Methotrexate monotherapy and methotrexate combination therapy with traditional and biologic disease modifying anti‐rheumatic drugs for rheumatoid arthritis: A network meta‐analysis

Glen S Hazlewood; Cheryl Barnabe; George Tomlinson; Deborah Marshall; Daniel JA Devoe; Claire Bombardier

doi:10.1002/14651858.CD010227.pub2

. 2016 Aug 29;2016(8):CD010227. doi: 10.1002/14651858.CD010227.pub2

Methotrexate monotherapy and methotrexate combination therapy with traditional and biologic disease modifying anti‐rheumatic drugs for rheumatoid arthritis: A network meta‐analysis

Glen S Hazlewood ^1,^2,^3,^✉, Cheryl Barnabe ^2,^4,⁵, George Tomlinson ⁶, Deborah Marshall ^2,⁵, Daniel JA Devoe ⁵, Claire Bombardier ^7,^8,⁹

Editor: Cochrane Musculoskeletal Group

PMCID: PMC7087436 PMID: 27571502

Abstract

Background

Methotrexate is considered the preferred disease‐modifying anti‐rheumatic drug (DMARD) for the treatment of rheumatoid arthritis, but controversy exists on the additional benefits and harms of combining methotrexate with other DMARDs.

Objectives

To compare methotrexate and methotrexate‐based DMARD combinations for rheumatoid arthritis in patients naïve to or with an inadequate response (IR) to methotrexate.

Methods

We systematically identified all randomised controlled trials with methotrexate monotherapy or in combination with any currently used conventional synthetic DMARD , biologic DMARDs, or tofacitinib. Three major outcomes (ACR50 response, radiographic progression and withdrawals due to adverse events) and multiple minor outcomes were evaluated. Treatment effects were summarized using Bayesian random‐effects network meta‐analyses, separately for methotrexate‐naïve and methotrexate‐IR trials. Heterogeneity was explored through meta‐regression and subgroup analyses. The risk of bias of each trial was assessed using the Cochrane risk of bias tool, and trials at high risk of bias were excluded from the main analysis. The quality of evidence was evaluated using the GRADE approach. A comparison between two treatments was considered statistically significant if its credible interval excluded the null effect, indicating >97.5% probability that one treatment was superior.

Main results

158 trials with over 37,000 patients were included. Methotrexate‐naïve: Several treatment combinations with methotrexate were statistically superior to oral methotrexate for ACR50 response: methotrexate + sulfasalazine + hydroxychloroquine (“triple therapy”), methotrexate + several biologics (abatacept, adalimumab, etanercept, infliximab, rituximab, tocilizumab), and tofacitinib. The estimated probability of ACR50 response was similar between these treatments (range 56‐67%, moderate to high quality evidence), compared with 41% for methotrexate. Methotrexate combined with adalimumab, etanercept, certolizumab, or infliximab was statistically superior to oral methotrexate for inhibiting radiographic progression (moderate to high quality evidence) but the estimated mean change over one year with all treatments was less than the minimal clinically important difference of five units on the Sharp‐van der Heijde scale. Methotrexate + azathioprine had statistically more withdrawals due to adverse events than oral methotrexate, and triple therapy had statistically fewer withdrawals due to adverse events than methotrexate + infliximab (rate ratio 0.26, 95% credible interval: 0.06 to 0.91). Methotrexate‐inadequate response: In patients with an inadequate response to methotrexate, several treatments were statistically significantly superior to oral methotrexate for ACR50 response: triple therapy (moderate quality evidence), methotrexate + hydroxychloroquine (low quality evidence), methotrexate + leflunomide (moderate quality evidence), methotrexate + intramuscular gold (very low quality evidence), methotrexate + most biologics (moderate to high quality evidence), and methotrexate + tofacitinib (high quality evidence). There was a 61% probability of an ACR50 response with triple therapy, compared to a range of 27% to 64% for the combinations of methotrexate + biologic DMARDs that were statistically significantly superior to oral methotrexate. No treatment was statistically significantly superior to oral methotrexate for inhibiting radiographic progression. Methotrexate + cyclosporine and methotrexate + tocilizumab (8 mg/kg) had a statistically higher rate of withdrawals due to adverse events than oral methotrexate and methotrexate + abatacept had a statistically lower rate of withdrawals due to adverse events than several treatments.

Authors' conclusions

We found moderate to high quality evidence that combination therapy with methotrexate + sulfasalazine+ hydroxychloroquine (triple therapy) or methotrexate + most biologic DMARDs or tofacitinib were similarly effective in controlling disease activity and generally well tolerated in methotrexate‐naïve patients or after an inadequate response to methotrexate. Methotrexate + some biologic DMARDs were superior to methotrexate in preventing joint damage in methotrexate‐naïve patients, but the magnitude of these effects was small over one year.

Plain language summary

Methotrexate alone or in combination with other medications for rheumatoid arthritis

Researchers in the Cochrane Collaboration conducted a review of the effects of methotrexate either taken alone or with other disease‐modifying anti‐rheumatic drugs (DMARDs) for people with rheumatoid arthritis. After searching for all relevant studies up to January 19, 2016, they found 158 studies with over 37,000 people. These studies were published between 1985 and 2016 and were between 12 weeks and 2 years in duration. Their findings are summarised below:

In people with rheumatoid arthritis, compared to taking methotrexate alone:

‐The combination of methotrexate + sulfasalazine + hydroxychloroquine and methotrexate + most biologic DMARDs improves disease activity. Other treatment combinations (methotrexate + hydroxychloroquine, methotrexate + leflunomide, methotrexate + gold injections) may improve disease activity in people who do not respond to methotrexate alone.

‐The combinations of methotrexate + several biologic DMARDs (adalimumab, etanercept, certolizumab, or infliximab) reduces joint damage (as seen on x‐rays) slightly over one year in patients who have not taken methotrexate before.

‐The combinations of methotrexate + azathioprine, methotrexate + cyclosporine and methotrexate + tocilizumab (8 mg/kg) probably increases the chance of stopping the medication due to a side effect.

What is rheumatoid arthritis and what is methotrexate and other disease‐modifying anti‐rheumatic drugs?

When you have rheumatoid arthritis (RA) your immune system, which normally fights infection, attacks the lining of your joints. This makes your joints swollen, stiff and painful. There is no cure for RA at present, so the treatments aim to relieve pain and stiffness and improve your ability to move. Fortunately, there are many medications that can control the disease effectively. These medications are known as disease‐modifying anti‐rheumatic drugs, or DMARDs. Methotrexate is widely regarded as the preferred DMARD for most patients with RA as it works well for most patients and is generally well tolerated. Methotrexate can be used by itself or can be combined with other DMARDs. These other DMARDs include medications that have been available and used for many years (such as sulfasalazine and hydroxychloroquine), as well as newer more expensive treatments (biologic DMARDs and tofacitinib). It is important to understand how all of these treatments compare in terms of the benefits and side effects.

What happens to people with rheumatoid arthritis who take methotrexate combined with other disease‐modifying anti‐rheumatic drugs?

