Round 39: TNF Antagonists and Infection: The Pharmacoepidemiologist’s Perspective

By: Daniel Solomon, MD, MPH
Associate Professor of Medicine
Chief, Section of Clinical Sciences
Division of Rheumatology, Immunology, and Allergy
Division of Pharmacoepidemiology
Brigham and Women’s Hospital
Harvard Medical School


Dr. Solomon has no financial disclosures to report.

Release Date: February 14, 2012
Expiration Date: October 31, 2013



  • Discuss the validity of the current evidence regarding the association between TNF antagonists and infections
  • Describe the specific methodologic issues that contribute to the discrepancies in the literature regarding TNF antagonists and infection



In 2001, several years after the TNF antagonists were on the market, The New England Journal of Medicinepublished an article from some folks at the FDA as well as FDU based on MedWatch data.  These are data that are spontaneous adverse event reports that go to the FDA. There were 70 reported cases of tuberculosis (TB) and 40 of these cases were extra pulmonary with 12 deaths, 64 were in non-endemic countries.

Although there are deficiencies with MedWatch, they played an important role in signaling a potential relationship between these agents and opportunistic infections such as TB.  TB is very dramatic; cases are not seen every day, so the type of exposure is important—travel, drug, etc. However, not all adverse events are obvious. This article focuses on more typical infections, as opposed to TB.


A meta-analysis by Bongartz et al, published in JAMA (2006;295:2275) looked at nine published randomized trials.  Infliximab and adalimumab were included; etanercept was not. Across these 9 trials, there were 5,000 patients included in the active treatment arm and 1,400 patients from the placebo and what they found was a twofold risk of serious infection from these trials. Overall, there were 126 serious infections—12 were granulomatous (approximately 10 to 1 ratio of bacterial infections to granulomatous).

Sometimes a meta-analysis of randomized trials is available, but observational data is needed too.  There are reasons why the trial data in a meta-analysis may not be top-notch information.

Methodologic Considerations

  • New TNFA users compared with chronic MTX users
  • Exclusion of etanercept is noticeable
  • Analysis based on persons not person-time
  • Half of the trials showed a higher drop-out rate among placebo-patients; thus, follow-up for infection would be shorter for placebo treated patients

Observational pharmacoepidemiology
Major epidemiologic studies and results

Reference Study Population N Endpoint Adjusted Relative Risk
(95% CI)
Listing 2005 RABBIT 1,529 Hospitalized infection ETA: 2.16 (0.9 – 5.4)
IFX: 2.13 (0.8 – 5.5)
Wolfe 2006 NDB 16,788 Hospitalized pneumonia ADA: 1.1 (0.6 – 1.9)
ETA: 0.8 (0.6 – 1.1)
IFX: 1.1 (0.9 – 1.4)
Dixon 2006 BSRBR 8,973 Hospitalized infection, death, IV ADA: 1.07 (0.67 – 1.72)
ETA: 0.97 (0.63 – 1.50)
IFX: 1.04 (0.68 – 1.61)
Curtis 2007 United Health 5,326 Hospitalized infection or IV TNFA: 1.9 (1.3 – 2.8)
Schneeweiss 2007 Medicare + Rx 15,597 Hospitalized infection TNFA: 1.0 (0.60 – 1.67)
Dixon 2007 BSRBR 10,829 Hospitalized infection, death, IV TNFA: 1.3 (0.9– 1.8)*
TNFA: 4.6 (1.8 – 12)**
Curtis 2007 United Health 5,195 Hospitalized infection, IV ETA: 1.6 (0.7 – 3.3)**
IFX: 2.4 (1.2 – 4.7)**

These are all from papers published in Arthritis & Rheumatism. They come from different populations, some of them registries, some of them claims data. They’re all of large size. The end points are somewhat similar, hospitalized infection for most of them. Interestingly, there’s some discrepancy across the studies about what the answer is. Did certain studies get it wrong?  Did certain studies get it right?  Why do we have such a heterogeneous results?”

