To assess the association between the initiation of anti–tumor necrosis factor α (anti-TNFα) therapy and the risk of serious bacterial infections in routine care.
To assess the association between the initiation of anti–tumor necrosis factor α (anti-TNFα) therapy and the risk of serious bacterial infections in routine care.
This was a cohort study of patients with rheumatoid arthritis (RA) in whom specific disease-modifying antirheumatic drugs (DMARDs) were initiated. Patients were Medicare beneficiaries ages 65 years and older (mean age 76.5 years) who were concurrently enrolled in the Pharmaceutical Assistance Contract for the Elderly provided by the state of Pennsylvania. A total of 15,597 RA patients in whom a DMARD was initiated between January 1, 1995 and December 31, 2003 were identified using linked data on all prescription drug dispensings, physician services, and hospitalizations. Initiation of anti-TNFα therapy, cytotoxic agents other than methotrexate (MTX), noncytotoxic agents, and glucocorticoids was compared with initiation of MTX. The main outcome measure was serious bacterial infections that required hospitalization.
The incidence of serious bacterial infections was, on average, 2.2 per 100 patient-years in this population (95% confidence interval [95% CI] 2.0–2.4). Glucocorticoid use doubled the rate of serious bacterial infections as compared with MTX use, independent of previous DMARD use (rate ratio [RR] 2.1 [95% CI 1.5–3.1]), with a clear dose-response relationship for dosages >5 mg/day (for ≤5 mg/day, RR 1.34; for 6–9 mg/day, RR 1.53; for 10–19 mg/day, RR 2.97; and for ≥20 mg/day, RR 5.48 [P for trend < 0.0001]). Adjusted models showed no increase in the rate of serious infections among initiators of anti-TNFα therapy (RR 1.0 [95% CI 0.6–1.7]) or other DMARDs as compared with initiators of MTX.
In a large cohort of patients with RA, we found no increase in serious bacterial infections among users of anti-TNFα therapy compared with users of MTX. Glucocorticoid use was associated with a dose-dependent increase in such infections.
The Food and Drug Administration has approved 3 anti–tumor necrosis factor α (anti-TNFα) medications, etanercept (Enbrel; Amgen, Thousand Oaks, CA, and Wyeth, Philadelphia, PA), infliximab (Remicade; Centocor, Malvern, PA), and adalimumab (Humira; Abbott, Abbott Park, IL), for the treatment of rheumatoid arthritis (RA) and other inflammatory arthritides. These agents are among the most widely prescribed biologic immunomodulating drugs and have been used to treat a range of disorders, including RA, psoriasis and psoriatic arthritis, ankylosing spondylitis, and Crohn's disease.
Despite their increasing use, important concerns remain regarding the short- and long-term safety of anti-TNFα therapies, in particular, the risk of infections. While rare but serious opportunistic infections have been linked to anti-TNFα therapy, the association with serious bacterial infections remains a subject of controversy. Even without these treatments, patients with RA, as a result of their underlying condition and their use of other immunosuppressive medications, have ∼2 times the risk of infections compared with non-RA populations (1, 2). Serious bacterial infections (i.e., those requiring treatment with parenteral antibiotics and admission to the hospital), including sepsis, have been reported in patients receiving anti-TNFα therapy; some of these infections have been fatal (3–11).
A meta-analysis of randomized trials assessing the risk of serious infections in patients receiving infliximab or adalimumab found a doubling of the risk of serious nongranulomatous infections (odds ratio 2.0 [95% confidence interval 1.3–3.1]) among users of those 2 anti-TNF agents (12). Two epidemiologic studies of anti-TNFα therapy and infections based on disease registries or biologic agent user registries have found no increase in infection rates (13, 14). Investigators in another study observed a doubling of the risk of severe infections (15).
To help clarify these controversial findings, we sought to assess whether anti-TNF therapy and other disease-modifying antirheumatic drugs (DMARDs) are associated with an increased risk of serious bacterial infections in elderly, low-income patients with RA.
We conducted a cohort study of Medicare beneficiaries ages 65 years and older with RA who initiated use of a DMARD, including anti-TNFα and glucocorticoids, between 1995 and 2003. Patients were concurrently enrolled in the Pharmaceutical Assistance Contract for the Elderly (PACE) provided by the state of Pennsylvania. To be eligible for PACE, annual income must be less than $13,000 if single and less than $16,200 if married, but income must not be low enough to qualify for Medicaid. The PACE program provided full coverage of all study medications, with a $6 copayment. The PACE database provides detailed information on prescription drug use and was linked to the corresponding Medicare claims data for all study patients.
