Using the afore mentioned search terms and platforms, we identified 600 citations (details in Additional file 1 section). After excluding duplicates, studies not including TCZ, ABA, or RTX therapies and for being observational studies, we found 19 trials to be included for qualitative synthesis and meta-analysis. The PRISMA flowchart used for selecting studies can be found in Additional file 1: Fig. S1.

For our primary outcome meta-analysis, we included 19 RCTs with 7,835 patients (9,402 patient-years) randomized to the intervention arm or to the control arm. The mean age of patients included in the present study was 52.3 years; 77.2% of these patients were women; the mean disease duration was 8.7 years; and the mean follow-up was 1.2 years. Details of the basic trial characteristics are presented in Additional file 1: Table S2.

Among the 19 selected trials, 31.5% (six) included TCZ as intervention arm, 31.5% (six) included RTX as active arm, and 31.5% (six) included ABA as intervention arm. Only one study compared ABA with RTX. Additional file 1: Fig. S2 shows the network of direct comparisons among the treatment arms of included studies.

All studies had a low risk of bias according to the Cochrane Collaboration tool (Additional file 1: Table S3) and were of high quality according to the GRADE system (Additional file 1: Table S4). As shown in Additional file 1: Figure S3A-S3C, there was no significant publication bias in funnel charts and no significant small-study bias according to the Egger tests.

In the meta-analysis of random effects for the chance of achieving an ACR70 response at six months, six studies [20,21,22,23,24,25] with ABA therapy were selected. In these studies, 437 (25.6%) patients in the intervention arm and 250 (19.8%) patients in the control arm achieved an ACR70 response at six months and the observed HR was 1.35 (95% CI 1.17–1.55 p: 0.013, I2 = 41% p for heterogeneity = 0.013). For RTX, six studies were selected [26,27,28,29,30,31], where 350 (23.5%) patients in the intervention arm and 107 (12%) patients in the control arm achieved an ACR70 response at six months, indicating a HR of 2.43 (95% CI 1.99–2.96 p < 0.001, I2 = 64% p for heterogeneity < 0.001). For TCZ, six studies were selected [32,33,34,35,36,37], with 294 (18.5%) patients in the intervention arm and 97 (11%) patients in the control arm achieving an ACR70 response at six months, with a HR of 1.53 (95% CI 1.24–1.89 p = 0.002, I2 = 76% p for heterogeneity < 0.001) (Additional file 1: Fig. S4A).

As shown in Table 1 and, direct and indirect comparisons suggest a superiority of ABA, TCZ, or RTX vs. placebo, with TCZ showing a slightly lower magnitude of effect (HR 2.765, 95% CI 1.240–6.692 p = 0.009) compared to ABA (HR 3.423, 95% CI 1.422– 8.709 p < 0.001) and to RTX (HR 3.494, 95% CI 1.530–8.658 p < 0.001) in achieving ACR70 response at six months. There was no statistical difference between these drugs and the TNFi therapy, considering 4 studies of ABA vs TNFi (HR 1.043 (95% CI 0.422, 2.242), p = 0.59), 1 study of TCZ vs TNFi (HR 0.839 (95% CI 0.373, 3.804), p = 0.45) and 1 study of RTX vs TNFi (HR 1.059 (95% CI 0.339, 2.581), p = 0.51). For all analyses, there was a slight inconsistency between direct and indirect measures (Table 1, Indirect comparisons among bDMARDs), but they showed a substantial heterogeneity among RCTs (> 40%) (Additional file 1: Fig. S4A).

For the chances of achieving an ACR50 response, we selected the same studies as selected for the ACR70 response in six months. For ABA, 730 (43%) patients in the intervention arm and 439 (34.7%) patients in the placebo arm were involved, with a HR of 1.28 (95% CI 1.17–1.41 p ≤ 0.001, I2 = 54% p for heterogeneity = 0.005). For RTX, 586 (39.4%) patients in the intervention arm and 201 (22.6%) patients in the control arm were involved, with a HR of 1.94 (95% CI 1.70–2.20 p ≤ 0.001, I2 = 67% p for heterogeneity ≤ 0.001). For TCZ, 537 (34%) patients in the intervention arm and 178 (20.3%) patients in the control arm were involved, with a HR of 1.75 (95% CI 1.52–2.02 p ≤ 0.001, I2 = 79% p for heterogeneity ≤ 0.001) (Additional file 1: Fig. S4B).

Indirect comparisons among the intervention arms showed that ABA vs. RTX present a HR = 1.052 (95% CI 0.514–2.131, p = 0.986), TCZ vs. RTX present a HR = 0.833 (95% CI 0.378–1.835 p = 0.93), and TCZ vs. ABA present a HR = 0.879 (95% CI 0.393–1.961 p = 0.91) for achieving ACR50 in six months.

