Observational Ob ti lD Data t A Analysis l i ffor Comparative p Effectiveness Research Sebastian Schneeweiss, MD, ScD J Jeremy R Rassen, S ScD D Division of Pharmacoepidemiology and Pharmacoeconomics, Department of Medicine, Harvard Medical School Schneeweiss S, Clin Pharm & Ther 2007 1 Potential conflicts of interest PI of the Brigham & Women’s Hospital DEcIDE Research Center on Comparative Effectiveness Res Res. Co-investigator of the Mini Sentinel System funded by FDA No paid consulting or speaker fees from pharmaceutical manufacturers Consulting/ board membership in past year: HealthCore; H lthC Th The LLewin i Group; G RTI; RTI ii4sm; ii4 WHISCON Investigator-initiated research grants from Pfizer, H lthC HealthCore Multiple grants from NIH to study all sorts of things 2 Objective of Comparative Effectiveness Research Placebo comparison Efficacy y Effectiveness* Effectiveness (Can it work?) (Does it work in routine care?) Most RCTs for drug g approval pp (or usual care) Active comparison (head-to-head) Goal of CER * Cochrane A. Nuffield Provincial Trust, 1972 3 As much as we all love randomized effectiveness trials It is an unrealistic expectation that we will have head-to-head randomized trails ffor every intervention and its combinations i every patient in ti t subgroup b that exactly mimic routine care We need Effectiveness evidence in a timely manner manner. Randomized studies may take some time to conduct About Ab t 85% off the th CER evidence id is i from f nonexperimental data!* * Academy Health Report June 2009 Opportunities pp for non-randomized CE research with electronic healthcare data: Representative of routine care Spectrum of disease severity Spectrum of co-morbidities Co-medications R l world Real ld adherence dh Very large size Infrequent f exposure, recently l marketed k d medications d Many subgroups to study treatment effect heterogeneity Long L follow-up f ll With hard clinical endpoints Produce P d results l fast f 5 A Preferred Data Structure for fast, improved CER Example: Older adults using Medicare data 2006 2007 2008 Medicare Part A: Hospitalizations Medicare Part B: Medical services Medicare Part D: Pharmacy dispensings Laboratory results data Medicare Current Beneficiary Survey plus+ Ongoing disease registries, e.g. cancer registry, RA registry New study‐specific registries 2009 Claims data describe the sociology of health care and its recording practice in light of economic interests Schneeweiss, J Clin Epi 2005 7 Electronic health care information in each Center Constant flow of data with little delay and at low cost Millions of patients with defined person–time denominator Data reflect routine care Generalizable to large population segments HIPAA compliance protects patient privacy Claims Data • • • • • Member ID Plan Gender A Age Dates of Eligibility Administrative Data • Member ID • Prescribing physician h i i • Drug dispensed (NDC) • Quantity and date dispensed • Drug strength • Days supply • Dollar amounts Pharmacy Cl i Claims Data • Member ID • Physician or Facility identifier • Procedures (CPT-4, revenue codes, ICD-9) • Diagnosis (ICD-9CM, DRG) • Admission and discharge dates • Date and place of service • Dollar amounts Physician and F ilit Claims Facility Cl i Data Supplemental Data • Member ID • Lab Test Name • Result Lab Test R Results lt Data • • • • • • • Member ID Income Net Worth Education Race & Ethnicity Life Stage Life Style St le Indicators Consumer Elements • • • • • • Member ID Subspecialty notes Endoscopy reports Histology reports Radiology reports Free text notes Electronic M di l Medical Records Computerized Linked Longitudinal Dataset 8 Challenges that come to mind Can we handle confounding by indication? Can we reliably assess the relevant outcomes? Can we identify relevant subgroups? Can we really study long-term outcomes? …and don’t forget the basics 9 What to do when RCTs find different treatment effects than nonexperimental studies? 