Treatment sequence network meta-analysis in Crohn’s disease: a methodological case study

Abstract

Objective: Several biologic therapies are available for the treatment of mild-to-moderate Crohn’s disease (CD). This network meta-analysis (NMA) aimed to assess the comparative efficacy of ustekinumab, adalimumab, vedolizumab and infliximab in the maintenance of clinical response and remission after 1 year of treatment.

Methods: A systematic literature search was performed to identify relevant randomized controlled trials (RCTs). Key outcomes of interest were clinical response (CD activity index [CDAI] reduction of 100 points; CDAI-100) and remission (CDAI score under 150 points; CDAI < 150). A treatment sequence Bayesian NMA was conducted to account for the re-randomization of patients based on different clinical definitions, the lack of similarity of the common comparator for each trial and the full treatment pathway from the induction phase onwards.

Results: Thirteen RCTs were identified. Ustekinumab 90 mg q8w was associated with statistically significant improvement in clinical response relative to placebo and vedolizumab 300 mg. For clinical remission, ustekinumab 90 mg q8w was associated with statistically significant improvement relative to placebo and vedolizumab 300 mg q8w. Findings from sub-population analyses had similar results but were not statistically significant.

Conclusions: The NMA suggest that ustekinumab is associated with the highest likelihood of reaching response or remission at 1 year compared with placebo, adalimumab and vedolizumab. Results should be interpreted with caution because this is a novel methodology; however, the treatment sequence analysis may be the most methodologically sound analysis to derive estimates of comparative efficacy in CD in the absence of head-to-head evidence.

Keywords:

• Network meta-analysis
• Crohn’s disease
• biologics
• adaptive trial design
• study design
• indirect treatment comparisons
• ustekinumab

Introduction

Crohn’s disease (CD) is a chronic inflammatory condition of the gastrointestinal tract that is most prevalent among adolescents and young adults between the ages of 15 and 35^1,2. Patients with CD experience social stigma, impinging on their ability to attend work or school, and have impaired quality of life^1,2. The treatment goal for patients with CD is therefore to achieve long-term remission, which provides a number of benefits for patients, their families, care providers and employers³. Disease severity of patients with CD is typically measured with a Crohn’s Disease Activity Index (CDAI) score³.

There are several biologic treatment options available for CD⁴, including adalimumab, infliximab, vedolizumab and ustekinumab. The efficacy and safety of ustekinumab in the treatment of moderate-to-severe CD was demonstrated in two induction trials, UNITI-1⁵ and UNITI-2⁶ (patients received intravenous ustekinumab 6 mg/kg), and in the IM-UNITI⁶ maintenance trial (patients received subcutaneous injections of ustekinumab 90 mg every 12 or 8 weeks). UNITI-1⁵ included patients who failed conventional therapy, whereas UNITI-2⁶ included patients who failed anti-tumor necrosis factor alpha (anti-TNF-α) biologic therapy, and IM-UNITI⁷ included patients who failed either conventional or anti-TNF-α therapy.

No direct head-to-head comparisons are currently available for ustekinumab and other biologics in the treatment of moderate-to-severe CD. The present network meta-analysis (NMA) therefore sought to compare the efficacy of ustekinumab with other biologic treatment options during the maintenance phase. The primary objective was to compare different biologics that have never been directly compared in existing clinical trials for moderate-to-severe CD after 1 year of treatment to assess their relative effects on clinical response and remission.

Previous NMAs have not adequately accounted for heterogeneity in patient characteristics and study design^8–11. Specifically, these NMAs did not account for carry-over effects from induction, re-randomization and variations in placebo response rates. Therefore, this analysis was performed using a treatment sequence Bayesian NMA to account for the re-randomization of patients based on different clinical definitions, the lack of similarity of the common comparator (placebo) for each trial, and the full treatment pathway from induction phase (using clinical response: CDAI-70). Given that a treatment sequence Bayesian NMA accounts for differences in patient and study characteristics, it should yield more reliable estimates of comparative efficacy in moderate-to-severe CD.

Methods

A study protocol was drafted as a priori and followed during the process of this review. The protocol is available upon request. We also adhered to PRISMA reporting guidelines for network meta-analysis, (Appendix 13).

Literature search strategy

A systematic literature search was performed to identify English-language randomized controlled trials (RCTs) of biologic therapies used to treat moderate-to-severe CD published from 1997 to 2015. An information specialist conducted searches of electronic medical databases (MEDLINE, EMBASE and the Cochrane Library) using a predefined search strategy (Appendix 1) on 2 July 2015. Unpublished studies up to July 2015 were searched on Clinicaltrials.gov. The search strategy was independently peer reviewed by a second information specialist.

Eligibility criteria

The assumption of exchangeability of studies^12,13 is important for ensuring the validity of NMAs. Exchangeability suggests that the treatments are unique competing interventions that could be included in a clinical trial. In accordance with these principles, RCTs were evaluated to identify studies that had similar patient characteristics and study designs. Eligibility criteria were established using the following PICOS (Population, Intervention, Comparators, Outcomes, Study design) framework.

Population

The population comprised adult patients diagnosed with moderate-to-severe CD. Two patient sub-populations were also of interest: patients who failed conventional therapy, and patients who failed anti-TNF-α therapy.

Intervention and comparators

Relevant comparators included biologic therapies licensed for the treatment of moderate-to-severe CD: ustekinumab, infliximab, adalimumab and vedolizumab. Conventional therapy and placebo treatment were also included, given their anticipated role as a key source of indirect evidence in the network. Eligible doses of interest for each intervention are detailed in Appendix 2.

Outcomes

The outcomes measured were: studies that reported CDAI-70 (clinical response), and CDAI-100 (clinical response) or CDAI < 150 (clinical remission).

Study design

Eligible studies were randomized controlled trials.

Study selection

Two reviewers independently reviewed the study records, citation titles and abstracts identified in the literature search to assess study eligibility. Citations considered to describe potentially eligible articles were independently reviewed in full-text form to determine final inclusion status in the review. Disagreements between reviewers were resolved during a consensus meeting. A flow diagram documenting the process of study selection is shown in Figure 1. Information on study selection and lists of included and excluded studies are presented in Appendix 3.

Figure 1. PRISMA flow diagram.

Data extraction and risk of bias assessment

Data was extracted independently by two reviewers using a standardized data extraction form designed in Microsoft Excel (Microsoft Corporation, Seattle, USA). The form was used to gather data related to general publication characteristics (e.g. authors, study location, year and journal of publication), intervention details (e.g. treatment, dose and frequency of administration), patient traits (e.g. disease duration, CDAI score, C-reactive protein levels, concomitant therapies and subgroup information) and study endpoints. Discrepancies in collected data were resolved by discussion.

Eligible studies were assessed for risk of bias by a single reviewer. Quality assessment was conducted using the checklist provided in the National Institute for Health and Care Excellence (NICE) single technology appraisal template¹⁴. This checklist was used to assess each study in terms of selection bias (systematic differences between the comparison groups), performance bias (systematic differences between groups in the care provided, apart from the intervention under investigation), attrition bias (systematic differences between the comparison groups with respect to loss of participants) and detection bias (bias in how outcomes are ascertained, diagnosed or verified). Results are presented in Appendices 4 and 5.

Analysis considerations – comparison of randomized controlled trial designs in moderate-to-severe Crohn’s disease and variations in placebo response rates in maintenance phase

Appendix 6 summarizes the study design of included maintenance trials. Considerable variability in study designs was found amongst the maintenance trials for patient selection and blinding.

In the included maintenance trials, entry into the maintenance phase required achievement of a clinical response (CDAI-70 or CDAI-100) at a trial-specific time point during the induction phase. Re-randomization into the maintenance phase occurred at different time points (2 to 8 weeks) and was based on achieving different outcomes (CDAI-70 or CDAI-100).

Re-randomization of patients in most trials was based on open label induction phases. Open label trials are known to have higher relative treatment effects than blinded trials^15,16. Therefore, trials that are open label in the induction phase have a larger number of patients who achieve clinical response and who are eligible for the maintenance phase. Therefore, a carry-over effect occurs, which causes differences in placebo response and remission rates, and relative effect estimates, as shown in Figure 2 and Appendix 7. Figure 2 shows that maintenance placebo rates are lower in trials that were open label during the induction phase. Appendix 7 shows that as placebo response (Panel A) and remission (Panel C) rates increase, the treatment response and remission rates increase. For maintenance phase clinical response and remission, higher placebo rates produce lower odds ratios, and lower placebo rates produce higher odds ratios (Panel B and D).

Figure 2. Placebo response/remission rates for clinical response and clinical remission for maintenance phase trials^{7,25–27,31,32}. Abbreviations. ADA, Adalimumab; CDAI, Crohn’s Disease Activity Index; IFX, Infliximab; NA, Not available; VDZ, Vedolizumab; UST, Ustekinumab.

Treatment sequence analysis to address issues with heterogeneity in randomized controlled trial design

A treatment sequence analysis was conducted to address the issue of heterogeneity in placebo rates in maintenance trial data. The treatment sequence analysis utilized data from both the induction and maintenance phases to calculate relative effect estimates that account for the carry-over effects from the induction phase. To obtain relative effects, it was necessary to calculate the “probability of obtaining and maintaining response” by the end of the maintenance phase. This probability was calculated by multiplying the “probability of achieving response” during induction by the “probability of maintaining response” during the maintenance phase (Figure 3). Clinical response (CDAI-70) during the induction phase was used as induction input. Clinical response (CDAI-100) and remission (CDAI < 150) at the end of maintenance were used as maintenance inputs for the treatment sequence calculation.

Figure 3. Treatment sequence calculation.

To address the non-comparability of placebo arms, maintenance data for placebo arms was imputed. This approach was used to reduce any bias that would occur if an NMA was conducted using unadjusted maintenance data. Placebo-to-placebo response rates at the end of maintenance were calculated using IM-UNITI individual patient level data, which was adjusted for the proportion of responders at the end of induction in each trial.

