Observational Studies of Infections in Rheumatoid Arthritis: A Metaanalysis of Tumor Necrosis Factor Antagonists

MATERIALS AND METHODS

We conducted a systematic literature search to identify observational studies of TNF antagonists and serious infections in RA. For our metaanalysis, inclusion criteria were an observational study design in RA, providing adjusted relative risk (RR) estimates for overall serious infection with anti-TNF exposures. The definition for serious infection was based primarily on need for hospitalization, but was not otherwise specified (i.e., could include mortality or disability).

We used the text-based computerized search and retrieval system of the US National Center for Biotechnology Information, which accesses over 18 million citations from Medline as well as other life science journal citations. Our search strategy used the terms rheumatoid arthritis, infection, risk, tumor necrosis factor (or TNF or biologic), excluding case reports and clinical trials. We also searched abstracts presented 2002–2008 at annual meetings of the American College of Rheumatology and the European League Against Rheumatism. A review of references from primary and review articles was performed to identify additional studies. One reviewer screened all citations (titles/abstracts); those of relevance were subsequently reviewed in full text to identify the final set of studies.

For the final set of studies, we assessed the quality of each (population sampled, outcome definition, analytical approach), and constructed study summaries and a funnel plot of RR versus precision (precision being the inverse of the RR standard errors).

Our metaanalysis computed a summary RR with 95% CI using a random-effects model. This represented a weighted mean of the adjusted RR from the studies; weights were the inverse of total variance (the sum of between-studies and within-study variances)¹. We assessed the consistency of treatment effect across studies using Cochran’s Q test for heterogeneity².

RESULTS

Our electronic search produced 230 citations. Excluding guidelines, commentaries, and studies limited to a single type of infection or on an unrelated topic, we were left with 15 observational studies published as full papers and 2 as abstracts only. The 2 abstracts were excluded, 1 because it included all infections, not just serious infections, and the other because it did not provide adequate summary information about adjusted RR. Of the 15 articles, 1 was excluded because it studied polyarthritis other than RA, and 5 were excluded because they did not provide adjusted RR. Also, in 2 cases, multiple reports came from the same study base. First, Dixon, et al produced 2 cohort studies on the same registry population; since more sophisticated methods and updated data were used in the more recent study, that one was included in our metaanalysis³. Second, Curtis, et al produced 2 reports within the same year⁴, but the later report was a secondary analysis, so only the first report was included.

We thus included 5 cohort^1,3,5–7 and 2 nested case-control^8,9 studies (Table 1). One study presented RR estimates separately for the agents infliximab and etanercept⁶; here, we included a single weighted average in our metaanalysis. All studies (Table 1) were of large size with appropriate design and methodology (e.g., uniform outcome ascertainment). Methods used to control for confounding by indication (i.e., RA disease activity) varied, but all included some adjustment for demographics, disease duration, and other drug exposures, including the important potential confounder of corticosteroid use (potentially correlated with both other medication use and infection risk). A plot of RR versus precision suggested a symmetric, inverted funnel shape, implying no publication bias (Figure 1).

• Download figure
• Open in new tab
• Download powerpoint

Figure 1.

Funnel plot of observational studies of serious infection risk after tumor necrosis factor (TNF) antagonist exposure. Does not include study by Listing, et al⁶, which produced RR estimates separately for different TNF antagonists. Precision: the inverse of the RR standard error.

The Q test of heterogeneity for the RR was nonsignificant; however, we felt the random-effects model was still appropriate, since the test has limited power, and the estimate of variance between studies suggested by our methods¹ was not null. The summary RR suggested about a 40% increased risk of serious infections in patients with RA exposed to TNF antagonists (RR 1.37, 95% CI 1.18, 1.60). Sensitivity analyses, stratifying by whether the studies used administrative data versus patient registries, showed very similar results for administrative data (RR 1.36, 95% CI 1.02, 1.82) compared to patient registries (RR 1.41, 95% CI 1.17, 1.71).