To be eligible for the cohort, patients had to demonstrate use of the health care system by filling at least 1 prescription for any drug and having at least 1 physician service in each of 2 consecutive 6-month periods in addition to being enrolled in the PACE program. Patients were identified as having RA if, at 3 physician visits, they had a diagnosis of RA according to the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) code, a definition that resulted in prevalence estimates similar to population-based studies using primary data collection to confirm a diagnosis of RA (16). Patients were considered for this study only after the third diagnosis of RA. Patients with any cancer (except nonmelanoma skin cancer) or human immunodeficiency virus/acquired immunodeficiency syndrome, as identified by at least 2 physician services with a corresponding ICD-9-CM diagnosis, were excluded from the analysis.
Within this large cohort of patients with RA, we identified each patient's date of initiating DMARD therapy (index date). To be categorized as initiating a specific DMARD therapy, at least 6 prior months must have elapsed without the use of that agent. Followup time started on the date of the index dispensing of the DMARD. Followup ended at the earliest of 1) discontinuation of the drug, 2) first occurrence of a study infection, or 3) end of the study period. Patients could initiate use of >1 DMARD as well as multiple courses of the same DMARD during the study period. Such patients could contribute a maximum of 1 outcome during each course of therapy.
To protect patient confidentiality, all traceable personal identifiers were removed from the data set prior to analysis. The Institutional Review Board of the Brigham and Women's Hospital approved this study, and signed data use agreements were obtained.
Initiation of methotrexate (MTX) therapy served as the reference drug exposure. Other drug categories we studied included anti-TNFα therapy (etanercept, infliximab, and adalimumab), other cytotoxic DMARDs (leflunomide, cyclosporine, and azathioprine), noncytotoxic DMARDs (sulfasalazine, gold, penicillamine, hydroxychloroquine, and minocycline), and oral glucocorticoids. The index date for each drug was the date of first dispensing. If patients started 2 of the study drugs on the same day, each was considered an episode of treatment initiation and the other drug would be considered a concurrently used DMARD.
To allow for continued immunosuppressive effects of these drugs even after dosing may have stopped, we calculated the end of exposure based on approximate half-lives. All oral DMARDs, including MTX and glucocorticoids, were assigned 3 weeks of exposure beyond the number of days of the supply (17). For injectable drugs, discontinuation dates were calculated as follows. Within the exposure category of anti-TNFα therapy, we assumed that each infliximab infusion provided exposure for 18 weeks, each dispensed adalimumab syringe provided exposure for 6 weeks, and each etanercept dose provided exposure for 10 days. Similarly, each gold injection was assumed to provide exposure for a maximum of 3 months after the last dispensing.
To assess whether the duration of use modifies the effect of DMARD treatment on serious infections, we stratified the analysis into short-term effects (1–90 days after initiation) and longer-term effects (>90 days). For this analysis, all infection outcomes were combined to increase the number of outcomes in each time period.
The oral dose of glucocorticoids was grouped into 4 categories, as follows: very low dose (≤5 mg of prednisone equivalents per day), low dose (6–9 mg/day), medium dose (10–19 mg/day), and high dose (≥20 mg/day).
Concomitant use of other DMARDs in addition to the index drug was assessed on the index date. If, for example, a patient added infliximab to an ongoing MTX regimen, then the initiation of infliximab would be the index exposure while concomitant MTX therapy would be adjusted in the analysis. Conversely, if a patient added MTX to ongoing infliximab therapy, then the initiation of MTX would be considered the index exposure.
The study outcome was the first occurrence of a hospitalization for a serious bacterial infection during a course of therapy, with the following diagnoses as the primary or first-listed diagnosis: septicemia or bacteremia, pneumonia, osteomyelitis, pyelonephritis, meningitis, encephalitis, or endocarditis. Due to the small number of events, only septicemia or bacteremia, pneumonia, and osteomyelitis were considered as separate event categories. All events were also grouped into a composite outcome “any serious bacterial infection.” For each end point, the event date was the date of hospital admission.
In a validation study of 158 patients who were hospitalized for any of the above diagnoses in Department of Veterans Affairs hospitals in the northeastern US between 2001 and 2004, we found a positive predictive value of >80% for a set of predefined ICD-9-CM codes (Table 1), using detailed medical records as the gold standard method. The positive predictive value increased to >90% if the outcome definition was “any bacterial infection leading to hospitalization,” rather than a specific diagnosis (18).
|Discharge diagnosis||ICD-9-CM code|
|Septic arthritis||711.0x, excluding 711.08*|
|Osteomyelitis||730.0x, 730.1x, 730.2x|
|Septicemia or bacteremia||038.x, 790.7|
|Other opportunistic infections|
|Pneumocystis jiroveci pneumonia||136.3|
An additional outcome was hospitalization because of opportunistic infections, including pulmonary tuberculosis, atypical mycobacteria, cryptococcosis, aspergillosis, histoplasmosis, listeriosis, leishmaniasis, and Pneumocystis jiroveci pneumonia. The ICD-9-CM codes for this set of diagnoses are also shown in Table 1. Validation data indicated good validity of these diagnoses as well, with a positive predictive value of 79%.
