Title: Research Designs, Statistical Strategies for Dealing with Selection Bias in Treatment Delivery, and
1Research Designs, Statistical Strategies for
Dealing with Selection Bias in Treatment
Delivery, and Limitations
- Naihua Duan
- UCLA and RAND
- May 2000
- Selection bias in treatment assignment/delivery
- Research designs
- Mitigating for overt selection bias
- Dealing with hidden selection bias
- Discussions
2Selection Bias in Treatment Delivery
- In naturalistic settings
- Pre-treatment health ? treatment delivered
- Pre-treatment health ? outcome
- Treated group dissimilar from untreated group
- Direct comparison of treated vs. untreated
results in biased - estimate for treatment effect
- Need to mitigate selection bias in order to
assess treatment effect more appropriately
3Selection Bias in Treatment Delivery Typology
- Overt selection bias
- Treatment related to covariates
- T ? X
- Given covariates, treatment independent of
outcome - T ? Y X (ignorability)
- Like a stratified randomized experiment
- Hidden selection bias
- Given covariates, treatment still related to
outcome - T ? Y X
- Rosenbaum (1995) Observational Studies,
Springer-Verlag
4Rubin Causal Model
- Potential outcome
- Y1T Y1C Y1T - Y1C
- ..
- YmT YmC YmT - YmC
- ----------------------------------------------
- Ym1T Ym1C Ym1T - Ym1C
- ..
- YmnT YmnC YmnT - YmnC
- Challenging missing data problem
- Missing at random (ignorable missingness)
5Research Designs
- Ideal randomized clinical trial (RCT)Imperfect
RCT with noncomplianceRandomized encouragement
design (RED)Observational studies - Settings controlled vs. naturalistic
- Treatment assignment/delivery mandated vs.
choice - Treated vs. untreated groups balance vs.
imbalance - Research questions efficacy vs. adoption,
program effect, and efficacy - Analytic strength interval validity vs. external
validity -
6Randomized Clinical Trial
Intensive efforts made to mandate assignment
7Randomized Encouragement Design
Encouragement training, providing information,
case management, reducing barriers (child care,
transportation, flexible hours, reducing
co-payment), decorate waiting room,...
8Randomized Encouragement Design Features
- Analogous to marketing experiment
- Encouragement ? higher adoption rate?
- ? better overall outcomes?
- ? better outcomes for new users?
- Naturalistic, incorporate user preferences,
facilitate choice - Broader participation, external validity,
dissemination - Zelen (1979 NEJM, 1990 Stat. in Medicine
randomized consent design), Holland (1988) in
Clogg CC, ed. Sociological Methodology, Hirano et
al. (2000, Biostatistics), Wells et al. (2000,
JAMA), Duan et al. (2000, manuscript)
9Mitigating Overt Selection Bias
- Assume overt selection bias T ? X
- Assume no hidden selection bias T ? Y X
- Covariate adjustment through ANCOVA
- Stratification (through propensity score method)
- Matching (through propensity score method)
10Covariate Adjustment
- Y ? T b X g ( T X d) e
- Extrapolation can be risky when imbalance is
substantial - Y
T 1
X Pre-Tx health
T 0
11Limitations for Covariate Adjustment
- Extrapolation can be risky when imbalance is
substantial - Compare apples and oranges, rely on model to
adjust - Careful model diagnosis is essential
- Multivariate imbalance might be more problematic
- Why so popular?
- Ease of push-botton analysis
- Almost always gives an answer
- Could be a bad answer!
-
12Stratification When Covariate Is Univariate
- Stratify, then compare by stratum
- Compare apples and apples, oranges and oranges
- Y
T 1
T 0
X Pre-Tx health
13Stratification Procedure
- Stratify, then compare treated vs. untreated by
stratum - Two-sample comparison within each stratum
- ANCOVA within each stratum
- Assess interactions across strata
- Synthesize treatment effects across strata
- Weighted average
- Overall intervention effect on treated
- Overall intervention effect on untreated
- Overall intervention effect on entire pool
- Can be specified as ANCOVA with interactions
- Nonparametric regression of Y on X, stratified
by T
14Covariate Adjustment, Nonparametric Version
- OK for low dimension X
- Curse of dimensionality for high dimension X
- Y
T 1
X Pre-Tx health
T 0
15Stratification Features
- Why not used as widely as ANCOVA?
