OBSERVATIONAL Large-scale regularized regression for identifying appropriate treatment comparisons for comparative effectiveness research Patrick Ryan, Alec Walker, Paul Stang, Martijn Schuemie, David Madigan, Marc Suchard, Marc Overhage, Jesse Berlin August 26, 2013
Disclosures I am employee and shareholder of Janssen Research and Development Research presented is part of the Observational Medical Outcomes Partnership (OMOP) is a public-private partnership, which recieved funding support from multiple pharmaceutical companies I am a principal investigator for OMOP, but do not recieve any funding for my contributions
http://www.youtube.com/watch?v=jdgua7-urfw
Walker et al. CER 2013 Key idea: fit propensity score model as a diagnostic to determine if two treatments are meaningfully comparable Propensity score is probability of treatment assignment, conditional on baseline covariates Probabilities are dependent on the relative proportion of treatments Preference score is probability of choosing one treatment over another if two options were equally available
Walker et al. CER 2013 Evaluate equipoise by normalizing propensity to preference score and assess proportion of treated in each group Preference score near 0.5 means near indifference for choosing between two treatments Clinical equipoise criteria: >50% of patients in BOTH cohorts have preference score between 0.3 and 0.7
Building on Walker et al s proposal Objective: To design and implement a standardized approach to determine whether treatments in a healthcare database are sufficiently comparable to be candidates for CER. Methods: We implemented an automated heuristic that compares two treatments to determine whether they appear to be used similarly We fit a propensity score model by applying Lasso logistic regression with 35,249 baseline covariates representing patient demographics, comorbidities, concomitant medications, and health service utilization behaviors. We evaluated the comparability of two treatments by assessing the proportion of patients in each cohort near clinical equipoise based on the propensity score distribution. We applied the heuristic within a claims database to 138 drugs across 8 indications.
LASSO logistic regression ref: Genkin Technometrics 2007 7
Using all medical history in propensity score model All comorbidities Indication Target Drug All drugs Comparator drug All procedures >= 6mo washout period Index date In total, 35,249 covariates were constructed for potential inclusion in the propensity score model: Demographics (age, gender) Conditions (binary variable for all SNOMED concepts and MedDRA PT/HLT/HLGT groupings) Drugs (binary variable for all RxNorm ingredients and ATC3/5 class groupings) Procedures (binary variable for each procedure code) Number of drugs in last 6mo, number of concomitant drugs at index Number of drugs used within each indication Number of inpatient/outpatient visits in last 1mo/6mo Number of service days with indication diagnosis recorded Time-to-first diagnosis of indication Proposed solution: Apply the same comprehensive model consistently to all drug-drug comparisons across all indications 8
Equipoise plots among drugs indicated for Atrial fibrillation Flecainide and propafenone provide quintessential example of clinical equipoise Model coefficients can help explore what variables likely causing lack of equipoise: Pulmonary thrombotic/embolic conditions Beta-blocker use Deep venous thrombosis Heart valve stenosis Prostethic knee arthroplasty Device-related complications Drugs used to treat different aspects of disease (anticoagulants vs. anti-arrythmics) may provide useful unexpected comparisons Dabigatran and warfarin do not achieve equipoise but dabigatran does achieve equipoise with propafenone, solalol, and dronedarone
Equipoise plots among drugs indicated for diabetes Glipizide and glyburide show strong equipose Liraglutide (GLP1) does not show equipoise with pioglitazone or sulfonylureas, but more comparable with saxagliptin (DPP4) and glinides
Equipoise plots among drugs indicated for rheumatoid arthritis TNF inihibitors show equipose with each other but less comparability with NSAIDs and other DMARDs
12 OBSERVATIONAL Comparisons of alternative treatment across different indications Indication Drugs studied Pairwise comparisons conducted Comparisons achieving equipoise % of comparisons achieving equipoise Atrial fibrillation 8 27 10 37% Diabetes mellitus 11 34 24 71% Essential hypertension 30 341 102 30% Major depression 13 78 52 67% Migraine 9 36 22 61% Pain 29 374 54 14% Psoriasis 18 153 72 47% Rheumatoid arthritis 20 190 91 48%
Conclusions Large-scale regularized regression can be used to evaluate the appropriateness of studying alternative treatments through systematic estimation of propensity scores and assessment of clinical equipoise By applying an objective heuristic consistently across all treatments, this approach can provide empirical evidence to support the selection and prioritization of CER questions, and identify potential threats to validity which could be considered when designing the study Proposed next step: Develop open-source tool for comparator selection and produce public repository of results so that the community can share their findings across multiple data sources and users can explore comparisons of interest Learn more at OMOP Symposium, Nov 5-6, 2013 http://omop.org 13