A set of potential confounders was measured based on health care utilization data within 180 days prior to the initiation of the index drug (index date). These included sociodemographic characteristics (age, sex, race, nursing home residence), generic markers of comorbidity that have shown good validity (19, 20) (hospitalization for any reason, number of physician visits, number of distinct prescription drugs excluding the DMARDs listed above, Charlson comorbidity score ), markers of RA severity (number of orthopedic procedures, number of intraarticular procedures, number of tests for markers of inflammation ordered, number of extraarticular manifestations of RA) (22), as well as independent predictors of serious infections (diabetes mellitus, numbers of previous outpatient or inpatient visits because of infections), any previous antibiotic use, influenza vaccination, and pneumococcal vaccination.
To address the potentially nonlinear age dependency of serious infections (23), we also included a quadratic age term in the model. Calendar time was modeled by including multiple indicator terms to adjust for possible changes in DMARD choice during the study period and possible changes in the prevention of serious infections in this population (24).
We calculated incidence rates of serious bacterial infections leading to hospital admission for all incident users of specific DMARDs. We used MTX exposure as a common reference group for calculating rate ratios (RRs) and their 95% confidence intervals (95% CIs) for 4 drug exposure contrasts: 1) anti-TNFα versus MTX, 2) other cytotoxic DMARDs versus MTX, 3) noncytotoxic DMARDs versus MTX, and 4) glucocorticoids versus MTX. RRs were adjusted for age and sex by fitting a multivariate Cox proportional hazards regression (25). Further adjustment for the baseline confounders listed above was attempted using backward-selection algorithms with an alpha level of 0.2 as the inclusion criterion.
In studies with very few outcomes, multivariate regression analyses tend to overfit the data, resulting in biased risk estimates (26). We therefore developed propensity score models for more efficient estimation (27, 28). Propensity scores were estimated using logistic regression models predicting the probability of using a specific DMARD compared with MTX as a function of all potential confounders listed above (C statistics, a marker for predicting treatment choice, varied between 0.65 and 0.75). Propensity score estimates for each patient were entered into the Cox regression as a continuous variable. Including propensity score quintiles did not improve the model fit (P > 0.1). The proportional hazards assumption was evaluated by visual inspection of log-cumulative hazard curves.
Less than 10% of the patients were members of >1 drug-initiator cohort, since patients were allowed to initiate >1 DMARD during the study period and, for that reason, <1% of events were attributed to the same patient twice. We therefore repeated the analyses with an adjustment of standard errors for repeated observations (29).
Important predictors of RA severity may not be fully captured in health care utilization databases. If RA severity were associated with an increased rate of infection and biologic DMARD users were more likely to have severe disease, then this could result in an overestimation of the rate ratio of serious infections among users of biologic agents. We used sensitivity analyses to quantify the extent of such residual confounding as a function of these associations (30).
We identified 1,900 initiations of MTX therapy, 469 initiations of anti-TNFα, 654 initiations of other cytotoxic DMARDs, 1,957 initiations of noncytotoxic DMARDs, and 10,617 initiations of glucocorticoids. Together, these 15,597 courses of therapy contributed 5,676 patient-years of exposure. The average followup time varied among drug exposure groups, from 0.20 years to 1.29 years (Table 2); 4.4% of patients were censored because of disenrollment from the PACE plan. Patients initiating anti-TNFα therapy were most likely to have had a previous intraarticular procedure, a previous orthopedic surgery, and measurement of markers of inflammation during the previous 180 days (Table 2), all of which are indicators of the likely presence of more serious RA. The same patients were also more likely to have received an influenza vaccination and less likely to reside in nursing homes, both of which are markers of susceptibility for infections.
