of
regulators? Treatment effect…
– if I adhere to regimen?
– observed in study population?
– taking into account that not all subjects
completed the course of treatment?
– if none had stopped treatment early
-
What do you want to estimate?
• What estimand would serve the needs of
regulators? Treatment effect…
– if I adhere to regimen?
– observed in study population?
– taking into account that not all subjects
completed the course of treatment?
– if none had stopped treatment early
-
What do you want to estimate?
• What estimand would serve the needs of
regulators? Treatment effect…
– if I adhere to regimen?
– observed in study population?
– taking into account that not all subjects
completed the course of treatment?
– if none had stopped treatment early
Perhaps
some form
of this
estimand?
-
What do you want to estimate?
• What estimand would serve the needs of
regulators? Treatment effect…
– if I adhere to regimen?
– observed in study population?
– taking into account that not all subjects
completed the course of treatment?
– if none had stopped treatment early
Often referred to as
“effectiveness estimand” in the
literature
Perhaps
some form
of this
estimand?
-
What do you want to estimate?
• What estimand would serve the needs of
Often referred to as
“effectiveness estimand” in the
literature
Perhaps
some form
of this
estimand?
regulators? Treatment effect…
– if I adhere to regimen?
– observed in study population?
– taking into account that not all subjects
completed the course of treatment?
– if none had stopped treatment early
Could also be“Treatment policy”
estimand
-
From estimand to estimate
• Suppose we choose estimand “the
treatment effect taking into account that
not all subjects completed the course of
treatment”.
• For example, choose “effect attributable to
study treatment” (Mallinckrodt et al. 2014).
• Historic data can help us choose plausible
assumptions for missing outcomes.
• Two neuroscience examples….
-
Plausible assumptions
• Scenario: superiority study; control arm is
standard of care or placebo; study
estimand is effectiveness: “effect
attributable to study treatment”.
• If subject withdraws early, this estimand
implies no continuing benefit from
experimental treatment over and above
control except what would occur due to
residual drug in the body.
-
Assumptions using historic data
• Can we identify likely trajectory of efficacy
after withdrawal, assuming no continuing
benefit from experimental treatment over
control except what would occur due to
residual drug in the body?
• Can we identify likely trajectory of
withdrawals, relative to that of control?
• If yes: we have justifiable assumptions for
this effectiveness estimand.
-
Example 1: major depressive
disorder: MDD
-
Historic data: typical trajectories
for MDD
Brannan et al. (2005)
-
Historic data: typical trajectories
for MDD
-
Historic data: typical trajectories
for MDDA variety of assumptions based
on trajectories of control group
are implementable via multiple
imputation
-
Historic data: typical trajectories
for MDD
?
“Jump to reference” (J2R)
-
Historic data: typical trajectories
for MDD
?
Imputations at mean of
reference
-
Historic data: typical trajectories
for MDD
?
Model experimental arm
using reference data only:
“Copy reference” (CR)
-
Using post-withdrawal data
• Motivation to collect post-withdrawal
efficacy in early studies
• If post-withdrawal data not available from
standard studies, may be available from
randomized withdrawal studies
Apr2014
Perahia et al. (2006)
-
Ranomized withdrawal, MDD
-
Ranomized withdrawal, MDD
-
Ranomized withdrawal, MDD
Rapid worsening
to placebo-like efficacy
-
Select suitable control-based
trajectory?
-
Select suitable control-based
trajectory?
Imputations at mean of
reference may be
clinically reasonable,
based on the historic data
-
Select suitable control-based
trajectory?
“Jump to reference”
assumption could be a
conservative choice, also
clinically justifiable.
-
Using post-withdrawal data
• Based on historic data, clinically
interpretable (and clinical reasonable)
assumptions for missing data could be
– Imputations for experimental arm have the
mean of the placebo group at each visit, or
– Imputations for experimental arm follow the
“jump to control” assumption: withdrawal from
experimental arm, usually with poor
HAMDT17, will be imputed to perform
similarly poorly in the control arm.
Apr2014
-
Example 2: Insomnia
• Efficacy score total sleep time (TST)
– higher is better.
• About 20% withdrawals expected in each
treatment group.
• Experimental treatment has a mechanism
of action similar to the orexin receptor
antagonists (ORAs).
T ds and Innovations in Clinical Trial Statistics: Missing data, developments in practice, copyright MOKelly 2014Apr2014
-
• Data on post-withdrawal TST in an ORA is
available in recent randomized withdrawal
study available as FDA Briefing
Document.
T ds and Innovations in Clinical Trial Statistics: Missing data, developments in practice, copyright MOKelly 2014Apr2014
Example 2: Insomnia
-
ORA randomized withdrawal
FDA Peripheral & Central Nervous System Drugs Advisory Committee Meeting (2013)
-
ORA randomized withdrawal
-
ORA randomized withdrawal
-
ORA randomized withdrawal
?Imputations at mean of reference
-
ORA randomized withdrawal
?Imputations at mean of reference˟
-
ORA randomized withdrawal
?“Jump to reference” (J2R)
-
ORA randomized withdrawal
“Jump to reference” (J2R)
Take account of apparent early lower TST by subtracting 10 minutes here (“delta adjustment”)?
-
ORA randomized withdrawal
“Jump to reference” (J2R)
Take account of apparent early lower TST by subtracting 10 minutes here (“delta adjustment”)?
-
Plausible assumptions
• Based on historic data, strategy could be
– Imputations for experimental arm follow the
“jump to control” assumption: withdrawal from
experimental arm, usually with poor TST, will
be imputed to perform similarly poorly in the
control arm, OR
– As above, but subtract 10 minutes from
imputed values for first post-withdrawal visit
(“delta-adjust” with δ = 10) to take account of
apparent early lower TST in withdrawals.
T ds and Innovations in Clinical Trial Statistics: Missing data, developments in practice, copyright MOKelly 2014Apr2014
-
Conclusions
• Estimands and their assumptions about
missing data will be most useful if they
reflect clinical scenarios of interest:
statisticians and clinicians can work
together to identify these.
– Provide justification for the chosen strategy
for missing data in the protocol or statistical
analysis plan.
• Sensitivity analyses still essential, even for
scientifically justified MNAR estimands.
Apr2014
-
References
Brannan et al. (2005) Onset of action for duloxetine 60 mg once daily: double-blind,placebo-controlled
studies J. Psychiatr Res. 39 161-72.
FDA Peripheral & Central Nervous System Drugs Advisory Committee Meeting (2013) Suvorexant
Tablets Insomnia indication Advisory Committee Briefing Document for NDA204569.
Mallinckrodt et al. (2014) Recent Developments in the Prevention and Treatment of MissingData.
Therapeutic Innovation & Regulatory Science 48 68-80.
Perahia et al. (2006) Duloxetine in the prevention of relapse of major depressive disorder:Double-
blind placebo-controlled study. Br J Psychiatry. 188 346-53.
Rubin (1978) Multiple imputation in sample surveys – a phenomenological Bayesian approach to
nonresponse. Proceedings of the Survey Research Methods Section of the American Statistical
Association 1 20–34.
-
Back up slides
-
Questions?
-
Historic data: typical trajectories
for MDD
Brannan et al. (2005) Onset of action for duloxetine 60 mg once daily: double-blind, placebo-controlled studies
?
Match slope of reference:
“Copy increment from reference”
(CR)