Stratification of “remitters” (defined as patients who achieved a CDAI <150) and “responders non-remitters” (defined as patients who have achieved a CDAI reduction from baseline >70 and an absolute CDAI score >150) for IM-UNITI at the end of induction was performed. A weighted average of placebo response rates for the other biologic treatments (adalimumab, infliximab and vedolizumab) was estimated using the proportion of remitters and “responders non-remitters” at the end of induction. The weighted placebo response and remission rates were used as inputs for the placebo arms during the maintenance phase (Figure 4).

Figure 4. Treatment sequence analysis strategy. Abbreviations. CDAI, Crohn’s Disease Activity Index; UST, Ustekinumab.

Summary of measures of treatment effect

A Bayesian NMA was conducted using the inputs estimated with the treatment sequence analysis. To obtain odds ratios (ORs), the number of patients estimated with the treatment sequence analysis was rounded up.

For all pairwise comparisons estimated using NMA, ORs with 95% credible intervals (CrIs) were used as a measure of the association between each treatment and its efficacy. As clinical response and remission both represent positive events for patients with CD, ORs greater than 1 corresponded to a beneficial treatment effect of the first treatment compared with the second treatment; estimates with CrIs that did not include 1 were considered to reflect statistically significant differences. Additional measures of effect that are commonly presented for Bayesian NMAs were generated, including the probability of ustekinumab being better than a comparator for a given pairwise comparison (p[better]) and Surface Under the Cumulative RAnking curve (SUCRA). The SUCRA measure is expressed as a percentage, and summarizes the relative probability of an intervention being among the best options¹⁷. For interpretation, both p(better) and SUCRA values range between 0 and 1; values closer to 1 indicate preferred treatments.

Network meta-analysis of unadjusted and adjusted study data using treatment sequence analysis

Analyses were conducted on the overall patient population (i.e. patients who failed conventional and/or anti-TNF-α therapy), as well as on the two sub-populations of interest. Unadjusted analyses utilized data inputs that were reported directly from the maintenance phase of included RCTs, while adjusted analyses utilized data inputs derived from the treatment sequence methodology described above (which used data from the induction and maintenance phases of included clinical trials).

Primary analyses for both clinical endpoints studied were based upon a Bayesian hierarchical model for NMA recently described by Owen et al.¹⁸. This model simultaneously analyzes the treatments at the class level and at the dose level, therefore identifying the most efficacious treatment at the optimal dose. This overcomes certain limitations inherent with more commonly used approaches that separately analyze treatments at the dose level and class level¹⁸. For the Bayesian hierarchy NMA models, it was not assumed that that higher doses were no worse than lower doses. Secondary analyses using well established approaches for binary endpoints described by Dias et al.^13,19–21 were also performed to inform comparisons between the hierarchical approach and more commonly used approaches; for the latter, analyses at both the drug-dose level and the drug-class level (pooling doses together) were performed. Fixed effects (FE) models were implemented for all outcomes, given that the majority of comparisons were informed by single studies, an approach which has been previously used²². Random effects (RE) models were also conducted.

For all analyses, adequacy of model fit was assessed by comparing the posterior residual deviance with the corresponding number of unconstrained data points (i.e. the number of intervention arms across studies), which should be approximately equal if the fit is adequate²³. The deviance information criterion (DIC) was also assessed, where a lower DIC implies a better model fit if the difference in DIC between the two models is 3 or more. Network meta-analyses were performed using WinBUGS software (version 1.4.3, MRC Biostatistics Unit, United Kingdom) and R (version 3.3.1, R Development Core Team, University of Auckland, New Zealand), and were based on burn-in samples of ≥20,000 iterations and subsequent sampling iterations of ≥20,000 iterations. Trace plots and Gelman–Rubin plots were reviewed to assess model convergence.

Another key assumption behind NMA is that the analyzed network is consistent; that is, there is no conflict between direct and indirect evidence²¹. To assess inconsistency, deviance and DIC statistics were compared in fitted consistency and inconsistency models²¹. The posterior mean deviance of the individual data points in the inconsistency model was also plotted against the posterior mean deviance in the consistency model to identify any loops where inconsistency was present²¹.

Results

Studies included in unadjusted and adjusted treatment sequence analysis

A total of 5375 study records were identified in the search strategy and then screened. Of these records, 5127 were excluded because they did not meet the inclusion criteria and 248 were subjected to full-text screening. This resulted in 13 studies^{5–7,24–33} that met the eligibility criteria. A list of the studies that were included in the overall population and sub-population NMAs is provided in Appendix 3. The Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) flow diagram for the selection of these studies is presented in Figure 1.

Patient characteristics

Although an induction phase NMA was not conducted, available data from induction phase trials was utilized for treatment sequence calculations. For induction trials, mean age ranged from 31 to 38 years, the percentage of male patients ranged from 43% to 58%, disease duration ranged from 9 to 12 years and CDAI score ranged from 298 to 324. Among the maintenance trials, mean age ranged from 35 to 38 years, the percentage of males ranged from 38% to 47%, disease duration ranged from 7.5 to 10 years and CDAI score ranged from 299 to 320. Tables summarizing patient characteristics by study are presented in Appendix 8.

Evidence networks

Figure 5 presents network diagrams that summarize the available evidence for clinical response (CDAI-100) and clinical remission (CDAI < 150) in the overall study population. For clinical response, 9 comparisons were informed by head-to-head trials, 8 of which were informed by single studies. For clinical remission, 12 comparisons were supported by data from head-to-head trials, 11 of which were informed by single studies. Additional network diagrams corresponding to the failed conventional and failed anti-TNF-α sub-populations are also presented (Figure 5).

Figure 5. Evidence networks. Treatment nodes are sized to proportionately reflect the number of patients randomized to each of the interventions in the treatment networks. Widths of lines joining treatment nodes are sized to proportionately reflect the number of studies informing each comparison^{7,24,25,26,27,31,32}. Abbreviations. ADA, Adalimumab; CDAI, Crohn’s Disease Activity Index; IFX, Infliximab; TNF-α, Tumor necrosis factor alpha; UST, Ustekinumab; VDZ, Vedolizumab.

Calculation of adjusted rates

The treatment sequence calculations for the overall population, and the failed conventional and failed anti-TNF-α sub-populations are shown in Table 1 and Appendix 9, respectively.

Table 1. Treatment sequence calculations for the overall population.

Induction studies	Maintenance studies	Treatment	N Induction	Induction CDAI-70 response rate	% CDAI-100 maintenance	% CDAI < 150 maintenance	% CDAI-100 sequence	% CDAI < 150 sequence	1 Year responders (CDAI - 100)	1 Year remission (CDAI < 150)
UNITI 1^5 & UNITI 2^6 & CERTIFI	IM-UNITI^7	Placebo–Placebo	588	33%	51%	42%	17%	14%	100	82
		UST 6–UST 90 q12w	589	52%	58%	49%	30%	26%	177	151
		UST 6–UST 90 q8w	589	52%	59%	53%	31%	28%	181	163
GEMINI II^32 & GEMINI III^29	GEMINI II^32	Placebo–Placebo	354	34%	49%	37%	16%	13%	59	46
		VDZ 300–VDZ 300 q8w	426	47%	44%	39%	21%	18%	88	79
		VDZ 300–VDZ 300 q4w	426	47%	46%	36%	22%	17%	92	71
Classic I^26 & GAIN^28	CHARM^24	Placebo–Placebo	240	35%	49%	37%	17%	13%	42	32
		ADA 160/80–ADA 40 eow	235	54%	41%	36%	23%	20%	53	46
		ADA 160/80–ADA 40 weekly	235	54%	48%	41%	26%	23%	61	53
Watanabe 2012^31	Watanabe 2012^31	Placebo–Placebo	23	30%	49%	37%	17%	13%	4	3
		ADA 160/80–ADA 40 eow	33	70%	38%	38%	27%	27%	9	9
EXTEND^27	EXTEND^27	Placebo–Placebo	65	34%	NA	37%	NA	14%	NA	9
		ADA160/80–ADA eow	64	56%	NA	33%	NA	19%	NA	12
Targan 1997^30	ACCENT I^25	Placebo–Placebo	24	17%	NA	58%	NA	13%	NA	3
		IFX 5	27	81%	NA	32%	NA	26%	NA	7
		IFX 5 & 10	27	81%	NA	38%	NA	33%	NA	9

Abbreviations. ADA, Adalimumab; CDAI, Crohn’s Disease Activity Index; eow, Every other week; IFX, Infliximab; NA, Not available; q4w, Every 4 weeks; q8w, Every 8 weeks; q12w, Every 12 weeks; UST, Ustekinumab; VDZ, Vedolizumab.

Results from treatment sequence network meta-analysis

Table 2 presents a summary of findings from the Bayesian hierarchical NMA model at the drug-dose level for ustekinumab 90 mg every 8 weeks (q8w) relative to other competing regimens. For clinical response (CDAI-100), ustekinumab 90 mg q8w was associated with numerical improvements compared to all other treatments for each population. For the overall population, statistical significance was observed for ustekinumab 90 mg q8w relative to placebo (OR 2.15; 95% CrI 1.66–2.80) and to vedolizumab 300 mg q4w (OR 1.58; 95% CrI 1.03–2.43) and q8w (OR 1.61; 95% CrI 1.04–2.47). The probability of treatment being better was in favor (≥50%) of ustekinumab 90 mg q8w relative to all other treatments.

Table 2. Results from treatment sequence network meta-analysis for the overall population, failed conventional and failed anti-TNF-α sub-populations.