Methodologic issues

Some of the methodologic issues that might underlie these differences are largely public health concerns when thinking about epidemiology:

  • Exposure risk window
  • Comparator drug
  • Confounding (by disease severity)
  • Multiple exposures
  • Time-variant hazard
  • New user (vs prevalent user) design
  • Endpoint assessment
  • Compare cohort types

Exposure Risk Window

The issue around an exposure risk window is really critical.  What is the appropriate “at risk” period for a given drug exposure? In other words, what is the biologic relationship between exposure to a TNF antagonist and infection?  Does it start the same day as drug use?  Does it end the day the drug use ends?  Does it continue after the drug use ends?  Is there a lag period?  The figure below shows 4 examples of the exposure risk window:


  1. Start to look for exposures immediately.
  2. If there were infections on day two, they couldn’t be related to the exposure.
  3. Should we just wait several half-lives because the drug is still having some pharmacodynamic relationship in the body with this endpoint or should we consider somebody once exposed, always exposed?
  4. One has to think about biologic relationships to approach this in an intelligent fashion and perhaps do sensitivity analysis the relationship is unclear.

Comparator Drug

In a randomized trial we often think about the placebo arm, but in real life, people don’t take placebo.  Is “no use” actually the right comparison? A lot of the analysis is based on a TNF versus nothing. This isn’t realistic; very few rheumatoid arthritis patients take no drugs. An appropriate comparator would probably produce ‘truer’ results.

Confounding (by disease severity)

Having active comparators helps to control for some potential confounding. When deciding the reference group, which is the comparator drug? Should it be methotrexate? Should it be somebody who failed methotrexate but somehow goes on to other non-biologic DMARDs? Should it be someone who goes to a different biologic DMARD? These are really important issues when you’re thinking about what’s the right reference or comparator group.

Patient factors become confounders if they are associated with treatment choice and are also independent predictors of the outcome. Severity of disease, the prognosis and co-morbidities might be important confounders when thinking about this relationship.

The randomized trial is so useful because it breaks the relationship between the confounder and the exposure.  That’s what randomization does.  Measurable confounders, although worrisome, are easier to deal with than unmeasured confounders because they can be adjusted for through regression.

There are ways to deal with confounding:


Multiple exposures 

Combination therapy is the norm, so how does one isolate the risk of a given treatment?

A. Separate exposures:

How can one isolate which drug causes infection?

  • Each one is clearly represented
  • But how does a doc “multiply” risks

How should the relative risks of these exposures be independently shown for a patient?

B. Explicit combination exposure:

There are dominant combinations of drugs in rheumatoid arthritis; methotrexate plus a TNF.  An explicit combination category can be very useful, but how many different explicit combination categories can be created?  Using this approach, there is no reason to multiply risk factors, but how many combinations is enough? A good strategy could be to represent the six dominant combinations and call the rest “other”––that may be acceptable in certain circumstances.

C. Hierarchical exposure:

A hierarchy of exposures reduces the number of comparisons, but makes strong assumptions. The dominant exposure is identified and the other exposures are adjusted in a regression.

Time-varying hazard – does the risk of infection vary during the exposure time?
Another very difficult issue is the possible time-varying nature of the risk of infection. Based on some of the biology and based on some prior literature, it appears that TNF antagonist and infection have a time-varying hazard. Early on there’s really low risk, and then it shoots way up and then it comes back down.

New user design 
A. RCT – Randomized Control Trial
The randomized trials of the TNF antagonist use methotrexate failures.  They are randomized either to placebo or TNF.  In (A) the subjects are chronic users of methotrexate plus a placebo.

