Study Designs Appropriate for Comparative Effectiveness Research

advertisement
2/20/2012
Study Designs Appropriate
for
Comparative Effectiveness Research
Matthew L. Maciejewski, PhD
Durham VA HSR&D and Duke University
Objectives of Session
•
•
•
•
Motivation
Incident vs. prevalent user cohorts
Counterfactuals (aka comparators, controls)
Understand definition of and threats to internal validity and external validity
– Fundamental to study design selection
• Appreciate most rigorous quasi‐experimental study designs
– Quasi‐experiments as “broken trials”
2
Motivation & Value of Quasi‐Experimental Studies
• Quasi‐experiment is 1st step to RCT
– Big secondary data yield sample estimates of outcome variation and differences
– May generate/support hypotheses of causes
• Quasi‐experiment is only way to assess change in policy or clinical practice
– Population effect in real world is parameter of interest
3
1
2/20/2012
Ideal Quasi‐Experiment
• Answers the research question
• Great internal validity
– Treatment effect estimated is “true” effect
• External validity
– Generalizable across settings, time
4
Competing Demands in CER
• As scientists, internal validity is #1 concern
– Why RCTs considered gold standard
– Why evidence from observational studies considered suspect or invalid
– Historically, this is of greatest concern
Hi t i ll thi i f
t t
• As policy/clinical experts, generalizability is of concern
– Want results to be definitive across space (& time?)
– In CER, this is more prominent and quasi‐
experimental studies can address this issue
5
Four Design Choices
• In absence of randomization, three study design choices determine internal validity of your study
– Incident or prevalent users
– Control group
– Pre‐period measurement
– Post‐period measurement
• The more thoughtfully these design elements are chosen a priori, the stronger the study’s internal validity and causal statements
6
2
2/20/2012
Concept of Internal Validity
• Internal validity refers to ability to make causal inferences (e.g., Y is caused by X)
– With successful randomization, X is only source of variation between groups
f i ti b t
– In quasi‐experimental studies, X is no longer only source of variation b/c tx & control not balanced as in RCT
• Internal validity is not generalizability, reproducibility or reliability
7
Conditions for Causality
• Treatment precedes outcome
– Temporal relationship
• Treatment and outcome covary
– Significant correlation
g
• No other plausible competing hypotheses
– Identification of “true” cause is process of eliminating all competing hypotheses but 1
– If all other explanations for outcome variation are eliminated but treatment, treatment is cause
8
Incident User & Prevalent User Designs
g
3
2/20/2012
Three Cohort Choices
• Incident (aka new) users
– Index date refers to first use with no history of prior use
– Useful for medication and surgery studies
• Prevalent users
– Index date refers to treatment use in reference to some other date (e.g., policy change)
– Represents majority of medication users at point in time
• Both incident & prevalent users (‘wedge design’)
10
Tradeoffs of Incident User Cohort
• Advantages
– Well suited to assess acute meds or initial risks of chronic meds
– Captures events that C t
t th t
occur early in therapy
– Can control for risk factors before they are altered by exposure
• Disadvantages
– Likely small sample
– Limited precision/efficiency
– Need to be careful about characterizing risk of adverse events if risk varies greatly over time in early period
Ray WA, 2003 Am J Epi, 158(9): 915‐920
Defining Pre‐Period for Incident Cohort Design
• Two reasons to define pre‐period carefully
– Assess covariates before patient exposed to tx
– Ensure no prior history of treatment exposure (wash out)
• Need to choose look back period carefully
– 6 months: 32 of 1000 pts were eligible
– 12 months: 27.5 of 1000
– 24 months: 23.5 of 1000
– 108 months: 17.2 of 1000
Gardarsdottir H, Heerdink ER, Egberts, 2006 PharmacoEpi Drug Safety 15(5): 338 12
4
2/20/2012
Tradeoffs of Prevalent User Cohort
• Disadvantages
• Advantages
– Can characterize event risk because risk likely stable over time
– Large sample
– Improved precision/efficiency
– Addresses experience of most current medication users
– Under‐ascertainment of events occurring early due to treatment
 You miss folks who stopped meds due to adverse event
– “Baseline” covariates may have been impacted by treatment exposure
– Treatment duration unknown and variable
– Patients who adhere to treatment over long term are not random sample
Ray WA, 2003 Am J Epi, 158(9): 915‐920
Defining Pre‐Period for Prevalent Cohort
• Need to choose lookback period and history of treatment exposure carefully
Blais 2003
Pre-Period Duration
(months)
70
Number of
Intervals
1
Number of Fills
per Interval
1+
Each Interval
Length
70
Roblin 2005
12
2
4+, 1+
6, 6
Landsman
2005
Schneeweiss et
al., 2007
Yin, 2008
24
2
1+
3, 24
12
2
0, 1+
6, 12
24
1
1+
24
Doshi 2009
24
2
1+
3, 24
Maciejewski
2010
12
2
1+
3, 9
14
The Value of a Control Group
for Internal Validityy
5
2/20/2012
Importance of Control Group
• Terminology
– Control group = counterfactual = comparator
• When observing treatment, how do we know what would have happened in the absence of
what would have happened in the absence of treatment?
