Janssen Research and Development Previous Eastern Hemisphere meeting Martijn Schuemie Study reproducibility 2 Future research topics 3 Obvious next developments Scifi Brainstorm figure out feasibility later ID: 810577
Download The PPT/PDF document "Future research topics Martijn Schuemie,..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Future research topics
Martijn Schuemie, PhD
Janssen Research and Development
Slide2Previous Eastern Hemisphere meeting
Martijn Schuemie: Study reproducibility
2
Slide3Future research topics?
3
Obvious next developments
Sci-fi
Brainstorm: figure out feasibility later
Slide4Methods research
= hypothesis of interest
(e.g. dabi vs warfarin for bleeding)
= effect size estimate for
= set of design choices:
Overall design (Cohort or case-control?)
Analysis choices within design (matching or stratification?)
Choice of database (CPRD or Truven CCAE?)
4
Slide5Evaluation
= True effect size for
Can the error distribution be explained by random error?
Often not: there is also systematic error
5
Slide6Candidate research topic 1:
Systematic Evaluation of methods
Things to evaluate:
Existing methods
New-user cohort design (inc. HDPS)Self-Controlled Case Series (SCCS)
Case-Control
…
Novel methods
New PS approaches (Alejandro, Yuxi)Comparative SCCS (Jamie)
What to evaluate on:Real negative controls + synthetic positive controlsReplication of RCTs6
Slide7Calibration
+
Purpose: to restore nominal characteristics
95% CI should contain truth 95% of times
7
Slide8Candidate research topic 2:
Combining calibrated estimates
If I run a study on
n
databases, calibrating each estimate, how do I combine these estimates?
Random error will be independent across databases: inverse variance weighting?
Systematic error will not
be
independent!Each database will have different population, so true effect size itself will be different: random effects model?
8
Slide9Candidate research topic 3:
Smooshed comparators
Compare new treatment to all old treatments for same indication?
Problem: propensity score may not be able to adjust for heterogeneous comparator group.
Alternative:
Combine across comparisons, taking
correlation
into account
9
Slide10Candidate research
topic 4:
Bayesian sensitivity analysis
To what extend does
depend on
?
10
Slide11Distribution of possible results
for one hypothesis
Stat
signif
> 1
Databases
Methods
OR
Stolen from George Hripcsak
Slide12Distribution of possible results
for one hypothesis
Stat
signif
> 1
Databases
Methods
OR
Stolen from George Hripcsak
Slide13Distribution of possible results
for one hypothesis
Stat
signif
> 1
Stat
signif
< 1
Databases
Methods
Stolen from George Hripcsak
Slide14Databases
Methods
Distribution of possible results
for one hypothesis
BMJ
Study #3
JAMA
Stolen from George Hripcsak
Slide15Distribution of possible results
for one hypothesis
OR
Databases
Methods
Stolen from George Hripcsak
Slide16Design and analysis choices
Epi community:
expert knows best (through unformalizable knowledge)
use some predefined sensitivity analyses.
If pretty stable:
mention
in
discussion
if not stable: mention in discussionOMOP: let the data decide (pick
that optimizes AUC)design choices matter (a lot)
16
Slide17Bayesian sensitivity analysis
= prior of design and analysis choices being ‘correct’.
Advantage: if uncertain choices (both could be ‘correct’) lead to large differences, posterior will be wide to reflect that uncertainty.
How to get priors?
Expert informed
Based on our systematic evaluation
17
Slide18Candidate research
topic 5:
Mechanistic modeling (effect modification)
= effect size
= treatment choice
Causal model: will tell you effect of changing
Could we estimate:
= effect modifiers
without pre-specifying specific effect modifiers?
18
Slide19Candidate research
topic 6:
Systematic informed priors
= Baseline characteristics
= Coefficients for baseline characteristics
Current state-of-the-art:
Uninformed prior for
Homogeneous prior for
Prior for diabetes and MI = Prior for otitis media and MI?
Mining of prior knowledge (e.g. MEDLINE and Wikipedia) to inform priors
19
Slide20Candidate research topics
Systematic Evaluation
of
methods
Combining calibrated estimates
Smooshed
comparators
Bayesian
sensitivity analysis
Mechanistic modeling (effect modification)Systematic informed priors
20
Slide21Next workgroup meeting
Eastern
hemisphere:
Februari 22
3pm Hong Kong / Taiwan4pm South Korea
5:30pm
Adelaide
8am
Central European time7am UK timeWestern hemisphere:
Februari 166pm Central European time12pm New York9am Los Angeles / Stanford
http://www.ohdsi.org/web/wiki/doku.php?id=projects:workgroups:est-methods