Objective
The objective of this appendix is to summarize and critically appraise the indirect treatment comparisons for insulin degludec (IDeg) in adults with type 1 or type 2 diabetes mellitus (T1DM or T2DM). Head-to-head studies were available comparing IDeg to insulin glargine (IGlar) and insulin detemir (IDet), but not to other basal insulins. Thus, a review of indirect comparisons was warranted.
In addition to the systematic search (see Appendix 2), a literature search was conducted by CADTH (MEDLINE, EMBASE, up to July 7, 2017) to identify potentially relevant indirect comparisons that included IDeg. Two published network meta-analyses (NMAs) were identified.62,63 The manufacturer also provided an NMA as part of the CADTH Common Drug Review (CDR) submission.12
Description of Indirect Comparisons Identified
The manufacturer-submitted NMA included patients with T1DM and T2DM, whereas the report by Freemantle et al.63 included T2DM patients and the report by Dawoud et al.62 included T1DM patients (). The manufacturer’s NMA was limited to three treatments (IDeg, IGlar, and neutral protamine Hagedorn [NPH] insulin) and focused on hypoglycemia events as the primary outcome. The other two NMAs included all the basal insulins of interest to this CDR review and examined efficacy outcomes (e.g., glycated hemoglobin [A1C], body weight) as well as hypoglycemia.
Freemantle et al. and Dawoud et al. conducted systematic review to identify studies for inclusion in the NMA. The manufacturer-submitted NMA was based on data from two other meta-analyses: an analysis conducted by Novo Nordisk and published as a poster in 201264 that compared IDeg and IGlar, and a meta-analysis conducted by CADTH published in 2008 that compared IGlar and NPH. The NMA authors conducted an update to the CADTH meta-analysis and searched for any randomized controlled trials (RCTs) comparing IGlar with insulin NPH that had been published since the CADTH review. This update identified six relevant trials, of which three were included in the NMA (two pediatric were excluded, along with one other study that reported data as dichotomous events only).
All three indirect comparisons conducted Bayesian NMAs with Markov Chain Monte Carlo simulations. A summary of the indirect comparison methods has been included in . Freemantle and Dawoud analyzed continuous outcomes using a generalized linear model with normal likelihood and identity link function; hypoglycemia rates were analyzed using a Poisson process and a log-link function. The manufacturer-submitted NMA analyzed hypoglycemia rate ratio data from individual treatment comparisons on a log scale using a logistic regression model and a normal likelihood.
Separate analyses were conducted in patients with T2DM based on the patients’ treatment history (e.g., basal insulin, basal + bolus insulin, or insulin-naive). Freemantle et al. and Dawoud et al. conducted several sensitivity analyses or used meta-regression to explore potential sources of heterogeneity or model assumptions. Dawoud et al. analyzed both fixed-effects and random-effects models and selected the random-effects model based on model fit. Freemantle et al. ran random-effects models (no justification provided or exploration of model fit). The authors of the manufacturer-submitted NMA stated that a fixed-effects model was selected due to the limited number of studies available. However, there were issues with model fit, and in at least two cases, a random-effects model was reported to have a better fit. The authors did not report residual deviance or deviance information criterion values to allow the reader to compare models.