- Does not always give an answer
- Provides warning where imbalance is too severe
- Not a push-button operation, but not difficult
- How to stratify?
- Clinical judgement
- Usually not critical sensitivity analysis
recommended - Cochran-Rubin-Rosenbaum recommend 5 strata
- How to stratify with multi-dimensional
covariates? - Curse of dimensionality
- Use propensity score method to reduce
dimensionality
16Propensity Score Method
- Assume overt selection bias, no hidden selection
bias - T ? Y X
- p p(X) P(T 1 X) is the propensity score
- Example logit(p(X)) a X b
- p(X) is a balancing score (most parsimonious)
- T ? X p(X)
- Given p(X), treatment independent of outcome
- T ? Y p(X)
- Need only stratify by propensity score
- Other dimensions of X can be neglected in
assessing treatment effect
17Propensity Score Method Procedure
- Estimate p(X) P(T 1 X)
- Logistic regression of T on X
- Stratify sample (X, T, and Y) by estimated p(X)
or Xb - Sort out apples and oranges
- Analyze each stratum, compare treated vs.
untreated - Two sample comparison within stratum
- ANCOVA within stratum
- Assess interactions across strata
- Synthesize treatment effects across strata
- Weighted average...
18Propensity Score Method Stratification for X
- Stratify, then compare by stratum
- Compare apples and apples, oranges and oranges
- Xk
Xb
X1
19Propensity Score Method Stratification for Y
- Stratify, then compare by stratum
- Compare apples and apples, oranges and oranges
- Y
T 0
T 1
Xb
20Propensity Score Method Model Specification
- Specification of propensity score model
- Lean towards over-fitting vs. under-fitting?
- Model diagnosis are the covariates balanced
across treatment groups within each stratum? - Stratify by propensity score and key covariates
(one or two)? - Model misspecification less serious than ANCOVA?
- Only rank of estimated propensity score is used
- Stratification not sensitive to minor
perturbations in model - Limited empirical evidence (Drake 1993
Biometrics, Dehejia and Wahba 1999 JASA)
21Propensity Score Method Options
- Stratification
- Matching (case-control)
- Curse of dimensionality relevant, less critical
- Mahalonobis distance matching
- Match on propensity score ( a few key
covariates?) - Design stage vs. analysis stage
- Primary vs. secondary data collection
- ANCOVA regress Y on T and propensity score ( a
few key covariates? interactions?) - Nonparametric regression? Stratified by T?
22Dimension Reduction
- Fundamental challenge in ANCOVA
- Valid assessment of treatment effect can be
obtained using nonparametric regression of Y on
X, stratified by T - Curse of dimensionality
- No obvious way to reduce dimensionality?
- Propensity score method is an elegant way to
reduce dimensionality - Alternative dimension reduction methods?
- Slicing regression (Duan and Li 1991 Annals of
Statistics, Li 1991 JASA) use inverse regression
of X on Y...
23Propensity Score Method References
- Rosenbaum and Rubin (1983 Biometrika, 1984 JASA)
- Lavori, Dawson, and Mueller (1994 Stat. in
Medicine) - Rosenbaum (1995) Observational Studies,
Springer-Verlag - Rubin (1997) Annals of Internal Medicine
- DAgastino (1998 Stat. in Medicine)
- Normand et al. (2000 manuscript)
- Hirano et al. (2000 manuscript)
24Dealing with Hidden Selection Bias
- T ? Y X
- Very challenging problem, no easy solutions
- Given X, how does treatment depend on outcome?
- Overt selection bias can be made to look like
stratified randomized experiment - Hidden selection bias cannot be made to
- Rosenbaum-Rubins sensitivity analysis
- Instrumental variable analysis a la Rubin Causal
Model - Selection modeling
25Rosenbaums Sensitivity Analysis General
Principle
- How robust is the observed treatment effect
against hidden selection bias? - Analogous to pattern mixture model for missing
data - Formulate a family of plausible models for hidden
selection bias (from mild to severe) - Assess treatment effect under each model
- Determine how much hidden selection bias wipes
out treatment effect - Is this much hidden selection bias realistic?
