4G-04 - Temporal High Dimensional Propensity Score (thdPS): A Novel Approach to Improve Confounding Control by Incorporating Temporality to High Dimensional variables

Background: Pharmacoepidemiologic studies often control for baseline features (yes vs. no). High-dimensional propensity score (hdPS) uses intensity of codes for confounding control, but has not accounted for temporality of codes.

Objectives: Develop a novel algorithm, thdPS, to incorporate temporality of codes for confounding control.

Methods: We identified HD variables by ICD-10 (3 digits), CPT (5 digits), and ATC (4th level) codes for top 200 prevalent claim codes in 5 dimensions (in/outpatient diagnoses & procedure, prescription) in 1-year baseline period, then create 6 binary variables for each code: once, sporadic (<= median number of times), and frequent (<=75th percentile number of times) accounting for intensity of HD (iHD) variables; distal (only from 12 to 9 months before T0 [-12m, -9m]), proximal (in [-3m, 0m]), and intermediate (<=1 code in [-9m,-3m], but not in [-3m,0m]) accounting for temporality of HD (tHD) variables. In a random 20% sample of 2019 US Medicare beneficiaries aged 65+, we emulated trials before influenza (FLU) season (early starting date is 9/1) by randomly matching vaccinees on each day with non-vaccinees with an outpatient visit on the same day. Follow-up starts on 14 days post exposure date, ends on a non-vaccinees vaccinated, the start of flu season (11/27/2019), or death. PS, the probability of receiving vaccine, is computed by 3 sets of variables along with age, sex, and race: 1) 79 predefined variables including comorbidities, healthcare, and comedications; 2) top 500 iHD variables, 3) top 500 tHD variables. To test impact of iHD and tHD variables, we did not add any predefined variables besides age, sex, and race, both iHD and tHD variables were selected by Bross formula. Crude hazard ratio (HR) and inverse probability treatment weight (truncated at 10) adjusted HR for pre-FLU season mortality were assessed, expecting a null association. In sensitivity analyses, we categorized ATC codes as stopper ([-12m, 6m]), initiator ([-6m, 0m]), chronic use (entire baseline period).

Results: A total of 262 codes are identified by both iHD and tHD variables, iHD set has 5 unique codes and tHD set has 29 unique codes including 7 ATC (eg, macrolides systematic use), 7 CPT, and 15 ICD codes. Crude risk, rate, and HR for mortality in 1743380 vaccinees vs 1247387 non-vaccinees are 0.33 vs 0.32%, 3.4 vs 4.3/100 person-year, and 0.76 (0.72-0.80). The adjusted HR from predefined, iHD, and tHD variables are 1.05 (1.00-1.10), 1.01 (0.95-1.07), and 1.00 (0.94-1.07), respectively. Sensitivity analyses confirmed consistency in findings.

Conclusions: The thdPS produced an aHR comparable to hdPS. Incorporating temporality to HD variables may capture meaningful information. Further work is required to find patterns of codes and determine the most appropriate criteria to use.