Safety of Biologic and Targeted Synthetic Disease-Modifying Antirheumatic Drugs in Rheumatoid Arthritis: A Longitudinal Analysis

Abstract

Background

Biologic and targeted synthetic disease-modifying antirheumatic drugs effectively manage rheumatoid arthritis but are associated with varying risks of adverse events.

Objective

Comparing the risk of key clinical events across biologic and targeted synthetic disease-modifying antirheumatic drugs to guide safer prescribing behavior.

Methods

A population-based cohort study analysed patients with rheumatoid arthritis initiated on one of 12 biologic and targeted synthetic disease-modifying antirheumatic drugs (2009–22) in Hong Kong, with adverse events assessed over 5 years, including major adverse cardiovascular events, infections, malignancies, diabetes mellitus, depression, hospitalisation and all-cause mortality. Marginal structural models with inverse probability weighting addressed time-varying confounding and selection bias. Death was treated as a competing risk, and Bonferroni correction accounted for multiple comparisons.

Results

Among 2697 patients with rheumatoid arthritis (4276 treatment episodes), the estimated 5-year cumulative probability of adverse events were broadly comparable across biologic and targeted synthetic disease-modifying antirheumatic drugs: with hospitalisation that ranged from 0.48 (95% confidence interval 0.271, 0.636) for baricitinib to 0.726 (0.665, 0.785) for tocilizumab; mortality from 0.028 (0.002, 0.049) for adalimumab to 0.07 (0.03, 0.118) for abatacept; major adverse cardiovascular events from 0.025 (0.004, 0.046) for golimumab and 0.025 (0.007, 0.046) for adalimumab to 0.072 (0.035, 0.107) for abatacept; infection from 0.213 (0.15, 0.265) for etanercept to 0.389 (0.17, 0.664) for rituximab; malignancy from 0.012 (0, 0.028) for abatacept to 0.116 (0, 0.309) for baricitinib; diabetes from 0.018 (0.003, 0.03) for etanercept to 0.079 (0.015, 0.152) for golimumab; depression from 0 (no events recorded) for baricitinib to 0.029 (0, 0.09) for rituximab; and gastritis from 0.041 (0, 0.078) for baricitinib to 0.114 (0.075, 0.157) for tofacitinib. Risks were compared with etanercept to estimate risk differences, which is the most frequently prescribed agent in Hong Kong with a well-documented safety profile. Tocilizumab was associated with a higher hospitalisation risk (risk difference 0.205; 95% confidence interval 0.104, 0.299) and tofacitinib was associated with a higher gastritis risk (risk difference 0.055; 95% confidence interval 0.001, 0.105).

Conclusions

This study demonstrates broadly similar safety profiles among Biologic and targeted synthetic disease-modifying antirheumatic drugs after rigorous adjustment, alleviating concerns about severe adverse events. These findings provide clinicians with evidence to make informed treatment decisions for patients with rheumatoid arthritis.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s40264-026-01687-1.

Key Points

Using advanced statistical methods to account for biases in observational data, this study found that the long-term safety of commonly used biologic and targeted synthetic therapies for rheumatoid arthritis is broadly similar across agents.

Compared with etanercept, tocilizumab was linked to a notably higher rate of hospitalisation and tofacitinib was associated with an elevated risk of gastritis.

These findings may allow clinicians to prioritise efficacy and patient preference when selecting treatment, without being unduly constrained by safety concerns.

Introduction

Rheumatoid arthritis (RA) is a common, chronic autoimmune inflammatory disease that leads to irreversible functional disabilities [1]. Biologic and targeted synthetic disease-modifying antirheumatic drugs (b/ts DMARDs) are effective in managing RA after failure of standard therapies [2]. Depending on the therapeutic target, b/ts DMARDs can be categorised into multiple drug classes such as tumour necrosis factor inhibitors (TNFi), interleukin-6 inhibitors and Janus kinase inhibitors (JAKi), which exhibit comparable clinical treatment efficacy [3]. However, b/ts DMARDs are associated with significant safety concerns, including an increased risk of serious infections (e.g. bacterial infections, viral reactivation and opportunistic infections), cardiovascular events and malignancies [2, 4]. Notably, potential safety differences across drug classes have been suggested by individual studies. For instance, one study published in JAMA reported a significantly higher risk of serious infections with infliximab compared with etanercept and adalimumab [5], and the ORAL Surveillance trial detected elevated malignancy and cardiovascular risks with tofacitinib relative to TNFi [6], although subsequent analyses have yielded inconsistent findings regarding these associations [7]. Evidence from a meta-analysis suggests that tocilizumab may be associated with a reduced risk of cardiovascular events compared with TNFi [8]. These discrepancies highlight the ongoing uncertainty surrounding the comparative safety profiles of individual b/ts DMARDs. With the increasing number of approved b/ts DMARDs, there is an urgent need for robust and up-to-date comparative safety data to resolve these uncertainties, support individualised benefit-risk assessments and guide personalised therapeutic strategies.

Randomised controlled trials face inherent limitations in evaluating drug safety, including a limited follow-up duration, inadequate power to detect rare adverse events and restrictive comparison groups (e.g. placebo-controlled or limited head-to-head comparisons) [9]. These constraints leave clinicians without robust evidence when selecting among all available treatment options in real-world practice. This evidence gap underscores the need for well-designed observational studies. Electronic medical records enable comprehensive comparisons of multiple treatments in real-world settings, offering large homogeneous cohorts and a sufficient follow-up duration to evaluate post-marketing safety profiles. However, previous observational studies have been limited by conventional methodologies vulnerable to time-varying confounding and selection bias, which may distort true safety outcomes [10, 11]. These methodological constraints demonstrate the necessity for advanced analytical approaches to generate more reliable evidence for clinical decision making. More recent methodologies, including marginal structural models (MSM) [12] and G-method approaches [13], offer promising solutions to handle complex time-varying exposures and confounders, potentially providing more accurate estimates of drug safety in real-world populations.

This population-based study compares the safety profiles of different b/ts DMARDs, addressing conventional design limitations including time-varying confounding and selection bias through inverse probability weighting (IPW) to strengthen the evidence level. The findings may enhance evidence-based treatment decisions with improved robustness over existing literature.

Method

Data Source

Longitudinal clinical data including anonymised information on demographics, clinical diagnoses, prescription, laboratory tests, hospitalisation and mortality were extracted from a Hong Kong territory-wide electronic medical record database—the Clinical Data Analysis and Reporting System. This database is managed by the Hong Kong Hospital Authority, a statutory institution overseeing all public hospitals and clinics in Hong Kong that provides healthcare services to more than 7.3 million residents, accounting for around 80% of routine hospital admissions [14]. The Clinical Data Analysis and Reporting System is renowned for accurate and reliable coding and is widely used by researchers to publish in leading scientific journals [15, 16].

Population

The study population included adult patients (age ≥18 years) with RA who have been initiated on at least one b/ts DMARD during 2009–22. Identification of patients with RA was based on the International Classification of Diseases, Ninth Revision, Clinical Modification (714.0). Patients with other autoimmune diseases were excluded [Inflammatory bowel disease (555,556); Systemic lupus erythematosus (710.0); Spondyloarthritis (720.0); Psoriasis (696); Multiple sclerosis (340)]. Indications linked with rituximab other than RA were also excluded from the cohort [Non-Hodgkin’s Lymphoma (200, 202) Chronic Lymphocytic Leukaemia (204.1) Granulomatosis with Polyangiitis and Microscopic Polyangiitis (446)].

