Outline Propensity Score Adjustment in Survival Models Carolyn Rutter Group Health Cooperative rutter.c@ghc.org AcademyHealth, Seattle WA June 25, 2006 • 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? Propensity score: the probability that a person receives treatment, or is exposed, given a set of observed covariates, X. 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 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. June 25, 2006 AcademyHealth, Seattle WA Depression & Mortality among Type 2 Diabetics 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. June 25, 2006 Propensity Scores AcademyHealth, Seattle WA Depression is common in patients with type 2 diabetes 11% 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 June 25, 2006 AcademyHealth, Seattle WA 1 Depression & Mortality among Type 2 Diabetics 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 depression 3749 without major depression 3 year Mortality Outcome All-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 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 June 25, 2006 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 PH Model Results Method Survivor function: S(t) = Pr(T*>t)=1-F(t) T* event time Hazard function: instantaneous event rate λ (t ) = f (t ) − ∂S (t ) / ∂t = S (t ) S (t ) Se(estimate) HR P-value 0.14 1.40 <0.02 Minimum 0.77 Adjustment* 0.14 2.16 <0.001 Full 0.26 Adjustment† 0.16 1.30 0.09 * 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) Cox proportional λ (t ) = λ (t ) exp(Zβ ) 0 hazards model Unspecified Baseline hazard AcademyHealth, Seattle WA June 25, 2006 Estimate Unadjusted 0.34 AcademyHealth, Seattle WA June 25, 2006 mediator Z Depression X Self Care Disease Severity Y Death Z X Y mediator common cause Education Age, Sex common cause June 25, 2006 June 25, 2006 2 Step 1: Estimate propensity scores Propensity Score Adjustment: 3-Step Process 1. Estimate propensity score 2. Evaluate covariate balance given propensity scores 3. Incorporate propensity score in analyses to ‘synthetically balance’ the sample • • • • Stratification Regression Matching Weighting Use logistic regression (or other method, e.g., CART) to estimate P(Z=1|X) = πi, propensity score logit(Z) =Xα 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 πˆ i the estimated propensity score for the ith individual 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 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 Propensity score missing for 6.6% None missing propensity score AcademyHealth, Seattle WA June 25, 2006 Step 2: Check covariate balance Percent Sedentary Strata Strata 1 Strata 2 Strata 3 Strata 4 Not Depressed 826 22% 794 21% 775 21% 736 20% 339 9% 146 4% 133 4% 3749 88% Depressed 27 5% 58 11% 78 15% 116 23% 87 17% 67 13% 80 16% 513 12% 853 20% 852 20% 853 20% 852 20% 426 10% 213 5% 213 5% 4262 June 25, 2006 AcademyHealth, Seattle WA June 25, 2006 Propensity Strata Total AcademyHealth, Seattle WA June 25, 2006 Strata Strata Strata 5 6 7 AcademyHealth, Seattle WA all 1 2 3 4 5 6 7 Not depressed Depressed N 27.1 4.8 16.2 23.1 39.4 51.6 66.4 78.2 44.3 7.4 12.1 26.9 41.4 51.7 67.2 73.5 4262 853 852 853 852 426 213 213 June 25, 2006 3 Step 3: Incorporate Propensity Scores into Proportional Hazards Model Regression-adjustment in the PH model λ (t ) = λ0 (t ) exp( Zβ ) λ (t ) = λ0 (t ) exp(Zβ + πˆθ ) 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, Assume proportionality: check this assumption using Shoenfeld residuals. Z exp( Z β + πˆ β ) Zi = rˆiZ = Zi − Zi ∑ j j∈Ri j ∑ exp(Z β + πˆ β ) j j∈Ri rˆiπ = πˆi −πˆi πˆi = ∑ πˆ j∈Ri j AcademyHealth, Seattle WA j exp(Z j β + πˆ j β ) ∑ exp(Z β + πˆ β ) j∈Ri June 25, 2006 j j j AcademyHealth, Seattle WA June 25, 2006 Schoenfeld Residuals PH Model Results Little evidence for non-proportional hazards in propensity scores. Method Estimate Se(estimate) HR P-value Min Adj 0.77 0.14 2.16 <0.001 Full Adj 0.26 Regression 0.25 0.16 0.14 1.30 1.28 0.09 0.08 Correlation between Schoenfeld-residual and rank-time Depression: 0.02 Propensity: -0.06 June 25, 2006 AcademyHealth, Seattle WA June 25, 2006 PH Model Results Stratification-adjustment in the PH model λ (t ) = λ0m (t ) exp(Zβ ) mth strata Stratified likelihood ⎛ δ exp( z β ) ⎞ j j ⎟ L( β ) = ∏ Lm (β ) = ∏ ∏ ⎜ ⎜ ⎟ exp( z β ) m =1 m =1 j∈S m ∑k∈R k j ⎝ ⎠ M M 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 0.14 1.27 0.10 0.24 δj : censoring indicator (1 if death obs) June 25, 2006 AcademyHealth, Seattle WA June 25, 2006 