S-L Normand (FDA Non-Tech: 2004) A REVIEW OF CONCEPTS & APPROACHES FOR CAUSAL INFERENCE: PROPENSITY SCORES INNOVATIVE STATISTICAL APPROACHES IN HSR: BAYESIAN, MULTIPLE INFORMANTS, & PROPENSITY SCORES Sharon-Lise T. Normand Thomas R. Belin, UCLA Nicholas J. Horton, Smith College Sharon-Lise T. Normand, Harvard University (sharon@hcp.med.harvard.edu) Harvard Medical School & Harvard School of Public Health Thanks: Laura A. Petersen 1 Sharon@hcp.med.harvard.edu (ARM 2004) REGIONALIZING CARDIAC SERVICES Under FFS, incentive for hospitals to duplicate profitable cardiac service capabilities. Under non-FFS, incentive to regionalize costly services to referral institutions. Non-FFS systems provide lower rates of cardiac procedures than FFS systems. Availability of on-site cardiac procedures affects the use of such procedures. Should we regionalize invasive cardiac services? Examine use of clinically needed angiography between AMI patients treated within a regionalized system (VA) and those treated within a nonregionalized system (FFS). (Petersen, Normand, Leape, McNeil, NEJM: 2002) 3 Sharon@hcp.med.harvard.edu (ARM 2004) 4 Sharon@hcp.med.harvard.edu (ARM 2004) OUTLINE OF TALK WHAT IS A CAUSAL EFFECT? • What is a causal effect? Let Yi denote the “outcome” for the ith patient. The causal effect is: δi = Yi(treatment) – Yi(control) – Role of Randomization. • Statistical Adjustments. – Regression – Stratification – Both Receipt of Angiography if treated in a regionalized system (VA) • Concluding remarks. 5 Sharon@hcp.med.harvard.edu (ARM 2004) 2 Sharon@hcp.med.harvard.edu (ARM 2004) Receipt of Angiography if treated in a nonregionalized system (FFS) Outcomes under treatments not assigned are missing 6 Sharon@hcp.med.harvard.edu (ARM 2004) 1 S-L Normand (FDA Non-Tech: 2004) ROLE OF RANDOMIZATION ROLE OF RANDOMIZATION Properties of an RCT: 1. Experimenter determines the assignment of treatments to patients using a known mechanism. 2. Every participant has a non-zero “chance” of being assigned the treatment. 3. Outcome and treatment assignment are independent given the covariates. So what? Randomization (treatment assignment mechanism) permits: δi = Yi(treatment) – Treatment = Effect Y for those assigned to treatment Yi(control) Y for those assigned to control 7 Sharon@hcp.med.harvard.edu (ARM 2004) 8 Sharon@hcp.med.harvard.edu (ARM 2004) TYPES OF BIASES OBSERVATIONAL STUDY 1. Non-representative sample of target population. 2. Treatment assignment is non-ignorable, can not assume treatment assignment depends only on observed covariates. 3. All Measured Confounders: Properties of an Observational Study: 1. Investigator has no control over the assignment of treatments to patients so that the mechanism is unknown. 2. No longer true that every participant has a non-zero “chance” of being assigned the treatment. 3. Outcome and treatment assignment may be dependent given the covariates. • • Inexact matching on basis of covariates. Incomplete matches (discarding treated). 9 Sharon@hcp.med.harvard.edu (ARM 2004) 10 Sharon@hcp.med.harvard.edu (ARM 2004) ANALYTIC ADJUSTMENTS TYPES OF BIASES 1. Model-Based (Statistical) Adjustment: regress outcome on the confounders and estimate adjusted outcomes. 2. Matched Sampling: create categories from the confounding variables (X) such that within each category, there are both treated and control. 3. Adjust & Stratify: Replace confounders by a single summary, e(X), and construct strata/matches/weight. Estimate treatment differences within each stratum/pair etc. RCT has random assignment of treatments but not random selection of subjects from the target population. Observational study has random sampling of target population but not random assignment of treatments. 11 Sharon@hcp.med.harvard.edu (ARM 2004) 12 Sharon@hcp.med.harvard.edu (ARM 2004) 2 S-L Normand (FDA Non-Tech: 2004) SO THEY ARE DIFFERENT! (USUAL) REGRESSION CHARACTERISTICS OF ACC/AHA CLASS I HOSPITALIZED AMI PATIENTS VA FFS Stand. Characteristic Diff. (N = 1104) (N = 10464) Hx Hypertension 67% 59% 18% Black Prior MI Asthma/COPD ST-Elevation 13% 38% 31% 49% 4% 35% 23% 43% 35% 7% 17% 10% Ischemia 29% 12% 44% Angiography 44% 51% logit(P(Y=1|X)) = β0 + α(VA) + βTX Performs poorly when the variances of the confounders in the treated and control groups are unequal. Unequal variances of confounders are not uncommon in observational studies. . 