The risk of infection associated with tumor necrosis factor ╨Ю┬▒ antagonistsMaking sense of epidemiologic evidence.код для вставкиСкачать
ARTHRITIS & RHEUMATISM Vol. 58, No. 4, April 2008, pp 919–928 DOI 10.1002/art.23396 © 2008, American College of Rheumatology REVIEW The Risk of Infection Associated With Tumor Necrosis Factor ␣ Antagonists Making Sense of Epidemiologic Evidence Daniel H. Solomon,1 Mark Lunt,2 and Sebastian Schneeweiss1 full spectrum of benefits and risks of a drug in routine care. There have been several epidemiologic studies regarding the association of TNF␣ antagonists with infections, as summarized in Table 1 (3–9). On first glance, the studies seem similar in their population composition and in the choice of comparator drugs, but the results vary meaningfully. Three studies found no increase in the risk of bacterial infection associated with use of TNF␣ antagonists as compared with methotrexate, whereas 3 studies showed a statistically significant increase in infection risk. Does this seemingly inconsistent data from the literature suggest an inherent weakness in the methodologies used? Or, as we would suggest, are the studies different in subtle, but important, methodologic aspects, so that they are actually addressing different study questions? This review examines key methodologic issues in pharmacoepidemiologic studies of TNF␣ antagonists and the related risk of infections, and weighs the strengths and weaknesses of each study to provide a more coherent framework for understanding these issues. Introduction The data suggesting that tumor necrosis factor ␣ (TNF␣) antagonists are associated with certain opportunistic infections seem quite strong; however, the association of these agents with typical bacterial infections is less clear. Since bacterial infections are much more common than opportunistic infections, defining this potential risk is clinically important. Although one metaanalysis found an increased risk of infection with the use of these agents as compared with methotrexate, the limitations of that systematic review have been well described (1). Randomized controlled trials are the gold standard for defining the efficacy of a drug or for determining how beneficial a drug can be in ideal circumstances. However, trials often are conducted in highly selected populations for short durations and in very controlled settings. Not only can the benefits of a particular drug be much different in a routine care setting than in a randomized controlled trial, but also the full range of adverse events related to a drug is rarely observed in a randomized controlled trial during the premarketing stage (2). Thus, epidemiologic studies may offer important complementary information about the Exposure risk window 1 Daniel H. Solomon, MD, MPH, Sebastian Schneeweiss, MD, ScD: Brigham and Women’s Hospital, and Harvard Medical School, Boston, Massachusetts; 2Mark Lunt, PhD: University of Manchester, Manchester, UK. Dr. Solomon has served in the last year as an unpaid member of the advisory boards of Abbott and Amgen; he has received research grants in the last 2 years from Pfizer and Savient. Dr. Schneeweiss has received consulting fees, speaking fees, and/or honoraria (less than $10,000) from HealthCore and (more than $10,000) from i3 Drug Safety. Address correspondence and reprint requests to Daniel H. Solomon, MD, MPH, Brigham and Women’s Hospital, Division of Rheumatology, Immunology and Allergy, 75 Francis Street, Boston, MA 02115. E-mail: email@example.com. Submitted for publication October 22, 2007; accepted in revised form December 20, 2007. The exposure risk window refers to the period during which a given drug may be causing the (toxicityrelated) outcome of interest, either because the drug is physically still present in the body (pharmacokinetic presence) or because the drug is acting through more indirect physiologic pathways (pharmacodynamics). In our example, the characterization of the exposure risk window should be based on the presumed underlying biologic link between TNF␣ antagonist exposure and infection. This relationship is poorly understood, and assumptions need to be made about when to begin and when to end the exposure risk window. 919 National Data Bank for Rheumatic Diseases British Society of Rheumatology Biologics Register Commercial insurance beneficiaries Medicare beneficiaries 65 years and older British Society of Rheumatology Biologics Register Commercial insurance beneficiaries Wolfe, 2006 (4) 5,195 10,829 15,597 5,326 8,973 16,788 1,529 No. of patients Hospitalized with infection, death, or requiring IV antibiotics Hospitalized with infection or requiring IV antibiotics Hospitalized with infection, death, or requiring IV antibiotics Hospitalized with infection or requiring IV antibiotics Hospitalized with infection Hospitalized with pneumonia Hospitalized with infection End point Methotrexate Nonbiologic DMARD Methotrexate Methotrexate Nonbiologic DMARD Absence of drug of interest† Nonbiologic DMARD Comparator group Drug-specific adjusted relative risk (95% confidence interval) etan. 1.55 (0.73–3.34); inflix. 