×

CPRD StudyPrimary Care PASS Study

Abstract 868: Multi­Level Modelling (MLM) in Specialist Cohort Event Monitoring (SCEM) Studies

Abstract 868: Multi­Level Modelling (MLM) in Specialist Cohort Event Monitoring (SCEM) Studies

Deborah Layton, Debabrata Roy,Sarah E. Marley, Saad A. W Shakir

Background

Prescribing guidelines (PG) influence treatment choice based on patient (pt) factors. Clinical use is also influenced by non-pt factors. MLM can provide insight into sources of variability in healthcare, where hierarchical structures exist. A SCEM study is investigating the safety and use of rivaroxaban (R) in clinical use, with a warfarin (W) cohort for context.

Objectives

Ad-hoc MLM to explore influences on prescribing anticoagulants.

Methods

Data on 53 NHS acute trusts in England/Wales (e.g population size, PG) were linked to interim SCEM pt demographic and drug utilization data, and prescriber details (e.g degree, specialism). MLM was applied to 1006 (55%) R vs. 816 W (45%) adult pts nested in 514 prescribers, nested in trusts. The binary outcome was R or W treatment. Variance components were expressed as median ORs (median relative increase in odds of R treatment if pt changed prescriber/trust) and proportional change in variance (PCV) between models when successively adding fixed effects (pt/prescriber/trust). If treatment choice was dominated by pt factors, having accounted for their effects, variance between prescribers and trusts would be comparatively low.

Results

Differences between trusts and prescribers in trusts are important in treatment choice; trust being more influential (MORTRUST(T)=6.9; MORPRESCRIBER(P)=3.9). Some pt factors had a relatively large effect on odds of treatment choice, e.g cerebrovascular accident [OR 2.0 (95%CI 1.3,3.0)]. Adjusting for pt factors, MORT=7.8 (PCV=20%); MORP=4.5 (PCV=14%). Adjusting for prescriber factors, MORT=7.1 (PCV=-8%); MORP=3.8 (PCV=-19%). Adjusting for trust factors did not improve the model performance.

Conclusion

In this exploratory analysis, treatment variability appears dominated by differences between trusts and prescribers; most notably trusts. Some pt factors were important in treatment choice, but PCV between models suggest that accounting for pt differences does not fully explain the variance between prescribers and between trusts. This study highlights the utility of MLM in exploring non-pt factors. This interim analysis will be superseded on completion of the SCEM study.