(077) Large scale propensity score matching with time-stratification in older women with or without fractures in the UK

Background: Comparable groups need to be defined to estimate the resource use attributable to osteoporotic fractures.

Objectives: To evaluate the performance of data-driven large scale propensity score (PS) matching in patients with fractures and controls.

Methods: We included women aged ≥50 years with ≥2 years of prior history registered in a primary care database in the UK (CPRD Aurum). We included 3 cohorts: target (T) cohort - women with an index fracture and ≥1 previous fracture in the 2 years prior, making the index fracture the subsequent fracture; comparator cohort 1 (C1) - women with an index fracture and no history of fracture in the 2 years prior; comparator cohort 2 (C2) - women with no history of fracture at any time prior. Women could contribute to ≥1 cohort. The patient selection window (01 April 2010 to 31 March 2018) was divided into 6-month periods to account for seasonality of fractures. The index date for C2 was randomly assigned in the 6-month period. Large scale L1 regularised (LASSO) regression was used to identify confounders and covariates predictive of fractures from a large set of candidate covariates. Selected covariates were included in a logistic regression to compute the PS separately for comparison groups T-C1 and C1-C2. Age and PS matching (1:5) was performed for the first 6-month period, and this process was repeated on an iterative rolling basis over calendar time. We evaluated balance using absolute standardised mean difference (ASMD).

Results: We included 11,836 (T), 56,237 (C1), and 432,677 (C2) women. There were 7,479 candidate covariates. Before matching, there were more patient characteristics with ASMD >0.1 for C1-C2 group than T-C1. After LASSO regression, the median number of selected covariates varied across periods and was higher for T-C1 (53, range 22-101) as compared to C1-C2 (22, range 8-48). The matched cohorts included T (n = 10,790) - C1 (39,827) and C1 (55,767) - C2 (157,692). All selected covariates in both groups had ASMD < 0.1 after matching, and the distribution of patient characteristics was similar for both comparison groups. Although a smaller number of covariates was selected in the C1-C2 group, known confounders such as cardiovascular disease had ASMD < 0.1 although unspecified in the PS.

Conclusions: Data-driven PS matching selected different number of covariates across comparison groups and calendar time periods. Covariate balance was achieved after large number of proxy covariates was selected when comparing women with subsequent fractures to those with single fracture; after limited number of key covariates with clinical importance was selected when comparing women with fracture to those without. Further study is required to evaluate the impact on effect estimates when potential instrumental variables are omitted from the PS.