2.41 (1.23–4.70)§ TNF␣ 1.30 (0.93–1.78); TNF␣ 4.6 (1.8–11.9)‡ TNF␣ 1.0 (0.60–1.67) TNF␣ 1.94 (1.32–2.83) ada. 1.1 (0.6–1.9); etan. 0.8 (0.6–1.1); inflix. 1.1 (0.9–1.4) ada. 1.07 (0.67–1.72); etan. 0.97 (0.63–1.50); inflix. 1.04 (0.68–1.61) etan. 2.16 (0.9–5.4); inflix. 2.13 (0.8–5.5) * DMARD ⫽ disease-modifying antirheumatic drug; etan. ⫽ etanercept; inflix. ⫽ infliximab; ada. ⫽ adalimumab; IV ⫽ intravenous. † For example, adalimumab was compared with no adalimumab. ‡ The first TNF␣ study included a 90-day extension period, while the second assessed only the initial 90 days of exposure. § Assessed only the initial 180 days of exposure. Curtis, 2007 (9) Dixon, 2007 (8) Schneeweiss, 2007 (6) Curtis, 2007 (7) Dixon, 2006 (5) German Biologics Registry Study population Listing, 2005 (3) First author, year (ref.) Table 1. Summary of major epidemiologic cohort studies of the risk of infection associated with tumor necrosis factor ␣ (TNF␣) antagonists among patients with rheumatoid arthritis* 920 SOLOMON ET AL RISK OF INFECTION WITH TNF␣ ANTAGONISTS Figure 1. Exposure risk windows. The exposure risk window refers to the period during which a subject would be considered exposed and at risk for the outcome of interest. The start and end of the exposure risk window can be defined in many different ways, based on the biologic features of the potential drug–outcome association. A–D illustrate several options associated with the use of tumor necrosis factor ␣ (TNF␣) antagonists. A, An exposure risk window with no lag period would begin immediately with the first use of the medication. This may be a reasonable assumption for an outcome such as infection, but may not be reasonable for a malignancy with a longer induction time. B, An exposure risk window with a lag period would begin after a defined period of delay. This may be more appropriate if the relationship between drug and outcome is assumed to be of slower onset. C, An exposure “washout” risk window would have a brief period of extension after a drug has been discontinued, to account for the period when a drug is still biologically active in the body. D, An indefinite exposure risk window would be assumed to continue indefinitely after the cessation of the drug; this may be a useful assumption for an outcome such as malignancy. One might assume that the risk of infection begins with the first dosage of a TNF␣ antagonist (Figure 1A). This would seem to be a reasonable assumption, based on prior case series regarding mycobacterial infections among patients taking TNF␣ antagonists (10). However, in other drug–toxicity relationships, there may be a lag period between drug initiation and onset of risk (Figure 1B). In such instances, events that occur prior to the onset of the exposure risk window (i.e., before the end of the lag period) should not be considered to be outcomes related to the study treatment. When taking into account lagged exposure risk windows, it is presumed that one knows when a drug was initiated. This is not the case when patients on long-term treatment, without a known start date, are allowed in an analysis. There is a similar set of questions regarding the 921 end of the exposure risk window after the last dosage of drug. The risk might continue for some period after the last dosage is fully washed out of the body, i.e., several half-lives of a given drug (Figure 1C); this might be as long as 4–6 months for a drug like infliximab. In fact, with rituximab, this washout period might be much longer. Conversely, the risk might be indefinite, especially if one is considering indolent outcomes, such as osteomyelitis or cancers developing as a result of genetic damage caused by a given drug (Figure 1D). Defining an extended exposure risk window incorrectly results in falsely attributing nonexposed cases to the study exposure, leading to overestimated risk ratios. Conversely, if the time period of the window is too short, valid cases might be missed, leading to underestimation of a causal effect. In practice there is no algorithm to classify exposure 100% correctly. The choice of strategy depends on whether one needs to be more concerned about falsely classifying the person-time period as exposed or as unexposed, and also depends on the pharmacologic aspects of the hypothesized drug effect. Comparator drug Most epidemiologic analyses of drug safety focus on the relative risk associated with use of a given medication compared with a comparator drug with the same or similar indication. In the example of TNF␣ antagonists, most analyses use methotrexate or nonbiologic disease-modifying antirheumatic drugs (DMARDs) as the comparator drugs. Ideal comparator drugs are those that have the same indication as the study drug and that might be used interchangeably, so that the physician’s choice of drugs is almost random (11). If the population receiving the drug of interest (i.e., TNF␣ antagonists) and the population with the comparator exposure (i.e., nonbiologic DMARDs) differ in clinical characteristics that could be predictive of the study outcome, then confounding will bias the results. In such circumstances, statistical adjustments are required. If the reference group for patients taking TNF␣ antagonists is those who do not take TNF␣ antagonists (i.e., “nonusers”), then the patient populations being compared may not be similar. Such nonusers of TNF␣ antagonists may include patients who are taking only nonsteroidal antiinflammatory drugs (NSAIDs), since they would have very mild rheumatoid arthritis (RA). Because disease severity may be related to the risk of infection (12), and because this is difficult to measure and completely adjust for in multivariable models, non- 922 SOLOMON ET AL user comparison groups may cause intractable confounding. The use of methotrexate as a comparator drug may be more appropriate in cases in which, for example, RA patients taking methotrexate have disease activity that is similar to that in patients taking TNF␣ antagonists. Such a comparison group may still result in noncomparable exposure groups. Requiring that the comparators be between those who have switched from one DMARD to methotrexate and those who have switched from a DMARD to a TNF␣ antagonist might further improve the balance in disease activity between groups. Drug initiator versus ongoing user cohorts The relative risk of an unintended event often varies during the course of therapy. Several of the epidemiologic studies of TNF␣ antagonists and infection suggest that the relative risk of infections is highest shortly after treatment initiation and then drops over time (8,9). Such a time-varying hazard function is typical in settings in which adverse effects related to a drug are recognized quickly, which leads to patients being removed from the cohort because of drug discontinuation, resulting in a remaining cohort of “survivors” (13). This has implications in choosing the study cohort, which could include either patients who have recently started the drug (“drug initiators”) or patients who have been continuing to receive the treatment for some time (“ongoing users”) (Figure 2). Drug initiators are defined as patients for whom there is no recent use of a given drug, i.e., within the preceding 6–12 months, whereas ongoing users are those who have been taking a given drug for some time immediately prior to entering the study cohort. If infections tend to occur shortly after treatment initiation, then a cohort of ongoing (long-term) users would cause the risk of infection experienced by drug initiators to become mixed with that experienced by long-term users, diluting the effect of the early increased risk. Moreover, if patients in the TNF␣ antagonist cohort are initiators, while those in the comparison group are ongoing users, this may lead to an overestimation of the risk with TNF␣ antagonists. Drug initiator cohorts have the additional advantage of a clearly defined temporality of baseline characteristics and treatment. In such a design, patient characteristics are assessed before drug initiation, and are therefore not the consequence of treatment but predictors thereof. Because pure placebo-controlled trials in RA would be unethical, randomized controlled trials of Figure 2. Comparison of drug initiator and ongoing user designs. Because of the time-varying risk of most adverse events, the most valid way of assessing whether a drug may “cause” a specific adverse event is the drug initiator design. However, other designs for assessing the risk associated with use of tumor necrosis factor ␣ (TNF␣) antagonists are also available (A–D). A, Typical randomized controlled trial (RCT) design. Patients taking methotrexate (MTX) are randomized (R) to receive either placebo (Pbo) or a TNF␣ antagonist. This design is very useful for assessing the efficacy of TNF␣ antagonists in comparison with placebo. However, since initiators of a TNF␣ antagonist are being compared with long-term users of MTX, the design may not be as useful for examining the potential of a TNF␣ antagonist to have adverse effects. B, Epidemiologic version of the RCT design. Patients taking a disease-modifying antirheumatic drug at baseline (DMARDold) either continue on the DMARDold monotherapy or add or substitute a TNF␣ antagonist in the regimen. C, Drug initiator design. Patients taking a DMARD at baseline then either switch to monotherapy with a new DMARD (DMARDnew) or add or substitute a TNF␣ antagonist. The design in B compares drug initiators with drug initiators, whereas that in C compares drug initiators with new users and is therefore less subject to bias. D, A version of the drug initiator design. Patients with no prior use of a DMARD start treatment with a TNF␣ antagonist and/or a DMARD or start treatment with a comparator drug (DMARDnew). While the scenario in D may be less common, this possibility should be considered. C ⫽ cohort. TNF␣ antagonists frequently add active treatment or placebo to a background of methotrexate. Thus, initiators of a TNF␣ antagonist are compared with ongoing users of methotrexate (Figure 2A). Although epidemiologic studies can mimic the randomized controlled trial design (Figure 2B), recommendations for pharmacoepidemiologic studies suggest using cohorts of initiators of both the study drug (i.e., TNF␣ antagonists) and the RISK OF INFECTION WITH TNF␣ ANTAGONISTS 923 comparator drug (Figures 2C and D), to minimize bias (14). Combination therapy Combination therapy is the rule