Article Text

Download PDFPDF

Etanercept therapy in rheumatoid arthritis and the risk of malignancies: a systematic review and individual patient data meta-analysis of randomised controlled trials
  1. T Bongartz1,
  2. F C Warren2,
  3. D Mines3,
  4. E L Matteson1,
  5. K R Abrams2,
  6. A J Sutton2
  1. 1
    Division of Rheumatology and Department of Internal Medicine, Mayo Clinic College of Medicine, Rochester, Minnesota, USA
  2. 2
    Department of Health Sciences, University of Leicester, Leicester, UK
  3. 3
    Global Safety Surveillance and Epidemiology, Wyeth Research, Collegeville, Pennsylvania, USA
  1. Dr T Bongartz, Division of Rheumatology, Mayo Clinic, 200 1st Street SW, Rochester, MN 55905, USA; bongartz.tim{at}


Purpose: Tumour necrosis factor (TNF) plays an important role in inflammation and may affect tumour growth control. To assess the risk of malignancy with etanercept, a fusion protein that inhibits TNF action, a meta-analysis was performed using individual patient data from randomised controlled trials (RCT) in patients with rheumatoid arthritis (RA).

Methods: A search was conducted of bibliographic databases, abstracts from annual meetings and any unpublished studies on file with manufacturers of etanercept to December 2006. Only RCT of etanercept used for 12 weeks or more in patients with RA were included. Nine trials met the inclusion criteria. To adjudicate endpoints, the case narratives of potential cases were reviewed. Patient-level data were extracted from the clinical trials databases.

Results: The nine trials included 3316 patients, 2244 who received etanercept (contributing 2484 person-years of follow-up) and 1072 who received control therapy (1051 person-years). Malignancies were diagnosed in 26 patients in the etanercept group (incidence rate (IR) 10.47/1000 person-years) and seven patients in the control group (IR 6.66/1000 person-years). A Cox’s proportional hazards, fixed-effect model stratified by trial yielded a hazard ratio of 1.84 (95% CI 0.79 to 4.28) for the etanercept group compared with the control group.

Conclusions: In this analysis, the point estimate of malignancy risk was higher in etanercept-treated patients, although the results were not statistically significant. The approach of obtaining individual patient data of RCT in cooperation with trial sponsors allowed important insights into the methodological advantages and challenges of sparse adverse event data meta-analysis.

View Full Text

Statistics from

The question of whether the inhibition of tumour necrosis factor (TNF) alpha may increase the risk of malignant disease is still a matter of controversy.1 Our previous aggregate data meta-analysis of randomised controlled trials (RCT) using anti-TNF antibodies for the treatment of patients with rheumatoid arthritis (RA) showed a significantly increased risk of malignancy in anti-TNF antibody-treated patients compared with control patients.2 3 Etanercept, a fusion protein that is able to bind TNF and is also used in the treatment of RA, was deliberately excluded from this analysis due to differences in its molecular structure and mechanism of action within the anti-TNF class.4 Etanercept is an anti-TNF receptor fusion protein with unique properties that distinguish it from the anti-TNF antibodies infliximab and adalimumab. In contrast to anti-TNF antibodies, etanercept also neutralises lymphotoxin alpha, which has been associated with tumour growth control independent of TNF activity.5 6 The observation that etanercept is not beneficial in Crohn’s disease7 while anti-TNF antibodies are,8 9 suggests distinct biological properties of the two classes of anti-TNF treatment.

The potential for assessing the safety of etanercept based on single RCT in RA is limited. These trials are valuable tools to assess a drug’s efficacy but are limited in their assessment of safety. The sample size chosen on the basis of expected efficacy is usually insufficient to detect potential differences in sparse adverse events between treatment arms. Although observational studies offer a valuable approach to assess the risks of approved drugs, widespread use of a drug after approval for a significant amount of time is required to generate data that can be used for analysis. In addition, selection bias may be a limitation to safety assessments based on observational data.10 Meta-analyses of RCT, in contrast, may reveal important safety signals early and mitigate the effect of selection bias.11 12

We sought to explore further a potential association between anti-TNF therapy and malignancies by performing a systematic review and individual patient data (IPD) meta-analysis of RCT using etanercept in patients with RA.