Marginal Structure Model (MSM)

Observational cohort studies often involve confounders and exposures that are time dependent, where the time-dependent confounders are influenced by prior exposures and affect the subsequent exposures. When time-dependent confounders function as a risk factor for the outcome event and also predict subsequent exposure, traditional regression methods for estimating the effect of time-dependent exposure on the outcome event are subject to over-adjustment bias and selection bias, which has been elucidated elsewhere [17, 18]. Time-dependent exposure and time-dependent confounders widely exist in the context of RA management. For example, inflammatory markers such as the erythrocyte sedimentation rate (ESR) and C-reactive protein (CRP) are routinely used in clinical practice to assess RA disease activity. These markers occupy a dual role in the causal structure: they act as mediators on the pathway from treatment to clinical outcomes, as b/ts DMARDs exert their therapeutic and safety profiles in part through the suppression of systemic inflammation; at the same time, they function as time-varying confounders because clinicians may consider switching to alternative b/ts DMARDs when the inflammatory response deteriorates after initiation of the first b/ts DMARD, where the biomarker level was affected by the previous treatment but also predicts the next exposure. Meanwhile, these inflammatory biomarkers are also associated with a series of RA comorbidities [19]. Similarly, the use of adjunctive symptomatic therapies, such as analgesics, non-steroidal anti-inflammatory drugs (NSAIDs) and corticosteroids that act to control pain and inflammation often reflect inadequate disease control with current b/ts DMARDs, suggesting a higher likelihood of switching therapies. These adjunctive treatments are also linked to RA-related comorbidities [20, 21]. A MSM [12] is a regression model applying the inverse probability of a treatment and censoring weighting method to create a pseudo-population, where the effects of baseline confounders and time-dependent confounders can be theoretically removed, allowing for unbiased estimation of causal effects and valid inferences in observational studies.

Exposure and Outcomes

Approved b/ts DMARDs used for RA in Hong Kong were included and grouped according to the mechanism of action: TNFi: adalimumab, certolizumab pegol, etanercept, golimumab and infliximab; interleukin-6 inhibitors: sarilumab and tocilizumab; lymphocyte-targeting agents: abatacept and rituximab; and JAKi: baricitinib, tofacitinib and upadacitinib (Table 1 of the Electronic Supplementary Material [ESM]). Patients were followed up from the index date, defined as the first b/ts DMARDs prescription, to the occurrence of the first event, observation end date (5 years after index date), study end date (31-12-2023), death, treatment discontinuation or switch, whichever occurred earlier. Treatment discontinuations comprised scenarios where a treatment gap greater than 180 days from two consecutive records was detected (A 365-day window was chosen for rituximab, as it is typically recommended for administration every 6 months). We treated all biosimilar versions of each drug to be equivalent considering their comparable efficacy and safety to reference biological originators, meaning that switching between them was not regarded as a treatment switch. Given the complex treatment algorithm of RA, patients are permitted to transition between different b/ts DMARDs, thus each patient could contribute to multiple treatment episodes with different drugs.

We investigated frequently reported adverse events related to a b/ts DMARD prescription including major adverse cardiovascular events, infection, depression, gastritis, malignancy, diabetes, any hospitalisation and all-cause mortality [2, 22–24]. Individuals with documented adverse events within the preceding 3 years, or with any hospitalisation 1 year prior to the index date were excluded from analyses to ensure only incident events were captured, except for all-cause mortality. This exclusion criterion was applied on an outcome-specific basis, which means an individual excluded from the analysis of one outcome could still contribute to the analysis of other outcomes for which they had no prior history. The International Classification of Diseases, Ninth Revision, Clinical Modification codes used for identifying outcomes, covariates and exclusion criteria are detailed in Table 2 of the ESM. To capture adverse events that occurred shortly after a treatment switch or discontinuation, we applied a lag period of 90 days following the initial cessation date (183 days for rituximab).

Covariates

Time-fixed confounding variables including age, sex, number of b/ts DMARDs historically treated, number of conventional synthetic DMARDs historically treated (hydroxychloroquine, sulphasalazine, methotrexate, leflunomide, azathioprine, cyclosporin) and the Charlson Comorbidity Index score. Time-varying confounding variables include ESR and CRP levels, concurrent use of each conventional synthetic DMARD, oral anti-inflammation agents and analgesic agents: glucocorticoids (prednisolone), NSAIDs (celecoxib, etoricoxib, diclofenac, ibuprofen, mefenamic, naproxen, piroxicam, sulindac) and opioids (codeine, dextropropoxyphene, dihydrocodeine, methadone, morphine, oxycodone, tramadol). Time-varying variables were updated every 3 months. The rationale for a 3-month interval for time-varying confounders was because this aligns with clinical monitoring practices in RA [2], where disease activity and treatment adjustments are typically reassessed every 3 months. While a shorter interval could capture more rapid changes, it might introduce instability in IPW due to missing data. Conversely, a longer interval could miss important confounder changes. We believe this interval balances granularity and data availability while minimising noise from very short-term fluctuations.

For ESR and CRP laboratory tests, mean test results were applied if multiple test records were detected. Multiple imputation by chained equations [25] was applied to impute missing values if no laboratory tests were recorded. The missing rates of CRP and ESR were relatively stable, with a range from 36% to 44% throughout the entire follow-up period (Table 3 of the ESM). The descriptive metrics before and after imputation were detailed in Table 4 of the ESM. For prednisolone prescriptions, NSAIDs, and opioids, the drug strength (e.g. 5 mg), drug frequency (e.g. twice a day) and drug dosage (e.g. two tablets) along with the prescription period were extracted together to compute the daily dose over the predefined period. Furthermore, to facilitate comparability across different drugs, all computed daily doses were standardised by the World Health Organization Defined Daily Dose (Table 5 of the ESM), described as the estimated average daily maintenance dose of a drug for its primary indication in adults [26].

Statistical Analysis

Starting from the index date, we assessed whether patients continued b/ts DMARD therapy every 3 months, patients were censored once they deviated their initial b/ts DMARD strategy (whether discontinued or switched to other agents). To account for both baseline confounding by indication and potential bias induced by artificial censoring, we constructed stabilised IPWs as the product of a baseline inverse probability of treatment weight and a time-varying inverse probability of censoring weight (IPCW) [12, 27], estimated independently for each treatment arm. The stabilised inverse probability of treatment weight addressed baseline confounding. For each treatment arm, the denominator was the predicted probability of receiving the assigned b/ts DMARD conditional on baseline covariates, estimated from a multivariable logistic regression model; the numerator was the marginal (unconditional) probability of the assigned treatment. The stabilised IPCW addressed informative censoring due to treatment deviation. The unstabilised IPCW was computed as the cumulative product of the inverse probability of remaining on the initial b/ts DMARD given both baseline and time-varying covariates up to each visit. The numerator was the corresponding cumulative probability conditional on baseline covariates only, yielding a stabilised IPCW to reduce variance. The final stabilised weight at each visit was the product of the stabilised inverse probability of treatment weight and the stabilised IPCW. The variables and functional forms used to create the weights are detailed in Table 6 of the ESM. Last, weights were truncated at 99 percentiles for both sides to minimise the influence of outliers. Detailed weight diagnostics are provided in the Table 7 of the ESM. Across all treatment groups, the mean stabilised truncated weights ranged from 0.865 to 1.019 with standard deviations of 0.41–0.73. The proportion of weights at the upper truncation threshold was ≤1.75% across all arms; the proportion at the lower threshold was ≤2.2% in seven of eight arms, with the exception of rituximab (17.89%), reflecting its selective prescribing pattern. Density plots (Fig. 1 of the ESM) showed that the stabilised truncated weights were right-skewed with means close to 1.0 across all treatment arms. Love plots of the absolute standardised mean differences before and after weighting showed that the covariate balance was improved after applying the IPWs, with the majority of covariates falling below the 0.1 threshold (Fig. 2 of the ESM).

The truncated stabilised IPWs were applied to create a pseudo-cohort in which the effects of both time-fixed and time-varying confounders were eliminated. We then fitted a pooled logistic regression model to this pseudo-cohort, adjusting for b/tsDMARD agents, baseline covariates, visit and visit-by-agent interaction terms to account for non-proportional hazards. From this model, we estimated the 5-year risk of each clinical outcome under a strategy of continuous treatment with the index b/ts DMARD. Patients were censored at the time of switching to any other b/ts DMARD, including switches within the same drug class. Switches between an originator biologic and its biosimilar were not considered treatment changes and did not trigger censoring. We then calculated the risk difference (RD) by directly subtracting the absolute risk of the reference group (etanercept) from the absolute risk of each comparator drug. We selected etanercept as the reference drug because it is the most frequently prescribed TNFi agent in Hong Kong [28] and is generally considered to have a favourable safety profile among b/ts DMARDs. It has also been commonly used as the reference group in previous comparative safety studies [4]. We used bootstrap sampling with replicates 1000 times to obtain the pointwise 95% percentile confidence interval (CI). Particularly, products with treatment episodes lower than 200 were removed from the analysis to reduce the influence of extreme variability in small samples, ensuring more robust effect estimates.

