We used data from the French National Health Data System (SNDS); the SNDS covers almost the totality (> 99%) of the French population—68 million residents. Each person is identified by a unique and anonymous number. The SNDS records comprehensive outpatient (procedures and pharmacy deliveries of reimbursed drugs) and inpatient (pharmacy deliveries of expensive drugs, procedures performed during hospital stays, and discharge diagnoses coded according to the International Statistical Classification of Diseases and Related Health Problems, Tenth Revision [ICD-10]) reimbursement information since 2006. The SNDS also contains sociodemographic information on sex, age, place of residence, and vital status, among others. Patients’ status for 100% reimbursement of care related to a severe and costly long-term disease (LTD) is recorded and LTD diagnosis is coded according to the ICD-10. The SNDS has been extensively used to conduct pharmacoepidemiological studies, especially on the use, safety, and effectiveness of health products [20,21,22,23].

We included every patient who had at least one delivery of rituximab (Anatomical Therapeutic Classification [ATC] code L01FA01, brand names and additional information are presented in Online Resource Table 1) from 1 October 2017 (date of the commercialization of the first biosimilar of rituximab) to 31 December 2021. The term ‘originator’ was chosen to designate the reference medicine product, as opposed to the biosimilar product.

A first cohort of initiators was constituted, gathering originator or biosimilar rituximab new users (i.e. having no delivery of rituximab in the past 2 years before the considered delivery—the 2-year threshold was chosen to account for the long half-life of rituximab). The index date was set to the first rituximab delivery date. Initiations occurring with simultaneous deliveries of both biosimilar product and originator product were excluded. There were two groups of exposure—originator and biosimilar. The reference group was the originator group. Exposure was either studied at first injection, or over time.

A second cohort was created, gathering patients who had a least one delivery of originator rituximab within a 2-year time span before the study start date and still under treatment or during the study period, and who switched from originator rituximab to any biosimilar product of rituximab during the study period. The index date was set to the first biosimilar delivery date. Switches occurring with simultaneous deliveries of both biosimilar product and originator product were excluded. To take into account the history of originator deliveries and patient characteristics, and to build an appropriate comparator group for the switching patients, switchers were matched to non-switchers, with up to two patients still treated with the originator product on age (in categories), sex, number of deliveries of rituximab in the past 2 years (1–3 or ≥ 4) and pathology (hematology, inflammatory/immune dysfunction disorders, multiple pathologies, undetermined). Switchers that could be matched with only one non-switcher were kept, but non-matched switchers were excluded. The index date for the non-switcher was set on the date of the rituximab delivery that matched the switcher index date.

Schemes of inclusion and follow-up of both cohorts are schematized in Online Resource Fig. 1.

Sex and age (converted into 10- to 20-year categories) were collected at the index date. Patients without those characteristics were excluded. Corticosteroid delivery 30 days before rituximab injection was also described (excluding the 40 mg of prednisolone systematically administered just before rituximab injection).

Patients’ pathologies were defined at index date, in a two-step process. We first identified pathologies for which rituximab is indicated in France (i.e. non-Hodgkin lymphoma, chronic lymphoid leukemia, rheumatoid arthritis, Wegener granulomatosis, microscopic polyangiitis, and pemphigus vulgaris) thanks to ICD-10 codes for LTD status or any discharge diagnosis linked to a hospitalization within the 5 years preceding the index date. We also identified off-label indications of rituximab (including, among others, lupus [24], multiple sclerosis [25], and kidney disease [26]) with corresponding ICD-10 codes. ICD-10 codes used for these steps are presented in Online Resource Table 2. Finally, we categorized patients’ pathologies as hematology or inflammatory/immune dysfunction disorders according to the first list, and if no pathology was identified, we used the second list. The retained pathology in case of overlap between pathologies was the most recent (based on the start date of the LTD or the start date of hospitalization). A category ‘Multiple pathologies’ was set for non-distinguishable pathologies, and a category ‘Undetermined’ included patients with no diagnosis code in their medical background. After population size estimation, we grouped together the categories ‘Multiple pathologies’ and ‘Undetermined’ under the term ‘Undetermined’ in the analyses as these subgroups were very small.

The outcome was hypersensitivity reaction, which was defined by ICD-10 codes for anaphylactic shock (ICD-10 codes T78 and T886) or serum sickness (T805/T806). These ICD-10 codes have already been used in other studies [27,28,29,30]. The ICD-10 code for hypersensitivity reaction was retained if the rituximab injection had occurred during the same hospital stay. For the serum sickness code only, the hospital stay could begin no later than 15 days after the rituximab injection.

Duration of hospitalization, transfer to the intensive care unit, and death during hospitalization were documented for patients who experienced an outcome.

History of the outcome was defined as having had at least one hospitalization for an anaphylactic shock or serum sickness coded within the past year before rituximab injection.

First, the initiation cohort was analyzed. Multivariable logistic regression adjusted for age, sex, year of inclusion, pathology, corticosteroid delivery, and history of the outcome was used to estimate the adjusted odds ratios (ORs) and 95% confidence intervals (CIs) of biosimilar versus originator rituximab injection associated with the outcome at initiation. To assess the risk of hypersensitivity reaction associated with a rituximab injection over time, we used a multivariable Cox proportional hazards regression with time-dependent exposure and covariates, estimating adjusted hazard ratios (HRs) and 95% CIs of biosimilar versus originator rituximab injection. The model was adjusted for age, sex, year of inclusion and pathology as fixed covariates, and corticosteroid delivery, history of outcome and delivery number (taken as a linear continuous variable) as time-dependent covariates. As a post hoc sensitivity analysis, we changed the definition of initiation, taking a 10-year period without rituximab injection to be treatment-naïve, and repeated the analyses.

Second, we compared matched switchers and non-switchers at index date using a multivariable logistic regression adjusted for age, sex, pathology, number of past rituximab injections, corticosteroid delivery, and history of outcome, giving estimates of adjusted OR and 95% CI of switch versus non-switch associated with the outcome at first injection.

Multivariable Cox proportional hazards regression with time-dependent exposure and covariates was used to estimate the adjusted HR and 95% CI of biosimilar versus originator rituximab injection associated with the outcome during the study follow-up among switchers who remained on biosimilar (they were censored if they switched back to originator) treatment versus non-switchers who remained on originator treatment (they were censored if they switched to biosimilar). The model was adjusted for age, sex, number of past rituximab injections, and pathology as fixed covariates, and corticosteroid delivery, history of outcome, and delivery number as time-varying covariates. As a sensitivity analysis, conditional models were performed, to take into account the matching process.

Kaplan–Meier survival curves without the event were computed for each cohort. The linearity of continuous variables was checked comparing Akaike Information Criterion (AIC) of fractional polynomials of degree 2 with the simple linear model. The risk period modeled in the time-dependent Cox regression was fixed to the 15 days after each rituximab injection, to take into account the time window used for event identification. Patients were censored at switch (originator to biosimilar, or biosimilar to originator), death, 1 year after the last rituximab delivery, or on 31 December 2021, whichever came first. Follow-up stopped at first onset of the outcome. The proportionality of hazards was verified for the main exposure variable with Schoenfeld residuals. Figure 2a, b presented in the Online Resource show some examples of inclusions and follow-ups in the initiation cohort and in the switchers/non-switchers cohort.

All extractions from the SNDS were carried out using SAS Enterprise Guide software version 9.4 (SAS Institute, Inc., Cary, NC, USA); analyses were performed with R version 3.5.2 [31], using multiple packages, including dplyr [32], ggplot2 [33] and survival [34].