B. Observational Studies

  • The subjects are chronic users of methotrexate plus new users of TNF.  There are some inherent issues with people who are surviving on methotrexate; they’re not stopping their methotrexate––they’re just adding a drug. Adding a drug with TNF antagonist to have a new use versus a chronic use. However, there may be some issues with estimating the relative risks even in randomized trials.
  • Another observational study design might be to take people all on a DMARD, enter them in a cohort––time zero would be when they switched to a TNF antagonist. Another group may be one that just stays on their background DMARD. This might present some problems based on new users versus chronic users.  This is precisely what’s done in the British Biologics Registry:
  • They have people who come in for a visit––they could change therapy; some of them do and some of them don’t. The people who don’t stay on the non-biologic DMARD and they’re in the reference group.
  • The people who stay on the non-biologic DMARDs are completely different than the people who switched to TNFs.
  • There is a time course issue with chronic users versus new users.

Endpoint Assessment
Endpoints are also quite difficult to assess:

  • There is a period, a pre-diagnosis period, a possible infection and some treatment period.
  • Studies examine patients using different methods.
  • Insurance or claims data assessment is typical practice, and may not be structured.
  • Structured surveillance is not uniform.
  • Infections are defined differently in different cohorts.

Surveillance bias (detection bias)
The issue of surveillance bias is an important one. TNF blockers can cause upper respiratory type symptoms. Because of this, doctors are looking at patients differently if they’re on a TNF antagonist then if they’re not on a TNF antagonist.  This leads to potential detection bias: testing, diagnosis and treatment may be affected. Other DMARDs may not effect the same reaction. There is potential bias here in how people look at patients with these different symptoms based on exposures.

Cohort types

Cohort Type Strengths Weaknesses
Disease-based Registry (NDB, CORRONA, BRASS)
  • Diagnosis is usually very accurate
  • Disease severity measures
  • Medical records available
  • Patients may not be typical
Drug-based Registry (BSR, RABBIT)
  • Diagnosis is usually very accurate
  • Disease severity measures
  • Medical records available
  • “Controls” may not be comparable to exposed
  • Surveillance bias
Practice-based or Population-based registry (REP, GPRD)
  • Medical records are often available
  • Patients represent routine care
  • Often allows for linkage to prescriptions
  • Often links to other registries
  • Diagnosis may not be accurate
  • Outcomes may not be accurate
  • Disease specific info lacking
Healthcare utilization data (Medicare, United)
  • Patients represent routine care
  • Includes linkage to prescriptions
  • Very large cohorts can be assembled
  • Dx may not be accurate
  • Outcomes may not be accurate
  • Disease specific info lacking


Listing et al. Arthritis & Rheumatism 2005;52:3403

  • German biologics register (RABBIT): drug-based
  • 1,529 patients with rheumatoid arthritis; 2001 to 2003; one-year follow up for infection
  • Exposure risk window defined with no lag period and the duration was 365 days
  • Controls were non-biologic DMARDs
  • Drug initiators: DMARD – yes; NBDMARD – yes
  • Confounding control: Propensity score adjustments (dropped pts with score < 0.4)
    • Age, # DMARDs, DAS28, CRP, RF, and HAQ
    • Multivariable adjustments: Propensity score, prednisone dose at study entry, cumulative dose during one year follow-up, COPD, DM, psoriasis
  • Follow up: 12 month maximum with 74% completion
  • Endpoint assessment: reported by investigators

They found a doubling of the risk of serious infectious in both the etanercepts and infliximab.

Wolfe et al. Arthritis & Rheumatism 2006;54:628

  • National Data Bank for Rheumatic Diseases: disease-based
  • 16,788 patients with rheumatoid arthritis, 2001 to 2004
  • Exposure risk window defined with no lag period and the duration was unclear
  • Comparator drug: No prednisone
  • Drug initiators: DMARD – no; TNFA – no
  • Confounding control: HAQ, RA duration, prednisone, comorbidities
  • Follow up: 30-months median
  • Endpoint assessment: self-report with confirmation in some %

The bottom here: the TNF antagonists were not associated with an increased risk of infection.  Prednisone at increasing dosages was associated with increasing risk of infection.  Methotrexate had no increased risk.