This study was performed according to a protocol that prespecified study selection, eligibility criteria, data extraction and statistical analysis. The methodology was developed according to Cochrane collaboration guidelines ( and the manuscript was prepared in accordance with the QUOROM13 statement.

Search strategy

Our search strategy was divided into two major steps: the first step included an electronic database review performed by a librarian who was blinded to the study hypothesis. EMBASE, Medline, Cochrane Library and Web of Science were searched from inception to December 2006 for RCT of etanercept in patients with RA, using the keyword terms: “arthritis rheumatoid”; “etanercept”; “Enbrel”; “tumour necrosis factor fusion protein”; “randomised controlled trial”; “random allocation”; “clinical trials phase II”; “clinical trials phase III”; “clinical trials phase IV”.

The second step encompassed direct communication with the manufacturers of etanercept, Wyeth and Amgen, in order to locate unpublished trials and/or published trials that were missed with the electronic database search.

Trial selection

Trials were included in our analysis if they met the following criteria: study participants were diagnosed with RA according to American College of Rheumatology criteria,14 patients were randomly assigned to etanercept or control treatment and the study duration was at least 12 weeks. Assessment of eligibility criteria was performed independently by two investigators. Abstracts of all citations retrieved through the electronic database search were reviewed and potential candidates were further evaluated based on final publications. In addition, sponsors were contacted to obtain original trial protocols for in-depth review.

Study quality assessment

All included trials were reviewed for methodological features most relevant to issues of bias. Two independent reviewers assessed randomisation, random allocation concealment, masking of allocation, intent-to-treat analysis, completeness of follow-up, outcome assessment and attrition using original study protocols provided by sponsors as well as published reports. Disagreements were resolved by consensus-forming discussions.

Data extraction

The primary outcome of our analysis was (first) incident cancer, defined as a disease characterised by abnormal cells that divide without control and have the ability to invade other tissues. This definition did not include carcinoma in situ. Three investigators independently adjudicated potential malignancies based on a review of adverse event case narratives from which information about treatment assignment had been removed. Disagreements were resolved by consensus-forming discussions.

After the assessment of case narratives was completed, study sponsors provided data for every patient who participated in trials selected for the meta-analysis: demographic information; treatment assignment; dose of study drug; date of first and last dose of study drug; time point and reason of premature study discontinuation; date of last follow-up and concomitant disease-modifying antirheumatic drug therapy.

Data synthesis

All patients from eligible trials who were randomly assigned and received at least one dose of the study drug were included in the analysis (one patient who was lost to follow-up on the day of the first dose was excluded). The risk window for incident malignancies began with the date of the first dosing of study drug to the date of last follow-up in the respective RCT. A survival analysis of time-to-first-event using a Cox’s proportional hazards model stratified by trial and assuming a fixed treatment effect was performed. In addition, a meta-analysis of study-level hazard ratios (HR) based on a random effects model (an approximation of a Cox’s proportional hazards model using a Poisson generalised linear model) was conducted.

Sensitivity analyses entailed omitting cancers diagnosed within 6 weeks of trial entry and omitting all non-melanoma skin cancers (NMSC) from case definition. To evaluate any potential duration response, we conducted separate analyses for three non-overlapping periods of follow-up time (<6 months, 6–12 months, >24 months). As a secondary analysis, we performed an aggregate data-based meta-analyses using study-level odds ratios (OR). In contrast to the primary analysis, which uses a time-to-event approach, this analysis used the number of randomly assigned patients as the denominator of the incidence measure. For this secondary data synthesis, Mantel–Haenszel methods were used with a continuity correction inversely proportional to the relative size of the other treatment arm for that study.15

All analyses were performed using Stata version 9.2, with the exception of the random effects survival model, which was performed using R version 2.5.0.

Role of the sponsor

This study was sponsored by Wyeth, who together with Amgen markets etanercept in North America. Wyeth and Amgen provided data for the analysis, and Wyeth provided payment to support the costs of study preparation, data analysis and manuscript preparation. The current meta-analysis, designed to evaluate the risk of malignancy only, arose in the context of a request to Wyeth from a regulatory agency regarding cancer risk. It was in this context that DM, a Wyeth employee and one of the study co-authors, approached TB and ELM in January 2006.