A) People who have not taken methotrexate before:

ACR 50 (number of tender or swollen joints and other outcomes such as pain and disability)

‐61 out of 100 people who took methotrexate + sulfasalazine + hydroxychloroquine and 56 to 67 people out of 100 who took methotrexate + biologic DMARDs or tofacitinib experienced improvement in the symptoms of their rheumatoid arthritis, compared to 41 out of 100 people who took methotrexate alone.

X‐rays of the joints:

‐People who took methotrexate combined with adalimumab, etanercept, certolizumab, or infliximab had a small reduction in the progression of joint damage (Sharp‐van der Heijde score) over one year compared to oral methotrexate, but the estimated amount of damage even with oral methotrexate was very small (2.6 point increase).

Stopping the medication due to a side effect

‐36 out of 100 people who took methotrexate + azathioprine had to stop the medication due to a side effect, compared to 8 people out of 100 who took methotrexate alone.

B) People who have taken methotrexate before:

ACR 50 (number of tender or swollen joints and other outcomes such as pain and disability)

‐61 out of 100 people who took methotrexate + sulfasalazine + hydroxychloroquine and 27 to 64 people out of 100 who took methotrexate + biologic DMARDs or tofacitinib experienced improvement in the symptoms of their rheumatoid arthritis, compared to 13 out of 100 people who took methotrexate alone.

X‐rays of the joints:

‐No treatment resulted in a significant reduction in the amount of joint damage seen on x‐rays over one year.

Stopping the medication due to a side effect

‐21 out of 100 people who took methotrexate + cyclosporine and 12 out of 100 people who took methotrexate + tocilizumab (8 mg/kg) had to stop the medication due to a side effect, compared to 7 people out of 100 who took methotrexate alone.

Background

Description of the condition

Rheumatoid arthritis (RA) is a systemic autoimmune disease manifesting primarily as a symmetric and erosive polyarthritis, affecting between 0.5 and 1% of the adult population (Kvien 2004). Patients with RA experience pain, functional limitation and a significant decline in their health‐related quality of life (Kvien 2004). New treatment approaches with early and intensive treatment targeted to a goal of remission or low disease‐activity can significantly improve outcomes (Knevel 2010).

Description of the interventions

Disease‐modifying anti‐rheumatic drugs (DMARDs) target pathways of inflammation responsible for joint swelling and damage and are the cornerstone of treatment for RA. DMARDs can be classified based on their structure and mechanism of action (Smolen 2014). Conventional synthetic DMARDs are derived synthetically without a specific molecular target in mind and found to have activity in the treatment of RA. Methotrexate (MTX) is considered the preferred conventional synthetic DMARD, based on its excellent benefit to toxicity profile (Singh 2016; Smolen 2014a). The conventional synthetic DMARDs most commonly used in combination with methotrexate are hydroxychloroquine, sulfasalazine and leflunomide. Less commonly used conventional synthetic DMARDs include intra‐muscular gold, cyclosporine and azathioprine.

Biologic DMARDs, in contrast to conventional synthetic DMARDs, are derived through biologic processes and are designed to target specific cells or proteins involved in the inflammatory response. Biologic DMARDs are newer treatments for RA, with the first biologic DMARD (etanercept) approved for RA in 1998. Biologic DMARDs currently in use for RA include: abatacept, rituximab and tocilizumab and the anti‐tumour necrosis factor (TNF) inhibitors (adalimumab, certolizumab, etanercept, golimumab, infliximab). Targeted synthetic DMARDs are the most recent class of medications approved for use in RA. Like conventional synthetic DMARDs, they are developed through synthetic methods, and like biologic DMARDS, they are designed to target specific cellular processes. Tofacitinib is the first targeted synthetic DMARD in use for the treatment of RA.

How the intervention might work

The pathophysiology of RA is complex, involving an interplay between genetic risk factors and environmental triggers, and both innate and adaptive immune responses (McInnes 2011). Methotrexate likely works through multiple mechanisms, including the promotion of adenosine‐mediated anti‐inflammatory effects, increased apoptosis of T cells, and reduction of cell proliferation (Braun 2009; Tian 2007). Absorption of oral methotrexate varies between individuals and is improved with parenteral administration, particularly at doses > 15mg/week (Hoekstra 2004; Schiff 2014). The addition of other conventional synthetic DMARDs to methotrexate may improve control of disease activity through the targeting of complementary immunopathologic mechanisms, although little is known about the mechanism of action of many conventional synthetic DMARDs (Bingham 2001). Biologic DMARDs work through inhibition of cytokines involved in RA pathogenesis including TNF‐alpha (anti‐TNF therapy) and interleukin‐6 (tocilizumab), T‐cell co‐stimulation blockade (abatacept) and B‐cell depletion (rituximab). Tofacitinib inhibits Janus kinase 1 (JAK1) and Janus kinase 3 (JAK3), intracellular tyrosine kinases involved in signal transduction.

Why it is important to do this overview

Methotrexate‐based treatments form the core of rheumatoid arthritis (RA) treatment. Methotrexate is recommended as the first DMARD for most patients with RA, and methotrexate co‐prescription is generally recommended when using biologic DMARDs or the recently approved tofacitinib (Singh 2016; Smolen 2014a). Combining methotrexate with other conventional synthetic DMARDs, however, is more controversial. A trial of conventional synthetic DMARD combination therapy prior to biologic therapy is not currently recommended by either major rheumatology guideline, although each provides the option (Singh 2016; Smolen 2014a). Understanding the comparative benefits and harms of these treatments is essential to inform decision‐making, as biologic DMARD therapy and tofacitinib costs over 10‐20 times that of methotrexate and most conventional synthetic DMARDs. It is important to maximize the use of conventional synthetic DMARDs that are safe and effective, while at the same time avoiding unnecessary delays in administering biologic therapy by using treatments of no proven benefit over methotrexate monotherapy.

Network (mixed treatment) meta‐analyses are a natural avenue of comparative effectiveness research, as they combine all direct and indirect evidence to estimate treatment effects between all treatments of interest (Jansen 2011). If treatments A and B are in the same study, there is direct evidence linking A and B. If they are compared in separate studies to a common comparator C, then the A‐C and B‐C studies allow an indirect comparison of A and B. Longer chains of indirect comparisons (A‐B, B‐C, C‐D) are also possible. Considering indirect evidence is critical if a treatment decision must be made and the treatments have not been directly compared in a head‐to‐head trial. Indirect evidence is also important to consider when treatments have been directly compared, as it adds to the entire body of evidence and may help refine the precision in estimation of the treatment effect (Jansen 2011).

A previous Cochrane network meta‐analysis examined the relative effects of different biologic therapies through indirect comparisons, and found some differences between agents (Singh 2009). Our review expands on this, by including combination therapy with methotrexate + conventional synthetic DMARDs. A previous Cochrane traditional (non‐network) meta‐analysis did not find an additional overall benefit with combination therapy over methotrexate alone (Katchamart 2010). By including indirect evidence we expand the evidence base to draw from. For example, three trials have been published that have compared combination therapy with methotrexate + sulfasalazine + hydroxychloroquine versus methotrexate + anti‐TNF therapy (RACAT 2013; SWEFOT 2012; TEAR 2012). The inclusion of these trials in a network meta‐analysis adds indirect evidence on the relative effects of methotrexate + sulfasalazine + hydroxychloroquine compared to all other treatments in the network.