|MTX||TNFα antagonists||Cytotoxic DMARDs||Noncytotoxic DMARDs|
|No. of treatment episodes||1,900||469||654||1,957||10,617|
|Followup, mean years||0.58||1.29||0.64||0.73||0.20|
|Age, mean ± SD years||76 ± 6.2||75 ± 5.8||76 ± 5.8||76 ± 6.0||79 ± 6.4|
|No. (%) female||1,672 (88)||427 (91)||593 (91)||1,743 (89)||9,394 (88)|
|Ethnicity, no. (%)|
|White||1,745 (92)||430 (92)||610 (92)||1,816 (92)||9,830 (93)|
|Black||139 (7)||32 (7)||139 (7)||128 (7)||677 (6)|
|Other||16 (1)||7 (1)||6 (1)||13 (1)||7 (1)|
|Nursing home resident, no. (%)||125 (7)||20 (4)||29 (7)||95 (5)||749 (7)|
|Any hospitalization prior to index date, no. (%)||480 (25)||98 (21)||147 (22)||463 (24)||3,180 (30)|
|Charlson comorbidity score, mean ± SD||1.0 ± 1.5||0.5 ± 1.2||0.6 ± 1.4||1.1 ± 1.5||1.3 ± 1.9|
|No. of physician visits, mean ± SD||6.7 ± 5.3||7.7 ± 5.5||7.3 ± 6.0||6.7 ± 5.3||6.0 ± 4.9|
|No. of different non-DMARDs used before index date, mean ± SD||6.6 ± 4.4||8.2 ± 5.1||8.2 ± 4.9||6.8 ± 4.5||8.4 ± 4.8|
|No. of different DMARD classes used before index date, mean ± SD†||0.6 ± 0.7||1.4 ± 1.0||1.1 ± 0.9||0.5 ± 0.6||0.3 ± 0.5|
|Any intraarticular procedures, no. (%)||687 (36)||214 (46)||224 (34)||656 (34)||2,505 (24)|
|Any extraarticular manifestations, no. (%)||51 (3)||21 (4)||27 (4)||60 (3)||234 (2)|
|At least 1 marker of inflammation tested during previous 180 days, no. (%)||696 (37)||184 (39)||234 (36)||644 (33)||2,173 (20)|
|Any orthopedic surgeries, no. (%)||775 (41)||236 (50)||250 (38)||732 (37)||3,012 (28)|
|Diabetes mellitus, no. (%)||165 (8)||41 (9)||67 (10)||165 (8)||176 (9)|
|At least 1 antibiotic dispensed during previous 180 days, no. (%)||580 (30.5)||188 (40)||279 (43)||581 (30)||4,279 (40)|
|Previous hospitalization for serious bacterial infections, no. (%)||56 (3)||8 (2)||18 (3)||58 (3)||360 (3)|
|Vaccinations, no. (%)|
|Influenza||148 (8)||94 (20)||78 (12)||137 (7)||1,271 (12)|
|Pneumococcal||65 (3)||20 (4)||16 (2)||61 (3)||325 (3)|
There was no meaningful difference in other markers of health among the different drug treatment groups, including the Charlson comorbidity score, number of previous physician visits, and number of different non-DMARD drugs used (Table 2). Other markers of susceptibility for bacterial infections, including diabetes mellitus, use of antibiotics during previous 180 days, and previous hospitalization for serious bacterial infections, were similar among the treatment groups.
Concomitant use of other DMARDs was most frequently observed in patients initiating anti-TNFα therapy. Among these patients, 84% also took at least 1 other DMARD. In 54% of these patients, MTX was taken (Table 3), while glucocorticoids were more likely to be taken as monotherapy (75% of patients).
|Index DMARD||Concomitant medications on index date|
|None||MTX||TNFα antagonists||Glucocorticoids||Cytotoxic DMARDs||Noncytotoxic DMARDs||Total|
|MTX||979 (52)||–||68 (4)||580 (31)||69 (4)||432 (23)||1,900 (100)|
|TNFα antagonists||77 (16)||253 (54)||–||209 (45)||83 (18)||112 (24)||469 (100)|
|Glucocorticoids||7,939 (75)||1,554 (15)||177 (2)||–||272 (3)||1,140 (11)||10,617 (100)|
|Cytotoxic DMARDs||184 (28)||243 (37)||45 (7)||243 (37)||–||183 (28)||654 (100)|
|Noncytotoxic DMARDs||1,212 (62)||332 (17)||14 (1)||492 (25)||45 (2)||–||1,967 (100)|
Incidence rates of serious bacterial infections in patients who began any DMARD or a glucocorticoid was, on average, 2.2 per 100 patient-years (95% CI 2.0–2.4) and ranged between 0.3 per 100 patient-years to 4.2 per 100 patient-years (Table 4). The rate of septicemia or bacteremia was 3.9 per 100 patient-years; it was 0.7 for osteomyelitis, and 2.0 for pneumonia in all drug user categories combined.
