Skip to main content

Research Repository

Advanced Search

Application of propensity scores and marginal structural models evaluating the effect of allopurinol in gout using primary care medical records

Rathod-Mistry, Trishna

Application of propensity scores and marginal structural models evaluating the effect of allopurinol in gout using primary care medical records Thumbnail


Authors

Trishna Rathod-Mistry



Contributors

Milica Bucknall
Supervisor

Abstract

Background
Primary care electronic health records (EHR) capture real life patterns of healthcare utilisation over time. This provides the opportunity to estimate the effect of allopurinol on long term outcomes in people with gout. However, use of such data gives rise to confounding by indication which may change over time, a major impediment in treatment effect estimation.

Methods
A cohort of patients consulting for gout between 1997 2002 and not previously prescribed urate-lowering drugs were identified from the Clinical Practice Research Datalink GOLD and were followed up until the end of 2014. Effect of allopurinol vs. non-use was evaluated on reaching target serum urate (SU) level =360µmol/L, mortality, healthcare utilisation, vascular and renal diseases.
Three statistical approaches with differing complexities and assumptions imposed were considered: (1) baseline measurement of allopurinol and covariates with confounding controlled for using propensity score (PS) subclassification; (2) extending the methods in (1) to repeated measures where allopurinol and covariates were measured yearly; (3) using marginal structural models (MSM) within the repeated measures set-up. Survival models estimated hazard ratios with 95% confidence intervals. Robustness of estimated treatment effects to unmeasured confounding was evaluated.

Results
16,876 patients were eligible for analysis (mean age (standard deviation) 62 (14.1) years, 77% male). Baseline analysis found allopurinol was associated with higher chance of reaching target SU level (2.32 (1.97, 2.74)) and fewer gout consultations (0.70 (0.65, 0.75)), and with increased risk of mortality (1.10 (1.03, 1.17)), gout hospitalisation (1.82 (1.64, 2.02)), coronary heart disease (1.11 (1.02, 1.21)), and renal disease (1.19 (1.10, 1.28)).
In the repeated measures setting, issues with poor performance of PS estimation were identified in both time-varying PS subclassification and MSM. These were resolved by allowing associations between covariates and initiation and continuation of allopurinol to differ in MSM; larger treatment effect estimates were obtained for most outcomes compared with baseline analysis and statistical significance was lost for mortality. The treatment effect estimates for target SU level and gout hospitalisation were likely to be robust to unmeasured confounding however, unmeasured confounding may explain away the treatment effects for coronary heart disease and renal disease.

Conclusion
Fitting complex models to EHR is challenging and consideration needs to be given to both clinical and statistical assumptions made during data preparation and analysis. Associations of allopurinol with adverse outcomes persisted, regardless of statistical approach used. This may be due to remaining residual confounding and/or because allopurinol dosage and adherence is suboptimal in primary care. Nevertheless, the treatment effect estimates obtain are relevant to UK primary care and provide evidence that managing gout in the long term needs to be improved.

Thesis Type Thesis
Publicly Available Date Mar 28, 2024
Award Date 2021-06

Files




Downloadable Citations