Odds ratios (95% CrI) and probability better for UST versus comparator
Dose	UST 90 mg q8w	UST 90 mg q12w	ADA 40 mg weekly	ADA 40 mg eow	VDZ 300 mg q4w	VDZ 300 mg q8w	IFX 5 mg/kg	IFX 5 & 10 mg/kg	PBO
Overall population
CDAI-100	1 p(better) = NA SUCRA = 89%	1.01 (0.87 to 1.19) p(better) = 50% SUCRA = 87%	1.37 (0.85 to 2.19) p(better) = 85% SUCRA = 55%	1.42 (0.88 to 2.26) p(better) = 86% SUCRA = 48%	1.58 (1.03 to 2.43) p(better) = 88% SUCRA = 37%	1.61 (1.04 to 2.47) p(better) = 90% SUCRA = 34%	NA	NA	2.15 (1.66 to 2.80) p(better) = 96% SUCRA = 2%
CDAI <150	1 p(better) = NA SUCRA = 75%	1.03 (0.89 to 1.26) p(better) = 52% SUCRA = 69%	1.29 (0.78 to 2.12) p(better) = 56% SUCRA = 50%	1.34 (0.82 to 2.18) p(better) = 63% SUCRA = 44%	1.55 (0.98 to 2.43) p(better) = 75% SUCRA = 27%	1.60 (1.01 to 2.53) p(better) = 78% SUCRA = 32%	0.73 (0.14 to 2.69) p(better) = 35% SUCRA = 75%	0.72 (0.14 to 2.64) p(better) = 34% SUCRA = 77%	2.27 (1.71 to 3.03) p(better) = 95% SUCRA = 2%
Failed conventional
CDAI-100	1 p(better) = NA SUCRA = 83%	1.01 (0.83 to 1.25) p(better) = 55% SUCRA = 82%	1.32 (0.62 to 2.79) p(better) = 66% SUCRA = 59%	1.38 (0.64 to 2.92) p(better) = 69% SUCRA = 52%	1.66 (0.89 to 3.12) p(better) = 90% SUCRA = 39%	1.71 (0.92 to 3.24) p(better) = 97% SUCRA = 33%	NA	NA	2.60 (1.77 to 3.83) p(better) = 99% SUCRA = 2%
CDAI <150	1 p(better) = NA SUCRA = 80%	1.02 (0.86 to 1.31) p(better) = 52% SUCRA = 65%	1.21 (0.51 to 2.66) p(better) = 67% SUCRA = 60%	1.24 (0.53 to 2.75) p(better) = 69% SUCRA = 49%	1.42 (0.70 to 2.84) p(better) = 86% SUCRA = 46%	1.45 (0.71 to 2.90) p(better) = 89% SUCRA = 37%	0.57 (0.09 to 2.56) p(better) = 36% SUCRA = 78%	0.57 (0.09 to 2.52) p(better) = 35% SUCRA = 80%	2.89 (1.94 to 4.38) p(better) = 99% SUCRA = 1%
Failed anti-TNF-α
CDAI-100	1 p(better) = NA	1.01 (0.84 to 1.24) p(better) = 51%	1.14 (0.60 to 2.15) p(better) = 54%	1.15 (0.60 to 2.18) p(better) = 58%	1.52 (0.81 to 2.73) p(better) = 79%	1.61 (0.86 to 2.90) p(better) = 86%	NA	NA	2.02 (1.40 to 2.94) p(better) = 93% SUCRA = 5%
CDAI <150	1 p(better) = NA	1.01 (0.84 to 1.25) p(better) = 51%	1.03 (0.50 to 1.96) p(better) = 53%	1.05 (0.51 to 2.02) p(better) = 56%	1.40 (0.73 to 2.69) p(better) = 80%	1.40 (0.73 to 2.68) p(better) = 80%	NA	NA	1.71 (1.17 to 2.58) p(better) = 88% SUCRA = 8%

Abbreviations. ADA, Adalimumab; CDAI, Crohn’s Disease Activity Index; CrI, Credible interval; eow, Every other week; IFX, Infliximab; NA, Not available; PBO, Placebo; q4w, Every 4 weeks; q8w, Every 8 weeks; q12w, Every 12 weeks; TNF-α, Tumor necrosis factor alpha; UST, Ustekinumab; VDZ, Vedolizumab.

Bolded values are statistically significant odds ratios.

For clinical remission (CDAI < 150), ustekinumab 90 mg q8w was associated with numerical improvements compared with all other treatments for each population, except for infliximab. For the overall population, statistical significance was observed for ustekinumab 90 mg q8w compared with placebo (OR 2.27; 95% CrI 1.71–3.03) and vedolizumab 300 mg q8w (OR 1.60; 95% CrI 1.01–2.53). The probability of treatment being better was in favor (≥50%) of ustekinumab 90 mg q8w relative to all other treatments, except for infliximab.

Results from unadjusted network meta-analysis

Clinical interpretations were notably different for NMA results that used data not transformed with treatment sequence methodology, with summary estimates favoring adalimumab for both endpoints (Appendix 10).

Results from additional network meta-analyses

Several sensitivity analyses were performed to assess the robustness of findings derived from the primary analyses. Results from these analyses are summarized in Table 3, which presents comparisons of the other included active interventions relative to ustekinumab from these analyses. Based on comparison of DIC, model fit of the Bayesian hierarchical model was found to be better than the dose-stratified model for both endpoints (measures of fit with the dose-pooled analyses were not directly comparable because some studies had to be removed if their study arms were redundant when collapsed).

Table 3. Results from sensitivity analyses.