|Drug exposure, outcome||No. of patient-years||No. of events||Event rate per 100 patient-years (95% CI)||Unadjusted rate ratio (95% CI)†|
|Septicemia or bacteremia||1,091.5||24||2.20 (1.33–3.07)||–|
|Any of the above||1087.6||41||3.77 (2.64–4.90)||–|
|Any bacterial infection||1087.6||41||3.77 (2.64–4.90)||–|
|Pneumonia||600.7||14||2.33 (1.12–3.54)||1.59 (0.78–3.26)|
|Septicemia or bacteremia||602.2||13||2.16 (1.00–3.32)||0.98 (0.50–1.93)|
|Osteomyelitis||606.7||3||0.49 (0.00–1.05)||0.77 (0.20–2.99)|
|Any of the above||595.3||28||4.70 (3.00–6.40)||1.25 (0.77–2.02)|
|Any bacterial infection||593.3||29||4.89 (3.15–6.62)||1.30 (0.81–2.09)|
|Pneumonia||2,117.2||67||3.16 (2.41–3.91)||2.16 (1.25–3.72)|
|Septicemia or bacteremia||2,097.3||133||6.34 (5.30–7.38)||2.88 (1.87–4.45)|
|Osteomyelitis||2,123.4||17||0.80 (0.42–1.18)||1.25 (0.52–3.02)|
|Any of the above||2087.4||187||8.96 (7.73–10.2)||2.38 (1.69–3.33)|
|Any bacterial infection||2086.6||196||9.39 (8.14–10.6)||2.49 (1.78–3.49)|
|Pneumonia||419.3||6||1.43 (0.29–2.57)||0.98 (0.38–2.49)|
|Septicemia or bacteremia||409.4||15||3.66 (1.84–5.48)||1.67 (0.87–3.18)|
|Osteomyelitis||418.9||2||0.48 (0.00–1.14)||0.75 (0.15–3.59)|
|Any of the above||410.4||21||5.12 (3.00–7.25)||1.36 (0.80–2.30)|
|Any bacterial infection||410.2||22||5.36 (3.18–7.54)||1.42 (0.85–2.39)|
|Pneumonia||1,432.5||13||0.91 (0.42–1.40)||0.62 (0.30–1.29)|
|Septicemia or bacteremia||1,429.5||33||2.31 (1.53–3.09)||1.05 (0.62–1.78)|
|Osteomyelitis||1,432.2||9||0.63 (0.22–1.04)||0.98 (0.37–2.64)|
|Any of the above||1,414.8||52||3.68 (2.70–4.66)||0.98 (0.65–1.47)|
|Any bacterial infection||1,412.7||53||3.75 (2.76–4.74)||0.99 (0.66–1.49)|
With MTX therapy as the common reference group, glucocorticoid users had a 2-fold increase in the incidence of serious bacterial infections (RR 2.1); the risk of septicemia or bacteremia was particularly pronounced (RR 2.5) after adjusting for potential confounders using propensity score analyses (Table 5). We found no increased rate of serious bacterial infections for those who initiated anti-TNFα therapy (RR 1.0) or any other DMARDs compared with MTX. When compared with the 979 patients initiating MTX monotherapy, point estimates did not change meaningfully or in a systematic way. Standard errors changed by <5% after adjusting for repeated measures.
|Drug exposure, outcome||Compared with MTX initiation, as monotherapy or combined with other DMARDs (n = 1,900)||Compared with MTX monotherapy (n = 979), propensity score adjusted rate ratio (95% CI)‡|
|Age- and sex-adjusted rate ratio (95% CI)||Multivariate adjusted rate ratio (95% CI)†||Propensity score adjusted rate ratio (95% CI)‡|
|Pneumonia||1.38 (0.66–2.86)||0.88 (0.41–1.90)||0.72 (0.32–1.87)||0.78 (0.28–2.13)|
|Septicemia or bacteremia||1.03 (0.52–2.06)||1.16 (0.58–2.34)||1.28 (0.62–2.66)||1.36 (0.55–3.36)|
|Osteomyelitis||1.04 (0.25–4.30)||1.04 (0.25–4.30)||1.07 (0.24–4.77)||0.91 (0.14–5.79)|
|Any of the above||1.16 (0.70–1.91)||1.00 (0.60–1.67)||0.97 (0.57–1.65)||0.87 (0.45–1.68)|
|Any bacterial infection||1.21 (0.74–1.97)||1.04 (0.63–1.72)||1.01 (0.60–1.70)||0.94 (0.49–1.80)|
|Pneumonia||2.35 (1.32–4.19)||2.07 (1.15–3.73)||1.93 (1.07–3.47)||2.35 (1.04–5.31)|
|Septicemia or bacteremia||2.88 (1.83–4.53)||2.57 (1.63–4.08)||2.51 (1.58–3.97)||2.91 (1.57–5.38)|
|Osteomyelitis||1.53 (0.57–4.10)||1.35 (0.50–3.66)||1.29 (0.47–3.52)||1.56 (0.43–5.67)|
|Any of the above||2.50 (1.75–3.57)||2.20 (1.53–3.17)||2.11 (1.47–3.03)||2.27 (1.42–3.65)|
|Any bacterial infection||2.54 (1.79–3.62)||2.25 (1.57–3.22)||2.14 (1.50–3.06)||2.34 (1.46–3.75)|
|Pneumonia||0.98 (0.38–2.53)||0.69 (0.26–1.81)||0.69 (0.26–1.84)||0.59 (0.18–1.99)|
|Septicemia or bacteremia||1.50 (0.78–2.91)||1.75 (0.88–3.47)||1.71 (0.85–3.42)||1.76 (0.74–4.19)|
|Osteomyelitis||1.04 (0.21–5.27)||0.94 (0.19–4.74)||0.99 (0.19–5.25)||2.03 (0.27–15.2)|
|Any of the above||1.29 (0.75–2.22)||1.33 (0.76–2.33)||1.21 (0.69–2.12)||1.16 (0.58–2.31)|
