June 25, 2006 Propensity Score Adjustment in Survival Models Carolyn Rutter Group Health Cooperative [email protected] AcademyHealth, Seattle WA
Mar 27, 2015
June 25, 2006
Propensity Score Adjustmentin Survival Models
Carolyn RutterGroup Health Cooperative
AcademyHealth, Seattle WA
June 25, 2006
Outline
• Propensity Scores: General Ideas • Background: depression & mortality
among type 2 diabetics• Propensity Scores applied to
depression & mortality
June 25, 2006
Example: Is depression associated with increased mortality in type 2 diabetics?
Underlying question: Does depression increase the risk of
death ?
Estimate the causal effect of treatment on response
exposure outcome Z Y
AcademyHealth, Seattle WA
June 25, 2006
Propensity Scores
Propensity score: the probability that a person receives treatment, or is exposed, given a set of observed covariates, X.
Randomized Study: P(Tx)=0.5, the propensity score is independent of patient characteristics and the distribution of P(Tx) is the same across treatment groups.
Observational Study: P(Tx|X) depends on patient characteristics and differs between treatment groups (because Tx is associated with covariates), so that the treated group has a higher propensity for treatment than the untreated group.
AcademyHealth, Seattle WA
June 25, 2006
Basic Ideas behind Propensity Score Methods
Reduce bias by comparing treated and untreated individuals who have the same propensity for treatment/exposure
Key assumption: Strongly Ignorable Treatment Assignment
The outcome is conditionally independent of treatment assignment given observed covariates
Y P(Z|X)
After adjusting for observed covariates, treatment assignment doesn’t inform the response.
No unmeasured confounders.
AcademyHealth, Seattle WA
June 25, 2006
Depression & Mortality among Type 2 Diabetics
Depression is common in patients with type 2 diabetes11% to 15% meet criteria for major depression
Depressed diabetic patients tend to have
– poorer self-management (diet, exercise, blood glucose checks)
– more lapses in refilling prescribed medications (oral hypoglycemics, lipid lowering, anti-hypertensive)
– have cardiac risk factors (smoking, obesity, sedentary lifestyle)
Studies have linked depression to increased mortality among diabetics, but these used a small number of patients, with medical diagnoses based on self report
AcademyHealth, Seattle WA
June 25, 2006
The Pathways Study: a population-based epidemiologic study of over 4000 patients with diabetes enrolled in an HMO.
4262* included in following analyses
513 with major depression3749 without major depression
Katon, Rutter, Simon et al “The association of comorbid depression with mortality in patients with type 2 diabetes.” Diabetes Care. 2005 Nov; 28(11):2668-72.
AcademyHealth, Seattle WA
Depression & Mortality among Type 2 Diabetics
June 25, 2006
3 year Mortality OutcomeAll-cause mortality: May 2001(start recruitment) – May 2004
5/1/2001 – 12/31/2003 (first 31 months): GHC automated health care records + Washington State mortality data 90% of deaths in the State mortality data were in GHC records
1/1/2004 – 4/30/2004 (last 5 months): GHC data alone.