and not the exception in RA therapy. In fact, TNF␣ antagonists were approved for use in RA in combination with other DMARDs. How should one estimate the independent risk of infection from a TNF␣ antagonist when it is used so commonly in combination with other therapies that also might contribute to an infection risk? Moreover, how can the relative risk from different combinations be appropriately communicated to physicians and patients? Several different analytic approaches have been considered for patients taking a TNF␣ antagonist in combination with other treatments. The most straightforward approach analytically may be the least clear to interpret. This involves assigning each exposure, whether mono- or combination therapy, into separate exposure categories. Thus, for a patient taking a TNF␣ antagonist, methotrexate, and a glucocorticoid simultaneously, all 3 exposure categories would contribute to the risk estimates in the multivariable model, allowing one to estimate the independent effects of each exposure compared with all other exposures. This approach is complicated to interpret for readers, since it creates a “floating” reference group that changes with each relative risk estimate. Another relatively straightforward approach is to represent combination therapy explicitly as a specific exposure group. Thus, during the time that patients are simultaneously taking a TNF␣ antagonist, methotrexate, and a glucocorticoid, they would be represented in an exposure group for this specific combination. This allows the combinations to be compared with a fixed and well-defined comparator group. As many combinations as necessary could be defined. This approach is most useful if there are several dominant combinations of treatments, thus allowing a reader to identify the relative risk associated with a combination exposure group of interest. This is the case with TNF␣ antagonists and combination treatment with a nonbiologic DMARD such as methotrexate. Group sizes can get small when patients switch to different drugs often or when there are no dominant combination regimens. Another approach is to consider a hierarchy of exposures, such that patients taking a TNF␣ antagonist, methotrexate, and a glucocorticoid are only represented in the TNF␣ antagonists category. This approach is problematic, since patients in the TNF␣ antagonists Figure 3. Potential confounders in the relationship between tumor necrosis factor ␣ (TNF␣) antagonists and infection. A confounder is a variable that is associated with both an exposure of interest, such as a TNF␣ antagonist, and an outcome, such as infection. The confounder is not an intermediate variable on the causal pathway between the exposure and the outcome, but instead has an independent relationship with both. Important potential confounders in the TNF␣ antagonist–infection relationship include disease activity or severity, such as in patients with rheumatoid arthritis, as well as comorbid conditions, such as in patients with diabetes. category are actually heterogeneous in their exposures, with some of these patients taking TNF␣ antagonist monotherapy and others taking a variety of combinations that include a TNF␣ antagonist. Nevertheless, this approach might be useful if there are only a few cases in the treatment group of interest and one wishes to broaden the group. Control for potential confounding Confounding is the bane of pharmacoepidemiology. It refers to the potential for a third factor to be a predictor of treatment choice and to be an independent risk factor for the study outcome. If we consider the example of TNF␣ antagonists and infection, one can imagine several possible potential confounders, such as age, disease severity, and comorbidities (Figure 3). For example, patients with worse disease activity are more likely to be prescribed a TNF␣ antagonist, and they also appear more likely to develop infections independent of treatment (12). Thus, the relationship between TNF␣ antagonists and infection may be explained by disease severity, which could cause doctors to differentially prescribe (“channel”) these agents to patients at higher risk of infection. Another type of potential confounder is comorbid conditions. If, for example, doctors are less likely to treat patients with diabetes with a TNF␣ antagonist because of infection concerns, and preferentially prescribe a nonbiologic DMARD, then nonbiologic 924 DMARDs could appear to be associated with infection, if diabetes is strongly linked with infection. One can attempt to control for these potential confounding factors using several different methods. First, if one has some information on the confounder, it should be included in the multivariable model. Thus, information on comorbid conditions, such as diabetes, need to be adjusted for, when available. Second, the use of active comparators, such as methotrexate, helps to “match” patients according to disease severity. However, this may not completely control for this issue, since there may be residual confounding. Propensity scores are increasingly being used in pharmacoepidemiologic studies to control for confounding (15). The propensity score is calculated as the probability