Data source
The Corrona registry is an independent, prospective, observational cohort of patients with RA, who were recruited at >160 private and academic practice sites across 40 states in the USA; additional details have been published previously [17]. Data on approximately 39,950 patients with RA have been collected as of 31 March 2014. The Corrona database includes information about 285,726 patient visits and approximately 119,298 patient-years of follow up observation time, with a mean time of patient follow up of 3.6 years (median 2.8 years). For this national study, approvals for data collection and analyses were obtained from a central institutional review board (New England Institutional Review Board) for private practice sites participating within Corrona. For the <20 % of sites that are affiliated with an academic medical center, the local institutional review board was the Institutional Review Board of record.
Study population
Data were collected from patients with RA from the Corrona registry who initiated rituximab or a subsequent anti-TNF agent (adalimumab, etanercept, golimumab, infliximab, certolizumab) on or after 28 February 2006. The study population was limited to patients who had received ≥1 anti-TNF agent and had not previously received rituximab. Patients must have had the following data available to be included in the study: date of the first rituximab infusion or initiation of a subsequent anti-TNF agent; follow-up visit at 1 year (±3 months); ≥1 visit between baseline and 1-year follow up; and Clinical Disease Activity Index (CDAI) measurements at baseline and 1-year follow up. For patients whose anti-TNF initiation occurred between visits, a prior visit (within 4 months of initiation) was used. Patients with CDAI low disease activity (LDA) or remission at initiation, or with a diagnosis of lymphoma prior to initiation, were excluded from the study (Fig. 1). All patients provided written informed consent prior to participation.
Measures and data collection
Data from Corrona were collected during the study period (28 February 2006 to 31 October 2012) from physician and patient questionnaires completed during routine clinical encounters. Data on use of nbDMARDs and biologic DMARDs, 28-joint tender and swollen joint counts, physician and patient global assessments of disease activity, patient assessment of pain and modified Health Assessment Questionnaire (mHAQ) scores assessing physical function were recorded at the time of the clinical encounter [18]. Data on demographics, insurance status, comorbid conditions, RA disease characteristics and RA medications were available for ≥99 % of patients.
Drug exposure cohorts
To balance for predisposing factors that may increase a patient’s likelihood of receiving either rituximab or an anti-TNF agent, a propensity score - or the probability of treatment selection - was calculated for each eligible patient using baseline (at the time of drug initiation) patient demographics (age, sex, race and insurance type), disease characteristics (rheumatoid factor (RF) seropositivity, duration of RA, American Rheumatism Association functional class, tender and swollen joint counts, patient and provider global assessments, patient pain and functional status), comorbidities (cardiovascular disease, cancer and/or diabetes mellitus), past medication history (number of prior nbDMARDs, anti-TNF agents and/or non–anti-TNF agents) and concurrent medications (prednisone and/or MTX). The rationale for the methodology for this approach is provided in the supplementary materials (see Additional file 1). For the first cohort (trimmed population), patients who fell outside the overlap of the propensity score distributions were excluded (see Additional file 2: Figure S1). The second cohort (stratified-matched population) included rituximab-treated and anti-TNF-agent–treated patients who were stratified by prior treatment with one versus two or more anti-TNF agents, and then matched within each stratum based on propensity score estimated within each strata without replacement, using calipers of 0.01. The resulting stratified-matched population resulted in greater similarity between the two drug exposure groups.
Study outcomes
The primary outcome was the proportion of patients in each group who achieved CDAI LDA or remission (CDAI score ≤10) at 1 year [18]. Secondary outcomes included the proportion of patients who achieved modified American College of Rheumatology (mACR) 20/50/70 responses, which omit the acute-phase reactant laboratory components, and the proportion who achieved a clinically meaningful improvement in functional status, defined as a decrease of ≥0.25 from baseline in the mHAQ score, at 1 year [19–21].
Safety events reported by providers over the 12-month study were examined and included infections (all infections and serious infections), cardiovascular events and new malignancies. Infections identified in this analysis included cellulitis, sinusitis, diverticulitis, sepsis, pneumonia, bronchitis, gastroenteritis, meningitis, encephalitis, urinary tract infection, upper respiratory tract infection, tuberculosis, joint infection, bursal infection and all other hospitalized and ambulatory infections. Cardiovascular events included cardiac arrest, congestive heart failure, myocardial infarction, coronary artery disease, unstable angina, ventricular arrhythmia, cardiac revascularization stroke, transient ischemic attack and deep vein thrombosis. Cancer events included breast cancer, lung cancer, lymphoma, skin cancer (not specified, squamous cell and melanoma) and other cancer diagnoses.
Analysis and statistical methods
Patients were included regardless of retreatment with rituximab or persistence with anti-TNF therapy. Baseline patient demographics and clinical and disease characteristics were compared between the two drug-exposure cohorts, and standardized differences were estimated. Response was defined as achievement of primary and secondary outcomes at 1 year regardless of continuation of initial treatment. Nonresponse was imputed for patients who switched biologic agents. Descriptive statistics were used to examine rates of response at 1 year overall and by treatment pattern subgroup: (1) patients who remained on the drug, (2) those who were not retreated with rituximab or who discontinued anti-TNF therapy and did not initiate another biologic agent, and (3) those who switched to another biologic agent.
Multivariable logistic regression models were fit to estimate odds ratios (ORs) and 95 % CIs comparing response rates in rituximab users to anti-TNF agent users in the two populations. Covariates used in the multivariable logistic regression models of the trimmed population included baseline parameters with a standardized difference of >0.1 and four factors chosen a priori to ensure no residual confounding despite the propensity score methodology: baseline CDAI score, steroid use (current or not), number of anti-TNF agents previously used (1 vs ≥2) and concomitant MTX use. The resulting multivariable logistic regression models were adjusted for fixed and random effects. Both patient and provider random effects were examined; however, only patient-related random effects were included in the model because provider random effects did not have a significant impact on responses. In the stratified-matched population, all baseline characteristics had standardized differences <0.1 except for baseline CDAI score (standardized difference 0.14). Therefore, logistic regression models were fit to estimate ORs and 95 % CIs comparing response with rituximab to that with anti-TNF agents, including baseline CDAI score as a covariate in the model and random effect for matched pairs (i.e., patients clustered within matched pairs).
Of the patients with available data on RF status, RF seropositivity was reported in 82 % (141 of 173) and 72 % (315 of 435) of patients receiving rituximab and anti-TNF agents, respectively. Of note, a missing indicator was used when RF was included as a variable in the multivariate model as well as the propensity score model. Inclusion of RF status as a covariate resulted in <5 % variation in OR estimates. Due to the impact on power of limiting the study sample to only those for whom serologic data were available, the ORs reported do not include this covariate. Additionally, no significant differences were observed in prednisone and nbDMARD use between users of rituximab or anti-TNF agents overall, or between subgroups in the stratified-matched population. Because no significant differences were found, we chose not to include these factors in the trimmed population (because any difference would likely be controlled for in the multivariable model) and stratified-matched analyses.
Safety event rates were calculated based on the number of events reported by providers, divided by the duration of exposure. In the trimmed population, sex- and age-standardized adverse event (AE) rates among rituximab users were calculated based on the age and sex distribution in the users of anti-TNF agents. Safety events were also compared in the stratified-matched population. The ratio of rates for the rituximab users in relation to the users of anti-TNF agents was generated for both the trimmed and stratified-matched populations.