Censoring at the end of the study or disenrollment
Deaths over a 3-year period:336 ( 9.0%) in 3749 patients without major depression 60 (11.7%) in 497 patients with major depression
June 25, 2006
Proportional Hazards Model
Survivor function: S(t) = Pr(T*>t)=1-F(t)
T* event timeHazard function: instantaneous event rate
Cox proportionalhazards model
)exp()()( Ztt 0
UnspecifiedBaseline hazard
)()(
)(/)(
)(tStf
tSttS
t
AcademyHealth, Seattle WA
June 25, 2006
PH Model Results
AcademyHealth, Seattle WA
* Known confounders: gender, age, race/ethnicity, education† Potential behavioral and disease severity confounders &/or mediators: BMI, current smoker, sedentary lifestyle, HbA1c, use of oral hypoglycemics, use of insulin, complications of diabetes, (pharmacy-based) comorbidity measure (excluding depression meds)
Method Estimate
Se(estimate)
HR P-value
Unadjusted
0.34 0.14 1.40 <0.02
Minimum Adjustment*
0.77 0.14 2.16 <0.001
Full Adjustment†
0.26 0.16 1.30 0.09
June 25, 2006
ZDepression
XSelf Care
Disease Severity
Age, Sex
Education
YDeath
mediator
common cause
June 25, 2006
Z X Y
mediator
common cause
June 25, 2006
Propensity Score Adjustment:
3-Step Process1. Estimate propensity score 2. Evaluate covariate balance given
propensity scores3. Incorporate propensity score in analyses
to ‘synthetically balance’ the sample• Stratification• Regression• Matching• Weighting
AcademyHealth, Seattle WA
June 25, 2006
Step 1: Estimate propensity scores
Use logistic regression (or other method, e.g., CART) to estimate P(Z=1|X) = i, propensity score
logit(Z) =X
Focus is on prediction rather than estimation. – Include all potential confounders, but leave out factors
related only to the exposure or outcome (Brookhart et al, 2006, AJE)
– Include interaction effects as needed– ROC curve can be used to evaluate fit, but doesn’t provide
insight about appropriate covariates
the estimated propensity score for the ith individuali
AcademyHealth, Seattle WA
June 25, 2006
Step 1: Estimate propensity for depression
proc logistic descending;model major=age male smoke obese somecoll sedentary cardio outofcontrol treatint rxrisk2 /outroc=roc;run;
Estimated AUC=0.72
Propensity score missing for 6.6%
AcademyHealth, Seattle WA
June 25, 2006
Step 1: Estimate propensity for depression
proc logistic descending;
model major=age male smoke obese somecoll sedentary cardio outofcontrol treatint rxrisk2 + missing value indicators /outroc=roc;
run;
Estimated AUC=0.72
None missing
propensity score
AcademyHealth, Seattle WA
June 25, 2006
Propensity Strata
AcademyHealth, Seattle WA
Strata1
Strata2
Strata3
Strata4
Strata5
Strata6
Strata7
NotDepressed
82622%
79421%
77521%
73620%
3399%
1464%
1334%
374988%
Depressed
275%
5811%
7815%
11623%
8717%
6713%
8016%
51312%
Total 85320%
85220%
85320%
85220%
42610%
2135%
2135%
4262
June 25, 2006
Step 2: Check covariate balance
Strata Not depressed
Depressed
N
all 27.1 44.3 4262
1 4.8 7.4 853
2 16.2 12.1 852
3 23.1 26.9 853
4 39.4 41.4 852
5 51.6 51.7 426
6 66.4 67.2 213
7 78.2 73.5 213
Percent Sedentary
June 25, 2006
Step 3: Incorporate Propensity Scores into Proportional Hazards
Model
1. Regression: Proportional hazards across different levels of the propensity score
2. Stratification: Allow different baseline hazards across propensity strata
3. Matching: Allow different baseline hazards for each matched pair
4. Weighting: Assume a common baseline hazard,
)exp()()( 0 Ztt
AcademyHealth, Seattle WA
June 25, 2006
Regression-adjustment in the PH model
Assume proportionality: check this assumption usingShoenfeld residuals.
)ˆexp()()( 0 Ztt
AcademyHealth, Seattle WA
iiiZ ZZr ˆ
i
i
Rjjj
Rjjjj
i Z
ZZ
Z)ˆexp(
)ˆexp(
iiir ˆˆˆ
i
i
Rjjj
Rjjjj
i Z
Z
)ˆexp(
)ˆexp(ˆ
ˆ
June 25, 2006
Schoenfeld ResidualsLittle evidence for non-proportional hazards in propensity
scores.