of receiving one treatment compared with another, i.e., a TNF␣ antagonist versus a nonbiologic DMARD. The probability of treatment is estimated in a logistic regression model that includes a host of covariates that may be related to choice of treatment, mimicking a physician’s decision process. If more covariates can be included in the propensity score model as compared with the conventional outcomes regression model (which is limited by the number of outcomes ), then adjustments in the propensity score may provide better adjustment for otherwise incompletely measured confounders. However, in practice, propensity score analyses have almost never produced results significantly different from those produced by conventional outcomes models in pharmacoepidemiology (15). Other than better control for confounding, there are several other reasons to consider the use of a propensity score model. By plotting the propensity score distributions in patients taking the study treatment in comparison with control patients, one can gain a better appreciation of the distribution of covariates. At times, there is little overlap at the low and/or high end of the propensity score distributions, and trimming the tails of the distribution may be appropriate to focus the analysis on those patients with apparent treatment equipoise in routine care. Finally, when the relative risks differ by propensity score stratum, it can suggest an effect modification that would need further exploration. Another promising analytic method for improving control of confounding is the instrumental variable analysis that has been applied in safety studies of NSAIDs (17). An instrumental variable refers to a variable independently related to treatment choice (i.e., TNF␣ antagonists), but unrelated to confounders or the outcome, other than through the actual treatment (18). Once these assumptions are fulfilled, an instrumental SOLOMON ET AL variable serves as an unconfounded substitute for the actual treatment, which results in an unbiased treatment effect estimate, even if confounders remain unmeasured. The statistical inefficiency of instrumental variable estimates may limit their application in the study of TNF␣ antagonists and infrequent outcomes. Definition of the end point The definition of infection differs dramatically across studies. Different definitions of infection can arise because of variation in prediagnosis surveillance and in the criteria applied for infection. Some studies rely on patient-reported infections, with some attempt to confirm the diagnosis based on physician notes. Other studies use a physician report of infection, with only minimal confirmation based on primary records. Still others use a primary hospital diagnosis of infection, with or without a review of the medical records. None of the studies have optimal sensitivity and specificity for identifying infections. Even in an ideal study setting, infection can be difficult to define, since many types of infections have no standard definition. Generally, in epidemiologic studies with relative risk measures, it is preferred to use outcome definitions that have high specificity, since they result in less biased relative risk estimates even if sensitivity is low (19). The validity of a study’s results can be threatened when there are different surveillance measures or diagnostic criteria applied to patients in various exposure groups. One can imagine that differential surveillance might be performed for patients taking a TNF␣ antagonist versus those taking a nonbiologic DMARD. If a physician and/or patient believes that a TNF␣ antagonist may be related to infection, he or she may be more likely to follow up on symptoms that may or may not be related to infection. This may result in the application of dissimilar diagnostic algorithms being applied to patients taking a TNF␣ antagonist versus those taking a nonbiologic DMARD, leading to an artificial difference in the risk of infection. This potential for surveillance bias is minimized by studying only serious infections that may require hospitalization, a definition that is less discretionary. Time-varying confounding Over time, patients are reevaluated regarding the need for medications, and this can change with variation in disease activity over time. It can be assumed that before every new prescription, such an evaluation has RISK OF INFECTION WITH TNF␣ ANTAGONISTS 925 Table 2. Major cohort types used in studies of tumor necrosis factor ␣ antagonists and infection Cohort type Disease-based registry Drug-based registry Practice-based or populationbased registries Health care utilization data Strengths Weaknesses Diagnosis is usually very accurate; diseasespecific information is very rich; medical records are often available Diagnosis is usually very accurate; diseasespecific information is very rich; medical records are often available Medical records are often available; patients represent those in routine care; often allows for linkage to pharmacy data; often allows for linkage to other registries Patients represent those in routine care; includes linkage to pharmacy data; often very large cohorts can be assembled taken place. Therefore, many