Objectives

To compare methotrexate‐based DMARD treatments for rheumatoid arthritis in patients naïve to or after an inadequate response (IR) to methotrexate.

Methods

This is an overview of reviews, as some of the interventions of interest have been previously evaluated through Cochrane reviews (Katchamart 2010; Singh 2009). It differs from a traditional overview of reviews though, as we are considering all comparisons between any intervention of interest and will therefore include trials not previously reviewed.

Criteria for considering reviews for inclusion

We included RCTs or controlled clinical trials (CCTs) of at least 12 weeks duration that contained any intervention of interest (defined in detail below). After identifying all studies, trials that could not be linked within the network to another intervention of interest through a shared comparator were excluded. For example, if we identified a trial comparing methotrexate to hydroxychloroquine, the trial would be included if another trial was available that compared hydroxychloroquine to methotrexate + hydroxychloroquine (or any other treatment of interest). In this example, the two trials (methotrexate vs. hydroxychloroquine and hydroxychloroquine vs. methotrexate + hydroxychloroquine) allow an indirect comparison to be made between the treatments of interest (methotrexate, methotrexate + hydroxychloroquine). Similarly, if a trial contained more than 2 arms, each arm was included only if it provided direct or indirect evidence on the treatments of interest.

Trials were divided into 3 groups for all analyses, characterised by prior medication exposure: 1) Methotrexate‐naïve; 2) Methotrexate‐inadequate response (IR); 3) Anti‐TNF‐ inadequate response. Methotrexate‐IR and TNF‐IR trials are trials where the protocol required all patients to have tried and failed methotrexate or anti‐TNF therapy respectively. Trials that included a mix of methotrexate‐naïve patients and patients who had tried methotrexate previously were classified as ‘partial‐exposure’ trials and included in the methotrexate‐naïve analyses, unless subgroup data was available separately for methotrexate‐naïve and methotrexate‐IR patients. We subsequently excluded the TNF‐IR trials, as the identified studies formed a network that included only trials of biologic therapy (i.e. – no studies evaluated conventional synthetic DMARD combination therapy). The comparative benefits and harms of biologic therapy (with and without concomitant methotrexate) have been evaluated in 2 previous Cochrane Overview of Reviews (Singh 2009; Singh 2011). We had pre‐specified in the protocol to not report analyses that overlapped completely with this review.

As the networks of methotrexate‐naïve and methotrexate‐IR trials were analysed separately, the decision to include trials that provided only indirect comparisons between interventions of interest was specific to each network. For example, if studies of methotrexate vs. hydroxychloroquine and hydroxychloroquine vs. methotrexate + hydroxychloroquine were identified, they had to be in the same network (both methotrexate naïve or methotrexate‐IR) to provide an indirect comparison and be eligible for inclusion.

Types of studies

Randomised controlled trials (RCTs) or controlled clinical trials (CCTs). CCTs were defined as per the Cochrane Handbook, as trials where randomisation was not truly random (i.e. quasi‐randomised), or trials where double blinding was used but randomisation was not mentioned (Higgins 2011).

Types of participants

Adults (age > 18 years) with RA, according to 1958, 1987 or 2010 classification criteria (Aletaha 2010; Arnett 1988; Ropes 1958).

Types of interventions

The interventions considered in this review were:

Oral methotrexate monotherapy
Parenteral methotrexate monotherapy (subcutaneous or intra‐muscular)
Methotrexate combined with conventional synthetic DMARDs. Conventional synthetic DMARDs were limited to: anti‐malarials (hydroxychloroquine/chloroquine), sulfasalazine, leflunomide, cyclosporine, intra‐muscular gold and azathioprine.
Methotrexate combined with biologic DMARDs, including anti‐TNF inhibitors (adalimumab, certolizumab, etanercept, golimumab, infliximab), abatacept, rituximab, and tocilizumab.
Methotrexate combined with tofacitinib

No dose restriction was applied to conventional synthetic DMARDs, given the variability of dosing in clinical practice. Biologic DMARDs and tofacitinib were limited to currently recommended doses or dose equivalents, specifically:

adalimumab 40 mg subcutaneously every two weeks;
certolizumab 200 mg subcutaneously every two weeks after initial dosing of 400 mg subcutaneously at baseline, two, and four weeks;
etanercept 50 mg subcutaneously every week or 25 mg subcutaneously twice weekly;
golimumab 50 mg subcutaneously every four weeks;
golimumab 2 mg/kg IV every 8 weeks after initial dosing at baseline and 4 weeks;
infliximab 3 mg/kg intravenously every eight weeks after initial dosing at baseline, two and six weeks;
abatacept every four weeks intravenously at ˜10 mg/kg (500 mg in patients < 60 kg, 750 mg in patients 60 kg to 100 kg and 1000 mg in patients > 100 kg), after the initial dosing regimen at baseline, two and four weeks;
abatacept 125 mg subcutaneously with or without an intravenous loading dose of ˜ 10 mg/kg
rituximab, two 1000 mg IV doses two weeks apart;
tocilizumab every four weeks intravenously at 4 mg/kg or 8 mg/kg. The two doses of intravenous tocilizumab were analysed separately, as the approved dosing varies by country (Furst 2013).
tocilizumab 162 mg subcutaneously every week
tofacitinib 5 mg orally twice daily

We excluded trials that evaluated the effect of corticosteroids as an intervention, as this was not the objective of the review. We included trials that required or allowed corticosteroids as part of the treatment arm if the corticosteroids were applied equally across arms.

Types of outcome measures

Major outcomes

Benefits (efficacy)

American College of Rheumatology (ACR)‐50 response (Felson 1995)
Radiographic progression as a continuous variable, measured by Larsen, Sharp or modified Larsen/Sharp scores including Scott‐modified Larsen, Genant‐modified Sharp and van der Heijde‐modified Sharp (Ory 2003)

Harms (toxicity)

Withdrawals due to adverse events, including death.

Minor outcomes

Benefits (efficacy)

ACR‐20 and ACR‐70 responses (Felson 1995).
Disease activity score (DAS) (van der Heijde 1992) or DAS28 (Prevoo 1995). DAS28 in its original form includes 4 variables: ESR, patient global assessment of disease activity, tender and swollen joint counts (DAS28‐4‐ESR). It has been modified to include just 3 variables (excluding global assessment) and with the CRP instead of ESR. We extracted the original DAS28‐4‐ESR if it was reported in multiple formats.
DAS28 remission, defined as a DAS28 <2.6.
European League Against Rheumatism (EULAR) response (moderate or good) (van Gestel 1996)
Radiographic non‐progression, as defined by the study's definition of radiographic progression. If multiple definitions were used in a study, we extracted the result that used the highest threshold for defining progression. We therefore used values in the following order of preference: minimal clinically important difference > smallest detectable change > any progression.
Swollen joint count.
Withdrawals due to inefficacy.
Pain (Visual Analogue Scale).
Functional limitation, as measured by the Health Assessment Questionnaire‐Disability Index (HAQ‐DI) (Fries 1982) or modified HAQ (mHAQ) (Pincus 2005).
Fatigue, as defined by the study.

