Methods
Eligibility Criteria
We included all randomized controlled trials that assessed treatments (azathioprine/6-mercaptopurine, methotrexate, infliximab, adalimumab, certolizumab, and vedolizumab) alone or in combination in adult patients with Crohn's disease. We included trials assessing induction of remission of immuno-suppressants between 12 and 17 weeks. We included trials assessing the induction of remission of biologic therapy between 4 and 17 weeks because the onset of action of biologic therapies is more rapid. Trials assessing maintenance of remission had to be at least 24 weeks in duration.
Studies assessing natalizumab were excluded because prescription of natalizumab is restricted to individuals failing immunosuppressive and anti-TNF therapy.11 We also excluded trials studying only pediatric (age, <18 y) or postoperative patients, studies in which the treatment was not fixed (eg, standard of care), studies with a randomized withdrawal design, trials with a cross-over design, studies exclusively assessing fistulizing Crohn's disease, and studies that did not report remission as an outcome. After identification of eligible studies, trials that could not be linked within the network through a shared comparator were excluded.
The primary outcome was remission, which was defined as a Crohn's Disease Activity Index (CDAI) less than 150. When the CDAI was not reported, we used the remission criteria defined in the study. Secondary end points included total withdrawals and withdrawals due to adverse events (WDAEs). Total withdrawals were defined as the total number of patients who were withdrawn from the study for any reason after randomization. WDAEs were collected as defined by the study.
Literature Search and Study Selection
Trials were identified through existing Cochrane systematic reviews and a technical report from the American Gastroenterology Association.10,12–16 We updated the database search from January 2007 (ie, the earliest search date of the Cochrane reviews) to June 2014 in MEDLINE, Embase, and the Cochrane Central register of controlled trials. The database search strategy was adapted from the systematic reviews (the full search strategy is available in the Supplementary Materials and Methods http://dx.doi.org/10.1053/j.gastro.2014.10.011). We supplemented this with a search of trial registries (www.clinicaltrials.gov) and by screening all American College of Gastroenterology, Digestive Disease Week, United European Gastroenterology Week, and European Crohn's and Colitis Organization conference proceedings published from January 2007 through June 2014. For vedolizumab, a separate literature search was performed from 1966 through June 2014 on the earlier-noted research databases. Search results were screened by 2 independent reviewers (A.R., M.B.) first by title and abstract and then by full text. Disagreements were resolved through consensus and discussion with a third reviewer (G.G.K.). Selected studies were reviewed independently by 5 experts in the inflammatory bowel disease literature (R.P., S.G., C.H.S., G.Y.M., and C.A.S.) to confirm the inclusion and exclusion of trials.
Data Collection and Quality Appraisal
Data were abstracted for relevant study characteristics (Supplementary Table 1) and for all primary and secondary outcomes. For induction, remission was extracted at the time of the primary outcome specified in the trial, with the following exception: we used data at or closest to 12 weeks when the time point of the primary outcome was not specified or induction was not the primary goal of the study. For maintenance, remission was extracted at the end of the trial. Total withdrawals and WDAEs were extracted at trial end for both induction and maintenance trials. Data were extracted on a basis of intention-to-treat analysis.
In each trial arm, we abstracted the total number of patients randomized and the total number of patients who experienced the outcome. If only percentages were reported, the number of patients with the outcome was calculated and rounded to the nearest whole number. For data available only in graphic format, images in the highest resolution available were digitized and extracted using the software program Graphclick (version 3.0.2; Arizona Software; www.arizona-software.ch/graphclick/). Any disagreements were resolved through discussion and repeat extraction. The quality of trials was rated through the Cochrane Risk of Bias tool.
Synthesis of Results
For the main analysis we excluded trials with a high risk of bias. Treatment effects for remission and total withdrawals were determined using a random-effects Bayesian network meta-analysis (mixed treatment comparison). A random-effects model was preferred based on the clinical heterogeneity across trials. For WDAEs (a secondary outcome), a fixed-effects model was used because a random-effects model did not produce meaningful results because of the rarity of events.
Statistical heterogeneity was measured by the betweenstudy standard deviation (SD) in log odds ratio (OR). There is no definite threshold for values of SD of the log(OR) that indicates minimal or important heterogeneity, although values between 0.1 to 0.5 are considered small, values between 0.5 and 1.0 are considered fairly high, and values greater than 1.0 indicate extreme heterogeneity.17 Statistical analyses were performed using R statistical software version 2.15.2 with the rjags package version 3.3.0 (www.r-project.org). The Just Another Gibbs Sampler (JAGS) code for the Bayesian network meta-analysis has been published and is provided in the Supplementary Materials and Methods http://dx.doi.org/10.1053/j.gastro.2014.10.011.
Uninformative prior probability distributions were used for all variables. All chains were run with 10,000 burn-in iterations followed by 10,000 monitoring iterations. Convergence was assessed by running 3 chains, inspecting the sampling history plots, and calculating Gelman–Rubin–Brooke statistics.
Summary Measures
For each pair-wise comparison we calculated the odds ratio (OR) with a 95% credible interval (CrI) and the probability that each treatment was superior to the other. We considered a treatment as showing superiority (or inferiority) if the 2-sided 95% CrI of the OR excluded 1, which equates to a 97.5% probability that the treatment is superior.
Additional Analyses
Several sensitivity analyses were performed for induction and maintenance of remission in which we performed the following: (1) included trials with a high risk of bias; (2) excluded trials with 6-mercaptopurine, low azathioprine dosing (<2 mg/kg), and oral methotrexate; and (3) excluded trials in which patients had prior exposure to anti-TNF therapy. For maintenance of remission we performed the following additional sensitivity analyses: (1) included only those trials whose primary outcome was maintenance of remission at 1 year or longer; (2) included only trials that randomized patients after successful induction (in remission or after achieving a decrease in CDAI of 70 or 100 points); and (3) included only treat-through trials. Treat-through trials were defined as randomized controlled trials that randomized active patients at baseline and continued to treat patients for 24 weeks or longer. Also, a network meta-analysis was developed using a fixed-effects model rather than a random-effects model.
To assess whether the effect of treatment depended on the underlying response rate, we performed meta-regression where treatment effects relative to placebo were allowed to depend on the placebo response rates through a common regression coefficient. The meta-regression analysis was restricted to placebo-controlled trials because there were too few head-to-head trials to estimate regression coefficients for each pair-wise comparison. In addition, we performed a random-effects network meta-analysis without metaregression, using only placebo-controlled trials. By comparing results from this analysis with the meta-regression, we evaluated whether differences in the effect of a treatment were owing to the meta-regression or the study selection.
To examine the consistency of the evidence, we compared the estimated treatment effects from the network meta-analysis with standard direct treatment effects where head-to-head trials existed. Direct treatment effects were estimated using a random-effects meta-analysis of odds ratios, with the betweenstudy variance estimated around the Mantel–Haenszel fixed effect, using Review Manager 5 (Version 5.3. Copenhagen: The Nordic Cochrane Centre, The Cochrane Collaboration, 2014).