Both companies had the opportunity to comment on the study design and manuscript. However, all final decisions regarding study design, analysis, and reporting and interpretation of results rested with the academic investigators.


Trials included

A total of 82 publications was initially considered, from which 69 articles were excluded based on abstracts and content. Of the remaining 13 citations, eight full-text publications1623 and five poster abstracts2428 reported on eight RCT, which met our inclusion criteria. Figure 1 summarises the flow of eligible clinical trials into our analysis.

Figure 1

QUOROM-style flow diagram indicating selection of studies for this meta-analysis. RA, rheumatoid arthritis.

To ensure complete data acquisition, 30 RCT of etanercept on file with the manufacturers were assessed for eligibility based on a review of original study protocols. In addition to the eight trials already identified through the electronic database search, one unpublished study (study TNR 00102) that was eligible for analysis was identified. Therefore, nine RCT were finally included in our analysis.

Trial characteristics

The characteristics of included trials are displayed in table 1. Trial duration ranged from 12 to 180 weeks. Four trials extended beyond the observation period reported in the initial publication. For three of these trials (studies 0881309, 0881308, 160012), follow-up data were published in subsequent reports after initial publication. For one trial (study 160009), extension data were not published but were obtained from the sponsor.

Table 1 Characteristics of randomised controlled trials included in the meta-analysis

All but one trial excluded patients with a history of cancer with less than a 5-year disease-free state (except NMSC). Trial 1881300 excluded patients who had a history of cancer at any time (except NMSC).

Based on the review of original study protocols, all trials were judged to be of high quality with appropriate randomisation, random allocation concealment and intent-to-treat analysis. Completeness of follow-up and attrition were assessed using IPD: 574 of 2244 (25.6%) in the anti-TNF treatment arms discontinued study treatment early compared with 455 of 1072 patients (42.4%) in the control arms. Common reasons for early discontinuation in etanercept-treated patients were adverse events (31.5%) and lack of efficacy (32.6%). Similarly, common reasons for early discontinuation in control patients were adverse events (25.1%) and lack of efficacy (46.6%). For the majority of patients who discontinued study treatment prematurely, follow-up beyond the date of treatment discontinuation was available: 90.8% in the etanercept arms versus 92.7% in the control arms.

All trials were sponsored by Wyeth or Amgen.


A total of 3396 participants was randomly assigned in the nine trials we assessed. Eighty individuals were excluded from further analysis, 79 of whom had never received the allocated treatment and one who was lost to follow-up immediately after the first dose of study drug. Our data for analysis comprised 2244 participants who received etanercept (contributing 2484 person-years of follow-up) and 1072 participants who received control therapy (contributing 1051 person-years of follow-up). Dataset validation revealed eight patients who participated in two trials. A total of 3308 separate individuals thus generated a denominator of 3316 participants. None of these patients who transferred between trials had an incident malignancy.


Twenty-six patients with incident malignancies were identified in the treatment groups (incidence rate (IR) 10.47/1000 person-years) and seven patients in the control groups (IR 6.66/1000 person-years). A detailed summary of all incident malignancies is given in table 2. In three trials (TNR 00102, 160004 and 160014), no incident malignancies were observed.

Table 2 Summary of malignancies in the randomised controlled trials

For one trial (study 160012), three additional incident NMSC were identified based on individual case narratives when compared with the original publication, which did not report on this type of malignancy.

Data synthesis

Combined analysis according to our primary model (IPD survival analysis) yielded an HR of 1.84 (95% CI 0.79 to 4.28) for malignancies in patients using etanercept compared with control treatment. Using a random effects model resulted in a similar estimate, with an HR of 1.82 (95% CI 0.78 to 4.22).

For methodological comparison, an aggregate data meta-analysis was performed. When applying Mantel–Haenszel methods, the OR for malignancies in patients using etanercept compared with patients receiving control treatment was 1.93 (95% CI 0.85 to 4.38), using a continuity correction according to Sweeting et al.15 The results of using a random effects DerSimonian and Laird model were very similar (HR 1.71; 95% CI 0.73 to 4.01), using the same continuity correction, reflecting the observation that between-trial heterogeneity using I-squared was 0.0%.