This study simultaneously investigated nine different clinical events, including negative control (fracture) outcomes, which introduces the risk of multiple comparisons. To address this, the Bonferroni correction [29] was implemented by adjusting the confidence level of the bootstrap percentile CIs. Specifically, instead of using the conventional 95% confidence level, we constructed CIs at a level of (1–0.05/9), where 9 corresponds to the number of outcomes examined within each comparison. An RD that does not cross 0 was considered statistically significant.

Competing Risk of Death

Death was treated as a competing event for all clinical outcomes other than death itself. Patients who die are no longer at risk for the outcome of interest; therefore, we did not treat deaths as censored observations. Instead, we used a discrete-time competing risks (multi-state) approach [30] modelled on three mutually exclusive outcomes (event of interest, death, no event) to obtain accurate estimates of the cumulative incidence of clinical events over time.

Negative Controls

In observational studies, even after adjusting for known confounding variables to the greatest extent possible, there remains no guarantee that the exposure and non-exposure groups are fully comparable. A negative control refers to an exposure or outcome that is not expected to have a true association with the outcome or exposure of interest [31, 32]. An unexpected association indicates potential bias or confounding, signalling needed adjustments. In this study, we selected fractures (International Classification of Diseases, Ninth Revision, Clinical Modification codes: 805, 812, 813, 814, 820)—an event risk expected not to differentiate between different b/ts DMARDs as the negative control. We anticipated observing a similar fracture risk for all types of b/ts DMARDs to prove that systematic errors, bias and confounding are well maintained in our study.

Sensitivity Analysis

To further explore how variations in study design and analytical approach might affect our results and conclusions, we conducted a series of sensitivity analyses. First, we examined the 5-year incidence of adverse events across drug groups categorised by their mechanisms of action. This class-level aggregation was intended to assess potential mechanism-specific safety signals and to improve precision where individual agents had limited sample sizes. Second, to evaluate how the choice of lag period might influence our findings, particularly for adverse events that may require a longer latent period to manifest such as malignancy, we extended the lag period to 180 days (and to 365 days specifically for rituximab). Third, we reintroduced the four agents initially excluded owing to fewer than 200 treatment episodes (certolizumab, sarilumab, upadacitinib, infliximab) to determine whether any safety signals could be detected despite limited event counts. Fourth, to eliminate the potential impact of a within-population correlation arising from patients contributing multiple treatment episodes, we restricted the analysis to each patient’s first treatment episode only. Fifth, we re-fitted the identical pooled logistic regression model without applying IPW to contextualise findings from earlier studies that relied on conventional multivariable regression without IPW. Sixth, we computed CIs without applying the Bonferroni correction to examine whether additional signals might emerge, acknowledging that such signals would be more susceptible to a type I error. Finally, to assess sensitivity to non-linear associations, we replaced linear terms for age, ESR, and CRP with restricted cubic splines (3 degrees of freedom) [33]. Splines were not applied to other continuous covariates (e.g. opioid dose, Charlson comorbidity index) because of their limited number of unique values or high proportion of zeros, which would risk overfitting.

Statistical analyses were conducted using R Version 4.1.4 (R Foundation for Statistical Computing, Vienna, Austria). Study results were cross-checked by two researchers independently to ensure accuracy and consistency (KP and VKCY).

Results

Baseline Characteristics

This study analysed 4276 treatment episodes from 2697 unique patients with RA between 2009 and 2022, with baseline characteristics showing a marked variation across b/ts DMARD cohorts (Table 1). Patients initiating rituximab (mean age (SD) 57.20 ± 12.49 years), Charlson comorbidity index (0.78 ± 1.47), abatacept (58.75 ± 13.44, 0.54 ± 0.98) and baricitinib (60.32 ± 13.27, 0.44 ± 0.84) represented older populations with greater comorbidity burdens versus the overall cohort (56.38 ± 12.84, 0.38 ± 0.92). These groups exhibited clinically relevant comorbidity differences: rituximab users had higher rates of congestive heart failure (3.2 vs 0.9% overall) and malignancy (11.2 vs 2.9% overall), while abatacept users showed elevated diabetes without complications prevalence (12.8 vs 7.1% overall). Treatment history varied significantly, with rituximab [number of different b/ts DMARD agents prior to current therapy (SD) 1.59 ± 1.53] and abatacept (1.21 ± 1.24) users having more prior b/ts DMARD exposures than average (0.74 ± 1.06). Inflammatory markers were highest in tocilizumab (ESR (mm/hour mean (SD)) 53.55 ± 31.87, CRP (mg/dL mean (SD)) 2.01 ± 2.72) and rituximab (52.26 ± 35.18, 2.07 ± 2.55) cohorts versus overall means (48.58 ± 30.78, 1.70 ± 2.36), suggesting greater baseline inflammation activity in these groups. Concomitant medication use differed substantially, particularly for methotrexate (range 57.0% in rituximab to 79.7% in etanercept vs 70.1% overall) and prednisolone (highest in rituximab at 0.43 ± 0.44 vs 0.27 ± 0.34 overall), reflecting distinct therapeutic approaches across agent subgroups.

Table 1.

