Transcript ABSTRACT
Biostatistical Considerations
Nae-Yuh Wang, PhD
ICTR Clinical Registries Workshop
November 3, 2010
OVERVIEW
Descriptive vs Analytic Goals
Selection of Controls
Confounding
Measurement errors
Missing data
Purposes of Patient Registry
Document natural history of disease
Evaluate effectiveness of treatment
Monitoring safety
Measuring quality
Frequently multi purposes, addressing
scientific, clinical, and policy questions
Natural History of Disease
Document characteristics, management, and
outcomes
May be variable across subgroups
May be variable over time
Change after new guidelines or treatments
SMOKING PREVALENCE, %, AMONG U.S. MALE
ADULTS*AND 1,213 WHITE MALE
PHYSICIANS:THE PRECURSORS STUDY
70%
60%
PREVALENCE ,%
50%
US MALE POPULATION
40%
30%
20%
MALE PHYSICIANS,
THE PRECURSORS STUDY
10%
0%
1955
1965
1970
1975
1980
YEAR
*CDC, National Health Interview Surveys, 18 years and older,1965-1994
1985
1990
1994
Effectiveness of Treatment
RCTs usually have well defined populations
RCTs usually are short term
Clinical effectiveness, cost effectiveness
Comparative effectiveness --- indirect
comparisons on differences between
treatments
Safety Monitoring
Adverse event reporting relies on recognition
of AE by clinician, and clinician’s effort in
reporting --- frequently nonsystematic
Serves as active surveillance
Provides denominator to estimate incidence
Enables comparison to a reference rate
Health Care Quality
Compare performance measures (treatments
provided or outcomes achieved) against
evidence based guidelines or benchmarks
(adjusted survival, infection rates) between
provider or patient subgroups
Identify disparity in access to care
Demonstrate opportunities for improvement
Establish payment differentials
Types of Registry
Product registries (drug, device)
Health services registries (procedure)
Disease or condition registries
Patients defined by exposure to a product,
procedure, or disease/condition
Frequently combination of types
Design of Registry
Research questions, stakeholders, and
practical factors (regulatory, political, funding)
define purpose and type of registry, and
other design considerations such as
sampling plan, data collection, validity,
sample size, and analytic approaches
Types define the patient population
Purposes define the outcomes
Outcomes define the duration
Design of Clinical Research
Research questions (descriptive vs. hypothesis
based --- analytic)
Population; outcome and exposure
Sampling (recruitment); measurements, duration
and frequency of data collection
Internal / external validity (bias / generalizability)
Sample size (precision of estimates/degree of
association, feasibility, resources)
Analytic plan
Sampling Design
External validity (generalizability): all patients,
patients from tertiary medical center only? Single
center, multiple center? Which way is more
representative of the target population under
study?
Do I need controls? (CI in children, language
outcomes vs. meningitis)
Selection of controls
Match or not to match?
Cohort Design
Sampling based on predictor (exposure variables) of
interest (collect as many exposure variables as
possible). Good for rare exposure.
Follow up patients for outcomes, could study multiple
outcomes (long time for outcomes to develop?)
Census (not feasible when population and per capita
cost are large)
SRS, stratified RS (oversampling subgroup), Cluster
RS (cluster characteristics as the aim), multistage RS
Nonrandom: case series / consecutive sampling
Case-Control Design
Stratified RS based on case status
Oversampling cases, good for rare diseases
No long follow up for disease development
Study multiple exposure variables
Exposure ascertainment is key
Nested case-control study using existing registry
Selection of controls, match or not to match
Measurement & Data Collection
Data from clinically based electronic sources
only?
Linking from different sources (e.g., NDI
searches)
Measurement (different labs) and coding
consistency
Additional data collections --- potential
confounders, nonclinical outcomes (e.g., QoL,
QALY), medications
Measurement & Data Collection
Research versus clinical protocol (BP, busy
schedule)
New / changes in treatments and guidelines over
time
Changes / improvement in measurement
precision and generation of technology over time
Change of outcome definition over time (clinical
designation or collect and record raw measures)
Analytic corrections could only be done if needed
data / information are available
Internal Validity --- Sources of Bias
Information bias: AE under reported if reporter
(provider) will be viewed negatively on care
quality. Self reported weight
Selection bias: patients included not
representative (unintentional incentives for
provider / patient), loss to follow-up, common
exposure to unaccounted confound
Confounding by indication: newest drug to
patients with worse prognosis
Survival bias: live long enough with exposure to
be selected
Internal Validity --- Sources of Bias
Confounding:
CVD risk, age, gray hair
Controlled by matching through study design
Accounted for through stratification, covariate
adjustment, or propensity score adjustment
during analyses
Only work if data on confounders were
collected, need to consider at design stage
Internal Validity --- Sources of Bias
Measurement errors:
Mean of 3 repeatedly measured BP readings
used in RCT versus single BP used in clinic
Measured versus self reported body weight
Fruit / vegetable availability in an area used
as proxy measure of fruit / vegetable
consumption value
Areas measured by 2nd vs. 1st generation CT
Measurement Errors
Nondifferentiable ME in outcome causes no
bias. Greater variability in outcomes due to
ME reduces statistical power
Differentiable ME in outcome causes violation
of constant variance assumption in
regression.