Table 69Network Meta-Analysis Study Characteristics
View in own window
| Manufacturer-Submitted12 | Freemantle et al., 201663 | Dawoud et al., 201762 |
|---|
| SR Criteria | | | |
|---|
| Population | Adults with T1DM or T2DM | Adults with T2DM treated with basal insulin (with or without bolus insulin) Studies must include patients from at least one of the following countries: US, France, Germany, UK, Spain, or Italy | Adults with T1DM |
|---|
| Interventions |
(Doses not specified) |
(Doses not specified) |
Insulin at UK licensed doses |
|---|
| Outcomes | Severe hypoglycemia Nocturnal hypoglycemia Overall hypoglycemia
(No definitions for these events were provided) | Change from baseline in A1C (%) Change from baseline in body weight (kg) Documented symptomatic hypoglycemia events per PY (defined as symptoms of hypoglycemia were accompanied by measured plasma glucose level below a threshold; no limit placed on plasma glucose thresholds used in trials) Nocturnal hypoglycemia events per PY (confirmed or symptomatic event occurring overnight)
| |
|---|
| Study Design | RCTs > 48 hours in duration | English-language, published RCTs > 20 weeks in duration | English-language, published RCTs > 4 weeks in duration |
|---|
| Exclusions | Pediatric studies | None listed | Premixed insulin, studies in children or pregnant women, crossover studies, studies reporting data for mixed populations (T1DM and T2DM), studies comparing different dosages of the same insulin or those using different short-acting insulins across treatment groups, abstracts, letters, or review articles |
|---|
| SR Methods | Data from several sources: meta-analysis comparing IDeg and glargine conducted by Novo Nordisk and published as a poster in 2012; meta-analysis comparing IGlar and NPH conducted by CADTH in 2008; and an update to the CADTH meta-analysis conducted by Abacus International in 2012 Other than the PICO elements, no details provided on the methods to conduct the update to the CADTH report No details provided on the meta-analysis of IDeg/IGlar No quality assessment of individual trials
| Literature search (1980 to date not specified) of multiple databases for English-language RCTs Conference abstracts from specific diabetes congresses also searched (2011 to 2013) Abstracts and posters accepted if they were the terminal source document; clinical study reports for IGlar 300 U/mL trials included Study selection and data extraction by two independent researchers Individual studies assessed for methods of randomization, allocation concealment, blinding, and losses to follow-up
| Literature search (up to August 2014) of multiple databases for published English-language RCTs Reference lists of included studies searched for RCTs Study selection and appraisal by one researcher Study quality assessed using the Cochrane risk of bias tool Data extraction verified by a second reviewer Hypoglycemia rate calculated if necessary based on the number of events divided by the mean follow-up duration multiplied by the sample size Studies reporting zero events in both groups excluded from the NMA
|
|---|
| Analysis | | | |
|---|
| NMA Methods | Bayesian NMA using WinBUGS software NMA for continuous data was run using a fixed effect model (due to small number of trials). Rate ratio and 95% CI data for each treatment comparison from RCTs were converted to natural log and standard error calculated. A normal likelihood on the log rate ratio was used to calculate rate ratios based on a logistic regression model with MCMC simulation. Non-informative priors No adjustment for sparse data or rare events was reported to be necessary in the NMA Goodness of fit was tested with residual deviance and DIC. Burn-in of 20,000 iterations with 50,000 iterations was used to estimate posterior distribution. Separate analyses were planned for different subpopulations: T1DM, T2DM receiving basal insulin; T2DM receiving basal plus bolus insulin; all T2DM patients. No evaluation of consistency was possible, as there were no closed loops No methods to assess model convergence were described. The author stated that it was not possible to conduct metaregression or subgroup analyses due to the small number of studies. Some sensitivity analyses were run excluding some trials (with poor model fit) but it was unclear if these analyses were pre-planned or if the methods used to select trials for exclusion were appropriate.
| Direct meta-analysis was conducted using an inverse variance-weighted method. Bayesian NMA with MCMC using OpenBUGS (random-effects model) was conducted based on NICE guidance. Continuous outcomes were modelled assuming a normal likelihood and an identity link. Event rate data were modelled using a Poisson mixed likelihood and log link. Non-informative priors Base-case analysis included patients on BOT. Sensitivity analysis included all patients (BOT and basal/bolus insulin); patients on BOT excluding premixed insulin; insulin-naive patients; studies with week 24 to week 28 results; excluding IDeg three-times-a-week dosing. Meta-regression was conducted adjusting for baseline study A1C, diabetes duration, and basal-bolus population. Analyses using a broader definition of hypoglycemia were also conducted (no details provided). No information describing how convergence, goodness of fit, or inconsistency were assessed; no information on number of burn-in iterations.