10 The Mantra of Trialists: Non-randomized studies are inherently biased due to patient selection into treatment groups 11 Confounding Patient factors become confounders (C) if they are associated with treatment choice and are also independent predictors of the outcome: C Randomization Trt Severity Prognosis C Comorbidity bidit Outcome 12 A spectrum of effectiveness research C Coxib Potential for confounding by indication GI event e.g. Coxibs and gastric toxicity “Effectiveness Research” e.g. e g Coxibs and cardiac events C Coxib MI Unintended effects Intended effects Intentionality of treatment effect by the prescriber “Safety research” 13 A causal experiment p P ti t A: Patient A Drug I Improvement t R i d titime Rewind Patient A: No Drug No improvement Watterson B. The Authoritative Calvin and Hobbes, p67-2 14 Design choice by source of exposure variation Exposure variation within patients yes no Case-crossover study Exposure variation between patients yes no Crossover trial Cohort study Exposure variation between providers yes Randomized controlled trial Instrumental variable analysis Cluster randomized trial Dealing with confounding Confounding Measured C f Confounders d Design Unmeasured Confounders Analysis •Restriction •Standardization •Matching •Stratification •Regression P Propensity it scores •Marginal Structural Models 16 Schneeweiss, PDS 2006 0 30 0.30 only hdd‐PS 0.60 Mos st + hhd‐PS adjustm adment just. Mocified re + spec adjustm ment covaars Age‐s age‐sex‐ ‐sex‐ race‐y yearyear ‐race adjustm adjument sted Unadjus sted Unadjuusted log elative risk) risk) log (re (reelative Adjustment in non-randomized research Clopidogrel ‐ MI (2) C ib ‐ GI bleed Coxib bl d Statin ‐ death TCA suicide (1) 0.00 ‐0.30 ‐0.60 Dealing with confounding Confounding Measured C f Confounders d Design Analysis Unmeasured Confounders Unmeasured, but measurable in substudy •Restriction •Standardization •Matching •Stratification •2-stage sampl. •Regression P Propensity it scores •Marginal Structural Models Unmeasurable Design Analysis •Ext. adjustment •Cross-over •Imputation •Active comparator (restriction) •Instrumental variable •Proxy analysis •Sensitivity analysis 18 Schneeweiss, PDS 2006 Foot-in-Mouth Award (Economist ‘04): “… there are known knowns; there are things we know we know. We also know that there are known unknowns; that is to say we know that there are some things we do not k know. But B t there th are also l unknown k unknowns – the ones we don’t know we don’t don t know. know …, it is the latter category that tend to be the difficult ones.” (Wisely unknowing) Donald Rumsfeld 19 Case-crossover studies For studying transient drug effects on acute outcomes Time of index event 24 hours prior to index event -25 -24 0 -1 X Control time period Patient periods time p Ctrl. Case 0 1 Exposed 1 0 Non-exp. Case time period - SSRIs and risk of hip fracture: * X X 16,341 cases of hip Fx in GPRD Case-control study: RR= 6.1 1 0 0 1 1 1 0 0 Case-crossover: RR = 1.9 - Viagra and MI ORMH or ORCond. log. reg. *Hubbard et al. Am J Epi 2003 X = Case-defining event = Shaded case or control periods represent periods exposed to the study drug 20 Case-crossover studies Why is the CCS design not more frequently used in PE? Requires rapid onset outcomes Requires time-varying time varying exposure (treatment x-over) x over) Requires transient drug effects Is subject j to within-person p (between-time) ( ) confounding: g Decreasing health status correlates with increasing drug use Can be expanded to the case-time-control design 21 Case-time-control analysis Beta-2-agonist inhaler use and risk of fatal or near near-fatal fatal asthma: Cases Current B2A use Discordant use (case crossover) Discordant use (control crossover) Case time control Controls High 93 Low 36 29 9 29 9 High 241 65 65 Adjusted j Low 414 OR 4.4 25 25 33.22 2.6 1.2 OR 3.1 