Finally, for the chances of achieving an ACR20 response, 1,108 (65%) patients in the intervention arm and 644 (51%) patients in the placebo arm were included for ABA, with a HR of 1.63 (95% CI 1.53–1.75, p ≤ 0.001, I2 = 84%, p for heterogeneity ≤ 0.001). For RTX, 880 (59.2%) patients in the intervention arm and 343 (38.5%) patients in the control arm were involved, with a HR of 1.75 (95% CI 1.61–1.91 p ≤ 0.001, I2 = 71% p for heterogeneity ≤ 0.001). For TCZ, 843 (53.2%) patients in the intervention arm and 292 (33.4%) patients in the control arm were involved, with a HR of 1.90 (95% CI 1.73–2.10 p ≤ 0.001, I2 = 85% p for heterogeneity ≤ 0.001) (Additional file 1: Fig. S4C).

When indirectly compared to each other, ABA vs. RTX show a HR of 1.013 (95% CI 0.669–1.535 p = 0.890) and TCZ vs. ABA show a HR of 1.014 (95% CI 0.626–1.618 p = 0.92).

The spider chart shown in Fig. 1 presents an important imbalance among studies in the frequency of TNFi treatment in the control arm. RCTs with TCZ less often included TNFi agents in the control arm: ABA, 50%; RTX, 43%; and TCZ, 17% (p = 0.048). Besides, the baseline HAQ score was significantly lower (ABA 1.68 ± 0.15, RTX 1.68 ± 0.23, and TCZ 1.50 ± 0.17; p = 0.049). The mean follow-up time (total study duration) among RCTs was significantly longer in trials with ABA (25 ± 18.75 months) compared to those with RTX (9.43 ± 3.21) and TCZ (6.00 ± 0) (p = 0.015). Baseline 28-joint Disease Activity Score (DAS28) also showed an imbalance among trials, but this variable was excluded from the meta-regressions for presenting a high collinearity with the baseline HAQ score (Additional file 1: Fig. S5A).

The Spider chart describes the imbalance among RCTs characteristics included in the meta-analysis according to the non-TNFi treatment used in active arm (ABA, RTX or TCZ). The suffix "_posit.vs.ACR70" represent the independent variables that show positive association with ACR70 response, and the suffix "_neg.vs.ACR70" represent the independent variables that show negative association with ACR70 response. Example: the higher the follow-up time (suffix "_neg.vs. ACR70"), the lower the chance of achieving ACR70 response. Followup.t: follow-up time (total study duration); ACR70: 70% achieving ACR response; ACR50: 50% achieving ACR response; ACR20: 20% achieving ACR response. Background TNFi: percentage of individuals on tumor necrosis factor (TNF) inhibitors; HAQ: Health Assessment Questionnaire score; ABA: abatacept; RTX: rituximab; TCZ: tocilizumab; ACR: American College of Rheumatologists

Meta-regressions were conducted to assess the impact of treatments on the ACR70 response in a manner adjusted to the baseline HAQ score, study follow-up time (total study duration), and frequency of TNFi treatment in the control arm (Additional file 1: Fig. S5B). These 3 explanatory variables were found after exploring all potential predictors of ACR70 response such as percentage of women, glucocorticoid use, patients' age, MTX dose, duration of RA, baseline HAQ, baseline DAS28, study follow-up time and frequency of TNFi treatment in the control arm.

These explanatory variables were chosen in a forward stepwise process, and all the 3 variables were associated with the ACR70 response in univariate regressions and were imbalanced among treatment arms. As shown in Table 2, each 0.1 point over the mean baseline HAQ in each study was associated with a RR of 2.043 (95%CI 1.032–14.44, p for difference = 0.028, R2 = 13%, I2 = 89%, p for heterogeneity ≤ 0.001) for achieving an ACR70 response (Additional file 1: Fig. S4B). Additionally, the presence of an TNFi therapy in the control arm of any RCT was associated with a RR of 0.316 (95% CI 0.134–0.746, p for difference = 0.009, R2 = 27%, I2 = 85%, p for heterogeneity ≤ 0.001) (Additional file 1: Fig. S4A) and each additional month of follow-up was associated with a RR of 0.980 (95% CI 0.961–0.996, p for difference = 0.048, R2 = 9%, I2 = 71%, p for heterogeneity ≤ 0.001) for achieving an ACR70 response.

In a multivariate model with the ACR70 response as an outcome and the covariates treatments in the intervention arm (RTX vs. ABA and TCZ vs. ABA), baseline HAQ score, follow-up time, and frequency of TNFi treatment in the control arm (Table 2), the heterogeneity was reduced (I2 = 24%, p for heterogeneity = 0.27) and the explanatory power of the model increased (R2 = 85%). In this model, ABA did not modify the chance of achieving an ACR70 response when compared to RTX (1.773, 95% CI 0.113–10.21, p = 0.765). In contrast, ABA was associated with a RR of 2.217 (95% CI 1.554–3.161, p for difference ≤ 0.001) when compared to TCZ for achieving an ACR70 response. It means that, if the conditions of the RCTs were similar, ABA could increase the chance of achieving an ACR70 response by 2.22-fold when compared to TCZ. In this model, the meta-regression of RTX vs TCZ was not included in Model 4 as this covariate showed high collinearity with other covariates (variance inflation factor of 9 with ABA vs TCZ).

Sensitivity analyses were performed considering the exclusion of the RCTs with the longest exposure durations (highest number of patient-years), one study being excluded for each class of non-TNFi biological drug [20, 25, 33]. As shown in Table 1, the exclusion of such studies did not modify the results.