Baseline characteristics of patients under different b/ts DMARD regimens

Description	Overall	Etanercept	Adalimumab	Golimumab	Tocilizumab	Tofacitinib	Abatacept	Rituximab	Baricitinib	P-value
Number of patients	4276	706	692	472	800	768	336	251	251
Male [n (%)]	700 (16.4)	120 (17.0)	113 (16.3)	78 (16.5)	118 (14.8)	127 (16.5)	64 (19.0)	40 (15.9)	40 (15.9)	0.826
Female [n (%)]	3576 (83.6)	586 (83.0)	579 (83.7)	394 (83.5)	682 (85.2)	641 (83.5)	272 (81.0)	211 (84.1)	211 (84.1)
Age [mean (SD)]	56.38 (12.84)	54.27 (12.61)	54.87 (12.62)	54.85 (13.37)	56.71 (12.81)	57.67 (12.01)	58.75 (13.44)	57.20 (12.49)	60.32 (13.27)	<0.001
Charlson Comorbidity Index score [mean (SD)]	0.38 (0.92)	0.30 (0.82)	0.31 (0.80)	0.31 (0.79)	0.43 (1.01)	0.32 (0.77)	0.54 (0.98)	0.78 (1.47)	0.44 (0.84)	<0.001
Medical history [n (%)]
Myocardial infarction	49 (1.1)	8 (1.1)	8 (1.2)	2 (0.4)	12 (1.5)	6 (0.8)	6 (1.8)	3 (1.2)	4 (1.6)	0.584
Congestive heart failure	40 (0.9)	1 (0.1)	2 (0.3)	3 (0.6)	11 (1.4)	3 (0.4)	8 (2.4)	8 (3.2)	4 (1.6)	<0.001
Peripheral vascular disease	15 (0.4)	2 (0.3)	0 (0.0)	1 (0.2)	1 (0.1)	2 (0.3)	5 (1.5)	3 (1.2)	1 (0.4)	0.003
Cerebrovascular disease	118 (2.8)	12 (1.7)	21 (3.0)	17 (3.6)	24 (3.0)	14 (1.8)	13 (3.9)	7 (2.8)	10 (4.0)	0.187
Chronic obstructive pulmonary disease	246 (5.8)	37 (5.2)	39 (5.6)	16 (3.4)	50 (6.2)	40 (5.2)	30 (8.9)	19 (7.6)	15 (6.0)	0.055
Rheumatic disease other than rheumatic arthritis	35 (0.8)	3 (0.4)	4 (0.6)	3 (0.6)	12 (1.5)	4 (0.5)	1 (0.3)	7 (2.8)	1 (0.4)	0.004
Peptic ulcer disease	221 (5.2)	31 (4.4)	35 (5.1)	23 (4.9)	48 (6.0)	29 (3.8)	27 (8.0)	13 (5.2)	15 (6.0)	0.134
Mild liver disease	31 (0.7)	6 (0.8)	5 (0.7)	1 (0.2)	4 (0.5)	4 (0.5)	5 (1.5)	2 (0.8)	4 (1.6)	0.311
Diabetes without chronic complication	303 (7.1)	38 (5.4)	33 (4.8)	35 (7.4)	56 (7.0)	51 (6.6)	43 (12.8)	21 (8.4)	26 (10.4)	<0.001
Diabetes with chronic complication	21 (0.5)	4 (0.6)	4 (0.6)	0 (0.0)	4 (0.5)	4 (0.5)	2 (0.6)	1 (0.4)	2 (0.8)	0.872
Hemiplegia or paraplegia	16 (0.4)	2 (0.3)	2 (0.3)	4 (0.8)	4 (0.5)	1 (0.1)	2 (0.6)	1 (0.4)	0 (0.0)	0.530
Chronic renal failure	36 (0.8)	5 (0.7)	4 (0.6)	2 (0.4)	8 (1.0)	7 (0.9)	6 (1.8)	3 (1.2)	1 (0.4)	0.477
Malignancy	122 (2.9)	11 (1.6)	12 (1.7)	10 (2.1)	19 (2.4)	23 (3.0)	8 (2.4)	28 (11.2)	11 (4.4)	<0.001
Metastatic solid tumour	7 (0.2)	2 (0.3)	1 (0.1)	0 (0.0)	0 (0.0)	1 (0.1)	0 (0.0)	3 (1.2)	0 (0.0)	0.005
Moderate-to-severe liver disease	7 (0.2)	3 (0.4)	1 (0.1)	0 (0.0)	0 (0.0)	0 (0.0)	2 (0.6)	1 (0.4)	0 (0.0)	0.122
Medications
Number of b/ts DMARDs previously used [mean (SD])	0.74 (1.06)	0.34 (0.66)	0.48 (0.78)	0.48 (0.88)	0.91 (0.95)	0.77 (1.12)	1.21 (1.24)	1.59 (1.53)	0.95 (1.34)	<0.001
Number of csDMARDs previously used [mean (SD)]	3.14 (1.28)	3.07 (1.35)	3.02 (1.29)	3.19 (1.15)	3.22 (1.20)	3.02 (1.36)	3.50 (1.07)	3.36 (1.32)	3.03 (1.35)	<0.001
Prednisolone [mean (SD)]	0.27 (0.34)	0.30 (0.34)	0.22 (0.30)	0.26 (0.32)	0.31 (0.37)	0.23 (0.29)	0.30 (0.32)	0.43 (0.44)	0.20 (0.27)	<0.001
NSAID [mean (SD)]	0.71 (0.62)	0.71 (0.61)	0.67 (0.62)	0.76 (0.58)	0.70 (0.61)	0.71 (0.65)	0.75 (0.67)	0.70 (0.60)	0.70 (0.64)	0.410
Opioid [mean (SD)]	0.06 (0.17)	0.05 (0.19)	0.04 (0.17)	0.05 (0.17)	0.07 (0.18)	0.05 (0.15)	0.09 (0.22)	0.06 (0.17)	0.05 (0.15)	0.001
Hydroxychloroquine	969 (22.7)	201 (28.5)	159 (23.0)	100 (21.2)	176 (22.0)	168 (21.9)	67 (19.9)	65 (25.9)	33 (13.1)	<0.001
Sulphasalazine	1050 (24.6)	230 (32.6)	171 (24.7)	109 (23.1)	185 (23.1)	177 (23.0)	81 (24.1)	50 (19.9)	47 (18.7)	<0.001
Methotrexate	2999 (70.1)	563 (79.7)	535 (77.3)	362 (76.7)	546 (68.2)	490 (63.8)	208 (61.9)	143 (57.0)	152 (60.6)	<0.001
Leflunomide	1183 (27.7)	215 (30.5)	217 (31.4)	178 (37.7)	185 (23.1)	209 (27.2)	73 (21.7)	33 (13.1)	73 (29.1)	<0.001
Azathioprine	21 (0.5)	6 (0.8)	2 (0.3)	1 (0.2)	2 (0.2)	0 (0.0)	4 (1.2)	6 (2.4)	0 (0.0)	<0.001
Cyclosporin	127 (3.0)	31 (4.4)	18 (2.6)	15 (3.2)	15 (1.9)	19 (2.5)	7 (2.1)	17 (6.8)	5 (2.0)	0.001
Laboratory
ESR, mm/h [mean (SD)]	48.58 (30.78)	46.67 (29.80)	47.14 (30.65)	49.85 (29.82)	53.55 (31.87)	46.13 (29.82)	47.58 (29.66)	52.26 (35.18)	44.90 (29.85)	<0.001
CRP, mg/dL [mean (SD)]	1.70 (2.36)	1.81 (2.43)	1.58 (2.26)	1.58 (2.11)	2.01 (2.72)	1.46 (2.14)	1.64 (2.21)	2.07 (2.55)	1.50 (2.18)	<0.001

b/ts DMARDs biological or targeted synthetic disease-modifying anti-rheumatic disease drugs, CRP C-reactive protein, csDMARDs conventional synthetic disease-modifying anti-rheumatic disease drugs, ESR erythrocyte sedimentation rate, NSAID non-steroidal anti-inflammatory drug, SD standard deviation

Five-Year Cumulative Incidence and Risk Differences (RDs)

The estimated 5-year cumulative risks are summarised in Table 2 and the corresponding survival curves were visualised in Fig. 1 by individual b/ts DMARDs. We found that 5-year absolute risks varied across b/ts DMARDs: hospitalisation: [absolute 5-year cumulative risk 0.48 (baricitinib) to 0.726 (tocilizumab)]; mortality: [0.028 (adalimumab) to 0.07 (abatacept)]; major adverse cardiovascular events: [0.025 (adalimumab and golimumab) to 0.072 (abatacept)]; gastritis [0.041 (baricitinib) to 0.114 (tofacitinib)]; infection: [0.213 (etanercept) to 0.389 (rituximab)]; malignancy: [0.012 (abatacept) to 0.116 (baricitinib)]; diabetes: [0.018 (etanercept) to 0.079 (golimumab)]; depression: [0 (baricitinib) to 0.029 (rituximab)].

Table 2.

Characteristics and risk estimates for selected adverse events by B/ts DMARD agent versus etanercept