Harms (toxicity)

Serious adverse events (SAE), as defined by the study.
Serious infections, as defined by the study.
Gastrointestinal (GI) side effects, excluding liver and oral toxicity (e.g. – aphthous ulcers).
Elevated transaminases (ALT or AST). If multiple definitions were provided we used the lowest threshold for an abnormal value.
Hematological toxicity (low haemoglobin, leucopenia/neutropenia or thrombocytopenia). If multiple definitions were provided we used the lowest threshold for an abnormal value.

Combined (efficacy and toxicity)

Combined withdrawals due to inefficacy or adverse events

The three major outcomes were selected to encompass key treatment benefits (improvement in disease activity (ACR‐50), and inhibition of radiographic progression) and harms (withdrawals due to adverse events). Minor efficacy outcomes included other composite measures of disease activity and selected patient‐reported outcomes (pain, functional limitation and fatigue). Toxicity outcomes were limited to selected outcomes that can be evaluated within the context of a randomised trial (i.e. those that occur with sufficient frequency over short‐term follow‐up) and which were hypothesised to have clinically important differences between the interventions of interest. A low threshold for abnormal lab values was chosen to increase the ability to detect a safety signal. All outcomes were prespecified in the protocol.

Search methods for identification of reviews

An electronic database search was performed in MEDLINE (including in process and non‐indexed citations), EMBASE (including EMBASE classic) and Cochrane Central Register of Controlled Trials (CENTRAL) from inception to January 19, 2016. As each intervention contained methotrexate, the search strategies contained subject headings and keywords for "rheumatoid arthritis", "methotrexate" and "randomised controlled trial" (Appendix 1; Appendix 2; Appendix 3). The database search strategies were adapted from a previously published Cochrane review (Katchamart 2010). We also searched the trial registries ClinicalTrials.gov (http://clinicaltrials.gov/) and the International Clinical Trials Registry Platform (ICTRP) (http://apps.who.int/trialsearch/) using the search term "rheumatoid arthritis AND methotrexate". Finally, we performed hand‐searches for abstracts from American College of Rheumatology (ACR) and European League Against Rheumatism (EULAR) conferences (2009‐2015) and reviewed all existing Cochrane reviews to ensure no relevant trials were missed. All languages were included.

Data collection and analysis

Selection of reviews

Two review authors (GH, ChB) independently screened articles for inclusion by title or abstract and full‐text if necessary. Disagreements were resolved by consensus and if not possible, by discussion with a third review author (ClB).

Data extraction and management

Three review authors working in pairs (GH, ChB, DD) abstracted relevant data from included studies on an Excel spreadsheet. A detailed data extraction template was developed and piloted on 5 articles. Changes were made to address inconsistencies and the 5 articles were then re‐extracted. Trial characteristics and baseline patient characteristics were extracted by one author (GH) and confirmed by a second (ChB or DD); outcome data were extracted independently, with disagreements resolved through discussion.

Dichotomous efficacy outcomes were abstracted as the number of patients with an event and the total number of patients in each arm, on an intention‐to‐treat basis as the number of patients randomised to the arm who received at least one dose of medication or, if this was not reported, the total number of patients randomised to the arm. Continuous efficacy outcomes were abstracted as the mean, standard deviation, and the total number of patients for the final values and change in values from baseline. Toxicity outcomes were extracted as the number of events and treatment exposure (person‐years) in each arm. If the total number of events was not reported, we used the total number of patients with at least 1 event, which should be a close approximation for uncommon events.

Assessment of methodological quality of included reviews

The methodological quality of included trials was assessed using the Cochrane Collaboration's tool for assessing risk of bias (Higgins 2011). Studies were graded as having a "low risk", "high risk" or "unclear risk" of bias across the seven specified domains: sequence generation, allocation concealment, blinding of participants and personnel, blinding of outcome assessment, incomplete outcome data, selective outcome reporting and other sources of bias (Higgins 2011). Other sources of bias included baseline imbalances in co‐interventions (particularly corticosteroids) and other biases identified during the review (not pre‐specified). Outcomes were divided into three categories (radiographic outcomes, withdrawals, other clinical outcomes) as the risk of bias could differ across several domains, as discussed below. For each of the three outcome categories we also judged an overall risk of bias. The overall ROB rating included a judgment across all domains for that outcome, although not all domains were equally weighted. The main domains that affected the overall ROB were blinding and incomplete outcome data.

The domains “blinding of participants and personnel” and “blinding of outcome assessment” were assessed separately for radiographic and non‐radiographic outcomes (which included withdrawals and all other clinical outcomes). Blinding of the outcome assessor occurs separately for radiographic versus other outcomes, and the impact of not blinding participants on the risk of bias for radiographic outcomes was felt to be “unclear”, whereas it was judged to be “high” for all other outcomes. The domain “incomplete outcome data” was assessed separately for each of the 3 outcome categories. The proportion of missing data, the balance of missing data between treatment arms and the methods of handling missing data may vary between clinical and radiographic outcomes. The outcome ‘withdrawals’ was judged at low risk of bias even if the withdrawal rate from the trial was high and/or imbalanced between arms. An exception was early‐escape trials, where high or imbalanced rates of early escape resulted in a higher risk of bias.

To improve the consistency between raters we developed a template for the risk of bias (ROB) assessment, based on Cochrane guidance, but with specific context‐specific clarifications and examples. For trials that were previously included in Cochrane reviews and graded using the Cochrane risk of bias, one review author (GH) reassessed the risk of bias to ensure the results agree with those published. If there was a discrepancy in the ratings or if the prior Cochrane review did not distinguish outcome categories as we did, the ROB was assessed independently by a second reviewer (ChB or DD). For studies not included in existing Cochrane reviews, two review authors (GH, and ChB or DD) independently assessed the risk of bias. Any disagreements were resolved by consensus and if not possible, by discussion with a third review author (ChB or DD).

Data synthesis

Data analysis

Random‐effects Bayesian network meta‐analyses were fitted for each outcome measure. The models used account for the correlation in multi‐arm trials, and have been previously published (Ades 2006; Dias 2013). Random‐effects models allow the treatment heterogeneity we expect, given the clinical heterogeneity amongst the trials. Uninformative prior probability distributions were used for all parameters. Markov chain Monte Carlo sampling was used to obtain samples from posterior distributions, with 10,000 burn‐in iterations followed by 10,000 monitoring iterations. Convergence was assessed by running three chains, inspecting the sampling history plots and calculating Gelman–Rubin–Brooks (GBR) statistics (Brooks 1998). Model fit was assessed using residual deviance and Deviance Information Criterion (DIC). All data analyses were performed using R statistical software version 3.1.2 (www.r‐project.org) with rjags package version 3‐14 running Just Another Gibbs Sampler (JAGS) version 3.4.0.