- Specificity analysis
26Unobserved Confounder Model
- logit(p(Xi)) a Xi b Ui g 0 ? Ui ? 1
- g gt 0 maximum impact of unobserved hidden bias
- G exp(g) is the upper bound between p(Xi)s
X - Example 2 x 2 table (analyzed with Fishers
exact test) - Worst case scenario for hidden bias
- Unobserved health is a perfect predictor of
survival - Healthy patients are more likely to receive
treatment - Ui 1 for all survivors 0 for all deceaseds
- Null distribution is a tilted hypergeometric
distribution - Given g, derive P-value under tilted
hypergeometric distribution
27Rosenbaums Sensitivity Analysis Limitations
- Does not give THE answer (should we expect one?)
- Rosenbaums sensitivity analysis is based on
permutation test (tilted by hidden selection
bias) - Permutation test is the foundation for randomized
trials, but rarely used heavy computation burden - Used more in recent years, e.g., COMMIT
- Special software required for tilted permutation
test - Programming logic not difficult
- Very heavy computation burden
- Inertia for users to stay with familiar packages
-
28Instrumental Variable (IV) Analysis for RED,a la
Rubin Causal Model
- Encouragement intervention serves as instrumental
variable - Assume binary intervention (I 0, 1)
- binary treatment (T 0, 1)
- T(0) T(1) Category
- 0 0 Never takers
- 0 1 Compliers (new users)
- 1 0 Defiers (assumed to be absent)
- 1 1 Always takers
- Very likely different beyond observed
characteristics
29IV Analysis Observed Compliance Status
- I 0
- Untreated C or N
- Treated A or D
- I 1
- Untreated N or D
- Treated C or A
- Randomized encouragement design
- Compliance status distributed similarly across
intervention groups - (C) (treated I 1) - (treated I 0)
- (untreated I 0) - (untreated I 1)
30IV Analysis Intervention Effect by Subgroups
- Key assumption
- Effect of encouragement intervention mediated
entirely through treatment (exclusion
restriction) - Always takers and never takers no treatment
variation - ? no intervention effect exclusion restriction
- ? cannot assess treatment effect
- Intervention effect manifested entirely through
compliers
31Complier Average Causal Effect
- Treatment Efficacy on compliers
- CACE Program effect / Incremental adoption rate
- Program effect intent-to-treat effect for
encouragement intervention on outcome - Incremental adoption rate intent-to-treat
effect for encouragement intervention on
adoption - Distribute intervention effect on outcome over
compliers
32IV Analysis External Validity
- Treatment effect estimable only for compliers
(new users) - Intrinsic limitation of design (RED or imperfect
RCT) - Should we be concerned about treatment effect for
always takers and never takers? - Yes for efficacy trials, less so for RED
- Never taker might never adopt treatment
voluntarily - Mandate vs. choice
- Universal dissemination vs. practical
dissemination - Always takers more critical absent for new
treatments - Presence of defier likely to cancel some
intervention effect - IV estimate is conservative for true CACE
33IV Analysis Discussions
- Exclusion restriction needs to be entertained
carefully - Likelihood and Bayesian methods available under
weaker assumptions - Non-randomized encouragement design
(observational studies with instrumental
variables) - Example McClellan et al. JAMA 1994, distance to
alternative types of hospitals - IV analysis usually deflates precision
substantially - Bias-variance trade-off?
- Combine propensity score analysis with IV
analysis?
34IV Analysis References
- Sommer and Zeger (1991 Stat. in Medicine)
- Angrist, Imbens, and Rubin (1996 JASA)
- Imbens and Rubin (1997 Annals of Statistics)
- Little and Yau (1998 Psych Methods)
- Hirano, Imbens, Rubin, and Zhou (2000
Biostatistics) - Wells, et al. (2000, manuscript)
35Discussions
- Formulate research questions
- Treatment effect for whom? Adoption?
- Careful design usually more effective than
analytic solutions - Matching to avoid severe imbalance
- Promising methods for mitigating overt selection
bias - Careful modeling warranted
- Propensity score method worth exploring
- Nonparametric regression worth exploring
- Hidden selection bias very challenging
- Rosenbaums sensitivity analysis warranted
- IV analysis and selection model require careful
assessment