Nondifferentable ME in covariate causes
underestimation of association (bias towards
the null)
ME in Covariate
β* = λβ, where
2x
λ=
2x 2
4
ME in Covariate Models
E ( Y | X ) = μ ( Xβ )
Classical error model:
W = X + ε , X || ε (Note: non-differential)
i. X the measured weight, W the self reported
weight
ii. X the measured BMI, W the self reported BMI
Berkson error model:
X = W + ε , W || ε (Note: non-differential)
i. X the “true” F/V consumption, W the proxy value
ME in Covariate Models
Goal: E ( Y | X ) = μ ( Xβ )
Actual: E ( Y | W ) = μ ( Wβ* )
Need to correct the estimate of β* to get proper
estimate of β
Need to quantify ME so proper correction of β* is
possible:
Validation: a subsample with both X and W
Replications: repeated measures of W (e.g., BP)
Transportability: information from another study if valid
ME in Covariate
Non-differential ME key assumption, not testable without
validation data
When covariate with ME in the model, covariates w/o
ME may also be biased. Directions of such biases
depend on directions of association among Y and
covariates in the model
ME model could be complicated: combined classical &
Berkson’s error model, additive versus multiplicative ME
Differential ME: bias direction depends on how ME
relates to Y
Design Considerations for ME
Conduct periodic validation study on small
random sample of participants (e.g. self
report vs. measured weight, outcomes coded
by billing vs. coded under research protocol)
If not available from external sources, repeat
assessments using old and new instruments
in random sample of participants during
transition to collect calibration data.
Sources of external validation/calibration data
Missing Data
Inevitable in population research
Prevention is better than statistical
treatments
Too much missing information invalidates a
study
Validity of methods accommodating missing
data depends on the missing data
mechanism and the analytic approach
Missing Data Mechanism
Missing completely at random (MCAR):
Pr (missing) is unrelated to process under study
Missing at random (MAR):
Pr (missing) depends only on observed data
potential “ignorability”
Not missing at random (NMAR):
Pr (missing) depends on both observed and
unobserved data non-ignorable
Simulations
N = 100, repeated outcome: y0, y1
Group = 0, 1 (n = 50 / 50)
FV = 0:
y0 ~ N(0,1) if Group = 0
y0 ~ N(1,1) if Group = 1
FV = 1:
y1 ~ N(0,1) if Group = 0
y1 ~ N(1,1) if Group = 1
E( y0) = E( y1) = 0.5, SD( y0) = SD( y1) = 1.12
Corr( y0, y1 | Group) = 0.6, Corr( y0, y1 ) = 0.68
Analytic Approach
Likelihood approach
Mixed effects models
Mean model = Intercept + FV versus
Intercept + FV + Group
Correlation model: Working independent
(WI) versus Unstructured (UN)
Model-based versus robust SE
Simulations
Full Sample:
N
Sample mean
Sample SD
y0 (FV=0)
100
0.47
1.04
y1 (FV=1)
100
0.54
1.09
y1 – y0
100
0.069
0.79
MCAR: 25% random missing at FV1
N
Sample mean
Sample SD
y0 (FV=0)
100
0.47
1.04
y1 (FV=1)
75
0.56
1.13
y1 – y0
75
0.088
0.83
Simulations
(y1 – y0)
Mean Model
[ Corr ]
No Missing
MCAR
.069 (.079)
.088(.096)
ModelBased
Robust
ModelBased
Robust
Int. + FV
[ WI ]
.069 (.150)
.069 (.079)
.090 (.166)
.090 (.100)
Int. + FV /
[ UN ]
.069 (.079)
.069 (.079)
.089 (.094)
.089 (.094)
Int. + FV (+ GP) /
[ WI ]
.069 (.134)
.069 (.079)
.084 (.149)
.084 (.095)
Int. + FV (+ GP) /
[ UN ]
.069 (.079)
.069 (.079)
.087 (.093)
.087 (.093)
Simulations
Full Sample:
N
Sample mean
Sample SD
y0 (FV=0)
100
0.47
1.04
y1 (FV=1)
100
0.54
1.09
y1 – y0
100
0.069
0.79
MAR1: 25% missing in Group 0 at FV1
N
Sample mean
Sample SD
y0 (FV=0)
100
0.47
1.04
y1 (FV=1)