95%CI 1.8-5.4 11.5-6.8 5-6 8 1.6-4.1 0.5-3.0 22 Restriction 23 Example study (basic design) Cohort study of PA Medicare beneficiaries 65+ with drug insurance through the PACE program Medicare Part A, B, and D 1995 - 2002 Exposure: statin use Outcome: 1-year 1 year all-cause all cause mortality Covariates: Cardiovascular co-morbidities, comedications, other proxies for comorbidity Censoring: Death and at 365 days Analysis: CoxCox regression w/ EPS adjustment 24 0) Incident and prevalent drug users vs. non-users (matched by exact date) 1a) Incident drug users vs. non-users (matched by exact date) 1b) IIncident id t d drug users vs. non-users (matched ( t h db by d date t and d system t use)) 2) Incident drug users vs. incident comparison drug users 3) Incident drug users vs. incident comparison drug users without contraindications 4) Adherent incident drug users v. adherent incident comparison drug users without contraindications Restrict to incident drug users Match non-users on system use Restrict to incident comparison drug users Restrict to Restrict to pats w/o adherent contrapatients indications Restrict to RCT inclusion criteria RCT population 25 Summarized results of restriction Cohort Unadjusted RR Fully adjusted RR 0) Incident and prevalent statin users versus non-users 0.32 (0.30, 0.33) 0.62 (0.58, 0.66) 1a)) Incident d statin users versus non-users 0.39 (0.36, ( 0.42)) 0.65 (0.59, ( 0.72)) 1b) Incident statin users versus non-users matched on Rx or visit date 0.41 (0.38, 0.44) 0.68 (0.61, 0.75) 2) Incident statin users versus incident glaucoma medication users 0.56 (0.51, 0.62) 0.79 (0.70, 0.88) 3) PS trimmed incident statin users versus incident glaucoma medication users 0.62 (0.56, 0.69) 0.78 (0.69, 0.89) 4) Incident and adherent statin users versus incident and adherent glaucoma medication users 0 64 (0 0.64 (0.55, 55 0.74) 0 74) 0 80 (0 0.80 (0.67, 67 0.95) 0 95) 5) PROSPER Eligible 0 72 (0.57, 0.72 (0 57 0.92) 0 92) 0 79 (0.60, 0.79 (0 60 1.03) 1 03) 26 Schneeweiss et al Med Care 2007 Results* in comparison to RCTs 1.2 PROSPER M Mortality Rate Ra atio 1 (70-82, prim +sec prevention) P Pravastatin t ti 0.8 (pooling 65+, 4S LIPID, CARE) (65+, secondary prevention) 0.6 0.4 0.2 0 0 1a 1b 1c 2 3 4 5 Cohort restriction * Unadjusted mortality rate ratios 27 1st decision point: Prevalent and/or incident users? Problems with mixed prevalent/incident user cohorts Under Under-ascertainment ascertainment of events related to drug initiation In prevalent users covariates will be assessed after initial p and mayy be the consequence q of treatment exposure Prevalent users are by definition adherent to index drug (=survivor cohort) Duration of use needs to be adjusted 28 Time-varying hazards Hazard fu unction (instanta aneous ris sk) EXAMPLES: - HRT and CHD/MI - Statins and muscle pain p y p - CYP polymorphisms Time since start of exposure Reasons? Biology g in cohort composition p Change 29 Depletion of Susceptibles Past experience modifies current risk P Persons who h tend t d tto remain i on d drug are those th who h can tolerate it Susceptibles select themselves out of exposure “Susceptibles” Changes the population at risk Form of informative censoring Askling J et al., Ann Rheum Dis 2007 30 Ray AJE 2004 31 A basic cohort design in longitudinal healthcare claims data Fixed covariate assessment period Follow-up pp period Time Initiation of exposure with study and comparison drugs and start of follow-up 32 Dealing with switchers: comparing starters with starters and switchers with switchers First line use drug A First-line N d No drug use S First-line use drug B Common 1st line use drug A Switch to 2nd line drug C S Switch to 2nd line drug D Common 1st line use drug A Add drug C to A S Add drug D to A 33 Proportion of ne ew users in ation popula New drug g on the block 100% N l marketed Newly k t dd drug Calendar time Old drug Time off new drug Ti d