potential confounders (i.e., disease activity) can vary in terms of the amount of bias that may be introduced over time. Since disease activity may change over time in RA, either because of treatment or because of the typical oscillations in disease, and disease activity may affect the risk of infection, it may be desirable to control for such changes. However, if a TNF␣ antagonist reduces disease activity, then controlling for the subsequent change in disease activity may “control away” the benefit of the TNF␣ antagonist that may be on the causal pathway to infection. While there are techniques that consider time-varying confounding, there are only rare instances when such methods have been shown to produce substantially different results than those produced with models that include only baseline variables (20). Different data sources There are 4 dominant data sources used in studying the safety of medications: disease-based registries (e.g., the National Data Bank for Rheumatic Diseases or the Consortium of Rheumatology Researchers of North America), drug-based registries (e.g., the British Society of Rheumatology Biologics Register), practice- or population-based registries (e.g., the UK General Practice Research Database or Scandinavian patient cohorts), and health care utilization (“claims”) databases (e.g., Medicare). Each of these different data sources has a set of strengths and weaknesses, many related directly to the methodologic issues described above (Table 2). A disease-based RA registry includes patients with RA whose disease is often diagnosed by specialists. Thus, patients recruited into such databases represent a cohort who may have more severe disease that requires Patients may not represent “typical” cases “Unexposed” patients may not be similar Diagnosis may not be accurate; outcome assessment may not be accurate; diseasespecific information may be lacking Diagnosis may not be accurate; outcome assessment may not be accurate; diseasespecific information may be lacking consultation with a specialist. It is unclear how the sample of referring specialists is recruited and, in turn, how they select their patients. These factors do not affect the validity of the findings from such cohorts, but the generalizability may be more limited. Moreover, the report of specific comorbid conditions and adverse events occurring between scheduled visits may be imprecise, relying on a patient’s self-report. When investigators study such cohorts, they often will attempt to access medical records to confirm important diagnoses, such as an infection or cancer. Drug-based registries are generally established in conjunction with a governmental prescribing program, such as a registry designed to assess the value of TNF␣ antagonist use. Alternatively, they may be established by a drug manufacturer interested in postmarketing safety studies. Although these registries may enroll a fairly complete group of patients, as in registries comprising RA patients taking TNF␣ antagonists, the selection of comparator patients is less clear. In the case of the British Society of Rheumatology Biologics Register, patients who were candidates for a biologic DMARD but opted for a nonbiologic DMARD were eligible for recruitment. However, some of the patients taking nonbiologic DMARDs may have had very different levels of disease activity or a relative contraindication to use of a biologic DMARD, such as a prior malignancy. Confirmation of end points presents the same set of challenges in a drug-based registry as it does in a disease-based registry. Practice- or population-based registries combine the advantages of being large with providing representative populations and fairly detailed clinical data. Sometimes, the full medical record, including complete data on the disease of interest, can be accessed. Access No lag, duration not reported No lag, duration according to supply No lag, duration according to supply plus 90 days No lag, duration according to supply plus 3 half lives No lag, varied duration No lag, duration according to supply plus 90 days Wolfe, 2006 (4) Curtis (9) MTX Nonbiologic DMARD MTX MTX Nonbiologic DMARD No prednisone Nonbiologic DMARD Comparator drug Propensity score with disease severity measures, prednisone use, no comorbidities HAQ scores, disease duration, prednisone use, comorbidities HAQ score, DAS, prednisone use, comorbidities DMARD, TNF␣ No MTX, TNF␣ HAQ score, DAS, prednisone use, comorbidities Comorbidities, prednisone use, health system factors Propensity score, comorbidities, prednisone use, health system factors MTX, TNF␣ No DMARD, TNF␣ Comorbidities, prednisone use, health system factors No MTX, TNF␣ No DMARD, TNF␣ No DMARD, no TNF␣ Control for confounding Drug initiated Median 17 months Varied Mean 15 months with TNF␣, mean 7 months with nonbiologic DMARD Hospitalized with infections, death, or IV antibiotics Median 15 months with TNF␣, median 11 months with nonbiologic DMARD Median 17 months Hospitalized with infections defined by diagnosis codes with primary record confirmation Hospitalized with infections, death, or IV antibiotics Hospitalized with infections defined by validated primary diagnosis code