Measures of treatment effect

For the primary analyses, trials with a high risk of bias for that outcome category were excluded. Treatment effects on dichotomous outcomes were evaluated using odds ratios (ORs). For continuous outcomes that were measured on the same or very similar scale (DAS, DAS28, HAQ‐DI/mHAQ, pain VAS), treatment effects were evaluated as mean differences, with final values or change in values from baseline; we used final values if both final and change values were reported. For continuous outcomes measured on different scales (radiographic progression, swollen joint count, fatigue) treatment effects were evaluated using standardised mean differences (SMD). For SMD, we performed separate analyses for change and final values, as it is not recommended to combine the two in the same analysis (Higgins 2011). SMD were estimated by dividing the modeled change or final value in each arm by the within‐trial pooled standard deviation of the change or final value. We summarized withdrawals due to adverse events and other minor toxicity outcomes as rate ratios to allow for differences in exposure between arms in early escape and crossover trials.

For all outcomes we reported the posterior median (point estimate) and 95% credible interval for all treatments effects relative to oral methotrexate. For the major outcomes we reported the effects and probability of superiority for all pair‐wise comparisons. We considered an effect to be ‘significant’ if the 95% credible interval (CrI) excluded the null effect. This equates to a 97.5% probability of superiority (one‐tailed) of one treatment over another.

Dealing with missing data

For all trials, we sought data from clinical trial registries (www.clinicaltrials.gov, http://www.who.int/ictrp/en/), US Food and Drug Administration (FDA), European Medicines Agency (EMEA) and drug manufacturer web sites. For dichotomous outcomes, if only percentages were reported, the actual number of events was calculated from the percentage and total number of patients and rounded to the nearest whole number. For continuous measures, if standard deviations (SD) were not available, they were calculated from standard errors (se), 95% confidence intervals, or exact p‐values from t‐tests, using formulas published in the Cochrane Handbook (Higgins 2011). If only medians and inter‐quartile ranges (IQR) were presented, these were extracted, with the SD calculated as IQR divided by 1.34, which assumes a normal distribution of outcomes. If no variance data was available, we used the baseline standard deviation as the final standard deviation, if available. For toxicity outcomes, if the drug exposure was not available, it was calculated. Withdrawals were assumed to occur at a constant rate, unless specific information was available to permit a more accurate calculation.

All data conversions above were performed using R statistical software version 3.1.2 (www.r‐project.org). Automated functions for each calculation (e.g.‐ converting an exact p‐value to a standard deviation) were developed, validated against examples provided in the Cochrane Handbook (Higgins 2011) and then used by two independent reviewers working in pairs (GH, ChB, DD).

If data were presented only in graphical format it was extracted digitally. Images in the highest resolution available were digitised and extracted using the software program GraphClick (version 3.0.2, Arizona Software). If another outcome or another time point on the same graph was also reported as a numerical value, it was used as an internal validation of the data extraction procedure. All graphical data were extracted by two independent reviewers (GH, ChB) and averaged or corrected if an obvious discrepancy was apparent.

Assessment of heterogeneity

Analyses were performed separately for trials of methotrexate‐naïve and methotrexate‐IR patients, as we judged patients in these two types of trials too clinically heterogeneous to pool. We assessed statistical heterogeneity by calculating the between‐study variance in the random‐effects model. For the major outcomes, we evaluated the consistency of the direct and indirect evidence through ‘node‐splitting’, which separates the direct and indirect evidence for a comparison where both direct and indirect evidence exist (Dias 2010). Direct effects were determined through a Bayesian fixed‐effects model. A fixed‐effect model was used as there were typically too few trials to estimate a between study variance. Node‐splitting for most comparisons was carried out using the R package gemtc (version 0.6); comparisons using standardised mean differences were not supported by the package, so were programmed directly. The results of the node‐splitting analyses were used to inform the GRADE quality assessments, as described below.

Meta‐regression and sensitivity analyses

For ACR50 response we explored clinical heterogeneity through meta‐regression for the following variables separately for methotrexate‐naïve and methotrexate‐IR populations:

Response rate to oral methotrexate (post‐hoc)
Disease duration
MTX dose ≥ 15 mg/week
Duration of trial
Baseline swollen joint count
Baseline HAQ‐DI
Year of publication of trial
Time‐point of assessment

The meta‐regression models included a covariate for the treatment effect relative to oral methotrexate. We assumed that this was the same for all treatments as there were too few trials to specify separate covariate for each comparison. For example, the effect of trial duration on the OR comparing methotrexate + etanercept to methotrexate was assumed to be the same as its effect on the OR comparing methotrexate + adalimumab to methotrexate. If a trial did not contain an arm with oral methotrexate, the trial was included in the analysis, but the treatment effect was not adjusted. For the meta‐regression analysis of the response rate to oral methotrexate, we used the modelled response rate as opposed to the actual response rate in each trial to limit the effect of regression to the mean (Sharp 1996). To determine if there was an association between prior methotrexate use and ACR50 response, we performed an additional meta‐regression analysis where all trials were included and the prior methotrexate status (methotrexate‐naïve versus methotrexate‐IR) was defined through a covariate.

Additional post‐hoc sensitivity analyses were performed. We fitted fixed‐effect models for the major outcome ‘radiographic progression’ as there were few trials available to estimate a random‐effect. To further evaluate the effect of differing time‐points of assessment, we performed analyses using 6‐month and 12‐month data separately, using interim data from a trial if available. We also performed an analysis where we used ‘pre‐rescue’ data instead of end‐trial data for all ‘rescue’ trials; if pre‐rescue data was not reported, the trial was excluded. For the methotrexate‐naïve analysis we performed an additional analysis where we excluded trials that included some patients with prior methotrexate exposure. For the methotrexate‐IR network we included a sensitivity analysis where we included SWEFOT (SWEFOT 2012). SWEFOT was a major trial comparing triple therapy to methotrexate + infliximab that was excluded from our main analysis for ACR50 response because it had a high risk of bias for clinical outcomes, resulting from its open label design and high withdrawal rate.

We also did sensitivity analyses around several modeling assumptions. In the protocol, we had planned to use odds ratios to pool withdrawals due to adverse events. However, we changed the analyses to rate ratios given the differences in exposure between arms in early escape and crossover trials. As a sensitivity analysis, we compared the rate ratios with odds ratios, in which we used the total exposure (in patient months) in each arm as the denominator, instead of the number of patients. The model estimates the effect on the monthly odds of an outcome, assuming independence between months, and should approximate the rate ratio from a Poisson model.

The choice of prior distribution for the between study variance may affect the estimated treatment effects, although this effect has been found to be small in analyses of ten or more studies (Lambert 2005). For the primary analysis, we followed published guidance and chose a prior that was vague but realistic (Lambert 2005). We then did sensitivity analyses using an additional uninformative prior and potentially informative priors of Turner et al (odds ratio for ACR50 response and rate ratio for withdrawals due to adverse events) (Turner 2012) and Rhodes et al (radiographic progression) (Rhodes 2015).

Presentation of key results ("Summary of findings" table)

The three major outcomes were presented in the Summary of Findings (SoF) tables, separately for the methotrexate‐naïve and methotrexate‐IR analyses. We converted the average treatment effect for each outcome into an absolute response by using an assumed (baseline) value for oral methotrexate. For all analyses, the assumed baseline value was the median from a bayesian random effects model of the oral methotrexate arms. For ACR50 response, we used all trials in the analysis to estimate the assumed probability of response. For radiographic progression, we calculated the assumed mean over one year on the Sharp‐van der Heijde scale (van der Heijde 2000) from the trials that reported this outcome for oral methotrexate. We then calculated the absolute effect for each treatment by using this assumed value and the mean differences for each treatment relative to oral methotrexate on the Sharp‐van der Heijde scale, which we calculated by multiplying the standardized mean differences by the pooled within arm standard deviation for studies that used the Sharp‐van der Heijde scale. For withdrawals due to adverse events, we estimated the assumed rate at one year from the available trials and converted it to an absolute probability by using the rate ratio, assuming that the time to withdrawal over one year was exponentially distributed for each person.