approval Schneeweiss PDS 2010 34 Summary of restriction Powerful way to control bias “see” the amount of bias control achieved Need to start with a large source population Generalizable to routine care? 35 Back to confounding 36 The power of proxies Measured confounders (C) may serve as redundant proxies for unmeasured confounders (U): Comorbidity U Age C Trt Outcome The more proxies the better… 37 Limited clinical information in admin databases ---------- ID=********** dob=**/**/1948 sex=M eligdt=1/2000 indexdt=6/2001 ------------------- Service Site of ___________Drug or Procedure________ ________Diagnosis_____ Date Service Prov Type Code Description * Code Description ---------------------------------------------------------------------------------------------10/01/00 OFFICE Family Practice 90658 INFLUENZA VIRUS VACC/SPLIT V048 VACC FOR INFLUEN 10/01/00 Rx Pharmacy CIPROFLOXACIN 500MG TABLETS 10 11/05/00 OFFICE Family Practice 17110 DESTRUCT OF FLAT WARTS, UP 0781 VIRAL WARTS 11/07/00 R Rx Pharmac Pharmacy CIPROFLOXACIN 500MG TABLETS 10 01/15/01 Rx Pharmacy CIPROFLOXACIN 500MG TABLETS 10 06/25/01 OFFICE Emerg Clinic 99070 SPECIAL SUPPLIES * 84509 SPRAIN OF ANKLE E927 ACC OVEREXERTION 06/30/01 OFFICE Orthopedist 99204 OV,NEW PT.,DETAILED H&P,LOW * 72767 RUPT ACHILL TEND 06/30/01 OFFICE Internist/Gener 99202 OV,NEW PT.,EXPD.PROB-FOCSD * 84509 SPRAIN OF ANKLE OUTPT HP Anesthesiologis 01472 REPAIR OF RUPTURED ACHILLES * 84509 SPRAIN OF ANKLE Hospital 27650 REPAIR ACHILLES TENDON * 84509 SPRAIN OF ANKLE 85018 BLOOD COUNT; HEMOGLOBIN * 84509 SPRAIN OF ANKLE Orthopedist p 27650 REPAIR ACHILLES TENDON * 84509 SPRAIN OF ANKLE 06/30/01 OFFICE Orthopedist 29405 APPLY SHORT LEG CAST * 72767 RUPT ACHILL TEND 07/30/01 OFFICE Orthopedist 29405 APPLY SHORT LEG CAST * 72767 RUPT ACHILL TEND 08/13/01 OFFICE Orthopedist L2116 AFO TIBIAL FRACTURE RIGID * 72767 RUPT ACHILL TEND Can we make better use of this information ? 38 Multivariate adjustment 10 covariates 100 covariates 1000 covariates 39 MI Outcome (Unadjusted) Cumulative e Incide ence HR=2.11 (1.46-3.04) 111% (46%-204%) Risk Increase Statin Initiators Statin Non-Initiators Non Initiators Months of Follow-Up Seeger et al. PDS 2005 Propensity score analysis Goal: To identify patients with the identical likelihood of receiving treatment but some will actually receive treatment others will not. not Estimation: St Step 1 1: Estimate E ti t the th propensity it ffor ttreatment t t as a function of observed covariates: - mimic the p prescribers decision process p for treatment - if exposure is prevalent then little limitations to modeling - Predicted value is each patient’s “propensity score” St Step 2 2: U Use th the estimated ti t d propensity it score tto adjust dj t treatment model: - quintiles, qu t es, deciles dec es o of p propensity ope s ty sco score, e, ttrimming g - PS matching - Model adjustment 41 42 Fig. A: Propensity score matching Patients always treated with study drug Patients never treated with study drug % off subjects 0 0 05 0.5 1 Exposure propensity score = treated with study drug = treated with comparison drug 43 Fig. B: After matching Patients always treated with study drug Patients never treated with study drug % off subjects 0 0 05 0.5 1 Exposure propensity score = treated with study drug = treated with comparison drug 44 45 MI Outcome (After Propensity Score Matching) 31% (7%-48%) Risk Reduction Cumulative e Incide ence HR=0.69 (0.52-0.93) Statin Non-Initiators Statin Initiators Months of Follow-Up PS matching provides results in actionable and transparent ways Relative risk Risk difference Number needed to treat Intended treatment effect Adverse effect 1 Adverse effect 2 Subgroup 1 Subgroup 2 Benefit-risk balance: Values and utilities 3 10 Investigator-specified covariates Empirically Empiricallyspecified covariates 101 Confounding Factors New Therapeutic Health Outcome 48 1,000,000-s of data