Hospitalized with infections defined by diagnosis codes with primary record confirmation Patient self-report, with some confirmation Reported by study investigators End point assessment Median 30 months 12 months maximum, 74% completed the full 12 months Duration of followup * DMARD ⫽ disease-modifying antirheumatic drug; HAQ ⫽ Health Assessment Questionnaire; DAS ⫽ Disease Activity Score; IV ⫽ intravenous; MTX ⫽ methotrexate. Dixon, 2007 (8) Schneeweiss, 2007 (6) Curtis, 2007 (7) Dixon, 2006 (5) No lag, fixed duration of 365 days Exposure risk window Listing, 2005 (3) First author, year (ref.) Table 3. Summary of the methodology used in major epidemiologic cohort studies of the risk of infection associated with tumor necrosis factor ␣ (TNF␣) antagonists among patients with rheumatoid arthritis* 926 SOLOMON ET AL RISK OF INFECTION WITH TNF␣ ANTAGONISTS to medical records allows for easier confirmation of end points, more reliable information about comorbidities, and often direct pharmacy data about medication-filling histories. This type of cohort registry rarely contains information about the severity of diseases such as RA, unless it is collected routinely and recorded in the medical record. However, laboratory information, such as levels of inflammation markers, may be accessible. Health care utilization databases have been widely used for a variety of pharmacoepidemiologic studies. These databases comprise insurance-billing claims for outpatient medical, hospital, and pharmacy services. Included with such claims are diagnosis and procedure codes that allow one to construct a detailed “virtual” medical record. Since these databases contain routinely collected information, they are relatively easy to construct and may contain many thousands of patients, such as large cohorts of patients with RA who are taking TNF␣ antagonists. Their major drawback is in the limited clinical information available, such as indices of disease severity. Moreover, one must rely on diagnosis and procedure codes provided by physicians and hospitals during routine encounters. Although coding algorithms for many conditions, including infections, have been developed and assessed (21), some misclassification of outcomes is inevitable. Some investigators have augmented the information from these databases by obtaining medical records, to confirm or refute specific infection diagnoses (22). Health system and health care utilization database cohorts have the advantage of being collected in routine practice. This may avoid certain biases that arise through the use of disease- or drug-based registries, when investigators are unblinded to their hypotheses during the collection of data. Routine practice may also include a broader mix of patients, representing both mild cases of RA and cases involving more complex comorbid conditions. Nevertheless, disease- and drugbased registries utilize a specified data collection tool, and thus may have more subtle clinical information that would be important when controlling for confounding. 927 comparator groups, the potential confounders, the exposure risk window, and the study end points. No one study provides a complete picture of the risk of infections associated with TNF␣ antagonists. However, as a group, the epidemiologic studies do provide some important insights. Compared with patients starting methotrexate, those starting a TNF␣ antagonist and continuing treatment for ⬃1 year did not appear to have an associated increased risk of serious infection requiring hospitalization (6). Other studies have shown that during the first 6 months of treatment with a TNF␣ antagonist, patients did appear to be at an increased risk of severe infection requiring hospitalization when compared with patients who were receiving ongoing treatment with methotrexate and who seemed to tolerate this or other DMARDs (3,5,7,9). These studies test both the causal effect of TNF␣ antagonists and the effect of step-up therapy (i.e., adding a TNF␣ antagonist to a DMARD) on infection. Among patients receiving long-term treatment with TNF␣ antagonists as compared with patients receiving glucocorticoids and no DMARD, there was no observed increase in the risk of pneumonia requiring hospitalization (4). Randomized controlled trials and meta-analyses of trials give a limited view of a drug’s safety. Epidemiologic studies designed to assess the safety of a drug in routine care can complement the information provided by randomized controlled trials. The current observational studies of TNF␣ antagonists and infection have important methodologic differences that may explain their apparently discrepant results. As with all science, the “devil is in the details” with epidemiologic studies of the safety of TNF␣ antagonists. AUTHOR CONTRIBUTIONS Dr. Solomon had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. Study design. Solomon, Lunt, Schneeweiss. Acquisition of data. Solomon. Analysis and interpretation of data. Solomon. Manuscript preparation. Solomon, Lunt, Schneeweiss. Statistical analysis. Solomon, Lunt, Schneeweiss. Conclusions Epidemiologic studies on the outcomes of drug therapy are methodologically challenging. The apparent discrepancies in findings across studies of TNF␣ antagonists and infections can be explained by important differences in methodology (Table 3). In part, the various studies are addressing different questions by applying different definitions of the at risk population, the REFERENCES 1. Dixon W, Silman A. Is there an association between anti-TNF monoclonal antibody therapy in rheumatoid arthritis and risk of malignancy and serious infection? Commentary on the meta-analysis by Bongartz et al. Arthritis Res Ther 2006;8:111. 