Transcript slide set

Some comments on the 3 papers
Robert T. O’Neill Ph.D
Comments on G. Anderson
 WHISH is a nice example
 Randomiztion (Zelen) but using different
sources of data for outcome
 Outcome data: self reported, adjudicated for
medical records, Medicare claim (hybrid-ability
to estimate SE and SP
 Impact of outcome misclassification
 Event data not defined by protocol-you depend
on the health care system
 Claims data DO NOT provide standardized data
– see Mini-Sentinel and OMOP
Comments on A J Cook
 Key component is randomization at
patient or clinic level and use of electronic
health record for data capture (cluster
randomization addresse different issues)

Missing data, informative censoring, switching,
measuring duration of exposure (repeat Rx, gaps) ,
different answers depending upon definition

Validation of outcomes makes the pragmatic trial less
simple

Only some outcomes (endpoints) , populations,
questions are addressable before complexities of
interpretation overwhelm
Comments on M Gaffney
 Precision and Eagle are not large simple
trials – they are large difficult trials
 Outcome adjudication, monitoring strategies
 Non-inferiority poses significant challenges for
pragmatic trials – generally no assay sensitivity
 Margin selection based upon evidence vs.
based upon close enough but not sure if both
are equally good or bad
Other comments on NI studies
 Pre-specifying the margins – why and what is the
difference in these two situations
 What treatment difference is detectable and
credible with the playoff of bias and huge sample
size
 When pre-specification is not possible because there is
no historical information, the width of the confidence
interval makes sense – but two conclusions – both
treatments the same and comparably effective vs. both
the same but both ineffective
 What endpoints are eligible : Hard endpoints (y),
patient symptoms(n)
Other comments
 Are NI designs appropriate for claims data of EHR
without independent all case adjudication –
implications for poor sensitivity and specificity to
drive estimate to null – what does a null result mean
 Experience suggests that Exposure (drugs) has better
accuracy than diagnoses or procedure in claims data
bases(outcomes)
 Duration of exposure dependent upon
algorithms for repeat prescriptions – different
results depending upon definitions of gaps
between repeated RX
Can randomization overcome lack of
blinding and personal choices after
randomization
 Use of observational methods of
accounting for unmeasured confounding
of assigned treatment and time to event
outcomes subject to censoring
 Directed Acylic Graphs to explore the
confounding-censoring problem diagnostics
 Instrumental variables
Lessons learned from Mini-Sentinel and the
Observational Medical Outcomes Partnership (OMOP)

Distributed data models

Common data models

Limits of detectability of effect sizes of two or more
competing agents – calibration, interpretation of p-values
for non randomized studies

Not all outcomes and exposures can be dealt with in similar
manner

Know the limitations of your data base – is this possible in
advance of conducting the study – part of the intensive
study planning, protocol and prospective analysis plan
An example of Medicare data use but not a RCT
Some other views and opinions on
CER using the learning health care
system
Lessons Learned from OMOP and MiniSentinel About observational studies
using health care claims data or EHR –
but no randomization
 Lessons about the limitations of the data
bases, outcome capturing, ascertainment ,
missing data (confounders) are relevant to
the RCT use of the same data source
 Lessons about data models, and challenges
for data (outcome standardization)
http://www.mini-sentinel.org/
http://omop.org/
The Observational
Medical Outcomes
Partnership – many
findings
Some ideas on what to
evaluate about a given data
source before committing to
conducting a study – focus
on observational studies –
but also relevant to
pragmatic RCTs
How do these presentations relate to
pragmatic trials within a health care system
 Two or more competing therapies on a formulary – never
compared with each other
 Randomize patients under equipoise principle – do you
need patient consent , physician consent if health plan
and no data to think ‘I or We don’t know but want to find
out ‘
 Collect electronic medical record data, including
exposures and outcomes – and decide if any additional
adjudication is needed
 Analyze according to best practices – but with some
prospective SAPs – causal inference strategies ?