Odds ratios (95% CrI), probability better for UST versus comparator and SUCRA										Model fit statistics
Dose	UST 90 mg q8w		UST 90 mg q12w	ADA 40 mg weekly	ADA 40 mg eow	VDZ 300 mg q4w	VDZ 300 mg q8w	IFX 5 mg/kg	IFX 5 & 10 mg/kg
Class	UST			ADA		VDZ		IFX
CDAI-100
Reference case – Bayesian hierarchical NMA	1 p(better) = NA SUCRA = 89%		1.01 (0.87 to 1.19) p(better) = 50% SUCRA = 87%	1.37 (0.85 to 2.19) p(better) = 85% SUCRA = 55%	1.42 (0.88 to 2.26) p(better) = 86% SUCRA = 48%	1.58 (1.03 to 2.43) p(better) = 88% SUCRA = 37%	1.61 (1.04 to 2.47) p(better) = 90% SUCRA = 34%	NA	NA	DIC = 77.22 Totresdev = 8.11 vs. 11
	1 p(better) = NA SUCRA = 89%			1.38 (0.84 to 2.25) p(better) = 73% SUCRA = 10%		1.58 (1.02 to 2.49) p(better) = 87% SUCRA = 2%		NA
Dose-stratified NMA	1 p(better) = NA SUCRA = 89%	1.03 (0.81 to 1.32) p(better) = 53% SUCRA = 84%		1.32 (0.79 to 2.17) p(better) = 78% SUCRA = 62%	1.54 (0.92 to 2.56) p(better) = 81% SUCRA = 40%	1.57 (1.002 to 2.47) p(better) = 84% SUCRA = 39%	1.66 (1.05 to 2.62) p(better) = 85% SUCRA = 3%	NA	NA	DIC = 81.33 Totresdev = 10.10 vs. 11
Dose-pooled NMA	1 p(better) = NA SUCRA = 91%			1.38 (0.88 to 2.16 ) p(better) = 74% SUCRA = 59%		1.59 (1.05 to 2.37) p(better) = 89% SUCRA = 42%		NA		DIC = 58.89 Totresdev = 7.06 vs. 8
Failed conventional Dose-stratified NMA	1 p(better) = NA SUCRA = 84%		1.04 (0.71 to 1.54) p(better) = 59% SUCRA = 79%	1.21 (0.53 to 2.70) p(better) = 67% SUCRA = 68%	1.65 (0.72 to 3.70) p(better) = 88% SUCRA = 41%	1.91 (0.95 to 3.82) p(better) = 97% SUCRA = 29%	1.60 (0.80 to 3.18) p(better) = 91% SUCRA = 46%	NA	NA	DIC = 64.87 Totresdev = 9.05 vs. 9
Failed conventional Dose-pooled	1 p(better) = NA SUCRA = 74%			1.31 (0.64 to 2.68) p(better) = 76% SUCRA = 64%		1.62 (0.88 to 2.97) p(better) =93% SUCRA = 45%		NA		DIC = 44.93 Totresdev = 6.01 vs. 6
Failed conventional Bayesian hierarchical NMA	1 p(better) = NA SUCRA = 83%		1.01 (0.83 to 1.25) p(better) = 55% SUCRA = 82%	1.32 (0.62 to 2.79) p(better) = 66% SUCRA = 59%	1.38 (0.64 to 2.92) p(better) = 69% SUCRA = 52%	1.66 (0.89 to 3.12) p(better) = 90% SUCRA = 39%	1.71 (0.92 to 3.24) p(better) = 97% SUCRA = 33%	NA	NA	DIC = 61.12 Totresdev = 7.70 vs. 9
	1 p(better) = NA SUCRA = 86%			1.34 (0.62 to 2.85) p(better) = 76% SUCRA = 46%		1.68 (0.89 to 3.20) p(better) = 83% SUCRA = 18%		NA
Failed anti-TNF-α Dose-stratified NMA	1 p(better) = NA SUCRA = 79%		1.05 (0.74 to 1.49) p(better) = 55% SUCRA = 73%	1.13 (0.55 to 2.28) p(better) = 54% SUCRA = 66%	1.21 (0.59 to 2.45) p(better) = 65% SUCRA = 58%	1.36 (0.72 to 2.61) p(better) = 80% SUCRA = 49%	1.85 (0.96 to 3.63) p(better) = 93% SUCRA = 18%	NA	NA	DIC = 66.37 Totresdev = 9.03 vs. 9
Failed anti-TNF-α Dose-pooled NMA	1 p(better) = NA SUCRA = 86%			1.14 (0.59 to 2.14) p(better) = 58% SUCRA = 72%		1.61 (0.88 to 2.91) p(better) = 89% SUCRA = 35%		NA		DIC = 45.93 Totresdev = 6.03 vs. 6
Failed anti-TNF-α Bayesian hierarchical NMA	1 p(better) = NA SUCRA = 79%		1.01 (0.84 to 1.24) p(better) = 51% SUCRA = 76%	1.14 (0.60 to 2.15) p(better) = 54% SUCRA = 64%	1.15 (0.60 to 2.18) p(better) = 58% SUCRA = 62%	1.52 (0.81 to 2.73) p(better) = 79% SUCRA = 37%	1.61 (0.86 to 2.90) p(better) = 86% SUCRA = 27%	NA	NA	DIC = 62.70 Totresdev = 7.72 vs. 9
	1 p(better) = NA SUCRA = 78%			1.14 (0.59 to 2.19) p(better) = 56% SUCRA = 57%		1.56 (0.82 to 2.82) p(better) = 78% SUCRA = 14%		NA
CDAI <150
Reference case – Bayesian hierarchical NMA	1 p(better) = NA SUCRA = 75%		1.03 (0.89 to 1.26) p(better) = 52% SUCRA = 69%	1.29 (0.78 to 2.12) p(better) = 56% SUCRA = 50%	1.34 (0.82 to 2.18) p(better) = 63% SUCRA = 44%	1.55 (0.98 to 2.43) p(better) = 75% SUCRA = 27%	1.60 (1.01 to 2.53) p(better) = 78% SUCRA = 32%	0.73 (0.14 to 2.69) p(better) = 35% SUCRA = 75%	0.72 (0.14 to 2.64) p(better) = 34% SUCRA = 77%	DIC = 101.28 Totresdev = 12.51 vs. 16
	1 p(better) = NA SUCRA = 71%			1.29 (0.77 to 2.15) p(better) = 59% SUCRA = 37%		1.54 (0.95 to 2.47) p(better) = 76% SUCRA = 15%		0.71 (0.14 to 2.62) p(better) = 41% SUCRA = 77%
Dose-stratified NMA	1 p(better) = NA SUCRA = 76%		1.11 (0.86 to 1.44) p(better) = 52% SUCRA = 64%	1.22 (0.70 to 2.11) p(better) = 62% SUCRA = 57%	1.44 (0.86 to 2.41) p(better) = 69% SUCRA = 40%	1.53 (0.93 to 2.50) p(better) = 74% SUCRA = 23%	1.74 (1.05 to 2.88) p(better) = 85% SUCRA = 37%	0.62 (0.11 to 2.59) p(better) = 29% SUCRA = 84%	0.90 (0.16 to 3.97) p(better) = 48% SUCRA = 66%	DIC = 106.56 Totresdev = 14.58 vs. 16
Dose-pooled NMA	1 p(better) = NA SUCRA = 79%			1.35 (0.85 to 2.14) p(better) = 69% SUCRA = 49%		1.54 (0.99 to 2.40) p(better) = 78% SUCRA = 36%		0.67 (0.13 to 2.50) p(better) = 43% SUCRA = 85%		DIC = 79.49 Totresdev = 10.57 vs. 12
Failed conventional Dose-stratified NMA	1 p(better) = NA SUCRA = 68%		1.15 (0.78 to 1.71) p(better) = 60% SUCRA = 55%	1.16 (0.47 to 2.77) p(better) = 63% SUCRA = 56%	1.46 (0.59 to 3.52) p(better) = 81% SUCRA = 39%	1.59 (0.74 to 3.37) p(better) = 88% SUCRA = 33%	1.44 (0.67 to 3.05) p(better) = 80% SUCRA = 41%	0.48 (0.06 to 2.44) p(better) = 20% SUCRA = 71%	0.70 (0.08 to 3.67) p(better) = 34% SUCRA = 86%	DIC = 79.30 Totresdev = 12.12 vs. 12
Failed conventional Dose-pooled NMA	1 p(better) = NA SUCRA = 68%			1.20 (0.54 to 2.63) p(better) = 71% SUCRA = 53%		1.40 (0.69 to 2.75) p(better) = 85% SUCRA = 42%		0.52 (0.07 to 2.41) p(better) = 30% SUCRA = 87%		DIC = 54.72 Totresdev = 8.10 vs. 8
Failed conventional Bayesian hierarchical NMA	1 p(better) = NA SUCRA = 80%		1.02 (0.86 to 1.31) p(better) = 52% SUCRA = 65%	1.21 (0.51 to 2.66) p(better) = 67% SUCRA = 60%	1.24 (0.53 to 2.75) p(better) = 69% SUCRA = 49%	1.42 (0.70 to 2.84) p(better) = 86% SUCRA = 46%	1.45 (0.71 to 2.90) p(better) = 89% SUCRA = 37%	0.57 (0.09 to 2.56) p(better) = 36% SUCRA = 78%	0.57 (0.09 to 2.52) p(better) = 35% SUCRA = 80%	DIC = 73.55 Totresdev = 9.74 vs. 12
	1 p(better) = NA SUCRA = 58%			1.21 (0.51 to 2.68) p(better) = 72% SUCRA = 38%		1.41 (0.69 to 2.84) p(better) = 87% SUCRA = 23%		0.56 (0.09 to 2.50) p(better) = 34% SUCRA = 81%
Failed anti-TNF-α Dose-stratified NMA	1 p(better) = NA SUCRA = 74%		1.06 (0.72 to 1.55) p(better) = 54% SUCRA = 67%	0.98 (0.46 to 2.07) p(better) = 50% SUCRA = 73%	1.16 (0.54 to 2.48) p(better) = 66% SUCRA = 55%	1.43 (0.71 to 2.91) p(better) = 84% SUCRA = 35%	1.44 (0.71 to 2.91) p(better) = 84% SUCRA = 35%	NA	NA	DIC = 65.16 Totresdev = 9.05 vs. 9
Failed anti-TNF-α Dose-pooled NMA	1 p(better) = NA SUCRA = 79%			1.03 (0.51 to 2.00) p(better) = 53% SUCRA = 74%		1.40 (0.75 to 2.61) p(better) = 80% SUCRA = 38%		NA		DIC = 45.17 Totresdev = 6.03 vs. 6
Failed anti-TNF-α Bayesian hierarchical NMA	1 p(better) = NA SUCRA = 72%		1.01 (0.84 to 1.25) p(better) = 51% SUCRA = 69%	1.03 (0.50 to 1.96) p(better) = 53% SUCRA = 68%	1.05 (0.51 to 2.02) p(better) = 56% SUCRA = 64%	1.40 (0.73 to 2.69) p(better) = 79% SUCRA = 34%	1.40 (0.73 to 2.68) p(better) = 80% SUCRA = 34%	NA	NA	DIC = 60.54 Totresdev = 6.88 vs. 9
	1 p(better) = NA SUCRA = 68%			1.03 (0.49 to 2.00) p(better) = 53% SUCRA = 63%		1.39 (0.72 to 2.71) p(better) = 79% SUCRA = 19%		NA

Abbreviations. ADA, Adalimumab; CDAI, Crohn’s Disease Activity Index; CrI, Credible interval; DIC, Deviance information criterion; eow, Every other week; IFX, Infliximab; NA, Not available; q4w, Every 4 weeks; q8w, Every 8 weeks; q12w, Every 12 weeks; SUCRA, Surface Under the Cumulative RAnking curve; Totresdev, Total residual deviance; TNF-α, Tumor necrosis factor; UST, Ustekinumab; VDZ, Vedolizumab.

Bolded values are statistically significant odds ratios.

For clinical response (CDAI-100), summary estimates of effect were similar in magnitude relative to those derived from the dose-stratified analysis; however, precision was improved using the hierarchical model, a consequence of borrowing strength from this approach. Findings from the pooled-dose NMA showed very similar summary estimates compared to those estimated at the drug level from the hierarchical model, with minimal differences in precision. Results from a RE analysis for the hierarchical model had slightly larger model fit statistics values than the FE model (Appendix 11). No inconsistencies in the primary NMAs were detected (Appendix 12). Lastly, analyses of clinical response in the sub-populations of patients failing conventional therapy and patients failing anti-TNF-α therapy showed similar trends in terms of the benefits of ustekinumab; however, results were associated with more uncertainty due to a reduction in available sample size, and differences between therapies were not statistically significant. The probability of a treatment being better than a comparator for pairwise comparisons and SUCRA were in favor of ustekinumab.

For clinical remission (CDAI < 150), similar patterns were observed in terms of comparability in summary estimates from the different models as well as differences in precision of the effect estimates. With the exception of infliximab, the probability of a treatment being better than a comparator for pairwise comparisons and SUCRA were in favor of ustekinumab.

Discussion

The objective of the current study was to evaluate the efficacy of ustekinumab relative to infliximab, adalimumab and vedolizumab. This study represents the most up-to-date systematic review and NMA to synthesize data comparing biologic treatments for moderate-to-severe CD. In total, 13 RCTs were identified that compared treatments for the maintenance phase of clinical response and clinical remission in patients with moderate-to-severe CD. Due to variations in study designs (i.e. lack of similarity amongst control groups in terms of response and remission rates), a treatment sequence calculation was deemed appropriate, and this data was used to inform NMAs. Bayesian hierarchical models for network meta-analysis were employed to account for the presence of different doses of the therapies of interest.

For the overall patient population, who failed conventional and/or anti-TNF-α therapy, ustekinumab was associated with statistically significant improvement in clinical response (CDAI-100) relative to placebo and vedolizumab. It was also associated with the highest probability of being better than a comparator and had the highest SUCRA. For clinical remission (CDAI < 150), ustekinumab was associated with statistically significant improvement relative to placebo and vedolizumab. Ustekinumab was also associated with the highest probability of being better than a comparator and had the highest SUCRA, with the exception of infliximab.

The treatment sequence NMA was also run by stratifying the overall population into two sub-populations: failed conventional and failed anti-TNF-α. Ustekinumab showed statistically significant improvement in both CDAI-100 and CDAI < 150 relative to placebo and was also associated with the highest probability of being better than a comparator and had the highest SUCRA, except for infliximab for CDAI < 150 in the failed conventional sub-population.