Dixon et al. Arthritis & Rheumatism 2006;54:2368

  • BSR Biologics Register: drug-based
  • 8,973 patients with rheumatoid arthritis
  • Exposure risk window defined with no lag period and the duration was supply
  • Comparator drug: NBDMARD
  • Drug initiators: DMARD – no; TNFA – yes
  • Confounding control: HAQ, DAS, RA duration, prednisone, comorbidities
  • Follow up: 11 to 15 months
  • Endpoint assessment: hospitalized infections, IV Abx, death

They found in their adjusted analyses no increased risk with any of the three TNF antagonists.

Curtis et al. Arthritis & Rheumatism 2007;56:1125

  • United Health: administrative data
  • 5,326 patients with rheumatoid arthritis
  • Exposure risk window defined with no lag period and the duration was day supply + 90
  • Comparator drug: MTX
  • Drug initiators: DMARD – no; TNFA – yes
  • Confounding control: comos, prednisone, health system factors
  • Follow up: 17 months median
  • Endpoint assessment: diagnosis codes with record review

The risk for infection was doubled in this analysis.

Schneeweiss et al. Arthritis & Rheumatism 2007;56:1754

  • Medicare + prescription benefit plans: administrative data
  • 15,597 patients with rheumatoid arthritis
  • Exposure risk window defined with no lag period and the duration was day supply + 3 half lives
  • Comparator drug: MTX
  • Drug initiators: DMARD – yes; TNFA – yes
  • Confounding control: comorbidities, prednisone, health system factors
  • Follow up: 7 to 15 months mean
  • Endpoint assessment: diagnosis codes based on validated algorithms

This analysis showed no increased risk of infection.  There were no increased risks across the TNF antagonist for multiple different outcomes, pneumonia, sepsis, osteomyelitis, any of the above or any bacterial infection at all.

Dixon et al. Arthritis & Rheumatism 2007;56:2896

  • BSR Biologics Register: drug-based
  • 10,829 patients with rheumatoid arthritis
  • Exposure risk window defined with no lag period and the duration varied
  • Comparator drug: NBDMARD
  • Drug initiators: DMARD – no; TNFA – yes
  • Confounding control: HAQ, DAS, RA duration, prednisone, comos
  • Follow up: varied
  • Endpoint assessment: hospitalized infections, IV Abx, death

The time varying analysis found that in the first 90 days of exposure there was a three- to five-fold increase risk of infection.



  1. Compared with patients starting methotrexate, the TNF antagonists do not appear to be associated with an increased risk of serious infection requiring hospitalization among patients using medications for a median of 10 months. This is difficult to translate into clinical practice because of the many caveats on it.
  2. However, patients starting a TNF antagonist do appear to have an increased risk for severe infection requiring hospitalization for the first 6 months compared to patients who are chronic users of methotrexate or other DMARDs.
  3. Glucocorticoid use appears to be associated with an increased risk of infection, but the confounding by disease severity is difficult to disentangle from steroid use. There are many reasons why higher steroid doses are associated with increased risk of infection.
  4. Opportunistic infection seems to be elevated in TNF users.  However, these infections are much less common and much of these data come from case series (5 patients, 8 patients, 20 patients), making it difficult to interpret the data.


  1. The exposure risk window must be appropriately and explicitly defined.
  2. The comparator drug is critical. Which DMARD should be used, should it be a biologic response modifier?
  3. A new user design is the best way to assess comparative safety for infections (other designs may be appropriate for other endpoints).
  4. Confounding is a tough issue, there are always limitations on what can’t be measured.
  5. There are no perfect methods for endpoint assessment. There are different methods. Authors need to be pushed to make clarify how patients define an endpoint. It’s not always clear when you read papers.
  6. The meta-analysis shows the cohort types are very different. Different methods will give different answers because the questions are different.


Updated: August 10, 2012

Arthritis Center

About Arthritis Center

Founded in 1998, the Arthritis Center at Johns Hopkins is dedicated to providing quality education to patients and healthcare providers alike.