| Direct meta-analyses were conducted using Review Manager (random-effects model). Bayesian NMA was conducted using WinBUGS. NMA was conducted as per NICE Decision Support Unit recommendations. For the change from baseline in A1C (%) a generalized linear model with normal likelihood and identity link function was used. Parameters were estimated using MCMC simulation. Insulin NPH b.i.d. was used as the baseline treatment effect (standard of care in UK) with a mean change in A1C of −0.32% (95% CI, −0.49 to −0.15%) based on a single-arm meta-analysis of seven studies. For the severe hypoglycemia event rate, a Poisson process with a constant event rate was assumed, and a log-link function used to model the event rate. Baseline event rate of 0.35 events per PY (95% CrI 0.11, 0.95) based on single-arm Bayesian meta-analysis in the NPH b.i.d. trials. Non-informative priors (mean 0 SD 100, normal distribution for difference in A1C, not specified for hypoglycemia). 100,000 burn-in and 100,000 simulations for parameter estimates. Convergence assessed by examining the history, kernel density plots, and Brooks– -Gelman–Rubin plots. Goodness of fit was tested using residual deviance. Random-effects and fixed-effects models were run and DIC used to select model. Inconsistency assessed using Bucher test (A1C) and chi-square test for inconsistency (hypoglycemia). Sensitivity analyses were run excluding open-label studies, those with inadequate allocation concealment, using half-normal prior distribution for between-study heterogeneity instead of vague priors, and testing the impact of treatment effect estimate from the largest study on the pooled treatment effect.
|
|---|
| Included Studies | 20 RCTs | 86 studies were identified for data extraction (41 included in NMA)a | 29 RCTs included in SR (28 included in NMA) |
|---|
| Funding | Novo Nordisk | Sanofi | NICE |
|---|
A1C = glycated hemoglobin; b.i.d. = twice daily; BOT = basal insulin–supported oral therapy; CI = confidence interval; CrI = credible interval; DIC = deviance information criterion; IDeg = insulin degludec; IGlar = insulin glargine; IDet = insulin detemir; MCMC = Markov Chain Monte Carlo; NICE = National Institute for Health and Care Excellence; NMA = network meta-analysis; NPH = neutral protamine Hagedorn; PICO = patient/problem/population, intervention, comparison/control/comparator, outcome; PY = patient-year; q.d. = once daily; q.i.d. = four times daily; RCT = randomized controlled trial; SD = standard deviation; SR = systematic review; T1DM = type 1 diabetes mellitus; T2DM = type 2 diabetes mellitus.
- a
No details were provided regarding the reason for exclusion for the 45 studies, except a statement that the studies had to have at least two treatment groups with a relevant insulin in the network.
Summary of the Manufacturer-Submitted Network Meta-Analysis
A total of 20 RCTs were included in the manufacturer-submitted NMA, including seven RCTs comparing IDeg with IGlar (T1DM: studies 3583 and 3770 [total N = 958]; T2DM: Studies 3579, 3672, 3586, 3582, and 3668 [total N = 3,389]). For the comparison between IGlar and NPH, 13 RCTs were included (T1DM: seven RCTs, N = 1,739; T2DM: six RCTs, N = 2,813). The trials were 12 weeks to five years in duration; except for one single-blind study, all were open label. No evaluation of potential sources of bias in the included trials was reported. The mean age per treatment group of patients enrolled ranged from 31.3 to 44.5 years for those with T1DM and from 50.3 to 59.3 years for those with T2DM. The mean baseline glycated hemoglobin (A1C) was 7.7% per treatment group among patients with T1DM, and ranged from 8.2% to 8.5% among patients with T2DM in the trials comparing IDeg with IGlar. The duration of diabetes ranged from 18.2 to 20.0 years in the T1DM trials and from 8.0 to 13.6 years in the T2DM trials for IDeg versus IGlar. No information on the mean A1C or duration of diabetes was provided for trials comparing IGlar with NPH. In total, 637 patients with T1DM and 2,275 with T2DM who were treated with IDeg were included in the NMA. The network diagram is presented in .
The outcomes analyzed included severe hypoglycemia, nocturnal hypoglycemia, and overall hypoglycemia. No definitions for these events were provided. A Bayesian NMA was conducted using the log rate ratio data from each trial (details provided in ). Data from the flexible dosing groups in sStudies 3770 and 3668 were excluded from the NMA. Four different populations were proposed for analysis: T1DM (nine RCTs), T2DM receiving basal insulin (eight RCTs), T2DM receiving basal + bolus insulin (unable to assess due to insufficient data), and all T2DM patients (basal or basal + bolus insulin, 11 RCTs).
The results of the manufacturer-submitted NMA are included in . Among patients with T1DM, no statistically significant differences were detected between IDeg, IGlar, and NPH on the rate of severe hypoglycemia or overall hypoglycemia based on a fixed-effects model. The rate of nocturnal hypoglycemia was lower for IDeg versus NPH based on the fixed-effects model; however, the model fit was better with the random-effects model, which showed no statistically significant differences. No significant differences were detected in the rate of nocturnal hypoglycemia for IDeg versus IGlar based on the fixed-effects or random-effects models.
