bias - home - national academy of medicine · comparative effectiveness research (ahrq) pi, decide...
TRANSCRIPT
1
Bias
Introduction of issue and background papers
Sebastian Schneeweiss, MD, ScD
Professor of Medicine and Epidemiology
Division of Pharmacoepidemiology and Pharmacoeconomics, Dept of Medicine, Brigham & Women’s Hospital/ Harvard Medical School
Potential conflicts of interest PI, Brigham & Women’s Hospital DEcIDE Center for
Comparative Effectiveness Research (AHRQ) PI, DEcIDE Methods Center (AHRQ) Co-Chair, Methods Core of the Mini Sentinel System (FDA) Member, national PCORI Methods Committee No paid consulting or speaker fees from pharmaceutical
manufacturers Consulting in past year:
WHISCON LLC, Booz&Co, Aetion
Investigator-initiated research grants to the Brigham from Pfizer, Novartis, Boehringer-Ingelheim
Multiple grants from NIH
2
Objective of Comparative Effectiveness Research
3
Efficacy (Can it work?)
Effectiveness* (Does it work in routine care?)
* Cochrane A. Nuffield Provincial Trust, 1972
Placebo comparison (or usual care)
Active comparison (head-to-head)
Most RCTs for drug approval
Goal of CER
Effectiveness = Efficacy X Adherence X Subgroup effects (+/-)
Reality of routine care RCT
4
Baseline randomization
Non-randomized
Primary data
Secondary data
CER
Primary data
Secondary data
Challenges of observational research
Measurement /surveillance-related biases Informative missingness/ misclassification
Selection-related biases Confounding Informative treatment changes/discontinuations
Time-related biases Immortal time bias Temporality Effect window
(Multiple comparisons) 5
6 H-C C and CC, BMJ 2010
Informative missingness: Unintended effects of statins
Data source: QResearch EMR system, England & Wales
7
Confounding (Obs) and informative censoring (Obs and RCT) By indication By contraindicaton, healthy users Adherence
Severity, comorbidity, prognosis
Side effect, treatment
failure
Treat A
Treat B
Patient follow-up
Rx
Rx Rx Rx
Patients Rx
Rx
8
Immortal time bias in an event-based cohort
Cohort entry is defined by a new diagnosis “>” but time until 1st drug exposure is misclassified
Cohort entry is defined by - 1st drug use for exposed - an event (Dx) for the unexposed
Suissa S, AJE 2008
Opportunities: Confounding
Utilize naturally occurring variation in the healthcare system (Dylan Small) Between providers Between systems Between regions Between time periods
Measure naturally occurring variation and capture via propensity score analyses
Negative controls (Prasad & Jena)
9
Opportunities: Censoring
As treated vs. intention to treat analyses (Miguel Hernan)
Inverse probability of discontinuing weighting And similar methods that rely on characterizing the factors
leading to treatment change
10
Opportunities: Transparency about choices
No single study(design) will satisfy the information needs of a decision maker
Need to understand the desired study characteristics and transparent choose accordingly: Internal validity External validity Precision Timeliness Logistical constraints etc
11
4
Intrinsic Study Characteristics Internal validity (bias) External validity (generalizability, transportability) Precision Heterogeneity in risk or benefit (personalized evidence) Ethical consideration (equipoise) External Study Characteristics Timeliness (rapidly changing technology, policy needs) Logistical constraints (study size, complexity, cost) Data availability, quality, completeness
13
Helpful references include:
Rothwell PM Lancet 2005
Miler FG & Joffe S NEJM 2001
Concato J PDS 2012
YES-Prefer baseline randomization for:
• high validity in the presence of strong baseline confounding
• if no ethical issues prevent randomization
• if sufficient resources available
• if enough time available to await results
NO-Prefer observational study for:
• high representativeness for “routine care” by not perturbing the care system
• Need good reason to believe that confounding can be controlled through adjustment
Is baseline randomization indicated?
Below are general considerations that may be neither comprehensive nor applicable for every scenario. These considerations are meant to be amended and changed as more experience is gained with this tool.
Design
14
Intrinsic Study Characteristics Internal validity (bias) External validity (generalizability, transportability) Precision Heterogeneity in risk or benefit (personalized evidence) Ethical consideration (equipoise) External Study Characteristics Timeliness (rapidly changing technology, policy needs) Logistical constraints (study size, complexity, cost) Data availability, quality, completeness
Opportunities: Study portfolio
Often we need multiple studies with different data sources (prim/sec) and different designs
Is there an optimal way to arrange multiple studies so that they complement each other (speed, validity, generalizability) and collectively provide most valid and comprehensive
information for decision makers?
15
Opportunities: Investigator error
Training Guidance ‘Standards’
16
17
Subgroup Analysis ?
Basic Design Consideration
Subgroup definition
Prior pharmacology knowledge
Prior clinical Knowledge
Yes
Cohort study (case-control, case-cohort sampling)
Exposure/outcome considerations
Exposure definition Outcome Definition
Comparison group considerations Clinical meaningfulness
Incident user design considerations
Exposure risk window considerations Case validation necessary?
Specificity and sensitivity of measurement
Yes Consider case-crossover design
no
Meaningful exposure variation within patients?
Defining covariates based on clinical knowledge
Defining additional covariates empirically (high-dimensional proxy adjustment)
Demonstrate covariate distributions by treatment group with RDs and 95% CIs
Supplemental covariate information required that is not available in primary data source?
Collect additional information in subpop. • 2-stage sampling • External data source
-(PS Calibration) - Multiple imputation
Yes
Propensity score (PS) analysis Missing covariate values in EMRs? Multiple imputation
Estimating propensity score
Yes
Explore effect measure modification by PS: tabulate RR,
RD for each PS stratum
Graphically explore PS distribution by treatment group
Yes Effect measure modification by PS?
•Stratify by PS deciles •Match on PS (1:1, 1:n, 1:n:m)
Trim 5% of patients on each end of PS distribution or match by PS
Balancing Patient Characteristics
Demonstrate covariate balance by treatment group with RDs and 95% CIs
19
Repeat analyses after changes in: • Definition of “incident users” • Definition of exposure risk window • Outcome definition if appropriate
Explore changes in effect estimates after making structural assumptions about unmeasured confounders
Sensitivity Analyses
Statistical analysis*
*For illustration purposes only an analysis after PS matching is shown.
Calculate risk difference (RD) and risk ratio (RR);
95% confidence intervals (CIs) for main result. Report person-time (p-t), number of events
Subgroup analysis Calculate RR, RD for
each subgroup
Dose-response analysis
Include time since initiation as subgroup
Report
Opportunities: Transparency
Sharing of data vs. sharing of analytics environment Data stay were they are but analytics infrastructure allows
other investigators (in collaboration) to re-analyze data
20