Others have faced similar challenges in attempts to synthesize the available evidence on biologics in moderate-to-severe CD. For instance, Hazlewood et al.⁹ conducted an NMA in the overall population. A commentary on this NMA, published by Bonovas et al.⁸, pointed out that results should be interpreted with caution due to various non-negligible issues such as pooling dosages, different assessment times for each trial and differences in patients’ baseline characteristics. Another NMA conducted by Stidham et al.³⁴ raised similar concerns. Singh et al.³⁵ also conducted an NMA for patients in the failed conventional sub-population, raising issues such as variation in placebo response rates, study design differences and assessment times across studies. To address these issues, a treatment sequence analysis was performed. The treatment sequence analysis used a Bayesian hierarchical model to account for heterogeneity among the trials, but also the presence of different doses of the therapies of interest.

Despite the strengths of the treatment sequence analysis, there are also limitations with the current analysis that should be noted. Most trials evaluating comparators used in the treatment sequence analysis were open label during the induction phase. This would in turn bias results against RCTs that were double-blinded during the induction phase (Appendix 6). The re-randomization into the maintenance phase was also non-homogenous across the included trials, therefore contributing to variations in placebo response/remission rates between maintenance trials. Another potential limitation to this study was the systematic literature review (SLR) was conducted on 2 July 2015. New studies and treatments may have been approved after this date which may have not been included in the treatment sequence analysis. Certolizumab, for instance, was excluded since it was not available in Canada at the time of the SLR search. However, a methodological case study such as the present study is not required to include all treatments to illustrate the methodological application. Additionally, we did not consider a range of imputed values for IM-UNITI⁷ for our treatment sequence analysis, but it is reassuring that the sensitivity analyses to assess robustness were comparable to the primary analysis. A final limitation is that many NMAs have been published in this therapeutic area; however, given that the focus of this study was to illustrate the novel method of treatment sequence NMAs, these findings should be of interest and help to guide interpretation of future NMAs.

Conclusions

The results of the current NMA suggest that ustekinumab is associated with significantly improved clinical response and clinical remission compared with placebo and vedolizumab, and similar efficacy compared to adalimumab and infliximab. Results should be interpreted with caution because this is a novel methodology; however, treatment sequence analysis may be the most appropriate method given the lack of head-to-head evidence and the heterogeneity across maintenance trials.

Transparency

Declaration of funding

This study was funded by Janssen Inc. Cornerstone received financial support from Janssen Inc. for the conduct of this study.

Author contributions: A.V. was involved with analysis of the data, drafting the manuscript and approving the final version to be published. F.R.W., P.D., M.H., B.H. and C.C. assisted with interpretation of data, critically reviewed for important intellectual content for the work and approved the final version to be published.

Declaration of financial/other relationships

P.D. and M.H. have disclosed that they are employees of Janssen Inc. B.H. has disclosed that he is an employee of the Ottawa Hospital Research Institute and provides methodological advice for Cornerstone Research Group Inc. C.C. has disclosed that she/he is an employee and shareholder of Cornerstone Research Group Inc. A.V. and F.R.W. have disclosed that they are employees of Cornerstone Research Group Inc. Cornerstone consults for various pharmaceutical, medical device and biotech companies.

CMRO peer reviewers on this manuscript have received an honorarium from CMRO for their review work but have no relevant financial or other relationships to disclose.

Previous presentation

The work from this study has been previously presented as an abstract at the 2017 Canadian Digestive Diseases Week. Our work builds off initial concepts presented by Pacou and colleagues at ISPOR in 2016.

Acknowledgments

The authors would like to acknowledge Meaghan Bartlett for writing support, and the peer reviewers for their helpful comments.

Appendices

Appendix 1. Search strategy.MEDLINE and MEDLINE-IN-PROCESS (date of the search: 2 July 2015)

	#	Search terms	Hits
Crohn’s disease	1	“Crohn Disease”[MeSH] OR “Inflammatory Bowel Diseases”[MeSH]	62,276
	2	“Crohn Disease”[TW] OR “Crohns Disease”[TW] OR “Crohn's Disease”[TW] OR “Inflammatory Bowel Diseases”[TW] OR “Inflammatory Bowel Disease”[TW] OR IBD[TW]	63,787
	3	#1 OR #2	80,389
Intervention	4	“Biological Therapy”[MeSH] OR “Antibodies, Monoclonal”[MeSH] OR infliximab[TW] OR remicade[TW] OR adalimumab[TW] OR humira[TW] OR vedolizumab[TW] OR entyvio[TW] OR certolizumab[TW] OR cimzia[TW] OR natalizumab[TW] OR tysabri[TW] OR ustekinumab[TW] OR stelara[TW]	528,522
Study type	5	“Randomized Controlled Trials as Topic”[MeSH]	97,784
	6	“Random Allocation”[MeSH]	82,964
	7	“Double-Blind Method”[MeSH]	128,769
	8	“Single-Blind Method”[MeSH]	20,311
	9	“Clinical Trial, Phase I”[PT]	14,752
	10	“Clinical Trial, Phase II”[PT]	23,851
	11	“Clinical Trial, Phase III”[PT]	9736
	12	“Clinical Trial, Phase IV”[PT]	1024
	13	“Controlled Clinical Trial”[PT]	89,034
	14	“Randomized Controlled Trial”[PT]	390,357
	15	“Multicenter Study”[PT]	182,458
	16	“Clinical trial”[PT]	802,630
	17	“Clinical Trials as Topic”[MeSH]	286,758
	18	#5 OR #6 OR #7 OR #8 OR #9 OR #10 OR #11 OR #12 OR #13 OR #14 OR #15 OR #16 OR #17	1,074,702
	19	“clinical trial*”[TW]	597,932
	20	“single?blind”[TW] or “double?blind”[TW] or “treble?blind”[TW] or “triple?blind”[TW]	183,132
	21	Placebos[MeSH]	32,632
	22	Placebo?[TW]	168,262
	23	“randomly allocated”[TW]	18,944
	24	#19 OR #20 OR #21 OR #22 OR #23	729,317
	25	#18 OR #24	1,155,213
Exclusion	26	“Case Reports”[PT] OR Comment[PT] OR Editorial[PT] OR “Historical article”[PT] OR Letter[PT] OR “case report”[TI]	3,307,182
	27	Animals[MeSH] NOT Humans[MeSH]	4,016,826
	28	#26 OR #27	7,234,209
Combined	29	#3 AND #4 AND #25	1288
	30	#29 NOT #28	1195
Final results	31	Limit to: English	1097

EMBASE (date of the search: 2 July 2015)

	#	Search terms	Hits
Crohn’s disease	1	Crohn Disease'/ OR 'Inflammatory Bowel Diseases'/	63,132
	2	'Crohn Disease'.mp. OR 'Crohns Disease'.mp. OR 'Inflammatory Bowel Diseases'.mp. OR 'Inflammatory Bowel Disease'.mp. OR IBD.mp.	99,878
	3	1 OR 2	99,878
Intervention	4	Biological Therapy'/ OR 'Antibodies, Monoclonal'/ OR infliximab.mp. OR remicade.mp. OR adalimumab.mp. OR humira.mp. OR vedolizumab.mp. OR entyvio.mp. OR certolizumab.mp. OR cimzia.mp. OR natalizumab.mp. OR tysabri.mp. OR ustekinumab.mp OR stelara.mp	224,640
Study type	5	Clinical trial/	850,725
	6	Randomized controlled trial/	377,685
	7	Randomization/	66,831
	8	Single blind procedure/	20,468
	9	Double blind procedure/	123,874
	10	Crossover procedure/	43,327
	11	Placebo/	270,979
	12	Randomi?ed controlled trial$.tw.	118,784
	13	Rct.tw.	17,421
	14	Random allocation.tw.	1467
	15	Randomly allocated.tw.	22,896
	16	Allocated randomly.tw.	2048
	17	(allocated adj2 random).tw.	810
	18	Single blind$.tw.	16,234
	19	Double blind$.tw.	158,261
	20	((treble or triple) adj (blind$)).tw.	482
	21	Placebo$.tw.	221,397
	22	Prospective study/	296,032
	23	Or/5-22	1,493,693
Exclusion	24	'Case Report'.tw. OR 'Case study'/ OR Abstract report/ OR letter/	1,267,244
	25	animal/ NOT human/	1,259,266
	26	24 OR 25	2,512,180
Combined	27	3 AND 4 AND 23	4121
	28	27 NOT 26	4025
Final results	29	(english).lg.	23,356,142
	30	28 and 29	3776

The Cochrane Library (date of the search: 2 July 2015)

	#	Search terms	Hits
Crohn’s disease	1	MeSH descriptor: [Crohn Disease] explode all trees	1029
	2	MeSH descriptor: [Inflammatory Bowel Diseases] explode all trees	2082
	3	(Crohn Disease):ab,ti,kw OR (Crohns Disease):ab,ti,kw OR (Inflammatory Bowel Diseases):ti,ab,kw OR (Inflammatory Bowel Disease):ti,ab,kw OR (IBD):ti,ab,kw	2501
	4	#1 OR #2 OR #3	3201
Intervention	5	MeSH descriptor: [Biological Therapy] explode all trees	10,978
	6	MeSH descriptor: [Antibodies, Monoclonal] explode all trees	5412
	7	(infliximab):ti,ab,kw OR (remicade):ti,ab,kw OR (adalimumab):ti,ab,kw OR (humira):ti,ab,kw OR (vedolizumab):ti,ab,kw OR (entyvio):ti,ab,kw OR (certolizumab):ti,ab,kw OR (cimzia):ti,ab,kw OR (natalizumab):ti,ab,kw OR (tysabri):ti,ab,kw OR (ustekinumab):ti,ab,kw OR (stelara):ti,ab,kw	1866
	8	#5 OR #6 OR #7	17,084
Final result	9	#4 AND #8	551
	10	Limit: Clinical trials	399

Appendix 2. Treatment doses in moderate-to-severe Crohn’s disease.