In the analyses in patients with T2DM who received basal insulin, the authors stated that both the fixed-effects and random-effects models for overall hypoglycemia rates showed poor model fit with high residual deviance values. The model fit was better for the random-effects model than the fixed-effects model in the analysis of severe hypoglycemia. The random-effects model found no statistically significant differences between treatments, whereas the fixed-effects model suggested a lower rate of severe hypoglycemia for IDeg than for IGlar or NPH. The rates of nocturnal hypoglycemia events were lower for IDeg versus IGlar or NPH based on the fixed-effects model, which also showed some evidence of poor model fit. Residual deviance and deviance information criterion values were not reported for any NMA.
The NMA that included all T2DM patients (who received basal insulin or basal + bolus insulin) found no statistically significant differences between IDeg and IGlar on the incidence of severe hypoglycemia; however, the rates of nocturnal and overall hypoglycemia events were lower with IDeg. The indirect data suggested that the rates of severe, nocturnal, or any hypoglycemia events were lower for IDeg than for NPH. No information on model fit was provided.
Table 70Results from Manufacturer-Submitted Network Meta-Analysis — T1DM and T2DM
View in own window
| Severe Hypoglycemia Median Rate Ratio (95% CrI)a | Nocturnal Hypoglycemia Median Rate Ratio (95% CrI)a | Overall Hypoglycemia Median Rate Ratio (95% CrI)a |
|---|
| T1DM | | | |
|---|
| IDeg versus IGlar | 1.12 (0.68 to 1.85)b | Fixed: 0.83 (0.69 to 1.00)
Random: 0.85 (0.29 to 2.55)c | 1.10 (0.97 to 1.26)b |
|---|
| IDeg versus insulin NPH | 1.09 (0.64 to 1.87)b | Fixed: 0.57 (0.47 to 0.69)
Random: 0.57 (0.15 to 2.19)c | 1.11 (0.97 to 1.27)b |
|---|
| IGlar versus insulin NPH | 0.98 (0.82 to 1.17)b | Fixed: 0.69 (0.65 to 0.72)
Random: 0.62 (0.31 to 1.46)c | 1.00 (0.98 to 1.03)b |
|---|
| T2DM, Basal | | | |
|---|
| IDeg versus IGlar | Fixed: 0.14 (0.03 to 0.68)
Random: 0.14 (0.01 to 2.89)c | 0.63 (0.47 to 0.85)b | 0.83 (0.70 to 0.99)d |
|---|
| IDeg versus insulin NPH | Fixed: 0.07 (0.01 to 0.37)
Random: 0.07 (0.00 to 2.16)c | 0.31 (0.23 to 0.41)b | 0.68 (0.57 to 0.82)d |
|---|
| IGlar versus insulin NPH | Fixed: 0.53 (0.33 to 0.84)
Random: 0.50 (0.10 to 2.50)c | 0.48 (0.45 to 0.52)b | 0.83 (0.79 to 0.87)d |
|---|
| T2DM, All Patients | | | |
|---|
| IDeg versus IGlar | 0.84 (0.46 to 1.51)e | 0.69 (0.58 to 0.83)e | 0.83 (0.74 to 0.93)e |
|---|
| IDeg versus insulin NPH | 0.49 (0.24 to 0.97)e | 0.33 (0.27 to 0.41)e | 0.69 (0.61 to 0.78)e |
|---|
| IGlar versus insulin NPH | 0.58 (0.40 to 0.84)e | 0.48 (0.45 to 0.52)e | 0.83 (0.79 to 0.87)e |
|---|
CrI = credible interval; DIC = deviance information criterion; IDeg = insulin degludec; IGlar = insulin glargine; NPH = neutral protamine Hagedorn; T1DM = type 1 diabetes mellitus; T2DM = type 2 diabetes mellitus.
- a
Fixed-effects model unless otherwise specified.
- b
Some evidence of poor model fit. Sensitivity analyses excluding specific trials with poor fit were run and showed results similar to the base case. No residual deviance values were reported, and no random-effects models were analyzed.
- c
The model fit was better for the random-effects model than the fixed-effects model (no DIC values were reported).
- d
Model fit for fixed-effects and random-effects models were poor with high residual deviance values.
