POS0379 WHEN SHOULD WE USE MARGINAL STRUCTURAL MODELS FOR DRUG STUDIES IN RHEUMATOLOGY? THE EXAMPLE OF CORTICOSTEROIDS AND INFECTIONS IN RHEUMATOID ARTHRITIS

Background The use and dosing of pharmacotherapy for chronic inflammatory disease is commonly changed over time, adapting to previous treatment response, current activity, tolerability, onset adverse events. This makes evaluating the treatments difficult, as received time becomes entangled with time-varying confounding factors, even if initial choice was randomized. To naïvely adjust these factors in regression models have long been recognized by epidemiologists introduce bias most realistic settings, marginal structural (MSM) has proposed a less biased alternative. Despite this, MSMs are used very sparingly drug studies rheumatology. Objectives objective this work explain rationale why recommended context what may go wrong they not used. method will be demonstrated comparing incidence serious infections between different oral glucocorticoid (GC) dose patterns first three years after RA diagnosis, adjusting confounders MSM or through conventional models. Methods For demonstration study, we followed 9654 recently diagnosed patients from their rheumatology visit Swedish register (SRQ), 2007 2018. Follow-up divided into 3-month periods. An average prednisone daily-dose calculated each period categorized “no use”, “low” (≤10mg/day) “high” (>10mg/day) doses. implemented sequential inverse probability weighting across periods, accounting demographic DAS28 HAQ, co-medication history comorbidity. (hospitalization infection identified National Patient Register) modelled pooled logistic allowing separate effects recent past GC. We compared adjustment via versus MSM. Results increased associated higher lower doses more Although theory results were similar those regression. Compared no glucocorticoids, exposure low during year added 1.9 cases per 100 (95%CI 0.9-2.9) at years, while high 4.1 (2.5-5.9) cases. Conclusion Our broadly agree observational showing dependent risk (recent) glucocorticoids. theoretically potential than ordinary models, two almost indistinguishable example, implementation led limitations study design (including categorization exposure), consuming both programming computation widely option. similarity guaranteed however, biases that designed avoid should kept mind all drugs change response possibly factors. REFERENCES: NIL. Acknowledgements: Disclosure Interests Andrei Barbulescu Consultant of: Employee Parexel, Arvid Sjölander: None declared, Bénédicte Delcoigne: Johan Askling Grant/research support from: JA PI ARTIS project, which supported agreements Karolinska Institutet listed entities, mainly national safety monitoring immunomodulators Sweden (ARTIS): Abbvie, BMS, Eli Lilly, Galapagos, MSD, Pfizer, Roche, Samsung Bioepis, Sanofi., Thomas Frisell: declared.