2. Strom BL. Pharmacoepidemiology. Chichester (PA): John Wiley; 2000. 3. Listing J, Strangfeld A, Kary S, Rau R, von Hinueber U, Stoyanova-Scholz M, et al. Infections in patients with rheumatoid 928 4. 5. 6. 7. 8. 9. 10. 11. 12. SOLOMON ET AL arthritis treated with biologic agents. Arthritis Rheum 2005;52: 3403–12. Wolfe F, Caplan L, Michaud K. Treatment for rheumatoid arthritis and the risk of hospitalization for pneumonia: associations with prednisone, disease-modifying antirheumatic drugs, and anti–tumor necrosis factor therapy. Arthritis Rheum 2006;54: 628–34. Dixon WG, Watson K, Lunt M, Hyrich KL, Silman AJ, Symmons DP, on behalf of the British Society of Rheumatology Biologics Register. Rates of serious infection, including site-specific and bacterial intracellular infection, in rheumatoid arthritis patients receiving anti–tumor necrosis factor therapy: results from the British Society for Rheumatology Biologics Register. Arthritis Rheum 2006;54:2368–76. Schneeweiss S, Setoguchi S, Weinblatt ME, Katz JN, Avorn J, Sax PE, et al. Anti–tumor necrosis factor ␣ therapy and the risk of serious bacterial infections in elderly patients with rheumatoid arthritis. Arthritis Rheum 2007;56:1754–64. Curtis JR, Patkar N, Xie A, Martin C, Allison JJ, Saag M, et al. Risk of serious bacterial infections among rheumatoid arthritis patients exposed to tumor necrosis factor ␣ antagonists. Arthritis Rheum 2007;56:1125–33. Dixon WG, Symmons DP, Lunt M, Watson KD, Hyrich KL, British Society for Rheumatology Biologics Register Control Centre Consortium, et al, on behalf of the British Society for Rheumatology Biologics Register. Serious infection following anti–tumor necrosis factor ␣ therapy in patients with rheumatoid arthritis: lessons from interpreting data from observational studies. Arthritis Rheum 2007;56:2896–904. Curtis JR, Xi J, Patkar N, Xie A, Saag KG. Drug-specific and time-dependent risks of bacterial infection among patients with rheumatoid arthritis who were exposed to tumor necrosis factor ␣ antagonists. Arthritis Rheum 2007;56:4226–7. Keane J, Gershon S, Wise RP, Mirabile-Levens E, Kasznica J, Schwieterman WD, et al. Tuberculosis associated with infliximab, a tumor necrosis factor ␣–neutralizing agent. N Engl J Med 2001;345:1098–104. Schneeweiss S. Developments in post-marketing comparative effectiveness research. Clin Pharmacol Ther 2007;82:143–56. Doran MF, Crowson CS, Pond GR, O’Fallon WM, Gabriel SE. 13. 14. 15. 16. 17. 18. 19. 20. 21. 22. Predictors of infection in rheumatoid arthritis. Arthritis Rheum 2002;46:2294–300. Moride Y, Abenhaim L. Evidence of the depletion of susceptibles effect in non-experimental pharmacoepidemiologic research [published erratum appears in J Clin Epidemiol 2004;57:111]. J Clin Epidemiol 1994;47:731–7. Ray WA. Evaluating medication effects outside of clinical trials: new-user designs. Am J Epidemiol 2003;158:915–20. Sturmer T, Joshi M, Glynn RJ, Avorn J, Rothman KJ, Schneeweiss S. A review of the application of propensity score methods yielded increasing use, advantages in specific settings, but not substantially different estimates compared with conventional multivariable methods. J Clin Epidemiol 2006;59:437–47. Cepeda MS, Boston R, Farrar JT, Strom BL. Comparison of logistic regression versus propensity score when the number of events is low and there are multiple confounders. Am J Epidemiol 2003;158:280–7. Schneeweiss S, Solomon DH, Wang PS, Rassen J, Brookhart MA. Simultaneous assessment of short-term gastrointestinal benefits and cardiovascular risks of selective cyclooxygenase 2 inhibitors and nonselective nonsteroidal antiinflammatory drugs: an instrumental variable analysis. Arthritis Rheum 2006;54:3390–8. Brookhart MA, Wang PS, Solomon DH, Schneeweiss S. Evaluating short-term drug effects using a physician-specific prescribing preference as an instrumental variable. Epidemiology 2006;17: 268–75. Kelsey JL, Whittemore AS, Evans AS, Thompson WD. Methods in observational epidemiology. New York: Oxford University Press; 1996. Hernan MA, Brumback B, Robins JM. Marginal structural models to estimate the causal effect of zidovudine on the survival of HIV-positive men. Epidemiology 2000;11:561–70. Schneeweiss S, Robicsek A, Scranton R, Zuckerman D, Solomon DH. Veterans Affairs hospital discharge databases coded serious bacterial infections accurately. J Clin Epidemiol 2007;60:397–409. Curtis JR, Martin C, Saag KG, Patkar NM, Kramer J, Shatin D, et al. Confirmation of administrative claims–identified opportunistic infections and other serious potential adverse events associated with tumor necrosis factor ␣ antagonists and disease-modifying antirheumatic drugs. Arthritis Rheum 2007;57:343–6.