Active substance	Induction	Maintenance
Ustekinumab (Stelara)	6 mg/kg	90 mg every 8 weeks (q8w) 90 mg every 12 weeks (q12w)
Infliximab (Remicade)	5 mg/kg	5 mg/kg every 8 weeks (q8w) 5 & 10 mg/kg every 8 weeks (q8w)
Adalimumab (Humira)	160/80 mg and 80/40 mg	40 mg every other week (eow) 40 mg weekly
Vedolizumab (Entyvio)	300 mg IV at 0,2 and 6 weeks	300 mg every 8 weeks (q8w) 300 mg every 4 weeks (q4w)
Conventional therapy	Monotherapy with a conventional glucocorticosteroids or budesonide or 5-ASA	Azathioprine or mercaptopurine as monotherapy when previously used with a conventional glucocorticosteroids or budesonide Methotrexate

Abbreviations. 5-ASA, 5-Aminosalicylic acid; CD, Crohn’s disease; eow, Every other week; IV, Intravenous; q4w, Every 4 weeks; q8w, Every 8 weeks; q12w, Every 12 weeks.

Appendix 3. Detailed rationale for included and excluded studies.Included studies for primary dosage network meta-analysis

Study	Comparison used in NMAs	Analysis
IM-UNITI^7	Maintenance: placebo versus ustekinumab 90 mg q12w vs. ustekinumab 90 mg q8w	Unadjusted NMA Treatment sequence NMA
GEMINI II^32	Induction: placebo versus vedolizumab 300 mg Maintenance: placebo versus vedolizumab 300 mg q8w versus vedolizumab 300 mg q4w	Unadjusted NMA Treatment sequence NMA
CHARM^24	Maintenance: placebo versus adalimumab 40 mg eow versus adalimumab 40 mg Weekly	Unadjusted NMA Treatment sequence NMA
Watanabe 2012^31	Induction: placebo versus adalimumab 160/80 mg Maintenance: placebo versus adalimumab 40 mg eow	Unadjusted NMA Treatment sequence NMA
EXTEND^27	Induction: placebo versus adalimumab 160/80 mg Maintenance: placebo versus adalimumab 40 mg eow	Unadjusted NMA Treatment sequence NMA
ACCENT I^25	Maintenance: placebo versus infliximab 5 mg/kg versus infliximab 5 & 10 mg/kg	Unadjusted NMA Treatment sequence NMA
UNITI-1^5	Induction: placebo versus ustekinumab 6 mg/kg	Treatment sequence NMA
CERTIFI^33	Induction: placebo versus ustekinumab 6 mg/kg	Treatment sequence NMA
UNITI-2^6	Induction: placebo versus ustekinumab 6 mg/kg	Treatment sequence NMA
GEMINI III^29	Induction: placebo versus vedolizumab 300 mg	Treatment sequence NMA
Classic I^26	Induction: placebo versus adalimumab 160/80 mg	Treatment sequence NMA
GAIN^28	Induction: placebo versus adalimumab 160/80 mg	Treatment sequence NMA
Targan 1997^30	Induction: placebo versus infliximab 5 mg/kg	Treatment sequence NMA

Included studies for failed conventional sub-population primary dosage network meta-analyses

Unadjusted NMA	Treatment sequence NMA
IM-UNITI GEMINI II CHARM ACCENT	UNITI-2 IM-UNITI GEMINI II GEMINI III Classic I Watanabe 2012 CHARM Targan 1997 ACCENT I

Included studies for the failed anti-TNF sub-population primary dosage network meta-analyses

Unadjusted NMA	Treatment sequence NMA
IM-UNITI GEMINI II CHARM	UNITI-1 CERTIFI IM-UNITI GEMINI II GEMINI III GAIN CHARM Watanabe 2012

Excluded studies from primary dosage network meta-analysis

Study	Reason for exclusion
D’Haens 1999^36	Lack of reported outcomes of interest & IFX doses not in line with recommendations
Classic II^37	Only included patients in remission at week 4 & 8 (dissimilar to other trials)
SONIC^38	Comparator not of interest
Rutgeerts 1999^39	Retreatment with infliximab (maintenance dissimilar to other trials)
Feagan 2008^40	Comparator not of interest
Hanauer 2013^41	Abstract of GEMINI II
Schreiber 2005^42	Comparator not of interest
Winter 2014^43	Comparator not of interest
PRECISE 1^44	Comparator not of interest
PRECISE 2^45	Comparator not of interest
Rutgeerts 2008^46	Comparator not of interest
WELCOME^47	Comparator not of interest
Feagan 2011^48	Comparator not of interest
Sandborn 2011^49	Comparator not of interest
ENACT-1^50	Comparator not of interest
ENACT-2^50	Comparator not of interest
ENCORE^51	Comparator not of interest
Ghosh 2003^52	Comparator not of interest

Appendix 4. Quality assessment of included studies.

Author(s)	Year	Was randomization carried out appropriately?	Was the concealment of treatment allocation adequate?	Were the groups similar at the outset of the study in terms of prognostic factors?	Were the care providers, participants and outcome assessors blind to treatment allocation?	Were there any unexpected imbalances in drop-outs between groups?	Is there any evidence to suggest that the authors measured more outcomes than they reported?	Did the analysis include an intention-to-treat analysis? If so, was this appropriate?	Were appropriate methods used to account for missing data (LOCF, MMRM)?
Sandborn et al. (CERTIFI)	2012	Yes	Yes	No	Yes	No	No	Yes	Yes
UNITI-1	2015	Yes	Yes	Yes	Yes	No	No	No	Yes
UNITI-2	2015	Yes	Yes	Yes	Yes	No	No	No	Yes
IM-UNITI	2015	Yes	Yes	Yes	Yes	No	No	No	Yes
Hanauer et al. (CLASSIC I)	2006	Yes	Yes	Yes	Yes	Unclear	No	No	Yes
Colombel et al. (CHARM)	2007	Yes	Yes	Yes	Yes	No	No	No	Unclear
Sandborn et al. (GAIN)	2007	Yes	Yes	Yes	Yes	No	Yes	Yes	Unclear
Sandborn et al. (CLASSIC II)	2007	Yes	Yes	Yes	Yes	No	No	No	Unclear
Watanabe et al.	2012	Yes	Unclear	Yes	Yes	No	No	No	Yes
Targan et al.	1997	Yes	Unclear	No	Yes	No	Yes	No	Unclear
Hanauer et al. (ACCENT I)	2002	Yes	Unclear	Yes	Yes	No	No	Unclear	Unclear
Rutgeerts et al. (EXTEND)	2012	Yes	Unclear	Yes	Yes	No	No	Yes	Unclear
Sandborn et al. (GEMINI II)	2013	Yes	Yes	Yes	Yes	No	No	No	Yes
Sands et al. (GEMINI III)	2014	Yes	Yes	Yes	Yes	No	No	Yes	Unclear

Appendix 5. ARisk of bias assessment.

Ref ID	Trial	Random sequence generation	Allocation concealment	Blinding of participants and personnel	Blinding of outcome assessment	Incomplete outcome data	Selective reporting
5	UNITI-I^5	+	+	+	+	?	+
6	UNITI-II^6	+	+	+	+	?	+
7	IM-UNITI^7	+	+	+	+	?	+
23	CHARM^24	+	+	+	+	+	+
24	ACCENT I^25	?	?	?	?	+	+
25	CLASSIC I^26	+	+	+	+	+	+
26	EXTEND^27	+	?	?	?	+	+
27	GAIN^28	+	+	+	+	+	+
28	GEMINI III^29	+	+	+	+	+	+
29	Targan 1997^30	+	+	+	+	+	+
30	Watanabe 2012^31	+	?	?	?	+	+
31	GEMINI II^32	+	+	+	+	+	+
32	CERTIFI^33	+	+	+	+	+	+

Abbreviations. +, Low risk of bias; -, High risk of bias; ?, Unclear risk of bias; RCT, Randomized controlled trial.

Appendix 6. Study design of maintenance phase trials.

Treatment	Study	Patient selection	Study design
Ustekinumab	IM-UNITI	Ustekinumab responders at week 8	DB for induction and maintenance
Adalimumab	CHARM Watanabe 2012 EXTEND	Adalimumab responders (CDAI-70) at week 4 Adalimumab responders (CDAI-70) at week 4 Adalimumab responders (CDAI-70) at week 4	OL for induction and DB for maintenance DB for induction and maintenance OL for induction and DB for maintenance
Infliximab	ACCENT I	Infliximab responders (CDAI-70) at week 2	OL for induction and DB for maintenance
Vedolizumab	GEMINI II	Vedolizumab responders (CDAI-70) at week 6	Mix of OL and DB for induction and DB for maintenance

Abbreviations. CDAI, Crohn`s Disease Activity Index; DB, Double blind; OL, Open label.

Appendix 7. Placebo response/remission rates versus treatment response/remission rates and odds ratios.

Abbreviations. ADA, Adalimumab; CDAI, Crohn’s Disease Activity Index; eow, Every other week; IFX, Infliximab; q4w, Every 4 weeks; q8w, Every 8 weeks; q12w, Every 12 weeks; UST, Ustekinumab; VDZ, Vedolizumab.