We used recently published GRADE (Grading of Recommendations Assessment, Development and Evaluation) guidance for assessing the quality of evidence from a network meta‐analysis (Puhan 2014). First, the quality of evidence for each direct comparison was evaluated using the GRADE domains of study limitations, inconsistency, imprecision, indirectness and publication bias (Guyatt 2008). Second, the quality of the indirect evidence (determined through ‘node‐splitting’ (Dias 2010)) was evaluated by considering the precision of the indirect estimate, the quality of the direct comparisons that formed the indirect evidence and the likelihood of ‘intransitivity’. Intransitivity exists if there is heterogeneity in the trials that form the different comparisons within the indirect evidence. The indirect evidence can be complex. A ‘first‐order’ indirect comparison is formed by 2 trials that share a common treatment arm. A ‘second‐order’ comparison will have 2 intermediary treatments (i.e.‐ the indirect comparison of treatments A and D in trials of A vs B, B vs C and C vs D). For the quality assessment we focused on first‐order comparisons, as recommended by GRADE (Puhan 2014). We then rated the quality of evidence for the network meta‐analysis, based on the quality ratings for the direct and indirect evidence.

Results

This review is also published as an abridged version that presents the results for the major outcomes (Hazlewood 2016). As each article underwent a separate peer‐review process, there are slight differences in the text of the two manuscripts, but the results in both are identical. The abridged version also has several open‐access appendices, which we refer to in this review when appropriate.

Search results

Our search identified 9817 unique records. After title/abstract and full‐text review, 412 potentially eligible records remained (Figure 1). 204 were further excluded because they could not be linked within any network; the comparators for these trials were most commonly experimental (not approved) DMARDs (Figure 1). After excluding the 11 trials in a TNF‐IR population that overlapped with a prior Cochrane review (Table 1), 197 records remained, representing 158 unique trials. Two records were pending classification and not included, as the full‐text article was not available.

1. Characteristics of excluded studies.

Study ID	Reason for exclusion
Kim 2012	RCT that compared ETN+methotrexate vs. ETN+DMARD; DMARD comparator not standardized
Kremer 2010	Wrong dose interval of IV golimumab (given Q 12 weeks)
De Stefano 2010	Unclear study design; compared methotrexate+TNF to LEF+TNF (n=120), but also appears to have randomised pts to 3 different TNF within each subgroup ; no response from e‐mail to corresponding author to confirm
Saunders 2008	Strategy trial that compared triple therapy to step‐up therapy (starting with SSZ). No pre‐switch data available
Taylor 2006	Wrong dose of IFX (5 mg/kg)
Keystone 2004	PBO only given for 8 weeks
Cohen 2001	Optional extension to 2 years of ULTRA
Drosos 2000	Extension study of Drosos 1998
Calguneri 1999	Compared triple therapy to methotrexate‐based 2‐drug therapy or monotherapy; 2‐drug combination therapy and monotherapy not standardized
Willkens 1996	Open‐label extension of Wilkens 1992 (Optional switch at 24 weeks to open‐label trial)
Usova 1993	Interim report of Sigidin Ya 1994
Peterfy 2011	Abstract with unclear methods; uncertain is PBO patients crossed over to RTX at 6 months. Only outcome available is SAE, which is only reported at study end.
Gao 2004	Abstract with no outcome of interest
Beals 2013	Abstract only with IFX dosing intervals not reported
Dougados 2014	52 week results of ACT‐RAY; ACT‐RAY required open‐label addition of DMARDs beyond 24 weeks in a treat‐to target approach; therefore considered a strategy trial beyond 24 weeks.
Burmester 2013	TNF‐inadequate response trial (all patients required to have failed anti‐TNF therapy)
Cohen 2006	TNF‐inadequate response trial (all patients required to have failed anti‐TNF therapy)
Combe 2013	TNF‐inadequate response trial (all patients required to have failed anti‐TNF therapy)
Emery 2008	TNF‐inadequate response trial (all patients required to have failed anti‐TNF therapy)
Furst 2007	TNF‐inadequate response trial (all patients required to have failed anti‐TNF therapy)
Keystone 2008	TNF‐inadequate response trial (all patients required to have failed anti‐TNF therapy)
Kume 2012	TNF‐inadequate response trial (all patients required to have failed anti‐TNF therapy)
Strand 2012	TNF‐inadequate response trial (all patients required to have failed anti‐TNF therapy)
NCT01283971 2014	TNF‐inadequate response trial (all patients required to have failed anti‐TNF therapy)
Strand 2015	TNF‐inadequate response trial (all patients required to have failed anti‐TNF therapy)
Manders 2015	TNF‐inadequate response trial (all patients required to have failed anti‐TNF therapy)

Open in a new tab

Description of included reviews

Overall trial characteristics

The 158 trials included over 37,000 patients across the arms included in this review (Table 2, with full study details available in Web appendix 2 of abridged review, Hazlewood 2016). Seventeen articles were available only as an abstract, although additional data were available through www.clinicaltrials.gov for 9 of these. An additional trial was available only as a trial register with results available through www.clinicaltrials.gov. Ten articles were in languages other than English.