|Any bacterial infection||1.32 (0.78–2.23)||1.47 (0.85–2.54)||1.29 (0.74–2.23)||1.31 (0.66–2.60)|
|Pneumonia||0.61 (0.29–1.28)||0.58 (0.27–1.24)||0.67 (0.32–1.40)||0.65 (0.26–1.67)|
|Septicemia or bacteremia||1.05 (0.36–2.89)||0.95 (0.56–1.62)||0.94 (0.33–2.68)||1.06 (0.54–2.06)|
|Osteomyelitis||1.02 (0.48–3.47)||1.02 (0.36–2.89)||1.31 (0.48–3.53)||1.03 (0.27–3.89)|
|Any of the above||0.84 (0.55–1.30)||0.80 (0.52–1.23)||0.84 (0.54–1.29)||0.78 (0.46–1.31)|
|Any bacterial infection||0.96 (0.64–1.45)||0.90 (0.60–1.37)||0.95 (0.63–1.44)||0.91 (0.55–1.53)|
Since anti-TNFα therapy is often reserved for the most severe cases of RA, incompletely adjusted RA disease severity would bias these results toward a higher rate ratio, assuming that RA severity would be associated with a higher incidence of serious infections. A quantitative sensitivity analysis showed that it is unlikely that the lack of an association between anti-TNFα therapy and serious bacterial infections can be explained by residual confounding. Assuming that the prevalence of severe RA, which is imperfectly measured in our data, was 20% higher in patients exposed to anti-TNFα and if RA severity were to increase the rate of serious bacterial infections 3.5-fold, then the fully adjusted RR would have been 0.9, in a direction away from a harmful effect of anti-TNFα therapy (Appendix A).
The increased rate of serious infections was more pronounced during the first 90 days after initiation of treatment with glucocorticoids (RR 2.99 [95% CI 1.60–5.60]) and cytotoxic DMARDs (RR 2.99 [95% CI 1.18–7.52]) than after that period (Table 6). Of the 654 treatment initiations with a cytotoxic DMARD, 388 were leflunomide. Initiation of leflunomide was not associated with serious bacterial infections after propensity score adjustment (RR 0.7 [95% CI 0.3–1.6]). Similarly, the effect of leflunomide within the first 90 days after initiation showed an RR of 1.0, with a wide 95% confidence interval of 0.2–5.2.
|Drug exposure||Propensity score adjusted rate ratio (95% CI)|
|Short-term effects (1–90 days after treatment initiation)||Longer-term effects (>90 days after treatment initiation)|
|TNFα antagonists||1.39 (0.40–4.83)||0.90 (0.52–1.56)|
|Glucocorticoids||2.99 (1.60–5.60)||1.20 (0.79–1.84)|
|Cytotoxic DMARDs||2.99 (1.18–7.52)||0.95 (0.51–1.79)|
|Non-cytotoxic DMARDs||1.11 (0.50–2.50)||1.51 (0.96–2.36)|
Among the 10,617 initiations of glucocorticoids, 17% were at an average dosage of ≤5 mg/day, 14% were at 6–9 mg/day, 42% were at 10–19 mg/day, and 27% were at ≥20 mg/day (Table 7). Very low glucocorticoid dosages of ≤5 mg/day were not associated with serious bacterial infections, but higher doses showed a dose-dependent increase in the rate ratio up to an RR of 5.5 for users of ≥20 mg/day (P for trend < 0.0001).
|Mean daily dose of glucocorticoids (no. of treatment episodes), outcome||Propensity score adjusted rate ratio (95% CI)|
|≤5 mg (n = 1,781)|
|Septicemia or bacteremia||1.68 (0.98–2.87)|
|Any of the above||1.37 (0.85–2.21)|
|Any bacterial infection||1.34 (0.85–2.13)|
|6–9 mg (n = 1,510)|
|Septicemia or bacteremia||1.61 (0.84–3.09)|
|Any of the above||1.64 (1.00–2.69)|
|Any bacterial infection||1.53 (0.95–2.48)|
|10–19 mg (n = 4,435)|
|Septicemia or bacteremia||3.15 (1.76–5.65)|
|Any of the above||2.86 (1.80–4.56)|
|Any bacterial infection||2.97 (1.89–4.68)|
|≥20 mg (n = 2,891)|
|Septicemia or bacteremia||6.83 (3.68–12.7)|
|Any of the above||5.32 (3.18–8.90)|
|Any bacterial infection||5.48 (3.29–9.11)|
Since physicians may have different thresholds for hospitalization because of bacterial infections based on the medication regimens prescribed, we also assessed the association between DMARD use and the initiation of outpatient antibiotic therapy. After adjusting for previous antibiotic therapy and other risk factors, we observed a doubling in the incidence of outpatient antibiotic use among glucocorticoid users compared with MTX users (RR 1.9 [95% CI 1.8–2.1]), but we did not observe such a relationship for other DMARDs.