Correlation between Schoenfeld-residual and rank-time
Depression: 0.02 Propensity: -0.06
June 25, 2006
PH Model Results
AcademyHealth, Seattle WA
Method Estimate
Se(estimate)
HR P-value
Min Adj 0.77 0.14 2.16 <0.001
Full Adj 0.26 0.16 1.30 0.09
Regression
0.25 0.14 1.28 0.08
June 25, 2006
Stratification-adjustment in the PH model
Stratified likelihood
j : censoring indicator (1 if death obs)
)exp()()( Ztt m0 mth strata
M
m Sj Rk k
jjM
mm
mj
z
zLL
11 )exp(
)exp()()(
AcademyHealth, Seattle WA
June 25, 2006
PH Model Results
AcademyHealth, Seattle WA
Method Estimate
Se(estimate)
HR P-value
Min Adj 0.77 0.14 2.16 <0.001
Fully Adj
0.26 0.16 1.30 0.09
Regression
0.25 0.14 1.28 0.08
Stratified
0.24 0.14 1.27 0.10
June 25, 2006
Matched Propensity Score Analysis
1. Use the full sample to estimate propensity scores
2. Identify matched pairs based on linear predictor from the propensity model. Matching within ±0.25*SE(X) is recommended by Rosenbaum & Rubin (1983, 1985)
3. Assess matching: differences between matched and unmatched individuals; balance within matched sample.
4. Analyze data, accounting for matching.
AcademyHealth, Seattle WA
June 25, 2006
Matching-adjustment in the PH model
only 2/513 depressed excluded
Within each matched pair, only the first death contributes to the likelihood leading to additional loss of information.
)exp()()( Ztt m0 mth pair
AcademyHealth, Seattle WA
M
m Sj Rk k
jjM
mm
mj
z
zLL
11 )exp(
)exp()()(
June 25, 2006
PH model results
AcademyHealth, Seattle WA
Method Estimate
Se(estimate)
HR P-value
Min Adj 0.77 0.14 2.16 <0.001
Full Adj 0.26 0.16 1.30 0.09
Regression
0.25 0.14 1.28 0.08
Stratified
0.24 0.14 1.27 0.10
Matching
0.26 0.21 1.30 0.21
June 25, 2006
)exp()()( Ztt 0
AcademyHealth, Seattle WA
Weighting-adjustment in the PH model (IPW)
Weighted partial Likelihood Function
N
i Rj jj
iii
izw
zwL
1 )exp(
)exp()(
up-weight individuals with ‘unexpected’ exposure
1)ˆ1)(1(ˆ iiiii zzw
Limits optionsfor handling ties
Performs best when weights are estimated (Qi, Wang, Prentice, JASA ,2005)
June 25, 2006
PH model results
AcademyHealth, Seattle WA
Method Estimate
Se(estimate)
HR P-value
Unadjusted
0.34 0.14 1.40 <0.02
Min Adj 0.77 0.14 2.16 <0.001
Full Adj 0.26 0.16 1.30 0.09
Regression
0.25 0.14 1.28 0.08
Stratified
0.24 0.14 1.27 0.10
Matching
0.26 0.21 1.30 0.21
IPW 0.36 0.09 1.43 <0.005
June 25, 2006
Z X Y
Covariatemodels
Estimate the effect of Z on Y conditional on X
June 25, 2006
IPW does not depend on estimating effects of Y | (Z and X)
Z X Y
Covariatemodels
Propensity
Synthetically
balancesX across
Z
Propensity models: P(Z|X)
June 25, 2006
Combined Adjustments
Regression adjust and weight.
AcademyHealth, Seattle WA
)ˆexp()()( 0 Ztt
June 25, 2006
PH Model Results
AcademyHealth, Seattle WA
Method Estimate
Se(estimate)
HR P-value
Min Adj 0.77 0.14 2.16 <0.001
Full Adj 0.26 0.16 1.30 0.09
Regression
0.25 0.14 1.28 0.08
Stratified
0.24 0.14 1.27 0.10
Matching
0.26 0.21 1.30 0.21
IPW 0.36 0.09 1.43 <0.005
IPW+Reg
0.36 0.09 1.43 <0.005
June 25, 2006
Doubly Robust
RegressionModel True
Propensity Model True
No Yes
No
Yes
An approach that is robust to misspecification of theregression model OR the propensity model.