2. Summary of trial characteristics.

Medication	Studies (N)	Patients (n)	Year published, median (range)	Early escape design, % of studies (n pts)	*Trial duration, wks, median (range)**	methotrexatedose >15mg/wk, % of studies (n pts)**	Disease duration, yrs, median (range)	Swollen joint count, median (range)	Low, % (n pts)	Unclear, % (n pts)	High, % (n pts)
MTX + biologic DMARDs/tofacitinib
MTX + etanercept	10	2448	2007 (1999‐2014)	0%	38 (12‐52)	50% (n=1833)	2 (0.5‐13)	13.9 (11‐24)	10% (n=424)	70% (n=1965)	20% (n=59)
MTX + infliximab	13	2806	2006 (2000‐14)	8% (n=264)	26 (13‐54)	46% (n=1990)	7.6 (0.4‐10.5)	15 (5‐21.5)	31% (n=898)	54% (n=1824)	15% (n=84)
MTX + adalimumab	16	4465	2010 (2003‐15)	38% (n=1936)	24 (12‐104)	50% (n=1809)	2.5 (0.1‐11.7)	16.3 (8.7‐22.5)	25% (n=2142)	69% (n=2258)	6% (n=65)
MTX + rituximab	4	1262	2008 (2004‐12)	25% (n=342)	24 (24‐104)	50% (n=683)	8.6 (0.9‐11.5)	20.9 (20.2‐21.6)	25% (n=80)	75% (n=1182)	0%
MTX + abatacept	10	3612	2012 (2005‐15)	0%	25 (17‐52)	60% (n=3014)	6.4 (0.5‐9.3)	17.1 (10‐22.4)	60% (n=2496)	40% (n=1116)	0%
MTX + tocilizumab	10	4859	2012 (2006‐16)	50% (n=3671)	24 (16‐52)	60% (n=2729)	6.6 (0.4‐9.2)	13.7 (6.4‐20.1)	10% (n=553)	60% (n=2765)	30% (n=1541)
MTX + certolizumab	7	2680	2012 (2008‐15)	71% (n=1561)	24 (24‐52)	29% (n=1119)	6 (0.3‐9.6)	21 (16.4‐22.5)	0%	29% (n=1192)	71% (n=1488)
MTX + golimumab	6	1640	2012 (2008‐13)	83% (n=1570)	24 (16‐52)	50% (n=1132)	6.9 (3.2‐8.7)	13.5 (11.6‐15.4)	0%	100% (n=1640)	0%
MTX + tofacitinib	4	749	2012 (2011‐15)	50% (n=621)	24 (12‐52)	50% (n=621)	8.7 (0.7‐9.1)	14.7 (14.1‐14.9)	0%	75% (n=268)	25% (n=481)
SUBTOTAL	80	24 521	2011 (1999‐2016)	31% (n=9965)	24 (12‐104)	50 (n=14930)	6.3 (0.1‐13)	16.4 (5‐24)	21% (n=6593)	61% (n=14210)	18% (n=3718)
Comparative effectiveness trials
Head to head biologic DMARDs/ tofacitinib	4	1658	2010 (2006‐2014)	25% (n=501)	27 (26‐104)	50% (n=1077)	7.8 (1.8‐11.3)	16.6 (15.9‐20.6)	25% (n=431)	50% (n=1147)	25% (n=80)
MTX + biologic DMARDs vs. MTX + conventional synthetic DMARDs	4	1382	2012 (2012‐2013)	0%	63 (16‐104)	25% (n=353)	0.5 (0.3‐5.2)	12 (11.2‐12.8)	50% (n=1108)	0%	50% (n=274)
SUBTOTAL	8	3040	2012 (2006‐2014)	12% (n=501)	27 (16‐104)	38% (n=1430)	5.2 (0.3‐11.3)	15.9 (11.2‐20.6)	38% (n=1539)	25% (n=1147)	38% (n=354)
MTX + conventional synthetic DMARDs
MTX + azathioprine	1	209	1992	0%	24	0%	8.6	17.3	0%	100% (n=209)	0%
MTX + hydroxychloroquine/ chloroquine	7	452	2005 (1993‐2012)	0%	26 (12‐52)	0%	4 (0.3‐7.7)	10.7 (8.2‐27.3)	14% (n=82)	14% (n=40)	71% (n=330)
MTX + sulfasalazine	6	639	2000 (1994‐2007)	0%	52 (24‐76)	0%	0.4 (0.2‐5)	16.7 (9.8‐22.6)	0%	67% (n=515)	33% (n=124)
MTX + cyclosporine	9	1100	2003 (1995‐2008)	11% (n=120)	48 (16‐104)	22% (n=64)	1.1 (0.2‐10.3)	13.6 (11‐19)	0%	89% (n=1076)	11% (n=24)
MTX + sulfasalazine + hydroxychloroquine	4	503	2005 (1996‐2013)	0%	91 (13‐104)	0%	4.4 (0.5‐8.6)	17.1 (9.5‐29.8)	0%	75% (n=463)	25% (n=40)
MTX + leflunomide	3	921	2006 (2002‐15)	0%	24 (16‐36)	67% (n=455)	6.2 (0.7‐11.6)	14.3 (10.7‐18)	0%	33% (n=263)	67% (n=658)
MTX + IM gold	1	65	2005	0%	48	100% (n=65)	3.2	11	0%	100% (n=65)	0%
SUBTOTAL	31	3889	2003 (1992‐2015)	3% (n=120)	48 (12‐104)	16% (n=584)	1.1 (0.2‐11.6)	13.6 (8.2‐29.8)	3% (n=82)	61% (n=2631)	35% (n=1176)
MTX vs conventional synthetic DMARD monotherapy
Placebo	5	324	1985 (1985‐93)	20% (n=52)	14 (12‐18)	0%	9.5 (4.8‐14)	27.5 (24‐30.9)	0%	100% (n=324)	0%
Azathioprine	5	257	1991 (1987‐94)	0%	24 (24‐52)	0%	12 (8.7‐13.9)	14.6 (9.5‐21.9)	0%	40% (n=106)	60% (n=151)
IM gold	5	489	1991 (1988‐2001)	0%	48 (26‐52)	20% (n=99)	4 (1.2‐6)	14 (13.9‐15.2)	0%	60% (n=249)	40% (n=240)
sulfasalazine	2	211	1998 (1995‐2002)	0%	24	0%	4 (1.4‐6.6)	9.3	0%	50% (n=126)	50% (n=85)
Cyclosporine	2	203	1999 (1998‐2000)	0%	65 (26‐104)	50% (n=100)	3.8 (2.2‐5.5)	13.1 (12.2‐14)	0%	0%	100% (n=203)
Leflunomide	16	3258	2002 (1999‐2014)	12% (n=567)	24 (12‐52)	25% (n=927)	3.9 (0.5‐6.8)	11.8 (8.2‐16.5)	0%	50% (n=1598)	50% (n=1660)
Hydroxychloroquine	2	409	2006 (2000‐12)	0%	24	0%	1.5 (1‐2.1)	NA	0%	0%	100% (n=409)
sc vs. oral MTX	2	467	2009 (2008‐10)	50% (n=383)	24	100% (n=467)	0.2	15	0%	50% (n=383)	50% (n=84)
SUBTOTAL	39	5618	2000 (1985‐2014)	10% (n=1002)	24 (12‐104)	21% (n=1593)	4.5 (0.2‐14)	14 (8.2‐30.9)	0%	51% (n=2786)	49% (n=2832)
TOTAL	158	37 068	2008 (1985‐2016)	20% (n=11 588)	24 (12‐104)	35% (n=18537)	4.8 (0.1‐14)	15.1 (5‐30.9)	13% (n=8214)	57% (n=20774)	30% (n=8080)

Open in a new tab

Abbreviations: DMARD, disease‐modifying anti‐rheumatic drug; IM, intra‐muscular; methotrexate, methotrexate; pts, patients; sc, subcutaneous; SJC, swollen joint count

Trials are grouped by comparator and sorted by the year of the first trial for within each class, for illustrative purposes. Patient characteristics, including the number of patients, only include the arms considered in the review. *Trial duration for efficacy outcomes; some studies had longer follow‐up for safety outcomes **Studies where the dose of methotrexate could be confirmed as >= 15 mg/wk. In some studies, methotrexate was dosed across a range of values that included 15 mg/wk but the actual dose was not provided. See Web appendix 2 of Hazlewood 2016 for further details.