– Counterfactual evidence is needed
• Control group is needed to compare
– Outcomes under treatment – Outcomes NEVER under treatment
16
What is a Counterfactual?
• Counterfactual is measurement on subject that mirrors subject in every respect
– Observed and unobserved
• Perfect counterfactual represents
– What would have happened to treated patient had they not received treatment
– What would have happened to control patient had they received treatment
• Ideal counterfactual: two universes
– Same person in treatment & in control at same time
17
Ways of Estimating Counterfactuals
• Second best: Identical twins in RCT
• Third best: RCT
• Fourth best: Quasi‐experiment w/ equivalent controls
– Recommended by Shadish (2008)
• Fifth best: Quasi‐experiment w/ non‐
equivalent controls
– This is typical counterfactual in quasi‐experiment
18
6
2/20/2012
What is Effect of QI Program on Hospital Acquired Infections?
50
Treatment
40
30
20
10
0
Yr ‐1
Yr 1
Yr 2
19
Harris, Maciejewski, in press Health Affairs
What is Effect of QI Program on Hospital Acquired Infections?
50
Treatment
Control
40
30
20
10
0
Yr ‐1
Yr 1
Yr 2
Harris, Maciejewski, in press Health Affairs
20
Counterfactuals in Quasi‐Experiments
• Self‐selection in quasi‐experiments – Imbalance in observed covariates
– Can assume imbalance in unobserved factors
– Imbalance = nonequivalence (compared to RCT)
• Non‐equivalence complicates interpretation of treatment effect b/c several explanations
– Treatment? Selection bias? Both?
21
7
2/20/2012
Threats to Internal Validity
Biggest Internal Validity Threats in Quasi‐Experimental CER
• Ambiguous Temporal Precedence
• Selection Bias
• Regression
• History
23
Threats to Internal Validity
• Ambiguous Temporal Precedence
– Major concern in cross‐sectional studies
– Eg., Volume‐outcome relationship
• Selection
– “confounding of treatment effects with population differences”
– Observed (easier to fix) or Unobserved
– May require knowledge of how people get treatment, not just knowledge of outcome
24
8
2/20/2012
Example of Selection Bias on Observables (Bariatric Surgery)
Age (Mean, SD)
Male (%)
Caucasian (%)
( )
Non‐Caucasian (%)
Married (%)
Previously Married (%)
Never Married (%)
Unknown Marital (%)
BMI (Mean, SD)
Diagnostic Cost Group score (Mean, SD)
Surgical Non‐Surg
Statistical Cases
Controls
Significance
(N=850) (N=41,244) p‐value Std Diff
49.5 (8.3) 54.7 (10.2) <0.001 ‐55.92
73.9
91.7
<0.001 ‐48.54
77.9
67.8
<0.001 22.86
16.0
19.3
0.015
‐8.66
52.1
57.7
0.001 ‐11.27
30.4
27.3
0.048
6.85
16.4
13.8
0.035
7.00
1.2
1.2
0.957
0
47.4 (7.8) 42.0 (5.0) <0.001 82.43
0.60 0.47 (0.75) <0.001 15.49
(0.92)
25
Addressing Temporal Precedence and History Threats
• Ambiguous Temporal Precedence
– Design: Choose sample to identify precedence
– Statistics: Examine Y = f(X) & X = f(Y)
• Selection
– Design: Careful a priori choice of primary comparator
– Measurement: Additional data collection
– Statistics: Propensity score, selection methods
26
History Threat to Internal Validity
• History
– Co‐occurring events (e.g., pt in multiple RCTs) that could impact outcome of interest
– Attitudes about race, homosexuality change
Attitudes about race homosexuality change
• Address by understanding, measuring and adjusting for co‐occurring changes over time
27
9
2/20/2012
History Threat: Near Simultaneous Copay Changes
December 6, 2001
Specialty Care up from $15 to $50
Primary Care copay ($15) introduced
2000
2001
2002
2003
February 4, 2002
Medication copay up from $2 to $7
Maciejewski, et al., 2010 AJMC
Attempt to Account for History Threat (1 of 2)
Full Model includes lagged OP visits
Copay non‐exempt
β
0.030
S.E.
.075
Short‐term (ST) post
0.292*** .055
Long‐term (LT) post
‐0.078
.076
Copay*ST post
‐0.082
.060
Copay * LT post
‐0.150
.082
Wang, Liu & Maciejewski, HSR in press
29
Attempt to Account for History Threat (2 of 2)
Full Model includes lagged OP visits
Copay non‐exempt
β
0.030
Short‐term (ST) post
Reduced model (includes current OP visits)
β
0.032
S.E.