items and patterns Mining codes, t t patterns texts, tt with new and existing tools The hd-PS SAS macro. The hd-PS SAS macro can be downloaded at www.drugepi.org … links … downloads. EXAMPLE CODE th/t / /di t /hd %i l d "/ %include "/path/to/macro/directory/hdps.mcr"; " 10,000-s of potential confounding factors Prioritization according to potential for confounding 1000-s of confounders f d Statistical analysis using e.g. propensity score methods or shrinkage estimators Title1 'High-dimensional propensity score adjustment'; Title2 '(study description)'; %RunHighDimPropScore ( var patient id var_patient_id var_exposure var_outcome vars_demographic vars_force_categorical top_n k trim_mode percent_trim input_cohort input_dim1 input_dim2 input dim3 input_dim3 input_dim4 input_dim5 output_scored_cohort output_detailed results_estimates results diagnostic results_diagnostic ); = = = = = = = = = = = = = = = = = = = id, id exposed, outcome, age sex race, year, 200, 500, 500 BOTH, 5, master_file, drug_claims outpatient_diagnoses inpatient_diagnoses inpatient diagnoses inpatient_procedures outpatient_procedures scored_cohort, detailed_cohort, estimates, variable info variable_info generic_name, icd9_dx, icd9_dx, icd9 dx icd9_proc, cpt, www.drugepi.org 49 Flow chart for basic high-dimensional propensity score algorithm. 1. Specify data sources Define D fi p ddata t dimensions; di i use ddata t stream t off 180 days d up to t the th initiation i iti ti off study t d exposure. This Thi includes diagnoses on the day of initiation but no drugs or procedures on the day of initiation. Exclude selected codes from covariate adjustment Base case: p = 8 1 2 3 4 5 6 7 8 MC Part A* MC Part A MC Part A MC Part A MC Part A Part B Part B Drugs Hosp Dx Hosp proc Amb Dx Amb proc Nurse Home Dx Proc generic (ICD-9) (ICD-9) (ICD-9) (ICD-9) (ICD-9) (ICD-9) (CPT-4) entities * MC = Medicare, Hosp = hospital, Amb = ambulatory, Dx = diagnosis, proc = procedure, Rx = prescription dispensings 2. Investigator specified covariates Demographics Age, sex, race, yyear Predefined Hx, Dx, Rx, Procs Identify empirical candidate covariates Within each data dimension sort by prevalence of codes. Identify the n most prevalent codes. Base case: n= 200; Granularity = 3 digit ICD-9, 5 digit CPT, generic drug name. 3. Walker. Ch. 9, Observation and Inference 1991 Assess recurrence For each identified code create three covariates: CovX once = 1 if that code appeared at least once within 180 days CovX_once CovX_sporadic = 1 if code appeared at least twice CovX_frequent = 1 if code appeared at least three times. 4. Prioritize covariates Within each data dimension calculate for each covariate the possible amount of confounding it could adjust in a multiplicative model given a binary exposure and outcome after adjusting for demographic covariates: 1 Biasmult = PC1 ( RRCD 1) 1 if RRCD 1, PC1 ( RR 1) 1 otherwise. Sort in descending order. Bross. J Chron Dis 1966 CD PC 0 ( RRCD 1) 1 5. Select covariates PC 0 ( RR1CD 1) 1 Schneeweiss et al, Epidemiol 2009 Add d demographic covariates from step 1 and l predefined covariates in the top positions. Select top k 50 _ pp CovX_sporadic = 1 if code appeared at least twice CovX_frequent = 1 if code appeared at least three times. 4 4. y Prioriti e covariates Prioritize co ariates Within each data dimension calculate for each covariate the possible amount of confounding it could adjust in a multiplicative model given a binary exposure and outcome after adjusting for demographic covariates: 1 Biasmult = PC1 ( RRCD 1) 1 if RRCD 1, PC1 ( RR 1) 1 otherwise. Sort in descending order. CD PC 0 ( RRCD 1) 1 5. PC 0 ( RR1CD 1) 1 Select covariates Add d demographic covariates from step 1 and l predefined covariates in the top positions. Select top k empirical covariates from step 4. 