Of the 158 trials, 80 (51%) compared methotrexate + biologic DMARDs or tofacitinib to methotrexate monotherapy or made comparisons of different dosing formulations (subcutaneous or intravenous) of the same biologic DMARD (Table 2). There were only eight ‘comparative effectiveness’ trials, with four providing head‐to‐head comparisons of different biologic DMARDs/tofacitinib and four comparing methotrexate + biologic DMARDs to methotrexate + conventional synthetic DMARD therapy (methotrexate + sulfasalazine + hydroxychloroquine in three of the four trials). Eleven trials were a strategy or crossover design (Web appendix 2 of Hazlewood 2016). An early escape design was more common in trials of methotrexate + biologic DMARDs/tofacitinib with no active comparator (31%) than comparative effectiveness trials of methotrexate + biologic DMARDs/ tofacitinib (12%) or trials of conventional synthetic DMARD combination or monotherapy (3% and 10%) (Table 2). Trials of the four biologic DMARDs/ tofacitinib with the most recent year of first publication (certolizumab, golimumab, tocilizumab, tofacitinib) had high rates of early‐escape design, at ≥ 50% of trials (Table 2).

The trials ranged in duration from 12 to 104 weeks and had similar median disease duration across the medication classes (Table 2). The median baseline swollen joint count across the trials was high at 15, with a similar distribution across medication classes. Methotrexate dosing varied across studies and was variably reported. The dose of methotrexate could be confirmed as >= 15 mg/week in 50% of biologic DMARDs/ tofacitinib trials and only 16% and 21% in trials evaluating methotrexate + conventional synthetic DMARD combination therapy or conventional synthetic DMARD monotherapy. The risk of bias for the main outcome category (non‐radiographic outcomes excluding withdrawals) varied across trials and medication classes (Table 2) and is discussed further in the section “methodological quality of included studies.”

Two trials were performed on an early RA population and included some patients who did not meet established criteria for RA (EMPIRE 2014; tREACH 2013). Both trials had high percentages of patients meeting 2010 criteria (88%) (tREACH 2013) and 94% (EMPIRE 2014) and presented subgroup data for patients meeting 2010 criteria for their main outcomes. We included these trials, using data for the subgroup meeting 2010 criteria, if available.

Characteristics of trial networks

The connections between the trials for the methotrexate‐naïve and methotrexate‐IR populations are presented in Figure 2. The network diagrams include all trials; the actual number of trials for each analysis varied. Two trials had data available for both networks. One trial had subgroup available for methotrexate‐naïve and methotrexate‐IR patients for ACR responses (O'Dell 2002). The other trial was a four‐arm trial (TEAR 2012). Two arms of this trial compared methotrexate + sulfasalazine + hydroxychloroquine versus methotrexate + etanercept in patients’ naïve to methotrexate and were included in the methotrexate‐naïve analysis. The other two arms compared a strategy of adding sulfasalazine + hydroxychloroquine vs. etanercept to methotrexate in patients with an inadequate response to methotrexate and were included in the methotrexate‐IR network.

Networks of included studies for methotrexate‐naïve (A) and methotrexate‐inadequate response populations (B). Each line represents a direct comparison between two treatments from one or more trials. The line thickness is directly proportional to the total sample size of all trials for that comparison (line length has no meaning). Biologic/targeted synthetic DMARDs are shown on the left of each network and conventional synthetic DMARDs on the right. Treatments on the innermost circle (green hashed line) are treatments of interest, whereas treatments on the outermost circle (red hashed line) are other treatments that form links between treatments of interest. Comparisons to methotrexate are shown in blue. Two trials were included in both analyses.

*Abbreviations*: ABAT, abatacept; ADA, adalimumab; AZA, azathioprine; CTZ, certolizumab; CQ, chloroquine; CyA, cyclosporine; ETN, etanercept; GOL, golimumab; HCQ, hydroxychloroquine; IFX, infliximab; IM, intra‐muscular; IV, intravenous; LEF, leflunomide; , methotrexate; RTX, rituximab; sc, subcutaneous; SSZ, sulphasalazine; TCZ, tocilizumab; TOFA, tofacitinib

Methotrexate‐naïve network

The methotrexate‐naïve analysis included 93 trials with over 19000 patients (Figure 2). Most comparisons of methotrexate+ biologic DMARDs were to methotrexate, with no head‐to‐head comparisons between different biologic DMARDs. Trials evaluating conventional synthetic DMARD therapy were generally smaller than biologic trials, but were more inter‐connected. Only one trial connected biologic DMARDs to combination therapy with methotrexate + conventional synthetic DMARDs (TEAR 2012). Ten treatments contributed only indirect evidence to comparisons between the treatments of interest (shown in the outer circle in Figure 2). Eight of the 93 trials (9%) included patients with some prior exposure to methotrexate, and the remainder were methotrexate‐naïve (Web appendix 2 of Hazlewood 2016). Of these eight trials, the percentage of patients at baseline with prior methotrexate use at baseline ranged from 1% (Jaimes‐Hernandez 2012) to 64% (Takeuchi 2013b).

Methotrexate‐inadequate response network

The methotrexate‐IR analysis included 67 trials with over 17patients (Figure 2). As compared to the methotrexate‐naïve network, the connections between conventional synthetic DMARDs were fewer and smaller in size. Connections between methotrexate + biologic DMARDs and methotrexate, in contrast, were large in size. The four head‐to‐head trials of biologic therapy formed links between several biologic therapies and all four trials comparing methotrexate+ biologic therapy to methotrexate + conventional synthetic DMARDs were included in this network.

Methodological quality of included reviews

The risk of bias of the trials varied considerably across trials and across each domain (Figure 3, with details for each study available as an appendix on Cochrane Musculoskeletal web site). Random sequence generation and allocation concealment were rarely rated as high risk of bias, although many studies were rated as ‘unclear’ for failing to provide details beyond “randomised”. The risk of bias for ‘blinding of participants’ and ‘blinding of outcome assessment’ was higher for non‐radiographic outcomes than for radiographic outcomes. For non‐radiographic outcomes, 28% and 25% of trials were rated as high risk of bias for blinding of participants and blinding of outcome assessment, respectively.

The risk of bias for the domain ‘incomplete outcome data’ varied across the outcome categories. The outcome ‘withdrawals’ had a low or unclear risk of bias for most studies; studies rated as ‘unclear’ were typically early escape studies. In contrast, the risk of bias for incomplete outcome data was high in 50% and 34% of trials for radiographic and non‐radiographic outcomes respectively. The risk of bias for the domain ‘selective outcome reporting’ was generally unclear, as most studies did not have a protocol available. The risk of bias in the domain ‘other bias’ was generally low.

The overall risk of bias was high in 21%, 17% and 30% of trials for radiographic outcomes, withdrawals and non‐radiographic outcomes respectively. Some important differences in the risk of bias across medication classes were observed. Trials assessing biologic DMARDs or tofacitinib had the lowest percentage of trials with a high risk of bias (19%), although exceptions existed. Five of the 7 certolizumab trials were rated as high risk of bias for the main outcome category (non‐radiographic outcomes, excluding withdrawals), due to high rates of withdrawal and/or early escape, often with imbalance between treatment arms.

Two of the four trials that compared methotrexate + biologic DMARDs to methotrexate + conventional synthetic DMARDs were rated as high risk of bias for non‐radiographic clinical outcomes (excluding withdrawals). One was a small open‐label study published only as an abstract (Joo 2012). The other was a larger open‐label trial with relatively high rates of incomplete outcome data (SWEFOT 2012).