ANTI–TUMOR NECROSIS FACTOR (TNF) THERAPY HAS revolutionized the treatment of rheumatoid arthritis (RA) and other inflammatory diseases over the last decade. However, because TNF has important physiological roles such as host defense and tumor surveillance, anti-TNF therapy has been subject to rigorous postmarketing safety assessment. In this issue of JAMA, Grijalva and colleagues report results from an analysis of data combined from 4 large US administrative databases within the Safety Assessment of Biologic Therapy (SABER) project that question the accepted notion that anti-TNF therapy confers an increased risk of serious infection. Early randomized controlled trials (RCTs) of anti-TNF therapy did not report a significantly increased rate of serious infections, although there were early signals of concern such as a case of tuberculosis in the first infliximab trial. Subsequent RCTs led to the first meta-analysis of infection risk with anti-TNF therapy, which reported a 2-fold increased risk of serious infection. The strength of the trials was randomization, such that any observed difference should be due to the treatment under investigation, not confounders. Despite some differential loss to follow-up in the individual trials, this signal was nonetheless concerning. The inability of individual RCTs to address the safety of rare or long-term outcomes led to the establishment of European drug registers at the launch of anti-TNF therapy, along with other international studies. An analysis from the German biologics register reinforced the meta-analysis results by reporting a 2-fold increased risk of serious infection. However, subsequent studies reported risks ranging from no increased risk to a greater than 4-fold increased risk. Datathat initiallyappearedconflictingslowlyrevealedamore coherent pattern. A time-dependent risk of serious infection was consistent across studies, with an initially high rate of infection thatdeclinedwithprolongeduseofbiologic therapies. Subsequentstudieshavereplicatedthetime-dependentpattern, suggesting the pattern is largely explained by changes in glucocorticoiduse, improvements indiseaseseverity, andadepletion of susceptible patients (whereby high-risk patients have their infections early and discontinue treatment, thus leaving a cohort that is at lower risk of infections). In previous studies, the rate of serious infection was initially higher among patients treated with anti-TNF than the rate in the comparison group (ie, not taking anti-TNF medications), and the rate ratio declined with time. Grijalva et al present the infection rates in the treated and comparison cohorts from 4 large, US automated databases (Figure 2 of their article). Unlike the time-dependent association observed in other studies, the infection rate was comparable throughout the year of follow-up. This finding raises the question of why the time-dependent risk was not observed. The answer may lie in either the infection rates in the cohort receiving anti-TNF treatment or, importantly, the rates in the comparison group. Different rates of infection in comparison groups across studies have accounted for discrepancies in other anti-TNF studies. For their comparison group Grijalva et al selected new users of nonbiologic therapy who did not benefit from methotrexate in the RA cohort, the largest user subgroup in their study. This approach differs from other studies that derived comparison cohorts from prevalent users of diseasemodifying antirheumatic drugs (DMARDs). From a new user comparison cohort, the comparative risks of starting one treatment vs another can be examined. The clinical issue is whether the infection risk would be higher, for example, for a patient starting anti-TNF therapy compared with starting a DMARD such as leflunomide. In the study by Grijalva et al, the absolute rate of infection in the comparison cohort was much higher than in other studies of patients using DMARDs, perhaps reinforcing that the difference may be driven by differences in the comparator. This is possible because the comparison group, like the treatment group in this study, consisted of underserved, vulnerable patients, a population typically excluded from clinical trials. The next issue to consider is whether the observed relative risk is due solely to the difference in medication exposure, or whether the cohorts also differ in other characteristics that may influence their risk of infection. Selection or channeling bias is inevitable in pharmacoepidemiology, as physicians prescribe a particular treatment for a reason. For