×

CPRD StudyPrimary Care PASS Study

Multi-level modelling to investigate factors impacting prescribing variability.

Multi-level modelling to investigate factors impacting prescribing variability.

Roy D, Wise L and Shakir S

Background

Prescribing guidelines influence treatment choice based on patient (pt) and healthcare system factors. Multi-level modelling (MLM) can provide insight into sources of variability in healthcare, especially where nested hierarchical structures exist. A Specialist Cohort Event Monitoring study investigated the safety and use of rivaroxaban in clinical use, with a warfarin cohort for context.

Objectives

Study to investigate prescribing variability using MLM.

Methods

Data on NHS acute trusts in England/Wales (e.g. population size, trust type) were linked to study pt demographic and drug utilization data, and prescriber details (e.g. degree, specialty).

Using MLM we explored the influence of pt, prescriber and trust characteristics on prescribing variability in 2106 rivaroxaban (59%) vs. 1468 warfarin (41%) adult pt nested in 780 prescribers, nested in 73 trusts. The majority of pt had an indication of DVT/PE (56.4%) or non-valvular AF (AF) (41.2%). The binary outcome was rivaroxaban or warfarin treatment.

Variance components estimate the variability accounted for by each level in the model and are expressed as: Median Odd Ratios (MOR – median relative increase in odds of rivaroxaban treatment if pt changed prescriber (PR) or trust (T)); Proportional Change in Variance (PCV) between models when successively adding fixed effects.

Results

DVT/PE group:

Adjusting for pt factors, MORT=6.9 (PCV=-0.6%); MORPR=2.8 (PCV=3.8%).

Adjusting for pt and prescriber factors, MORT=6.8 (PCV=-2.9%); MORPR=2.6 (PCV=-15.4%).

Adjusting for pt, prescriber and trust factors, MORT=4.9 (PCV=-30.2%); MORPR=2.6 (PCV=3.4%).

Differences between trusts and prescribers (in trusts) are important in treatment choice; trust being more influential. Some pt factors had a relatively large effect on odds of treatment choice although the absolute number of pt impacted was often small. Trust type, was shown to be associated with the odds of treatment choice [Final model: foundation vs acute trusts OR 4.0 (95%CI 1.5, 9.9)].

Data on AF and all indications (combined) will be included in the final presentation.

Conclusion

This study highlights the utility of MLM in exploring pt and non-pt factors in nested hierarchical healthcare settings. Prescribing variability appears dominated by differences between trusts and prescribers (in 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 (in trusts) and between trusts.