75
0.75
1.16
y1 – y0
75
0.094
0.76
Simulations
(y1 – y0)
Mean Model
[ Corr ]
No Missing
MAR1
.069 (.079)
.094(.088)
ModelBased
Robust
ModelBased
Robust
Int. + FV
[ WI ]
.069 (.150)
.069 (.079)
.279 (.159)
.279 (.102)
Int. + FV /
[ UN ]
.069 (.079)
.069 (.079)
.137 (.086)
.137 (.086)
Int. + FV (+ GP) /
[ WI ]
.069 (.134)
.069 (.079)
.129 (.147)
.129 (.093)
Int. + FV (+ GP) /
[ UN ]
.069 (.079)
.069 (.079)
.103 (.085)
.103 (.085)
Simulations
Full Sample:
N
Sample mean
Sample SD
y0 (FV=0)
100
0.47
1.04
y1 (FV=1)
100
0.54
1.09
y1 – y0
100
0.069
0.79
MAR2: 25% missing depends on values of y0
N
Sample mean
Sample SD
y0 (FV=0)
100
0.47
1.04
y1 (FV=1)
75
0.29
1.04
y1 – y0
75
0.213
0.78
Simulations
(y1 – y0)
Mean Model
[ Corr ]
No Missing
MAR2
.069 (.079)
.213(.090)
ModelBased
Robust
ModelBased
Robust
Int. + FV
[ WI ]
.069 (.150)
.069 (.079)
-.187 (.158)
-.187 (.117)
Int. + FV /
[ UN ]
.069 (.079)
.069 (.079)
.154 (.090)
.154 (.090)
Int. + FV (+ GP) /
[ WI ]
.069 (.134)
.069 (.079)
-.194 (.138)
-.194 (.115)
Int. + FV (+ GP) /
[ UN ]
.069 (.079)
.069 (.079)
.106 (.091)
.106 (.091)
Simulations
Full Sample:
N
Sample mean
Sample SD
y0 (FV=0)
100
0.47
1.04
y1 (FV=1)
100
0.54
1.09
y1 – y0
100
0.069
0.79
NMAR: 25% missing depends on values of y1
N
Sample mean
Sample SD
y0 (FV=0)
100
0.47
1.04
y1 (FV=1)
75
0.11
0.83
y1 – y0
75
-0.127
0.72
Simulations
(y1 – y0)
Mean Model
[ Corr ]
No Missing
NMAR
.069 (.079)
-.127(.083)
ModelBased
Robust
ModelBased
Robust
Int. + FV
[ WI ]
.069 (.150)
.069 (.079)
-.367 (.141)
-.367 (.090)
Int. + FV /
[ UN ]
.069 (.079)
.069 (.079)
-.228 (.080)
-.228 (.080)
Int. + FV (+ GP) /
[ WI ]
.069 (.134)
.069 (.079)
-.360 (.120)
-.360 (.090)
Int. + FV (+ GP) /
[ UN ]
.069 (.079)
.069 (.079)
-.257 (.081)
-.257 (.081)
Simulations
Model
Mean / Corr
No
Missing
MCAR
MAR1
MAR2
NMAR
.069 (.079)
.088(.096)
.094(.088)
.213(.090)
-.127(.083)
Int. + FV /
WI (Model-based) .069 (.150)
.090 (.166)
.279 (.159)
-.187 (.158)
-.367 (.141)
Int. + FV /
UN (Robust)
.069 (.079)
.089 (.094)
.137 (.086)
.154 (.090)
-.228 (.080)
Int. + FV (+ GP) /
WI (Model based) .069 (.134)
.084 (.149)
.129 (.147)
-.194 (.138)
-.360 (.120)
Int. + FV (+ GP) /
UN (Robust)
.087 (.093)
.103 (.085)
.106 (.091)
-.257 (.081)
(y1 – y0)
.069 (.079)
Observations
MCAR:
» Requires only correct mean model for valid
inferences
» Complete case analysis is valid, but not
efficient for estimating fully observed
variables
» Approaches valid for MAR also valid under
MCAR
» Unlikely to be true in population based
research
Observations
MAR:
» Ignorablility of missing is possible but not
given
» Requires correct specification of likelihood
(both mean and covariance model) for the
observed data to achieve valid inferences
» Empirically cannot be confirmed without
auxiliary data
Observations
NMAR:
» Empirically cannot be ruled out without
auxiliary data
» Likelihood, multiple imputation, propensity
score, inverse weighting approach cannot
completely eliminate bias
» Need to conduct sensitivity analyses under
various plausible NMAR scenarios to
evaluate potential impacts on inferences
Observations
Observational studies face similar issues
as RCTs with missing data:
» Bias due to missing data selection bias
» Proper selection of analytic models may
eliminate bias if the “selection” is based on
observed data values, i.e. we have data to
adjust for selection
» Bias due to “selection” according to data
values not observed will be hard to correct
Sample Size Considerations
Descriptive: estimation precision
Hypothesis based: power to detect association
Design effects
Longitudinal correlations