4 Optional, Optional include multiplicative 2-way 2 way interactions for d demographic and l predefined covariates with the top 20 empirical covariates. Base case: d = 4 (age, sex, race, year); l = 14; k= 500 6 6. Estimate exposure propensity score Estimate propensity score using multivariate logistic regression, including all d + l + k covariates. Truncate 5% of patients on either end of PS distribution and form deciles. 7. Estimate outcome model Estimate exposure-outcome association adjusted for propensity score deciles as well as PS weighted. Schneeweiss et al., Epidemiol 2009 51 Table 1: Characteristics of 49,653 initiators of selective COX-2 inhibitors or Example: non-selective (ns) NSAIDs as defined during 6 months prior to first medication use. Coxibs vs. nsNSAIDs and risk of GI complications p in 180 days Initiators of Cox-2 selective NSAIDs N % Initiators of nsNSAIDs N % OR* 95% CI N 32,042 Age75 years or older 24,079 75% 11,496 65% 1.61 1.545-1.674 Sex, % female 27,528 86% 14,293 81% 1.42 1.348-1.487 Race: white 30,583 95% 15,808 90% 2.39 2.23-2.57 black 1,133 4% 1,580 9% 0.37 0.34-0.403 other 326 1% 223 1% 0.80 0.68-0.95 Charlson comorbidity score >= 1 24,343 76% 12,521 71% 1.29 1.233-1.340 Use of >4 distinct drugs in prior year 24,120 75% 11,852 67% 1.48 1.421-1.541 >4 physician visits in prior year 22,919 72% 11,363 65% 1.38 1.328-1.437 Hospitalized in prior year 9,804 31% 4,591 26% 1.25 1.200-1.303 Nursing home resident 2,671 8% 996 6% 1.52 1.407-1.635 Prior use of gastroprotective drugs 8,785 27% 3,600 20% 1.47 1.407-1.536 1.407 1.536 Prior use of warfarin 4,252 13% 1,153 7% 2.18 2.041-2.337 Prior use of oral steroids 2,800 9% 1,373 8% 1.13 1.059-1.211 History of OA 15,549 49% 5,898 33% 1.87 1.802-1.945 History of RA 1 602 1,602 5% 476 3% 1 90 1.90 1 707 2 102 1.707-2.102 History of peptic ulcer disease 1,189 4% 426 2% 1.55 1.389-1.739 551 2% 196 1% 1.55 1.319-1.831 23,332 76% 12,363 70% 1.14 1.092-1.184 History of congestive heart failure 9,727 30% 4,328 25% 1.34 1.283-1.395 History of coronary artery disease 5,266 16% 2,603 15% 1.13 1.078-1.193 History of gastrointestinal hemorrhage History of hypertension * OR = odds ratio; CI = confidence interval 17,611 52 Table 3: Variations in covariate adjustment and relative risk estimates for the association of selective cox-2 inhibitors Model # p within 180 days y of first medication use. and GI complications Covariates included in propensity p p y score model Number of covariates adjusted j Variables tested per data source Data source granularity g y Covariate prioritization algorithm g cstatistic of PS model Outcome model Relative risk 95% CI N = 49,653 1 Unadjusted - 1.09 0.91-1.30 2 Age, sex, race, year** d 4 d=4 0.61 1.01 0.84-1.21 3 + predefined covars (Tab1) d=4; l=14 0.66 0.94 0.78-1.12 4 + empirical covariates d=4;l=14;k=200 n=200 3-digit ICD Biasmult 0.69 0.86 0.72-1.04 5* + empirical covariates d=4;l=14;k=500 n=200 3-digit ICD Biasmult 0.71 0.88 0.73-1.06 Bootstrapped 95% CIs: 5b Only demographics + empirical i i l covariates i t d=4;; k=500 n=200 3-digit g ICD Biasmult 0.71 0.87 0.73-1.06 0.72-1.05 53 Small sample performance Example: Coxibs vs vs. nsNSAIDs and GI complications in 180 days Confounder prioritization now with 0 0-cell cell correction (+0.1) hd-PS2, SAS 9.2 or higher, substantially improved speed 20mins -> 2mins 27 56 83 110 277 events 54 55 56 Clopidogrel ‐ MI (2) C ib ‐ GI bleed Coxib bl d Statin ‐ death TCA suicide (1) 0.60 0 30 0.30 0.00 ‐0.30 only hdd‐PS + hhd‐PS ad just. + speccified covaars Age‐s age‐sex‐ ‐sex‐ race‐y yearyear ‐race adjustm adjument sted ‐0.60 Unadjus sted Unadjuusted log elative risk) log (re elative risk) Performance of different adjustment procedures, including hd hd-PS PS adjustment 57 0 30 0.30 only hdd‐PS 0.60 + hhd‐PS ad just. + speccified covaars Age‐s age‐sex‐ ‐sex‐ race‐y yearyear ‐race adjustm adjument sted Unadjus sted Unadjuusted log elative risk) log (re elative risk) … and hd-PS adjustment alone Clopidogrel ‐ MI (2) C ib ‐ GI bleed Coxib