The incidence rate of opportunistic infections requiring hospitalization was, on average, 0.2 per 100 patient-years (95% CI 0.1–0.3). There were 5 hospitalizations for aspergillosis, 4 events of Pneumocystis jiroveci pneumonia, and 2 hospitalizations for pulmonary tuberculosis. The very small number of events of serious bacterial infections precluded any comparison among DMARD groups.
Even though opportunistic infections have been linked to anti-TNFα therapy, the association with serious bacterial infections, a far more common cause of morbidity in this population, has not been thoroughly assessed. Since bacterial infections are so common and potentially serious and since patients with RA appear to be particularly susceptible to them, this potential adverse event warrants attention.
In a large cohort study of Medicare beneficiaries, we found no increase in the rate of bacterial infections among initiators of anti-TNFα therapy compared with initiators of MTX therapy. This finding was independent of concomitant DMARD use and persisted after adjusting for other predictors of infections. The rate of opportunistic infections that required inpatient treatment was much lower in this population and limited drug-specific analyses.
Consistent with the results of randomized trials (31), we found that initiators of glucocorticoids were at twice the rate of hospitalization for serious bacterial infections as compared with initiators of MTX, independent of the use of other DMARDs. The same doubling of the incidence rate was observed for initiation of outpatient antibiotic regimens. Very low doses of glucocorticoids had no observable effect on bacterial infections in our study, but once a threshold of 5 mg/day was surpassed, we started to observe a dose-response relationship, which further supports a causal relationship between glucocorticoid use and the incidence of serious bacterial infections. This threshold for safe low-dose glucocorticoid use with regard to outcomes of infection corroborates the findings of a meta-analysis of randomized trials (31).
Our results are also consistent with another nonrandomized study of glucocorticoids that found no increase in the rate of serious infections. The same study, however, reported an increase in the rate of cellulitis, an outcome not included in this study because it was not specifically recorded in the claims database (14). A recent nonrandomized study of >850 patients receiving anti-TNFα therapy as compared with a mixture of other DMARD users retrospectively evaluated adverse drug effects, including infections (15). In a propensity score–adjusted analysis, the authors reported a 2-fold increase in serious infections for etanercept (RR 2.2) and infliximab (RR 2.1). However, the outcome assessment was performed by the treating physician, who was aware of the patient's drug exposure status, and this may have contributed to a higher estimate (30).
A recent meta-analysis on the effects of anti-TNFα therapy (12) excluded etanercept and included 2 studies (32, 33) that compared adalimumab monotherapy with placebo (RR 6.3 and RR 15.3). Such an increase in the rate of serious infections is not unexpected when comparing an immunosuppressive agent with placebo; however, “no treatment” is an unrealistic strategy in clinical practice. The 4 largest trials alone, excluding placebo-controlled studies, had RR estimates of 0.76 (95% CI 0.30–2.18) (34), 0.66 (95% CI 0.14–2.83) (35), 2.68 (95% CI 1.11–7.81) (36), and 2.1 (95% CI 0.8–6.4) (37) and would result in a pooled estimate of RR 1.78 (95% CI 1.1–2.8). This is somewhat lower than the 2.0 RR value reported in the meta-analysis, but it is still higher than the effect observed in our study.
The rate of serious bacterial infections in our population is similar to the rates reported in a population-based study of hospitalizations for serious infections conducted by Doran et al (1), based on medical records abstraction. Those authors reported a rate of infections requiring hospitalization of 0.8 per 100 person-years for septicemia or bacteremia, 0.2 for osteomyelitis, and 3.1 for pneumonia, which is slightly lower than the rates we found for septicemia or bacteremia (3.9 per 100 person-years) and osteomyelitis (0.7) and slightly higher than the rate for pneumonia (2.0). This may be due to the older age of our population. We observed a high proportion of patients identified as having RA that were treated with glucocorticoids only. While this patient group could be contaminated by some patients who have osteoarthritis that was misclassified as RA, it is compatible with treatment patterns of RA observed in population-based studies outside of academic centers (38).