Adverse events	Agents	Number of patients	Number of events	Median follow-up days (IQR)	Crude incidence rate per 1000 PYs	Absolute risk (95% CI)a	Risk difference (95% CI)a
Any hospitalisation	Etanerceptb	450	169	498 (224, 1407.25)	178.1	0.521 (0.442, 0.593)	Ref
Any hospitalisation	Abatacept	150	71	326.5 (158, 730.5)	344.7	0.595 (0.446, 0.752)	0.074 (−0.09, 0.241)
Any hospitalisation	Adalimumab	496	151	446.5 (248, 673.75)	200	0.561 (0.402, 0.653)	0.04 (−0.108, 0.172)
Any hospitalisation	Baricitinib	161	56	446 (193, 903)	227.6	0.48 (0.271, 0.636)	−0.041 (−0.24, 0.135)
Any hospitalisation	Golimumab	299	108	404 (231.5, 918.5)	207.3	0.567 (0.454, 0.669)	0.046 (−0.081, 0.167)
Any hospitalisation	Rituximab	96	37	215.5 (196, 607)	321.7	0.528 (0.299, 0.72)	0.007 (−0.211, 0.196)
Any hospitalisation	Tocilizumab	455	267	341 (153, 824.5)	370.3	0.726 (0.665, 0.785)	0.205 (0.104, 0.299)
Any hospitalisation	Tofacitinib	531	195	615 (281, 1203.5)	176.3	0.544 (0.433, 0.622)	0.023 (−0.087, 0.123)
Death	Etanercept	706	10	776 (320.5, 1826)	5.4	0.042 (0.011, 0.071)	Ref
Death	Abatacept	336	16	518 (279.5, 1176.5)	23	0.07 (0.03, 0.118)	0.028 (−0.019, 0.092)
Death	Adalimumab	692	7	495.5 (325.5, 849.25)	5.3	0.028 (0.002, 0.049)	−0.014 (−0.051, 0.027)
Death	Baricitinib	251	13	682 (258.5, 996.5)	28.6	0.031 (0.011, 0.052)	−0.012 (−0.045, 0.027)
Death	Golimumab	472	8	544 (284, 1456.75)	7.5	0.047 (0.007, 0.1)	0.004 (−0.047, 0.065)
Death	Rituximab	251	4	399 (198, 837.5)	9.9	0.039 (0, 0.079)	−0.003 (−0.063, 0.054)
Death	Tocilizumab	800	21	817.5 (313, 1826)	9.9	0.045 (0.019, 0.068)	0.003 (−0.037, 0.046)
Death	Tofacitinib	768	19	955 (423.75, 1519.5)	9.3	0.05 (0.02, 0.083)	0.008 (−0.035, 0.051)
Depression	Etanercept	687	10	763 (323, 1826)	5.6	0.027 (0.008, 0.045)	Ref
Depression	Abatacept	325	2	518 (281, 1178)	3	0.017 (0, 0.037)	−0.01 (−0.041, 0.02)
Depression	Adalimumab	679	7	487 (319, 840.5)	5.5	0.021 (0.003, 0.038)	−0.006 (−0.034, 0.019)
Depression	Baricitinib	248	0	679 (259.25, 996.25)	0	0 (0, 0)	−0.027 (−0.045, −0.008)
Depression	Golimumab	458	5	551 (284, 1432)	4.8	0.017 (0, 0.03)	−0.009 (−0.035, 0.015)
Depression	Rituximab	247	3	393 (198, 822.5)	7.6	0.029 (0, 0.09)	0.002 (−0.039, 0.067)
Depression	Tocilizumab	785	12	804 (301, 1826)	5.8	0.023 (0.008, 0.036)	−0.004 (−0.03, 0.021)
Depression	Tofacitinib	753	9	947 (412, 1509)	4.5	0.02 (0.004, 0.033)	−0.006 (−0.033, 0.018)
Diabetes	Etanercept	668	8	776 (318.25, 1826)	4.6	0.018 (0.003, 0.03)	Ref
Diabetes	Abatacept	296	7	499.5 (271, 1158.75)	11.7	0.028 (0.006, 0.051)	0.01 (−0.02, 0.039)
Diabetes	Adalimumab	665	9	495 (335, 845)	7.2	0.026 (0.007, 0.044)	0.008 (−0.014, 0.032)
Diabetes	Baricitinib	233	8	676 (257, 997)	19.1	0.055 (0.01, 0.134)	0.037 (−0.018, 0.115)
Diabetes	Golimumab	442	11	544 (284.25, 1464.25)	10.9	0.079 (0.015, 0.152)	0.061 (−0.005, 0.137)
Diabetes	Rituximab	231	6	395 (198, 837.5)	16.3	0.078 (0, 0.162)	0.06 (−0.016, 0.143)
Diabetes	Tocilizumab	753	22	822 (313, 1826)	11.1	0.04 (0.02, 0.06)	0.022 (−0.002, 0.05)
Diabetes	Tofacitinib	727	17	953 (410, 1531)	8.8	0.034 (0.016, 0.05)	0.017 (−0.007, 0.04)
Gastritis	Etanercept	677	17	789 (327, 1826)	9.6	0.06 (0.022, 0.092)	Ref
Gastritis	Abatacept	312	11	509.5 (271, 1176.5)	17.2	0.09 (0.017, 0.151)	0.031 (−0.047, 0.105)
Gastritis	Adalimumab	654	15	494.5 (321, 835.75)	12.2	0.055 (0.018, 0.089)	−0.004 (−0.058, 0.047)
Gastritis	Baricitinib	228	6	633.5 (256.75, 996.25)	14.7	0.041 (0, 0.078)	−0.018 (−0.066, 0.038)
Gastritis	Golimumab	447	14	544 (280.5, 1460.5)	13.8	0.048 (0.013, 0.082)	−0.012 (−0.061, 0.04)
Gastritis	Rituximab	232	6	394 (198, 825.25)	16.1	0.042 (0.002, 0.095)	−0.018 (−0.072, 0.05)
Gastritis	Tocilizumab	758	29	826.5 (309.25, 1826)	14.5	0.067 (0.035, 0.095)	0.008 (−0.038, 0.059)
Gastritis	Tofacitinib	725	44	908 (411, 1485)	23.3	0.114 (0.075, 0.157)	0.055 (0.001, 0.105)
Infection	Etanercept	586	76	694 (301.75, 1741)	52.9	0.213 (0.15, 0.265)	Ref
Infection	Abatacept	241	32	474 (266, 1012)	69.9	0.268 (0.161, 0.369)	0.055 (−0.066, 0.167)
Infection	Adalimumab	601	70	481 (302, 790)	64.6	0.228 (0.16, 0.297)	0.015 (−0.062, 0.106)
Infection	Baricitinib	207	35	641 (264, 969)	94.6	0.273 (0.135, 0.459)	0.06 (−0.093, 0.25)
Infection	Golimumab	404	48	475 (264, 1187.25)	57.6	0.257 (0.153, 0.365)	0.044 (−0.069, 0.178)
Infection	Rituximab	187	26	393 (197, 764)	97	0.389 (0.17, 0.664)	0.176 (−0.057, 0.448)
Infection	Tocilizumab	650	109	646.5 (266.25, 1767)	69.6	0.25 (0.194, 0.295)	0.037 (−0.039, 0.126)
Infection	Tofacitinib	673	90	866 (390, 1451)	53.1	0.22 (0.169, 0.277)	0.007 (−0.066, 0.078)
MACE	Etanercept	692	19	758 (318.25, 1826)	10.6	0.07 (0.02, 0.114)	Ref
MACE	Abatacept	320	16	508.5 (271, 1141.5)	24.8	0.072 (0.035, 0.107)	0.002 (−0.061, 0.073)
MACE	Adalimumab	678	10	495 (322.5, 848)	7.8	0.025 (0.007, 0.046)	−0.045 (−0.095, 0.011)
MACE	Baricitinib	243	9	684 (267.5, 989)	20.3	0.069 (0.013, 0.157)	−0.001 (−0.083, 0.108)
MACE	Golimumab	461	8	523 (284, 1429)	7.7	0.025 (0.004, 0.046)	−0.045 (−0.091, 0.016)
MACE	Rituximab	242	4	394 (198, 821.25)	10.5	0.026 (0, 0.056)	−0.044 (−0.101, 0.013)
MACE	Tocilizumab	776	22	811.5 (316, 1826)	10.8	0.036 (0.018, 0.053)	−0.034 (−0.082, 0.018)
MACE	Tofacitinib	757	20	934 (412, 1502)	10	0.04 (0.02, 0.057)	−0.03 (−0.082, 0.029)
Malignancy	Etanercept	703	14	768 (316.5, 1826)	7.6	0.031 (0.011, 0.049)	Ref
Malignancy	Abatacept	335	4	518 (279, 1175)	5.8	0.012 (0, 0.028)	−0.019 (−0.043, 0.009)
Malignancy	Adalimumab	689	10	495 (321, 845)	7.7	0.036 (0.01, 0.064)	0.005 (−0.029, 0.044)
Malignancy	Baricitinib	248	2	679 (259.25, 996.25)	4.4	0.116 (0, 0.309)	0.084 (−0.045, 0.278)
Malignancy	Golimumab	471	5	540 (284.5, 1460.5)	4.7	0.028 (0, 0.053)	−0.003 (−0.039, 0.033)
Malignancy	Rituximab	236	3	391 (198, 836.25)	7.9	0.027 (0, 0.09)	−0.005 (−0.042, 0.062)
Malignancy	Tocilizumab	795	10	817 (313, 1826)	4.8	0.026 (0.007, 0.044)	−0.005 (−0.033, 0.026)
Malignancy	Tofacitinib	763	14	947 (416.5, 1506)	6.9	0.034 (0.013, 0.053)	0.003 (−0.027, 0.033)
Negative control
Fracture	Etanercept	695	10	763 (321, 1826)	5.5	0.03 (0.009, 0.05)	Ref
Fracture	Abatacept	326	5	519 (274.25, 1153.5)	7.5	0.022 (0, 0.045)	−0.008 (−0.038, 0.021)
Fracture	Adalimumab	680	6	488 (319.5, 839.25)	4.7	0.017 (0.001, 0.032)	−0.013 (−0.04, 0.015)
Fracture	Baricitinib	249	2	676 (257, 996)	4.5	0.026 (0, 0.202)	−0.005 (−0.046, 0.17)
Fracture	Golimumab	467	6	544 (283.5, 1449.5)	5.7	0.037 (0.003, 0.068)	0.007 (−0.039, 0.048)
Fracture	Rituximab	250	9	397 (198, 825.75)	22.7	0.066 (0.006, 0.129)	0.035 (−0.026, 0.105)
Fracture	Tocilizumab	780	12	811.5 (318, 1826)	5.8	0.044 (0.011, 0.071)	0.013 (−0.03, 0.046)
Fracture	Tofacitinib	757	10	950 (419, 1503)	5	0.038 (0.006, 0.066)	0.008 (−0.03, 0.05)