AcademyHealth, Seattle WA 4 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(Xα) 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 λ (t ) = λ0m (t ) exp(Zβ ) mth pair ⇒ only 2/513 depressed excluded ⎛ δ exp( z β ) ⎞ M M j j ⎟ L( β ) = ∏ Lm (β ) = ∏ ∏ ⎜⎜ ⎟ m =1 j∈S m ∑ k∈R exp( z k β ) m =1 j ⎝ ⎠ Within each matched pair, only the first death contributes to the likelihood leading to additional loss of information. AcademyHealth, Seattle WA June 25, 2006 Weighting-adjustment in the PH model (IPW) 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 0.25 0.14 1.28 0.08 Stratified 0.24 0.14 1.27 0.10 Matching 0.26 0.21 1.30 0.21 λ (t ) = λ0 (t ) exp(Zβ ) Weighted partial Likelihood Function N ⎛ δ i wi exp( z i β ) ⎞⎟ Limits options L( β ) = ∏ ⎜ for handling ties ⎜ ⎟ w exp( z β ) i =1 ∑ j∈R j j ⎝ ⎠ i up-weight individuals with ‘unexpected’ exposure wi = [z i πˆ i + (1 − z i )(1 − πˆ i )] −1 Performs best when weights are estimated (Qi, Wang, Prentice, JASA ,2005) AcademyHealth, Seattle WA June 25, 2006 AcademyHealth, Seattle WA June 25, 2006 PH model results Se(estimate) HR P-value Unadjusted 0.34 Method Estimate 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 AcademyHealth, Seattle WA Z Covariate models X Y Estimate the effect of Z on Y conditional on X June 25, 2006 5 Combined Adjustments λ (t ) = λ0 (t ) exp(Zβ + πˆθ ) Covariate models Z Y X Regression adjust and weight. Propensity Synthetically balances X across Z Propensity models: P(Z|X) IPW does not depend on estimating effects of Y | (Z and X) June 25, 2006 AcademyHealth, Seattle WA June 25, 2006 PH Model Results Doubly Robust 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 0.14 1.27 0.10 0.24 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 Propensity Model True No Yes No Regression Model True Yes An approach that is robust to misspecification of the regression model OR the propensity model. AcademyHealth, Seattle WA June 25, 2006 AcademyHealth, Seattle WA June 25, 2006 Score Adjustment, ϕi Doubly Robust Estimators Idea: weighted estimators use only observed outcomes. DR estimators incorporate unobserved outcomes through their expected values. ⇒ Increase efficiency, increase robustness n 1 ⎛ ⎜ ⎝ U A (β ) = ∑ ∑ Δiw i δ i ⎜ z i − i =1 Z = 0 ∑ j ∈R w j z j exp(z j β + x j γ ) ⎞⎟ ⎛ Δ i − w i−1 ⎞ ⎟ϕ i = 0 −⎜ ∑ j ∈R w j exp(z j β + x j γ ) ⎟⎠ ⎜⎝ w i−1 ⎟⎠ i i ϕi is an augmentation term that is a function of the regression model, M(Y|X, β,γ) where Y=(δ, T): Adjusted Score Function: n 1 ⎛ ⎜ ⎝ U A (β ) = ∑ ∑ Δiw i δ i ⎜ z i − i =1 Z = 0 ∑ j ∈R w j z j exp(z j β + x j γ ) ⎞⎟ ⎛ Δ i − w i−1 ⎞ ⎟ϕ i = 0 −⎜ ∑ j ∈R w j exp(z j β + x j γ ) ⎟⎠ ⎜⎝ w i−1 ⎟⎠ i i ⎡ ⎛ ϕi = E⎢δi ⎜ zi − ⎢ ⎜ ⎣ ⎝ weighted score ∑ ∑ ⎤ z j exp(z j β + x jγ ) ⎞ ⎟ M (Y X , β , γ )⎥ ⎟ ⎥ exp(z j β + x jγ ) j∈Ri ⎠ ⎦ j∈Ri Δ i indicates observing the ‘assigned’ (patient selected) treatment June 25, 2006 June 25, 2006 6 Doubly Robust Estimator Doubly Robust Estimates n 1 ⎛ ⎜ ⎝ U A ( β ) = ∑ ∑ Δ iw i δ i ⎜ z i − ⎛ ∑ w z exp(z j β + x j γ ) ⎞⎟ ⎛ Δ i − w i−1 ⎞ ⎟⎟ϕ i = 0 U A ( β ) = ∑ ∑ Δ i w i δ i ⎜ z i − j ∈R j j −⎜ −1 ⎜ i =1 Z = 0 ∑ j ∈R w j exp(z j β + x jγ ) ⎟⎠ ⎜⎝ w i ⎠ ⎝ n i =1 Z = 0 1 ∑ j ∈R w j z j exp(z j β + x jγ ) ⎞⎟ ⎛ Δ i − w i−1 ⎞ −⎜ −1 ⎟⎟ϕ i = 0 ∑ j ∈R w j exp(z j β + x j γ ) ⎟⎠ ⎜⎝ w i ⎠ i i i i Expected value is 0 if propensity model is true Can calculate DR estimates iteratively: 1. Calculate starting values using PH 2. Estimate ϕi via simulation given M(Y|X, β,γ) and current parameter estimates, including baseline hazard (e.g., Nelson-Aalen estimators) U A (β ) = ∑{wi δi (z i − z i (wi ) − E (z i − z i ))} + ∑ϕi = 0 observed all Expected value is 0 if regression model is true June 25, 2006 1. ϕˆi = ⎞ eβ ∑ Rik* zk exp(x jγˆ) 1 m * ⎛⎜ ⎟ δk ⎜ zi − β ∑ * m k =1 ⎝ e ∑ Rik zk exp(x jγˆ) + ∑ Rik* (1− zk ) exp(x jγˆ) ⎟⎠ ϕˆi = 1 m *⎛ eβ A* ⎞ δk ⎜⎜ zi − β * ik * ⎟⎟ ∑ e Aik + Bik ⎠ m k =1 ⎝ + TS approx Use Newton-Raphson to solve the adjusted score for β & γ 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 DR AcademyHealth, 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 Propensity Adjustment for Survival Models: Recent Work • 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) • Sturmer et al. AJE, 2005, Develop a regressioncalibration 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 June 25, 2006 7 Propensity Models Additional research: • More than two treatment/exposure groups Leon 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 June 25, 2006 AcademyHealth, Seattle WA 8