.075
β
0.033
S.E.
.075
0.292*** .055
0.288***
.055
0.288***
.055
Long‐term (LT) post
‐0.078
.076
‐0.082
.076
‐0.082
.076
Copay*ST post
‐0.082
.060
‐0.078
.060
‐0.078
.060
Copay * LT post
‐0.150
.082
‐0.145
.082
‐0.144
.082
Wang, Liu & Maciejewski, HSR in press
S.E.
.075
Reduced model (excludes lagged OP visits)
30
10
2/20/2012
Regression Threat to Internal Validity
• History
– Co‐occurring events (e.g., pt in multiple RCTs) that could impact outcome of interest
– Attitudes about race, homosexuality change
Attitudes about race homosexuality change
• Address by understanding, measuring and adjusting for co‐occurring changes over time
31
What is Effect of QI Program on Hospital Acquired Infections?
50
Treatment
40
30
20
10
0
Yr ‐1
Yr 1
Yr 2
32
Harris, Maciejewski, in press Health Affairs
What is Effect of QI Program on Hospital Acquired Infections?
50
Treatment
40
30
20
10
0
Yr ‐2
Yr ‐1
Harris, Maciejewski, in press Health Affairs
Yr 1
Yr 2
33
11
2/20/2012
What is Effect of QI Program on Hospital Acquired Infections?
50
Treatment
Control
40
30
20
10
0
Yr ‐1
Yr 1
Yr 2
34
Harris, Maciejewski, in press Health Affairs
What is Effect of QI Program on Hospital Acquired Infections?
50
Treatment
Control
40
30
20
10
0
Yr ‐2
Yr ‐1
Yr 1
Harris, Maciejewski, in press Health Affairs
Yr 2
35
Measurement Issues for Internal Validity
for Internal Validity
12
2/20/2012
Pre‐tests, Post‐tests & Non‐equivalent Dependent Variables
• Pre‐test measurements
– Can assess selection bias
• Post
Post‐test
test measurements
measurements
– Can reduce/eliminate ambiguity about temporal precedence of cause & effect
– The more the better to see pattern of association
• Non‐equivalent outcomes (post‐tests)
– Assesses degree of change due to other (unmeasured) contextual factors
37
Natural Experiments
• Definition: A study that contrasts a naturally occurring event, such as an earthquake, with a comparable condition (Shadish, Cook & p
(
,
Campbell)
– Intervention cannot be manipulated by researcher or research subject
38
Example of a Natural Experiment
• Value‐based insurance design (VBID) program implemented by BCBSNC in Jan 2008
– Lowered generic copays for some, brand copays for all
– Population based implementation, so tons of patients across multiple drug classes
lti l d
l
– Many employers opted out, so good controls
• Controls weren’t equivalent, but made so via propensity score matching
• Non‐equivalent outcomes
– ACEIs & CAIs only had brand Rx available
Maciejewski, Health Affairs Nov 2010
39
13
2/20/2012
Design of VBID Evaluation
• Pre‐tests
– One year prior to VBID implementation
• Post‐tests
– Two years after VBID implementation
Two years after VBID implementation
• Controls – Non‐equivalent, but made so via propensity score matching
• Non‐equivalent outcomes (ACEIs & CAIs) – Only brand Rx available
Maciejewski, Health Affairs Nov 2010
40
Unadjusted Trends in Diuretic Adherence (MPR)
Cases
100%
90%
79%
80%
Controls
Difference-in-difference=3%
79%
79%
76%
70%
60%
50%
40%
30%
20%
10%
0%
2007
2008
Adjusted Adherence Difference Associated with VBID Program
4.5%
4.0%
3.5%
3.0%
2.5%
2.0%
1.5%
1.0%
0.5%
0.0%
Diuretics
OHAs
ACEIs
Beta block
Statins
CalciumCBs
ARBs
CAIs
Note: GLM in unmatched sample, controlling for age, gender, Episode Risk Groups, count of unique medications filled, generic dispensing rate, total copay, and whether an enrollee filled one or more prescriptions with a 90‐day supply; Maciejewski et al. Nov 2010 Health Affairs
14
2/20/2012
Characteristics of Quasi‐Experiments with Strong Internal Validity
• At least one pre‐test measurement
• At least one post‐test measurements
– Outcomes of interest
Outcomes of interest
– Non‐equivalent outcomes
• Control group
– The more equivalent the better
• Choice of incident or prevalent user design depends on research question
43
External Validity
Concept of External Validity
• External validity = generalizability of results
– Persons, settings, treatments outside of study
• Fundamental question: Is response to treatment is homogeneous or heterogeneous?