AcademyHealth, Seattle WA
June 25, 2006
Doubly Robust EstimatorsIdea: weighted estimators use only
observed outcomes. DR estimators incorporate unobserved outcomes through their expected values. Increase efficiency, increase robustness
Adjusted Score Function:
iindicates observing the ‘assigned’ (patient selected) treatment
weighted score
01
1
01
1
n
i Zi
i
ii
Rj jjj
Rj jjjjiiiiA w
wxzw
xzzwzwU
i
i
)exp(
)exp()(
June 25, 2006
Score Adjustment, i
i is an augmentation term that is a function of the regression model, M(Y|X, ) where Y=(, T):
),,()exp(
)exp(
XYM
xz
xzzzE
i
i
Rj jj
Rj jjj
iii
01
1
01
1
n
i Zi
i
ii
Rj jjj
Rj jjjjiiiiA w
wxzw
xzzwzwU
i
i
)exp(
)exp()(
June 25, 2006
Doubly Robust Estimator
01
1
01
1
n
i Zi
i
ii
Rj jjj
Rj jjjjiiiiA w
wxzw
xzzwzwU
i
i
)exp(
)exp()(
0 all
iobserved
iiiiiiiA zzEwzzwU )()()(
Expected value is 0 if regression model is true
Expected value is 0 if propensity model is true
June 25, 2006
Doubly Robust Estimates
Can calculate DR estimates iteratively: 1. Calculate starting values using PH2. Estimate i via simulation given M(Y|X, ) and current
parameter estimates, including baseline hazard (e.g., Nelson-Aalen estimators)
1. Use Newton-Raphson to solve the adjusted score for
01
1
01
1
n
i Zi
i
ii
Rj jjj
Rj jjjjiiiiA w
wxzw
xzzwzwU
i
i
)exp(
)exp()(
m
k ikik
ikiki
m
k jkikjkik
jkikiki
BAe
Aez
m
xzRxzRe
xzRez
m
1**
**
1**
**
1ˆ
)ˆexp()1()ˆexp(
)ˆexp(1ˆ
+ TS approx
June 25, 2006
PH model results
Method Estimate
Se(estimate)
HR P-value
Min Adj 0.77 0.14 2.16 <0.001
Full Adj 0.26 0.16 1.30 0.09
Regression
0.26 0.16 1.30 0.09
Stratified
0.24 0.14 1.27 0.10
Matching
0.26 0.21 1.30 0.21
IPW 0.36 0.09 1.43 <0.005
DRAcademyHealth, Seattle WA
June 25, 2006
Propensity Adjustment Compared to Inclusion of
Covariates• Separate models for treatment assignment and
outcome. Focus on synthetic balance of sample.
• Maintain power while adjusting for many covariates– Need about 10-15 events per independent variable
examined
• Multiple ways to adjust, allowing different assumptions about proportionality of hazards
• Can no longer make inference about individual covariates
June 25, 2006
Propensity Adjustment for Survival Models
• Omitting covariates from PH models may result in attenuation of estimates for included covariates (Mitra & Heitjen, Stat in Med, 2006).
• Covariate adjustment in PH model may reduce bias in estimates of covariate effects (Lagakos & Shoenfeld, Biometrics, 1984) but has little effect on the variance of estimates. (Anderson & Flemming, Biometrika, 1995)
June 25, 2006
Propensity Adjustment for Survival Models: Recent
Work• Sturmer et al. AJE, 2005, Develop a
regression-calibration approach to adjust for error in estimated propensity scores.
• Mitra & Heitjen, Stat in Med, 2006, develop a method for determining the effect an umeasured confounder would need to have to explain observed differences.
June 25, 2006
Propensity ModelsAdditional research:
• More than two treatment/exposure groupsLeon AC, Mueller TI, Solomon DA, Keller
MB. 2001, Stat Med. Luellen JK, Shadish WR, & Clark MH.
2005, Evaluation Review, & references therein
Imbens G. Biometrika, 2000.• Continuous treatment/exposure measures
AcademyHealth, Seattle WA