Additional analyses

Four malignancies were diagnosed during the first 6 weeks after the first treatment dose. As these cancers were likely to be present yet undetected when patients began the trial, we excluded these four patients as part of our sensitivity analysis. With these exclusions, the HR for malignancies in patients treated with etanercept compared with the non-etanercept group was 1.87 (95% CI 0.75 to 4.62).

In the light of recent observational data,2931 which suggest a significantly increased risk of NMSC (but not other solid malignancies in patients treated with anti-TNF therapy), we decided to exclude as events all NMSC from our analysis. Using this approach, the results were essentially unchanged (HR 1.86; 95% CI 0.62 to 5.59).

To investigate whether there are any particular time periods in which etanercept treatment is associated with an increased incidence of cancer, the dataset was stratified according to three different time points: 0–6 months; 6–12 months and more than 12 months. This analysis did not reveal a time period in which the risk of cancer was significantly increased. We also performed an exploratory analysis (stratified fixed effects model) of the effect of dose, categorising etanercept dosing regimens into two groups, less than 50 mg/week and 50 mg/week or greater (or 25 mg twice weekly). The lower dosing range accounted for only 21.2% of etanercept follow-up time. Compared with the control arms, the relative risk estimates were similar for each dosing range (HR for the higher dose group was 1.92; 95% CI 0.80 to 4.62), and for the lower dose group it was 1.59 (95% CI 0.49 to 5.09), both using comparator as reference (table 3).

Table 3 Effect of etanercept therapy on the occurrence of malignancies in randomised controlled trials

Statistical power

We used a traditional sample size formula (log rank test, Freedman method)32 for a single study to obtain a statistical power approximation (ie, this ignores stratification by study and differential follow-up times between studies) in this meta-analysis context. With a probability of seven malignancies per 1072 patients in the control group, it would require at least 9305 participants to detect a HR of 2.0 (statistical significance level of 5% and a power of 80%) in a large RCT, assuming 32% of patients are allocated to control—reflecting the situation in the existing studies. The number of individuals in our dataset (3316) was substantially lower.

Based on the numbers derived from the existing studies, the probability of detecting a doubling in the risk of malignancy (HR of 2.0) between the two groups, should such a difference exist, was 39%.


Our analysis found a higher incidence estimate of malignancies in etanercept as compared with placebo-treated patients, although the results are statistically not significant. Therefore, this study does not provide sufficient evidence to establish an association of malignancies and etanercept treatment. However, given the wide confidence interval of the effect measure (HR 1.84; 95% CI 0.79 to 4.28), it also cannot exclude a clinically meaningful association.

This meta-analysis provides important insight into methodological issues of an IPD meta-analysis of sparse adverse event data. Publication bias is usually viewed as one of the more prominent threats to the validity of a systematic review and meta-analysis. Our approach of obtaining IPD in cooperation with primary investigators and trial sponsors allowed us to review the complete collection of manufacturer-sponsored RCT for etanercept, making publication bias very unlikely. A review of inclusion and exclusion criteria of candidate studies, based on original study protocol review, resulted in a more reliable eligibility assessment. Furthermore, the IPD approach allowed us to include follow-up data that extended beyond the published period of follow-up.

Clinical trials of biological use in RA often show imbalances in the percentage of withdrawals between treatment and control groups, as a result of a higher rate of treatment failure in placebo-treated patients. This carries the theoretical risk of false estimates due to a higher loss to follow-up in the control groups and a longer exposure to the study treatment in the active treatment arm. A major benefit of our IPD approach was the ability to perform a time-to-event analysis. This allowed the censoring of patients who discontinued treatment early or were lost to follow-up, thereby removing them from the denominator. Of note, our aggregate data analysis yielded similar results. This strengthens the validity of an aggregate data approach of analysing sparse events in clinical trials of RA. Our analysis was robust to the application of a wide variety of statistical methods for data synthesis, reflecting that our results did not depend upon the assumptions of any one particular method.