Our findings are consistent with at least 2 hypotheses: first, that in routine care, physicians seem to be able to manage RA well with anti-TNFα therapy with regard to the risk of serious infections; and second, that improvements in patient functioning with anti-TNFα and MTX therapy may counteract some of the immunosuppressive effects of these drugs. Studies specifically designed to test these hypotheses will be necessary to better understand any such mechanisms. Our data do not rule out the possibility that both MTX and anti-TNFα therapy produce an increased rate of infections as compared with no treatment, with neither treatment showing a higher rate than the other.
Epidemiologic studies using health care utilization data are particularly scrutinized for their limited control of confounding and their potential for misclassifying the diagnoses (39). We found little evidence for confounding; adjustments for age and sex produced little change in the rate ratio estimates as compared with adjustments for a range of covariates in traditional multivariate regression or propensity score analyses. All such changes were well within the 95% CIs. The potential for confounding was partially reduced by choosing as the common reference group the initiators of MTX therapy. These patients have similar distributions of the measured covariates as compared with other drug user categories, particularly anti-TNFα therapy. A quantitative sensitivity analysis showed that residual confounding by incomplete measurement of RA severity would bias our results toward an increased effect estimate.
We limited our definition of serious infection to primary diagnoses of an infectious disease that required hospitalization. This reduced subjectivity in the assessment of severity, although it does not fully rule out individual differences in physicians' thresholds for admitting such patients to the hospital for treatment. Our validation study of diagnostic codes showed a robust positive predictive value for specific diagnoses (18), which further increased when the definition of infection outcome was relaxed to include any infectious activity as the primary reason for the hospitalization. A high positive predictive value or a high specificity minimizes bias in ratio measures such as the reported rate ratios (40). The reported data were generated prospectively in routine care settings without knowledge of the study hypothesis, which minimizes the potential for observer bias.
The relatively short mean duration of followup for a chronic condition such as RA may appear to be a limitation of our study. Most patients dropped out of the study because of treatment discontinuation or because they developed a study outcome; only 4.4% of the patients dropped out because of disenrollment from the PACE plan. Studies of the effect of long-term DMARD treatment are necessarily limited to highly selective patient populations with good tolerance and adherence to DMARD regimens (39). We conclude that neither confounding nor outcome misclassification is a likely explanation for the results of our study. This conclusion is confirmed by the fact that we could reproduce randomized trial findings of a dose-related association between glucocorticoid use and serious infections above a threshold dosage of 5 mg/day.
In summary, we found no increase in the rate of serious bacterial infections in users of anti-TNFα as compared with users of MTX. We found a doubling of the rate with systemic glucocorticoid use, however, with a clear dose-response relationship.
Dr. Schneeweiss 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. Schneeweiss, Setoguchi, Weinblatt, Avorn, Solomon.
Acquisition of data. Schneeweiss, Avorn, Levin, Solomon.
Analysis and interpretation of data. Schneeweiss, Setoguchi, Weinblatt, Katz, Avorn, Sax, Levin, Solomon.
Manuscript preparation. Schneeweiss, Setoguchi, Weinblatt, Katz, Sax, Solomon.
Statistical analysis. Schneeweiss, Setoguchi, Levin, Solomon.
We thank Drs. Ari Robicsek, Richard Scranton, and Dan Zuckerman for their help in validating the diagnostic codes for serious infections.
We performed a sensitivity analysis of residual confounding as a function of the association of RA disease severity with anti-TNFα therapy and as a function of the association of RA disease severity with increased rates of infection. We used the following definitions (30) and applied the assumptions described:
ARR represents the apparent, or observed, relative risk of anti-TNFα use. Our study found an ARR of 1.0. For rare outcomes, we can assume that the relative risk estimate closely approximates the rate ratio.
RRCD represents the independent effect of the unmeasured confounder on the disease outcome. We varied the association between RA severity and serious infection outcomes from an RR of 1.0 (no association) to an RR of 5.5 (very strong association).
PC0 represents the prevalence of the confounder in the patients not exposed to drug. Among the RA patients who were not exposed to anti-TNFα, we assumed that 50% had severe disease and 50% did not.
PC1 represents the prevalence of the confounder in the drug-exposed patients. Among the patients who were exposed to anti-TNFα, we varied the prevalence of severe RA from 50% to 100%. This reflects a preference for prescribing anti-TNFα to patients with increasingly severe RA.
As demonstrated in Figure A, in this scenario the fully adjusted (or “corrected”) RR would be <1 if a binary marker for severe RA had not been adjusted for in our main analysis; that is, we observed an adjusted RR of 1, whereas the RR would have been <1 had we been able to better adjust for RA severity. The fully adjusted RR would have been 0.9 if the prevalence of severe RA were 20% higher in the patients exposed to anti-TNFα therapy and if RA severity were to increase the risk of serious bacterial infections 3.5-fold.