b/ts DMARs biological or targeted synthetic disease-modifying anti-rheumatic disease drugs, CI confidence interval, IQR interquartile range, MACE major adverse cardiovascular events, PY person-year, Ref reference

^aThe absolute risk and risk difference of 5-year event risk comparing eight b/ts DMARD products were obtained from fitting an outcome model to a pseudo population where all individuals were assumed to continue the assigned treatment to the end of 5 years unless an outcome event occurs. The outcome model was obtained from weighted pooled logistic regression adjusted for baseline variables, visits (every 3 months) and b/ts DMARD agents, with an interaction term between b/ts DMARD agents and visits to allow non-proportional hazards. Certolizumab pegol, infliximab, sarilumab and upadacitinib were excluded from the analysis because of treatment episodes less than 200

^bEtanercept was selected as the reference agent

Fig. 1.

Projected absolute risk for selected biological or targeted synthetic disease-modifying anti-rheumatic disease drug (b/ts DMARD) agents

Figure 2 visualises the RDs for individual b/ts DMARDs using a forest plot. Across outcomes, several agents showed numerical differences on absolute risk, but the 95% CIs for RDs spanned zero, indicating no definitive excess risk versus etanercept. Signals for diabetes were higher with multiple agents (e.g. golimumab RD 0.061; rituximab RD 0.060). A similar pattern was seen for infection, where rituximab had a higher risk versus etanercept (RD 0.176). Mortality was numerically higher with abatacept (RD 0.028). Malignancy appeared higher with baricitinib (RD 0.084), while depression had zero events with baricitinib, yielding a negative RD (−0.027) that reflects imprecision rather than protection. These estimates are based on a few events with wide CIs and should be interpreted cautiously. Our analysis detected two positive signals: tocilizumab was associated with a higher risk of any hospitalisation [RD (95% CI): 0.205 (0.104, 0.299)], translating to roughly 21 additional hospitalisations per 100 patients over 5 years; and tofacitinib was associated with an increased risk of gastritis [RD (95% CI): 0.055 (0.001, 0.105)].

Fig. 2.

Forest plot of risk differences for selected biological or targeted synthetic disease-modifying anti-rheumatic disease drugs (b/ts DMARDs) versus etanercept. MACE major adverse cardiovascular events, Ref reference

Negative Controls

No significant difference was observed for the negative control outcome (fracture) between groups, indicating minimal residual confounding.

Sensitivity Analysis

Stratification by Mechanism of Action

The estimated 5-year cumulative risks stratified by mechanism of action are presented in Table 8 and Fig. 3 of the ESM, with corresponding RDs shown in Fig. 4 of the ESM. Compared with TNFi, interleukin-6 inhibitors were associated with a statistically significantly higher risk of any hospitalisation [RD (CI): 0.189 (0.114, 0.271)]. Similarly, JAKi exhibited a statistically significantly higher risk of gastritis [RD (CI): 0.051 (0.009, 0.094)] compared with TNFi. The risks of all other studied adverse events were similar across the four mechanism-of-action classes.

Extended Lag Period

The results of extending the lag period to 180 days (and to 365 days for rituximab) are presented in Table 9 of the ESM, tocilizumab retained a higher risk of any hospitalisation compared with etanercept [RD (CI): 0.195 (0.105, 0.297)]. One additional signal emerged: tocilizumab was associated with a higher risk of diabetes [RD (CI): 0.031 (0.002, 0.056)]. Consistent with the main analysis, malignancy risk remained comparable to etanercept across all treatments.

Inclusion of Agents with Limited Sample Sizes

In the sensitivity analysis expanding the comparisons to include all 12 b/ts DMARDs, the baseline characteristics of the added four agents certolizumab, infliximab, sarilumab, and upadacitinib are presented in Table 10 of the ESM, and the results of cumulative risk are displayed in the Table 11 of the ESM. Among the four additionally included agents, infliximab was associated with a significantly higher risk of any hospitalisation compared with etanercept [RD (CI): 0.27 (0.066, 0.422)]. However, many estimates showed signs of instability. For example, upadacitinib yielded an implausibly high absolute risk estimate for depression [0.76 (CI: 0, 1)] and death [0.363 (CI: 0, 0.909)], with only three events and one event recorded among 145 patients, respectively. Additionally, several zero-event comparisons yielded statistically significant negative RDs, including infliximab for death [RD (CI): −0.04 (−0.066, −0.008)] and depression [RD (CI): −0.027 (−0.044, −0.007)], and upadacitinib for malignancy [RD (CI): −0.031 (−0.048, −0.012)].

Restriction to First Treatment Episode

After restricting the analysis to the first eligible treatment episode per patient to eliminate within-patient correlation, the number of treatment episodes decreased from 4276 to 2697 (Table 12 of the ESM). The results were materially consistent with the primary analysis (Table 13 of the ESM). Tocilizumab and tofacitinib again showed a significantly elevated risk of hospitalisation [RD (95% CI): 0.190 (0.082, 0.312)] and gastritis [0.055 (0.005, 0.099)], respectively. Abatacept appeared to have a lower malignancy risk [−0.034 (−0.057, −0.009)], though this was based on only one event captured.

Unweighted Analysis

Table 14 of the ESM presented the unweighted associations removing IPW. Two signals retained their statistical significance: tocilizumab for any hospitalisation [unweighted RD (CI): 0.200 (0.110, 0.292); MSM RD (CI): 0.205 (0.104, 0.299)] and tofacitinib for gastritis [0.066 (0.020, 0.110); 0.055 (0.001, 0.105)], while no new signals emerged. Nevertheless, several non-significant estimates shifted materially in magnitude after IPW adjustment: the RDs for abatacept and rituximab on any hospitalisation were attenuated by approximately 30% and toward the null, respectively [abatacept: 0.106 (−0.025, 0.267) to 0.074 (−0.090, 0.241); rituximab: 0.041 (−0.191, 0.254) to 0.007 (−0.211, 0.196)], and the RD for abatacept on death was attenuated by approximately 40% [0.049 (−0.003, 0.124) to 0.028 (−0.019, 0.092)].

Analysis Without Bonferroni Correction

Without the Bonferroni correction, five additional comparisons reached statistical significance at the 95% confidence level (Table 15 of the ESM): a higher risk of diabetes with golimumab [RD (95% CI): 0.061 (0.011, 0.113); Bonferroni-adjusted CI: (−0.005, 0.137)] and tocilizumab [0.022 (0.005, 0.041); (−0.002, 0.050)] versus etanercept, and lower risks of fracture with adalimumab [−0.045 (−0.078, −0.003); (−0.095, 0.011)], golimumab [−0.045 (−0.081, −0.002); (−0.091, 0.016)], and rituximab [−0.044 (−0.086, −0.001); (−0.101, 0.013)] versus etanercept. All five comparisons involved small numbers of events, and the bounds of the unadjusted 95% CIs barely excluded zero.

Analysis with Restricted Cubic Splines

Both signals from the primary analysis persisted after applying splines, with only minor shifts in estimates; tocilizumab for any hospitalisation (spline RD: 0.217 vs RD 0.205), and gastritis for tofacitinib (spline RD: 0.052 vs RD 0.055). No new signals emerged (Table 16 of the ESM).