The major weakness of our study is the lack of statistical power. Assessment of statistical power in a meta-analysis has been suggested but is rarely performed.3335 Our experience emphasises the importance of such an analysis to estimate the likelihood of missing an association given that it was true. The combination of several underpowered studies can still produce a meta-analysis that is inadequate to detect a clinically important effect size. Nonetheless, our meta-analysis does provide a more precise assessment of cancer risk than those available from the individual RA trials.

Comparing our results with published clinical data, the observed excess of malignancies in patients who received the TNF receptor fusion protein etanercept is not inconsistent with the results of a meta-analysis of anti-TNF antibody treatment in RA patients.3 An updated version of this analysis2 including 5788 patients yielded an OR of 2.4 (95% CI 1.2 to 4.8). The risk of malignancies in this study appeared to be more pronounced in patients who received higher doses of anti-TNF antibodies according to a subanalysis that stratified by dose. In the etanercept meta-analysis we did not observe clear evidence of a dose response, although the proportion of patients treated at doses lower than 50 mg per week was small and does not allow definite conclusions.

Of note, an RCT of etanercept in Wegener’s granulomatosis36 found a significantly increased risk of malignancy in the etanercept arm.

The results of two large observational studies29 31 that included a greater number of patients and had longer follow-up did not replicate the overall increase of malignancies seen with the synthesis of RCT data. However, they did reveal a significantly increased risk of NMSC in anti-TNF-treated patients.

Contrasting results of trial data and observational data provides a valuable stimulus to explore further not only the central clinical question of a potential association of anti-TNF therapy and malignancies in particular, but also the methodological strengths and weaknesses of both methods for drug safety assessment in general. Even the best-designed observational studies may produce inaccurate answers, as differences in patient characteristics across treatment groups can seldom be perfectly controlled. In contrast, with successful randomisation in RCT, the baseline risk of a subsequent adverse event should be similar in all treatment groups, and whereas meta-analysis combines data across studies, stratification by study preserves the benefit of randomisation. Given sufficient trial evidence, a meta-analysis of RCT data may be able to detect potential drug hazards early, before observational data become available after a drug is marketed. However, the ability of meta-analysis to assess sparse event data in randomised trials is often constrained by relatively short follow-up periods and finite cumulative trial enrolment, which translate to limited statistical power to detect differences between treatment groups for rare events.

Different strategies may be considered to improve statistical power in this context. Several RCT of anti-TNF treatment for indications such as ankylosing spondylitis, psoriatic arthritis, cardiomyopathy, inflammatory bowel disease and a variety of connective tissue disorders have been performed. Including all these trials in a comprehensive, IPD meta-analysis will have a higher chance of approaching an adequate sample size and delivering more precise estimates. As an additional or alternative step, different agents with a similar mode of action may be combined to improve statistical power, yet this gain may come at the price of validity, if the effects of study drugs are not homogeneous.

A large, comprehensive meta-analysis utilising these two steps to improve statistical power by including all three approved anti-TNF agents over a wide range of different indications has been requested by the European Medicine Evaluation Agency and is currently in progress.


View Abstract


  • Funding: This study was sponsored by Wyeth, who together with Amgen, markets etanercept in North America. Wyeth and Amgen provided data for the analysis, and Wyeth provided payment to support the costs of study preparation, data analysis and manuscript preparation.

  • Competing interests: Declared.

    TB, FCW and AJS received grant support from Wyeth. ELM served as an investigator for Amgen, Biogen-IDEC, Centocor, Genentech, Hoffmann-LaRoche, Human Genome Sciences, Wyeth. He received grant support from Amgen, Centocor/Johnson & Johnson, Genentech, Mayo Foundation and Wyeth. He served as a consultant/on scientific advisory boards for Abbott, Amgen, BiogenIDEC and Centocor. DM is employed by Wyeth and owns stock options in the company. KRA received grant support from Wyeth. He served as a consultant to United BioSource Corporation (UBC) regarding a “mixed treatment comparison” project, which UBC has conducted for Bristol-Meyers Squibb in relation to rheumatoid arthritis.

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.