Appendix 8. Detailed summary of randomized controlled trial characteristics.Induction

Study	Blinding	Analysis population	Total	Treatments	Prior anti-TNF-α (%)	Sub-population	Primary outcome	Age (mean)	Male (%)	Disease duration (years)	CDAI score
UNITI-1^5	DB	FAS	741	Placebo UST 130 mg UST 6 mg/kg	NR	Failed anti-TNF	Clinical response at week 6 (CDAI 100)	36	42.8	10.1	317
UNITI-2^6	DB	FAS	627	Placebo UST 130 mg UST 6 mg/kg	NR	Failed conventional	Clinical response at week 6 (CDAI 100)	NR	NR	NR	NR
GEMINI II^32	OL	ITT	1115	Placebo VDZ 300 mg	62	Failed conventional & failed anti-TNF	Clinical remission at week 6 (CDAI < 150)	36.1	46.6	9	324
GEMINI III^29	OL	ITT	416	Placebo VDZ 300 mg	76	Failed conventional & failed anti-TNF	Clinical remission at week 6 (CDAI < 150)	37.9	43.3	10.3	307.7
CLASSIC I^26	OL	ITT	299	Placebo ADA 160/80 mg ADA 80/40 mg	NR	Failed conventional	Clinical remission at week 4 (CDAI < 150)	38.25	44.8	NR	297.8
Watanabe 2012^31	DB	FAS	90	Placebo ADA 160/80 mg ADA 80/40 mg	NR	Failed conventional	Clinical remission at week 4 (CDAI < 150)	31.1	57.8	9.5	303.3
GAIN^28	OL	FAS	325	Placebo ADA 160/80 mg	48	Failed anti-TNF	Clinical response at week 4 (CDAI 70/100)	38	45	NR	313
Targan 1997^30	OL	FAS	108	Placebo IFX 5 mg/kg	NR	Failed conventional	Clinical remission at week 4 (CDAI < 150)	37.7	55.8	11.5	300
EXTEND^27	OL	FAS	135	Placebo ADA 160/80 mg	52	Failed conventional & failed anti-TNF	Complete mucosal healing at week 12	37.1	47.2	10.1	319.9

Abbreviations. ADA, Adalimumab; CDAI, Crohn’s Disease Activity Index; DB, Double blind; FAS, Full analysis set; IFX, Infliximab; ITT, Intention to treat; NR, Not reported; OL, Open label; TNF, Tumor necrosis factor; UST, Ustekinumab; VDZ, Vedolizumab.

Maintenance

Study	Blinding	Analysis population	Total	Treatments	Prior anti-TNF-α (%)	Sub-population	Primary outcome	Age (mean)	Male (%)	Disease duration (years)	CDAI score
IM-UNITI^7	DB	FAS	397	Placebo UST 90 mg q8w UST 90 mg q12w	NR	Failed conventional & failed anti-TNF	Clinical remission at week 52 (CDAI < 150)	36^a	42.8^a	10.1^a	317^a
GEMINI II^32	DB	ITT	778	Placebo VDZ 300 mg q4w VDZ 300 mg q8w	62.00	Failed conventional & failed anti-TNF	Clinical remission at week 52 (CDAI < 150)	37.9^a	43.3^a	10.3^a	307.7^a
CHARM^24	DB	ITT	778	Placebo ADA 40 mg eow ADA 40 mg Weekly	48	Failed conventional & failed anti-TNF	Clinical remission at week 56 (CDAI < 150)	36.7	37.7	NR	316.6
ACCENT I^25	DB	ITT	573	Placebo IFX 5 mg IFX 5 & 10 mg	0	Failed conventional	Clinical remission at week 52 (CDAI < 150)	35^b	39^b	7.5^b	299^b
EXTEND^27	DB	ITT	129	Placebo ADA 40 mg eow	52	Failed conventional & failed anti-TNF	Clinical remission at week 52 (CDAI < 150)	37.1^a	47.2^a	10.1^a	319.9^a

^aBaseline characteristics from induction phase.

^bMedian values.

Abbreviations. ADA, Adalimumab; CDAI, Crohn’s Disease Activity Index; DB, Double blind; eow, Every other week; FAS, Full analysis set; IFX, Infliximab; ITT, Intention to treat; NR, Not reported; OL, Open label; q4w, Every 4 weeks; q8w, Every 8 weeks; q12w, Every 12 weeks; TNF, Tumor necrosis factor; UST, Ustekinumab; VDZ, Vedolizumab.

Appendix 9. Treatment sequence calculations for sub-populations.

Studies induction	Studies maintenance	Treatment	N Induction	Induction CDAI-70 response rate	% CDAI-100 maintenance	% CDAI < 150 maintenance	% CDAI-100 sequence	% CDAI < 150 sequence	1 Year responders (CDAI-100)	1 Year remission (CDAI < 150)
UNITI 2	IM-UNITI	Placebo–placebo	209	39%	60%	47%	23%	18%	49	39
		UST 6–UST 90 q12w	209	65%	67%	57%	44%	37%	91	77
		UST 6–UST 90 q8w	209	65%	68%	63%	44%	40%	93	84
GEMINI II & GEMINI III	GEMINI II	Placebo–placebo	127	38%	58%	37%	22%	14%	28	18
		VDZ 300–VDZ 300 q8w	163	53%	60%	50%	32%	26%	52	42
		VDZ 300–VDZ 300 q4w	163	53%	53%	45%	28%	24%	46	39
Classic I & Watanabe	CHARM	Placebo–placebo	84	35%	58%	40%	20%	14%	17	12
		ADA 160/80–ADA 40 eow	90	62%	46%	41%	29%	26%	26	23
		ADA 160/80–ADA 40 weekly	90	62%	56%	48%	36%	30%	32	27
Targan 1997	ACCENT I	Placebo–placebo	24	17%	N/A	36%	N/A	8%	0	2
		IFX 5	27	81%	N/A	28%	N/A	26%	0	7
		IFX 5 & 10	27	81%	N/A	38%	N/A	33%	0	9

Abbreviations. ADA, Adalimumab; CDAI, Crohn’s Disease Activity Index; eow, Every other week; IFX, Infliximab; NA, Not available; q4w, Every 4 weeks; q8w, Every 8 weeks; q12w, Every 12 weeks; UST, Ustekinumab; VDZ, Vedolizumab.

Treatment sequence calculations for failed anti-TNF sub-population

Studies induction	Studies maintenance	Treatment	N Induction	Induction CDAI-70 response rate	% CDAI-100 maintenance	% CDAI < 150 maintenance	% CDAI-100 sequence	% CDAI < 150 sequence	1 Year responders (CDAI-100)	1 Year remission (CDAI < 150)
UNITI 1 & CERFIFI	IM-UNITI	Placebo–placebo	379	30%	39%	35%	12%	11%	45	40
		UST 6–UST 90 q12w	380	45%	47%	39%	21%	16%	79	62
		UST 6–UST 90 q8w	380	45%	48%	41%	22%	17%	82	65
GEMINI II & GEMINI III	GEMINI II	Placebo–placebo	227	31%	38%	34%	12%	11%	27	24
		VDZ 300–VDZ 300 q8w	263	44%	29%	28%	13%	13%	34	33
		VDZ 300–VDZ 300 q4w	263	44%	38%	27%	17%	13%	44	33
GAIN & Watanabe	CHARM	Placebo–placebo	179	34%	36%	30%	12%	11%	22	19
		ADA 160/80–ADA 40 eow	178	53%	36%	31%	19%	15%	34	27
		ADA 160/80–ADA 40 weekly	178	53%	38%	34%	20%	17%	36	31

Abbreviations. ADA, Adalimumab; CDAI = Crohn’s Disease Activity Index; eow, Every other week; IFX, Infliximab; NA, Not available; q4w, Every 4 weeks; q8w, Every 8 weeks; q12w, Every 12 weeks; UST, Ustekinumab; VDZ, Vedolizumab.

Appendix 10. Results from network meta-analysis based on data prior to treatment sequencing adjustments.

Odds ratios (95% CrIs) and probability better for UST versus comparator
Dose	UST 90 mg q8w	UST 90 mg q12w	ADA 40 mg weekly	ADA 40 mg eow	VDZ 300 mg q4w	VDZ 300 mg q8w	IFX 5 mg/kg	IFX 5 & 10 mg/kg	Model fit statistics
Class	UST		ADA		VDZ		IFX
CDAI-100
Reference case – Bayesian hierarchical	1 p(better) = NA	1.01 (0.87 to 1.19) p(better) = 50%	1.37 (0.85 to 2.19) p(better) = 85%	1.42 (0.88 to 2.26) p(better) = 86%	1.58 (1.03 to 2.43) p(better) = 88%	1.61 (1.04 to 2.47) p(better) = 90%	NA	NA	DIC = 77.22 Totresdev = 8.11
Unadjusted maintenance NMA	1 p(better) = NA	1.06 (0.65 to 1.75) p(better) = 53%	0.39 (0.19 to 0.78) p(better) = 20%	0.50 (0.25 to 0.99) p(better) = 25%	0.93 (0.48 to 1.83) p(better) = 44%	1.01 (0.51 to 1.99) p(better) = 50%	NA	NA	DIC = 80.22 Totresdev = 10.47
CDAI <150
Reference case – Bayesian hierarchical	1 p(better) = NA	1.03 (0.89 to 1.26) p(better) = 52%	1.29 (0.78 to 2.12) p(better) = 56%	1.34 (0.82 to 2.18) p(better) = 63%	1.55 (0.98 to 2.43) p(better) = 75%	1.60 (1.01 to 2.53) p(better) = 78%	0.73 (0.14 to 2.69) p(better) = 35%	0.72 (0.14 to 2.64) p(better) = 34%	DIC = 101.28 Totresdev = 12.51
Unadjusted maintenance NMA	1 p(better) = NA	1.20 (0.74 to 1.96) p(better) = 67%	0.37 (0.18 to 0.77) p(better) = 18%	0.46 (0.23 to 0.90) p(better) = 22%	0.99 (0.49 to 2.00) p(better) = 49%	0.89 (0.44 to 1.78) p(better) = 41%	0.81 (0.34 to 1.88) p(better) = 37%	0.52 (0.22 to 1.19) p(better) = 26%	DIC = 104.46 Totresdev = 14.29

Abbreviations. ADA, Adalimumab; CDAI, Crohn’s Disease Activity Index; CrI, Credible interval; DIC, Deviance information criterion; eow, Every other week; IFX, Infliximab; NA, Not available; q4w, Every 4 weeks; q8w, Every 8 weeks; q12w, Every 12 weeks; Totresdev, Total residual deviance; UST, Ustekinumab; VDZ, Vedolizumab.