Discussion

In this territory-wide observational study, we utilised MSM incorporating time-varying variables to comparatively assess the risk of adverse events associated with a wide spectrum of b/ts DMARDs, using a large and relatively homogeneous cohort. We observed no differences in 5-year cumulative event risk across the studied outcomes, except for any hospitalisation.

Interpretation of the Key Clinical Findings

Tumour necrosis factor inhibitors represent the anchor drug class for RA treatment, serving as first-line biologic therapy owing to their extensive safety profile and well-established efficacy, with alternative b/ts DMARDs reserved for TNFi inadequate responders per clinical guidelines [2]. Consequently, those advancing to non-TNFi therapies generally exhibit poor baseline health status, including higher disease activity scores, a longer disease duration and greater accumulated joint damage [34]. Our cohort confirmed these patterns; patients treated with non-TNFi agents such as abatacept, rituximab and tocilizumab were generally older, had more comorbidities, used higher baseline doses of NSAIDs and prednisone, and had experienced more conventional synthetic DMARD failures and b/ts DMARD failures, all of which are linked to increased disease severity and poorer prognosis. Additionally, studies have shown that compared with patients where TNFi treatment failed, patients new to any b/ts DMARDs show a lower risk of treatment-related adverse events [35]. Therefore, patients receiving non-TNFi agents appear to have more adverse events overall, likely reflecting channelling bias.

In our analysis, we found that abatacept shows a numerically higher 5-year mortality versus etanercept, this finding appears counterintuitive given abatacept is considered to have a favourable safety profile regarding serious infections and major adverse cardiovascular events [2]. The adjusted RD is not statistically significant (95% CI includes zero), suggesting the signal likely reflects channelling bias rather than a true excess risk. This interpretation aligns with prior studies [4] that observed higher mortality with abatacept (e.g. hazard ratio 1.32 vs etanercept), where the study’s authors attributed this observation to systematic differences in baseline health status among abatacept-treated patients. In our study, tocilizumab exhibits an exceptionally high rate of hospitalisation compared with other b/ts DMARDs. While an earlier study [34] also reported a higher rate of treatment-associated adverse events in the tocilizumab group (4.8 vs 3.2% in the TNFi group), suggesting that this pattern is not unique to our cohort. We consider residual confounding to be the more plausible explanation, as tocilizumab is frequently prescribed to patients with more refractory disease and a higher burden of comorbidities. Despite careful adjustment with MSMs, residual confounding may persist across treatment groups, complicating direct comparisons. Accordingly, the observed mortality and hospitalisation patterns should be interpreted within the context of indication bias and the differing baseline risk profiles of these clinically distinct patient populations.

Patients with RA are more likely to develop diabetes due to RA-related insulin resistance and the use of glucocorticoids [36]. Biologic and targeted synthetic DMARDs may mitigate this risk by dampening systemic inflammation and modulating relevant pathways [37–39]. Theoretically, any effective treatment that reduces the overall inflammatory burden could lower the risk of developing diabetes, although it is unclear whether certain DMARDs have a more pronounced or specific beneficial effect on glucose metabolism.

A US-based new-user cohort [40] reported higher incident diabetes with adalimumab (hazard ratio 2.00, 95% CI 1.11, 3.03) and infliximab (hazard ratio 2.34, 95% CI 1.38, 3.98) versus abatacept; the authors suggested improved insulin sensitivity as a mechanism for the protective effect of abatacept on diabetes but offered no convincing explanation for its superiority over TNFi, especially given Solomon et al. [38] also found a reduced diabetes risk with TNFi for improved insulin sensitivity. Methodological limitations, residual confounding and the absence of head-to-head randomised controlled trials with incident diabetes as a primary endpoint further temper causal inference. In our study, 5-year diabetes risks were modest and broadly comparable across JAKi, interleukin-6 inhibitors, and TNFi (adjusted absolute risk 0.018–0.079). Before multiplicity adjustment, we observed slightly higher risks for golimumab [RD (95% CI): 0.061 (0.011, 0.113)] and tocilizumab [0.022 (0.005, 0.041)] compared with etanercept. However, these signals were based on few events (11 and 22, respectively), did not survive multiplicity adjustment [Bonferroni-adjusted CIs (−0.005, 0.137) and (−0.002, 0.050)], and lack a clear biological rationale given that etanercept and golimumab share a tumour necrosis factor-inhibiting mechanism. These results suggest that the control of systemic inflammation may be more important for diabetes risk in patients with RA than the choice of a specific b/ts DMARD.

Our analysis identified tofacitinib was associated with a higher risk of gastritis compared with etanercept [RD (CI): 0.055 (0.001, 0.105)], a finding that remained robust after Bonferroni correction and across all sensitivity analyses. This observation aligns with a previous comprehensive review that gastritis has been explicitly listed as a recognised gastrointestinal adverse effect in a comprehensive review for JAKi but not for TNFi such as etanercept [24]. Interestingly, baricitinib, which selectively inhibits Janus kinase (JAK)1/JAK2 but has minimal activity against JAK3, was not associated with an elevated risk of gastritis in our analysis [RD (CI): −0.018 (−0.066, 0.038)]. Whether the differential gastritis risk between tofacitinib and baricitinib relates to JAK3 inhibition warrants further investigation. Although the underlying biological mechanism is under-investigated, this finding carries clinical relevance that gastric monitoring may be warranted when receiving tofacitinib for patients with pre-existing gastrointestinal comorbidities or concurrent use of NSAIDs or corticosteroids.

The risk of serious infections is elevated in patients with RA owing to immune dysregulation and frequent glucocorticoid use. How to tailor treatment for patients at a high infection risk remains a central concern in clinical decision making [9]. In our study, etanercept showed the lowest 5-year infection risk at 0.213 (95% CI, 0.15, 0.265) in accordance with previous studies [4, 41]. However, because etanercept users in our cohort had more favourable baseline health, we cannot exclude channelling bias as an explanation for this apparent advantage. Additionally, while prior studies have reported higher rates of herpes zoster and tuberculosis with tofacitinib compared with TNFi [4, 42], we did not observe such differences, possibly because our outcome captured all-type infections rather than pathogen-specific events. Overall, our findings suggest that agent-specific concerns about all-type infections with contemporary b/ts DMARDs may be less critical.

The systemic inflammation and functional disability associated with RA make incident depression a common comorbidity among patients with RA [43]. Effective disease control with b/ts DMARDs improves psychosocial well-being and lowers the risk of depression. Numerous studies have explored the potential direct antidepressant effects of TNFi and JAKi with doubts about their impact on cytokine pathways [44, 45]. Because no studies have directly compared incident depression across b/ts DMARDs and randomised controlled trials have not assessed it as an endpoint, our analysis offers preliminary evidence that depression incidence does not differ by b/ts DMARD, although confirmation with independent cohorts is warranted.

Sensitivity Analysis

We conducted a series of sensitivity analyses that collectively reinforced the robustness of the primary findings. Analytical variations span from stratification by mechanism of action, extension of the lag period, restriction to the first treatment episode, removal of the MSM framework without using IPW and exclusion of Bonferroni correction to fit restricted cubic splines for continuous variables. The elevated risk of any hospitalisation associated with tocilizumab and the elevated risk of gastritis associated with tofacitinib compared with etanercept remained the most consistent signals, supporting these as the most credible findings of our study. Importantly, malignancy risk was comparable across all treatments regardless of how the analysis was configured, which addresses a key concern frequently raised in the literature regarding the long-term oncological safety of b/ts DMARDs. The only notable exception to this overall consistency arose when agents with fewer than 200 treatment episodes were reintroduced. Several estimates for these newly included agents demonstrated clear signs of instability, including implausibly high absolute risk estimates and statistically significant negative RDs in the presence of zero events. These findings suggest that the IPWs may become unreliable in very small treatment groups, and the primary analysis decision to exclude such agents was therefore justified. Comparing MSM estimates with unweighted results highlights the impact of treatment-confounder feedback. Several borderline associations for non-TNFi agents including abatacept and rituximab with hospitalisation, and abatacept with death were attenuated toward the null after IPW adjustment. This suggests that the apparent excess of safety events in non-TNFi users is likely a result of channelling bias, and supports the use of an MSM framework to avoid erroneous risk attribution. While the removal of the Bonferroni correction revealed five additional borderline significant comparisons, all involved very small numbers of events and CIs that only marginally excluded zero. We therefore retain the Bonferroni-adjusted results in the main text to guard against spurious associations, while presenting the unadjusted results in the ESM for readers to evaluate. Finally, the consistency between the spline-based sensitivity analysis and the primary analysis indicates that our main findings are not driven by the linearity assumption imposed on age, ESR and CRP, and confirms that our results are robust to alternative specifications of the functional form.