bl d Statin ‐ death TCA suicide (1) 0.00 ‐0.30 ‐0.60 58 Kitchen sink models and the risk of collider-stratification bias U U C Trt IV U Trt Outcome Outcome M-bias confounding is usually considered weak Z-bias is a bit more likely: conditioning on treatment will open a back-door path and an IV-like variable will ill become b a confounder f d Do we need variable un-selection? Greenland, Epidemiol 2003 Brookhart, Schneeweiss et al. AJE 2006 59 Summary of Propensity score analyses Models treatment choice If exposure is frequent then very large numbers of covariates can be adjusted It fits nicely with the incident user cohort design hdPS further improves confounding adjustment and p easyy and less prone p to makes implementation investigator error p analysis y since the It is a veryy transparent improvement in balance can be demonstrated Provides difference measures for fair benefit-risk assessment 60 Sensitivity analyses 61 Basic design sensitivity analyses (1) Varying wash wash-out out periods for incident user definition (2) Follow-up until discontinuation vs. fixed follow-up Time (3) Varying length of exposure risk window Initiation of exposure with study and comparison drugs and start of follow-up 62 Exposure risk window Allergic reaction? Cancer? Bacterial infections? Cancer? There is no right and no wrong. rong Yo You need to arg argue e your o r case based on event of interest, the biology, PK and PD 63 Expl: Coronary revascularization Bare metal vs. vs drug-eluting drug eluting stents Death MI “Landmark analysis” 64 Sensitivity y analysis y (array ( y approach)) 6.0 50 5.0 RRadjusted 4.0 Fixed: ARR = 2.0 PC0 = 0.5 3.0 20 2.0 4.5 1.0 25 2.5 0.8 0.9 1.0 0.6 0.7 0.4 0.3 0 0.2 0 0.1 0.8 0.5 0.0 0. 0 RRCD PC1 65 Schneeweiss PDS 2006 Sensitivity y analysis y (rule ( out approach)) * Plotted is the strength of the associations between an unmeasured confounder and treatment choice (OREC) and the association between an unmeasured confounder and disease outcome (RRCD) that are required to fully explain the observed association (ARR = 3.38) or its lower 95% confidence limit (ARR = 1 88) Any factor (a single factor or combination of multiple factors) that has a combination of RRCD and 1.88). OREC values resulting in points higher than and to the right of the plotted lines will be able to fully explain our observed results. Conclusions on methods Case-crossover analyses are valuable but there are only few applications in CER Propensity score analyses are very valuable in claims data analyses for a wide range of CER Attention to basic epidemiology principles is paramount Sensitivity analyses will help you put results in perspective 67 Th k you very much Thank h 68 Recommended reading Overview: Schneeweiss S. Developments in Comparative effectiveness research. Clin Pharm & Th 2007. Ther 2007 Rubin DB. Estimating causal effects from large data sets using propensity scores. Ann Intern Med 1997 Propensity scores: Seeger JD et al. Analytic strategies to adjust confounding bias using exposure propensity scores and disease risk scores. PDS 2005 D’Agostino. Propensity score methods for bias reduction in the comparison of a treatment to a non non-randomized randomized control group. group Stat Med 1998 Sturmer T et al. Analytic strategies to adjust confounding bias using exposure propensity scores and disease risk scores. AJE 2005 Instrumental variables Brookhart MA et al. Evaluating short-term drug effects in claims databases using prescribing preferences as an instrumental variable. Epidemiology 2006 Brooks J et al. Heterogeneity and the interpretation of treatment effect estimates from risk adjustment and instrumental variable method. method Med Care 2007 Schneeweiss et al. Simultaneous assessment of short-term gastrointestinal benefits and cardiovascular risks of selective COX-2 inhibitors and non-selective NSAIDs: an instrumental variable analysis Arth & Rheum 2006 69