13 Sharon@hcp.med.harvard.edu (ARM 2004) 14 Sharon@hcp.med.harvard.edu (ARM 2004) HOW TO MAKE VALID CASUAL INFERENCE? REGRESSION Estimate: e(X) = P(VA = 1| X) Infer the assignment mechanism. Difficult to assess whether treated and control groups overlap enough on X to allow sensible estimation of an effect. Ensure everyone has a chance of being assigned the treatment. • If treatment groups have different covariate distributions, model-based adjustments depend on the form of model. • But, distn of X may differ for treated and control ⇒ must impose linearity and extrapolate over different regions of covariates. 0 < e(X) < 1 Guarantee we are comparing similar patients. Want comparable e(X) between treatment groups. Estimate effect of treatment (VA) on Y (angiography). E(Y| e(X),VA = 1) – E(Y| e(X),VA = 0) 15 Sharon@hcp.med.harvard.edu (ARM 2004) 16 Sharon@hcp.med.harvard.edu (ARM 2004) 200 TREATMENT ASSIGNMENT MECHANISM CLASS I VA PATIENTS (N = 1104) 50 100 150 PROPENSITY SCORE: TREATMENT ASSIGNMENT MECHANISM 0 P(VA = 1 | X) = e(X) = demographics (age, race) + severity on entry (ST-Elevation, chestpain duration, etc) + comorbidity (Hx of AMI, CHF, Diabetes, etc). 0.0 0.2 0.4 0.6 0.8 1.0 4000 6000 Estimated Propensity Score: e(X) = P(VA) 0 2000 CLASS I FFS PATIENTS (N = 10464) 17 Sharon@hcp.med.harvard.edu (ARM 2004) 0.0 0.2 0.4 0.6 0.8 Estimated Propensity Score: e(X) = P(VA) 18 3 S-L Normand (FDA Non-Tech: 2004) 1090 MATCHED PAIRS 200 300 REGRESS & STRATIFY (OR REGRESS & MATCH) 0 0.0 0.4 FFS Stand. Diff. Hx Hypertension 67% 66% 2% Black Prior MI Asthma/COPD ST-Elevation 12% 38% 30% 49% 12% 42% 30% 47% 1% -7% 1% 4% Ischemia 27% 26% 2% ROC Area = 0.80 300 200 0.0 0.2 0.4 0.6 0.8 1.0 Estimated Propensity Score: P(VA = 1) 20 ADJUSTED COMPARISON USING 1090 MATCHED PAIRS OF CLASS I PATIENTS Adjusted for Patient Characteristics Only No. (%) of Discordant Pairs No. (%) of Discordant Pairs in which VA Received Angiography 553 (51) 242 (44) P-Value 0.0038 Not adjusted for hospital clustering. OR from regression adjustment [95% CI] using all patients: 0.77 [0.66, 0.88]. 21 Sharon@hcp.med.harvard.edu (ARM 2004) 22 Sharon@hcp.med.harvard.edu (ARM 2004) ADJUSTED COMPARISON USING 1085 MATCHED PAIRS OF CLASS I PATIENTS: CONCLUSIONS FROM EXAMPLE Adjusted for Patient Characteristics and Hospital Availability of Angiography No. (%) of Discordant Pairs in which VA Received Angiography 255 (50) Under-use of angiography regardless of system of care. Under-use is higher within the regionalized system [VA]. Differences in under-use between VA and FFS associated with on-site availability of cardiac procedure technology. P-Value 0.86 Not adjusted for hospital clustering. OR from regression adjustment [95% CI] using all patients: 1.04 [0.89, 1.21]. 23 Sharon@hcp.med.harvard.edu (ARM 2004) 1.0 100 CHARACTERISTICS OF ACC/AHA CLASS I POST MATCHING (1090 PAIRS) VA 0.8 CLASS I FFS (N = 1090) Sharon@hcp.med.harvard.edu (ARM 2004) Characteristic 0.6 0 19 505 (47) 0.2 Estimated Propensity Score: P(VA = 1) Sharon@hcp.med.harvard.edu (ARM 2004) No. (%) of Discordant Pairs CLASS I VA (N = 1090) 100 Stratify (or match) patients based on the estimated propensity score. Within each stratum, distribution of observed confounders differs only randomly between treated and control subjects. Regardless: matching towards a population who looks like those who received the treatment. 24 Sharon@hcp.med.harvard.edu (ARM 2004) 4 S-L Normand (FDA Non-Tech: 2004) CONCLUDING REMARKS • Role for propensity score approaches in HSR. • Provides a principled and practical approach for assessing appropriateness and robustness of assumptions. • No Panacea – doesn’t overcome hidden bias problems. • Should you always use? 25 Sharon@hcp.med.harvard.edu (ARM 2004) 5