Strengths and Limitations

Our study possesses several unique strengths. First, this research represents one of the most comprehensive, long-term comparative safety evaluations of the full spectrum of b/ts DMARDs conducted in a real-world setting. By performing a head-to-head comparison of nine different agents with an extensive follow-up period, we provide critical evidence on long-term safety that extends far beyond the limited duration of pivotal clinical trials. This comprehensive scope equips clinicians with pragmatic comparative evidence essential for navigating treatment decisions including sequencing and switching over a patient’s entire long-term disease course. Second, our study is methodologically advanced, offering robust analysis with a complex real-world clinical scenario that mirrors actual patient journeys. We deliberately employed a prevalent new user design to ensure our findings are generalisable to routine care where patients move between various therapies over many years. By incorporating death as a competing risk, it accounts for potential biases that could arise from censoring events. The application of Bonferroni correction ensures that only robust variations are detected, minimising the likelihood of false-positive results and limiting misinterpretation. The use of fractures as a negative control outcome further supports the validity of our analytical approach. Glucocorticoid use is a well-established risk factor for fracture [46], and poorer disease control is associated with greater glucocorticoid reliance. Agents with comparatively weaker efficacy would therefore theoretically drive fracture risk upward through this mediated pathway. However, because glucocorticoid use was treated as a time-varying covariate within the MSM framework, this indirect association would have been adjusted for in our analysis. The absence of a detectable difference in fracture risk across agents thus proves that MSM have successfully accounted for this pathway, and may be taken as additional support for the validity of our analytical approach.

This study also has inherent limitations. First, treatment assignment was non-randomised in real-world practice, leaving some residual patient selection bias that may distort observed outcomes. For example, tofacitinib, often perceived as higher risk, is generally prescribed to younger patients with fewer comorbidities owing to safety warnings for elderly patients, potentially contributing to its favourable safety profile in our findings. Conversely, abatacept was more often given to sicker or refractory patients as it is generally considered safe. Despite MSM with IPW and competing-risk handling, residual confounding, channelling and differential follow-up may persist, particularly for therapies used in patients with a higher baseline risk. Second, because of the nature of the electronic medical record database, we were unable to capture validated composite measures of disease activity such as Disease Activity Score-28 or the Clinical Disease Activity Index, which directly reflect patients’ treatment response. The omission of these variables may have introduced residual confounding of uncertain direction and magnitude. Third, our analysis addressed informative censoring due to treatment switching through IPCW. Other forms of informative censoring, such as loss to follow-up were not explicitly modelled. However, the register-based nature of our data source ensures near-complete follow-up for the outcomes of interest, substantially mitigating the risk of bias from non-treatment-related informative censoring. Fourth, to capture more safety events, we defined events as the absence of a diagnosis record in the previous 3 years, rather than true incident (new-onset) cases. As a result, some events may represent recurrences or ongoing conditions rather than new occurrences. Despite our efforts, this large territory-wide study still faced challenges in capturing rare events such as malignancy and depression, the low event counts means that even sizable numerical differences may reflect imprecision rather than true drug effects. Future research could build upon this framework by integrating larger multi-database cohorts and incorporating registry or trial data where disease activity measures are available, thereby enhancing statistical power for rare outcomes and addressing the limitations associated with unmeasured key variables.

Conclusions

In this comparative safety analysis of b/ts DMARDs, we observed broadly similar risks of major adverse events after rigorous confounding adjustment. While isolated outcome-specific differences were detected, these were modest in magnitude and may reflect residual confounding rather than clinically meaningful disparities. Overall, our findings support the comparable safety profile of b/ts DMARDs in RA management and provide critical head-to-head evidence to inform decision making.

Supplementary Information

Below is the link to the electronic supplementary material.

Declarations

Conflict of interest

Xue Li received research grants or contracts from the Health and Medical Research Fund (HMRF) Main Scheme and HMRF Fellowship Scheme of the Hong Kong Special Administrative Region (SAR), China, and from the Research Grants Council Early Career Scheme, Research Impact Fund (Hong Kong SAR); was former nonexecutive director of Advanced Data Analytics for Medical Science Limited, Hong Kong SAR, China; received commission grants from Department of Health, Government of Hong Kong SAR, China; commission grants from Hospital Authority of Hong Kong SAR, China, internal funding from the University of Hong Kong, and research or education grants from Pfizer, Janssen and Bristol Myers Squibb (BMS), Novartis; received consultancy fees from Merck Sharp & Dohme, Pfizer, Open Health and The Office of Health Economics; and received honoraria for associate editorship from Nature Springer. None related to the work reported in the current manuscript. Nicole L. Pratt is an Editorial Board member of Drug Safety and was not involved in the selection of peer reviewers for the manuscript nor in any of the subsequent editorial decisions. Edward Chia‑Cheng Lai reports receiving research funding from Amgen, AbbVie, GW, IQVIA, J&J, MSD, Novartis, Pfizer, Prospection, Sanofi, and Taketa, outside the submitted work. Esther W.Y. Chan has received grants from Research Grants Council (RGC, HKSAR), Research Fund Secretariat of the Food and Health Bureau (Health and Medical Research Fund [HMRF], HKSAR), National Health and Medical Research Council (Australia), Narcotics Division of the Security Bureau of HKSAR, Innovation and Technology Commission of the Government of HKSAR, National Natural Science Fund of China, Amgen, AstraZeneca, Bayer, Bristol Myers Squibb, Eisai, Janssen, Pfizer, Takeda, and Novartis; and honorarium from Pfizer, Novartis, and the Hong Kong SAR Hospital Authority outside the submitted work. Kuan Peng, Vincent K.C. Yan, Shirley C.W. Chan, Ju-Young Shin and Chak-sing Lau have no conflicts of interest that are directly relevant to the content of this article.

Acknowledgements

We thank Lisa Lam for proofreading this paper, and we thank Dr. Deliang Yang for providing statistical support during the analysis.

Funding

The Enhanced Start-up Fund for new academic staff, LKS Faculty of Medicine, The University of Hong Kong, The University of Hong Kong Seed Fund for resubmission of General Research Fund, and the Internal Research Fund, Department of Medicine, School of Clinical Medicine, LKS Faculty of Medicine, The University of Hong Kong provided funding for the preparation of this article.

Ethics approval

This study was performed in accordance with the Declaration of Helsinki (1964) and its subsequent amendments. Ethical approval for the research was granted by the Institutional Review Board of The University of Hong Kong Hospital Authority Hong Kong West Cluster under reference numbers UW 21-338. The institutional review board granted a waiver of informed patient consent. This was justified as the studies were exclusively observational, meaning no practical interventions were administered, and all patient-level data utilised were de-identified by the data custodian at the time of data extraction.

Consent to participate

Not applicable.

Consent for publication

Not applicable.

Availability of data and material

The data supporting the findings of this study were obtained from the electronic health records database managed by the Hospital Authority of Hong Kong. Owing to data privacy regulations and licensing restrictions, the data cannot be shared publicly. Access to the data may be requested directly from the Hospital Authority of Hong Kong, subject to their approval and compliance with applicable regulations.

Code availability

The analytical code supporting this study is publicly available at https://github.com/kuanpeng420/BtsDMARDs-safety.

Authors contributions

KP and VKCY performed the data analysis. KP, SCWC and XL conceptualised the study. KP and XL wrote the manuscript. SCWC, NLP, CSL, ECCL,CSL, JYS and EWYC contributed to the interpretation of results. XL supervised the study and obtained funding. All authors reviewed and approved the final manuscript.