Bolded values are statistically significant odds ratios.

Appendix 11. Random effects model results.Results from primary analyses for random effects model

Odds ratios (95% CrIs) for UST versus comparator
Dose	UST 90 mg q8w	UST 90 mg q12w	ADA 40 mg weekly		ADA 40 mg eow		VDZ 300 mg q4w	VDZ 300 mg q8w	IFX 5 mg/kg	IFX 5 & 10 mg/kg	Model fit statistics
Class	UST		ADA				VDZ		IFX
CDAI-100
FE model: overall population	1	1.01 (0.87 to 1.19)	1.37 (0.85 to 2.19)		1.42 (0.88 to 2.26)		1.58 (1.03 to 2.43)	1.61 (1.04 to 2.47)	NA	NA	DIC = 77.22 Totresdev = 8.11 vs. 11
	1		1.38 (0.84 to 2.25)				1.58 (1.02 to 2.49)		NA
RE model: overall population	1	1.01 (0.82 to 1.25)	1.36 (0.63 to 2.74)		1.39 (0.64 to 2.81)		1.59 (0.78 to 3.36)	1.61 (0.78 to 3.38)	NA	NA	DIC = 78.76 Totresdev = 9.10 vs. 11
	1		1.37 (0.63 to 2.79)				1.60 (0.77 to 3.40)		NA
CDAI <150
FE model: overall population	1	1.03 (0.89 to 1.26)	1.29 (0.78 to 2.12)		1.34 (0.82 to 2.18)		1.55 (0.98 to 2.43)	1.60 (1.01 to 2.53)	0.73 (0.14 to 2.69)	0.72 (0.14 to 2.64)	DIC = 101.28 Totresdev = 12.51 vs. 16
	1		1.29 (0.77 to 2.15)				1.54 (0.95 to 2.47)		0.71 (0.14 to 2.62)
RE model: overall population	1	1.02 (0.84 to 1.28)	1.26 (0.61 to 2.40)		1.29 (0.63 to 2.44)		1.52 (0.76 to 3.07)	1.56 (0.77 to 3.13)	0.67 (0.13 to 2.51)	0.66 (0.13 to 2.46)	DIC = 102.45 Totresdev = 13.04 vs. 12
	1		1.25 (0.60 to 2.43)				1.51 (0.75 to 3.12)		0.65 (0.13 to 2.45)

Abbreviations. ADA, Adalimumab; CDAI, Crohn’s Disease Activity Index; CrI, Credible interval; DIC, Deviance information criterion; eow, Every other week; FE, Fixed effects; IFX, Infliximab; NA, Not available; q4w, Every 4 weeks; q8w Every 8 weeks; q12w, Every 12 weeks; RE, Random effects; Totresdev, Total residual deviance; UST, Ustekinumab; VDZ, Vedolizumab.

Bolded values are statistically significant odds ratios.

Appendix 12. Inconsistency models

Appendix 13. PRISMA network meta-analysis checklist.

Section/topic	Item #	Checklist item	Reported on page #
TITLE
Title	1	Identify the report as a systematic review incorporating a network meta-analysis (or related form of meta-analysis).	1
ABSTRACT
Structured summary	2	Provide a structured summary including, as applicable:	1
		Background: main objectives.
		Methods: data sources; study eligibility criteria, participants, and interventions; study appraisal; and synthesis methods, such as network meta-analysis.
		Results: number of studies and participants identified; summary estimates with corresponding confidence/credible intervals; treatment rankings may also be discussed. Authors may choose to summarize pairwise comparisons against a chosen treatment included in their analyses for brevity.
		Discussion/conclusions: limitations; conclusions and implications of findings.
		Other: primary source of funding; systematic review registration number with registry name.
INTRODUCTION
Rationale	3	Describe the rationale for the review in the context of what is already known, including mention of why a network meta-analysis has been conducted.	1
Objectives	4	Provide an explicit statement of questions being addressed, with reference to participants, interventions, comparisons, outcomes and study design (PICOS).	2
METHODS
Protocol and registration	5	Indicate whether a review protocol exists and if and where it can be accessed (e.g. web address); and, if available, provide registration information, including registration number.	2
Eligibility criteria	6	Specify study characteristics (e.g. PICOS, length of follow-up) and report characteristics (e.g. years considered, language, publication status) used as criteria for eligibility, giving rationale. Clearly describe eligible treatments included in the treatment network, and note whether any have been clustered or merged into the same node (with justification).	2
Information sources	7	Describe all information sources (e.g. databases with dates of coverage, contact with study authors to identify additional studies) in the search and date last searched.	2
Search	8	Present full electronic search strategy for at least one database, including any limits used, such that it could be repeated.	1
Study selection	9	State the process for selecting studies (i.e. screening, eligibility, included in systematic review and, if applicable, included in the meta-analysis).	2
Data collection process	10	Describe method of data extraction from reports (e.g. piloted forms, independently, in duplicate) and any processes for obtaining and confirming data from investigators.	2
Data items	11	List and define all variables for which data were sought (e.g. PICOS, funding sources) and any assumptions and simplifications made.	2
Geometry of the network	S1	Describe methods used to explore the geometry of the treatment network under study and potential biases related to it. This should include how the evidence base has been graphically summarized for presentation, and what characteristics were compiled and used to describe the evidence base to readers.	3
Risk of bias within individual studies	12	Describe methods used for assessing risk of bias of individual studies (including specification of whether this was done at the study or outcome level), and how this information is to be used in any data synthesis.	2
Summary measures	13	State the principal summary measures (e.g. risk ratio, difference in means). Also describe the use of additional summary measures assessed, such as treatment rankings and surface under the cumulative ranking curve (SUCRA) values, as well as modified approaches used to present summary findings from meta-analyses.	4
Planned methods of analysis	14	Describe the methods of handling data and combining results of studies for each network meta-analysis. This should include, but not be limited to:  • Handling of multi-arm trials;  • Selection of variance structure;  • Selection of prior distributions in Bayesian analyses; and  • Assessment of model fit.	5
Assessment of inconsistency	S2	Describe the statistical methods used to evaluate the agreement of direct and indirect evidence in the treatment network(s) studied. Describe efforts taken to address its presence when found.	5
Risk of bias across studies	15	Specify any assessment of risk of bias that may affect the cumulative evidence (e.g. publication bias, selective reporting within studies).	2
Additional analyses	16	Describe methods of additional analyses if done, indicating which were pre-specified. This may include, but not be limited to, the following:  • Sensitivity or subgroup analyses;  • Meta-regression analyses;  • Alternative formulations of the treatment network; and  • Use of alternative prior distributions for Bayesian analyses (if applicable).	6
RESULTS*
Study selection	17	Give numbers of studies screened, assessed for eligibility and included in the review, with reasons for exclusions at each stage, ideally with a flow diagram.	Figure 1
Presentation of network structure	S3	Provide a network graph of the included studies to enable visualization of the geometry of the treatment network.	Figure 5
Summary of network geometry	S4	Provide a brief overview of characteristics of the treatment network. This may include commentary on the abundance of trials and randomized patients for the different interventions and pairwise comparisons in the network, gaps of evidence in the treatment network, and potential biases reflected by the network structure.	10 6
Study characteristics	18	For each study, present characteristics for which data was extracted (e.g. study size, PICOS, follow-up period) and provide the citations.	6
Risk of bias within studies	19	Present data on risk of bias of each study and, if available, any outcome level assessment.	Figure 2
Results of individual studies	20	For all outcomes considered (benefits or harms), present, for each study: 1) simple summary data for each intervention group; and 2) effect estimates and confidence intervals. Modified approaches may be needed to deal with information from larger networks.	6
Synthesis of results	21	Present results of each meta-analysis done, including confidence/credible intervals. In larger networks, authors may focus on comparisons versus a particular comparator (e.g. placebo or standard care), with full findings presented in an appendix. League tables and forest plots may be considered to summarize pairwise comparisons. If additional summary measures were explored (such as treatment rankings), these should also be presented.	6 Table 2
Exploration for inconsistency	S5	Describe results from investigations of inconsistency. This may include such information as measures of model fit to compare consistency and inconsistency models, p values from statistical tests, or summary of inconsistency estimates from different parts of the treatment network.	10
Risk of bias across studies	22	Present results of any assessment of risk of bias across studies for the evidence base being studied.	Appendix 4
Results of additional analyses	23	Give results of additional analyses, if done (e.g. sensitivity or subgroup analyses, meta-regression analyses, alternative network geometries studied, alternative choice of prior distributions for Bayesian analyses, and so forth).	10 Table 3
DISCUSSION
Summary of evidence	24	Summarize the main findings, including the strength of evidence for each main outcome; consider their relevance to key groups (e.g. healthcare providers, users and policy-makers).	10–11
Limitations	25	Discuss limitations at study and outcome level (e.g. risk of bias), and at review level (e.g. incomplete retrieval of identified research, reporting bias). Comment on the validity of the assumptions, such as transitivity and consistency. Comment on any concerns regarding network geometry (e.g. avoidance of certain comparisons).	11
Conclusions	26	Provide a general interpretation of the results in the context of other evidence, and implications for future research.	11
FUNDING			1
Funding	27	Describe sources of funding for the systematic review and other support (e.g. supply of data); role of funders for the systematic review. This should also include information regarding whether funding has been received from manufacturers of treatments in the network and/or whether some of the authors are content experts with professional conflicts of interest that could affect use of treatments in the network.

Abbreviation. PICOS Population, intervention, comparators, outcomes, study design.

Text in italics indicates wording specific to reporting of network meta-analyses that has been added to guidance from the PRISMA statement.

*Authors may wish to plan for use of appendices to present all relevant information in full detail for items in this section.

Source: Hutton B, Salanti G, Caldwell DM, et al. The PRISMA extension statement for reporting of systematic reviews incorporating network meta-analyses of health care interventions: checklist and explanations. Ann Intern Med. 2015;162(11):777–784.