– Assessment by subgroup (defined by pt characteristic, site, time) can inform this issue
45
15
2/20/2012
Threat to External Validity due Heterogeneity in Units
• Due to Heterogeneity in Units: Response may differ across subgroups
– Age groups, genders, ethnicities
– Income levels
– Health systems (VBID example)
• Due to Treatment Variation
– Interventionist variation, if 2+ interventionists
– In multi‐component trials, intervention delivered may vary if some pts got components #2‐5 but others got components #6‐9?
46
Example of External Validity in VBID Results
6%
5%
Maciejewski
Choudhry
4%
Chernew
Sedjo
3%
Kelly
Chang
2%
Rodin
1%
Mahoney
0%
OHAs
ACEI
Beta block
Statin
External Validity & Treatment Variation in DPP
Original DPP
Lifestyle Arm
DEPLOY
East Harlem DPP
Study Design
RCT
RCT
RCT
Setting
AHCs
YMCA
Urban low income
Intervention
Individual
Individual lifestyle
16 group meetings ??
16 group meetings at YMCA
Weight loss compared to UC
6% greater
5.5% greater
1.9% greater
Intervention cost
$2,269 over 3 years
NR
NR
Cost/QALY
Cite
$1,100
NEJM 2002; Ann Int Med 05
NR
Am J Prev Med 2008
NR
Am J Pub Health 2010
16
2/20/2012
External Validity and CER/Translation Research
• Fundamental issue in disseminating RCTs to communities (we are struggling with)
• If intervention we test in RCT is implemented in community clinic, would results differ?
– Protocol: We shorten intervention intensity and/or change interventionist – Setting: Non‐academic
– Subjects: Lower severity & motivation
– Funding: None (unlike RCT)
49
Take‐Away Points
• Quasi‐experimental study designs are widely used b/c secondary data analysis is common
– Several threats to internal validity
– Greatest external validity
• Have more design/statistical issues than RCTs
– Require more data than RCTs
– More assumptions about data (not observed)
– Creativity in control group and statistics selection
• Carefully consider trade‐offs in incident & prevalent user designs
50
References & Resources
1.
2.
3.
4.
5.
6.
7.
Gardarsdottir, H., Heerdink, E. R., & Egberts, A. C. (2006). Potential bias in pharmacoepidemiological
studies due to the length of the drug free period: a study on antidepressant drug use in adults in the Netherlands. Pharmacoepidemiol Drug Saf, 15(5), 338‐343. doi: 10.1002/pds.1223
Harris, B. D., Hanson, C., Christy, C., Adams, T., Banks, A., Willis, T. S., & Maciejewski, M. L. (2011). Strict hand hygiene and other practices shortened stays and cut costs and mortality in a pediatric intensive care unit. [Research Support, U.S. Gov't, Non‐P.H.S.]. Health Aff (Millwood), 30(9), 1751‐1761. doi: 10.1377/hlthaff.2010.1282
Maciejewski, M. L., Bryson, C. L., Perkins, M., Blough, D. K., Cunningham, F. E., Fortney, J. C., . . . Liu, C. F. (2010). Increasing copayments and adherence to diabetes, hypertension, and hyperlipidemic medications. [Research Support U S Gov'tt, Non‐P.H.S.]. Am J Manag
[Research Support, U.S. Gov
Non‐P H S ] Am J Manag Care, 16(1), e20‐34. Care 16(1) e20‐34
Maciejewski, M. L., Farley, J. F., Parker, J., & Wansink, D. (2010). Copayment reductions generate greater medication adherence in targeted patients. [Research Support, Non‐U.S. Gov't Research Support, U.S. Gov't, Non‐P.H.S.]. Health Aff (Millwood), 29(11), 2002‐2008. doi: 10.1377/hlthaff.2010.0571
Maciejewski, M. L., Livingston, E. H., Smith, V. A., Kavee, A. L., Kahwati, L. C., Henderson, W. G., & Arterburn, D. E. (2011). Survival among high‐risk patients after bariatric surgery. [Research Support, U.S. Gov't, Non‐P.H.S.]. JAMA, 305(23), 2419‐2426. doi: 10.1001/jama.2011.817
Ray, W. A. (2003). Evaluating medication effects outside of clinical trials: new‐user designs. [Research Support, U.S. Gov't, Non‐P.H.S. Research Support, U.S. Gov't, P.H.S.]. Am J Epidemiol, 158(9), 915‐920. Wang, V., Liu, C. F., Bryson, C. L., Sharp, N. D., & Maciejewski, M. L. (2011). Does medication adherence following a copayment increase differ by disease burden? [Research Support, U.S. Gov't, Non‐P.H.S. Research Support, U.S. Gov't, P.H.S.]. Health Serv Res, 46(6pt1), 1963‐1985. doi: 10.1111/j.1475‐
6773.2011.01286.x
17
Download