DISCUSSION

To date, metaanalyses of infection incidence with TNF antagonists have focused on RCT data. One RCT metaanalysis stressed that increased infection was seen only when the dose exceeded current recommendations¹⁰. Another RCT metaanalysis demonstrated a pooled OR for serious infection of 2.0 (95% CI 1.3–3.1)¹¹; however, this generated considerable debate regarding methodological limitations, including the absence of data on etanercept (one of the most widely used anti-TNF agents) or adjustment for exposure duration, and that induction doses of infliximab were excessive in some cases. The metaanalysis of Alonso-Ruiz, et al¹² estimated the RR for serious infections with anti-TNF therapy at 1.4 (95% CI 0.8–2.2). This was interpreted as “no significant combined increases in risk,” although the point estimate and the upper limit of the interval obviously do not preclude a real increased risk.

Although our metaanalysis has limitations, we believe they are not greater than those of all recent syntheses (which, to date, have been restricted to RCT data). In all cases, the relevant studies are relatively few. In addition, observational research may include unmeasured confounders. The most troublesome of these, in our case, is RA disease activity. In well done registry studies, carefully collected data on disease activity is believed to be a great advantage over administrative database studies, where attempts to control for disease activity, including propensity scores, are laudable but not ideal. We note, however, that our results seemed relatively consistent across the 2 types of studies. Although most of the recent literature on serious infections in RA has based the outcome definition on requirement for hospitalization, this definition may underestimate the risk for infections that are serious but that do not necessarily require hospitalization; tuberculosis may be a case in point. Hence, the estimate produced in this metaanalysis might be considered conservative.

In metaanalyses, one must always consider the potential of publication bias. Specifically, research results showing an effect are more widely disseminated, and therefore easier to retrieve, than those showing no effect. A plot of the RR and precision estimates for the studies included in our metaanalysis did suggest a symmetric, inverted funnel shape, generally interpreted as implying no publication bias.

There has been great interest recently regarding the timing of infection risk following exposure to TNF antagonists. There are at least 3 studies that highlight high risk in early followup. In the study by Askling, et al⁷, the risk of infection after TNF antagonist exposure was 1.4 (95% CI 1.2, 1.7) in the first year, 1.2 (95% CI 0.9, 1.5) in the second year, and 0.8 (95% CI 0.6–1.1) thereafter. Dixon, et al³, in subanalyses, found that the relative risk for serious infections was 4.6 (95% CI 1.8, 11.9) during the first 90 days of treatment, and Curtis, et al found that any increased risk of serious infections with TNF antagonists was highest in the first 6 months of treatment, and then waned¹³. This phenomenon may represent a depletion of susceptibles, that is, those subjects who are likely to succumb to a serious infection will do so early on, so that RR will be highest in the early phase. Lower disease activity and less glucocorticoid exposure may also result in a real risk decrease over time. Unfortunately, we had inadequate summary data to assess for this phenomenon across studies. Regardless, one would expect if anything that we would be able to show an even higher summary RR, if we had been able to focus on early exposures (e.g., in the first 6 months) across studies.

Some data also suggest that infliximab may heighten infection risk in a more potent way than etanercept. In work by Curtis, et al¹³, infliximab was more strongly associated with serious bacterial infections in the first 6 months of treatment (RR 2.4, 95% CI 1.2–4.7) than etanercept (RR 1.6, 95% CI 0.8–3.8). However, the only registry publication examining the drugs separately contradicts this⁶.

A real advantage of our metaanalysis is that the subjects included are likely more reflective of the overall population with RA, compared to RCT data. Still, some of the results may reflect differential prescription, such that patients receiving TNF antagonists tend to have more severe RA, which itself may be a factor driving increased infection risk¹⁴. Also, our findings must be regarded in light of data emphasizing the very high infection risk conferred by corticosteroids⁸. The steroid-sparing effects of TNF antagonists may thus be important to consider when weighing the risks and benefits of biologic therapies for patients not responding to traditional DMARD.