- e
No information provided on model fit.
Source: CADTH Common Drug Review submission for Tresiba.12
Summary of the Network Meta-Analysis by Freemantle Et Al
The NMA by Freemantle et al.63 included 41 RCTs. Twelve of the studies (29%) had unclear allocation concealment and 40 (98%) were open label, with losses to follow-up ranging from 1.6% to 28.5%. Basal insulin–supported oral therapy was used in 25 (61%) trials (N = 15,746); patients in these studies had a mean age per study ranging from 52.4 to 61.7 years, duration of diabetes ranging from 8.2 to 13.8 years, and A1C ranging from 7.8% to 9.8%. Five of the included RCTs compared IDeg once daily (Studies 3582, 3672, 3668, and 3579)54,58,60,65 or three times per week (Study 3724)66 with IGlar (100 U/mL). The network diagram for A1C is presented in .
IDeg was associated with a statistically significant increase in A1C and body weight compared with IGlar 100 U/mL in the direct meta-analyses (). However, the clinical importance of the differences is unclear given the magnitude of the differences observed (mean difference in A1C: 0.13%; weight: 0.21 kg). No statistically significant differences in glycemic control or body weight were detected between IDeg and IGlar (100 U/mL or 300 U/mL) in the NMA.
Based on the direct meta-analysis, IDeg was associated with a statistically significantly lower rate of nocturnal hypoglycemia but a higher rate of symptomatic hypoglycemia compared with IGlar (100 U/mL). However, based on the NMA, the differences between treatments were not significant. The point estimates were similar in the direct meta-analysis and the NMA, but the direct meta-analysis had tighter confidence intervals.
The authors stated that results were generally similar across the numerous sensitivity analyses that were conducted. Data for all T2DM patients and insulin-naive patients are presented in , as well as the analyses that excluded the IDeg three-times-a-week dosing study. No results were reported comparing IDeg with NPH, IDet, or premixed insulin.
Table 71Results From Network Meta-Analysis by Freemantle 2016 (T2DM)
View in own window
| Change From Baseline in A1C (%) | Change From Baseline in Body Weight (kg) | Risk of Nocturnal Hypoglycemia | Risk of Symptomatic Hypoglycemia |
|---|
| Meta-Analysis | MD (95% CI) | MD (95% CI) | RR (95% CI) | RR (95% CI) |
|---|
| IDeg versus IGlar (100 U/mL) | 0.13 (0.06 to 0.20) | 0.21 (0.03 to 0.38) | 0.79 (0.67 to 0.93) | 1.35 (1.27 to 1.44) |
|---|
| NMA in Patients on BOT | MD (95% CrI)a | MD (95% CrI)a | RR (95% CrI)a | RR (95% CrI)a |
|---|
| IDeg versus IGlar (100 U/mL) | 0.14 (−0.03 to 0.30) | 0.18 (−0.35 to 0.70) | 0.88 (0.57 to 1.38) | 1.30 (0.75 to 2.24) |
|---|
| IGlar (300 U/mL) versus IDeg | −0.12 (−0.42 to 0.20) | −0.63 (−1.63 to 0.35) | 0.66 (0.28 to 1.50) | 0.55 (0.23 to 1.34) |
|---|
| Sensitivity Analyses | | | | |
|---|
| All T2DM patients | | | | |
|---|
| IGlar (300 U/mL) versus IDeg | −0.12 (−0.42 to 0.18) | −0.35 (−1.58 to 0.88) | 0.75 (0.41 to 1.34) | 0.64 (0.36 to 1.16) |
|---|
| Insulin-naive | | | | |
|---|
| IGlar (300 U/mL) versus IDeg | −0.12 (−0.62 to 0.37) | −0.46 (−1.71 to 0.80) | 0.61 (0.10 to 3.48) | 0.61 (0.17 to 2.25) |
|---|
| Excluding IDeg three times per week dosing | | | | |
|---|
| IGlar (300 U/mL) versus IDeg | −0.01 (−0.32 to 0.31) | −0.79 (−1.90 to 0.33) | 0.83 (0.42 to 1.46) | NR |
|---|
A1C = glycated hemoglobin; BOT = basal insulin–supported oral therapy; CI = confidence interval; CrI = credible interval; IDeg = insulin degludec; IGlar = insulin glargine; MD = mean difference; NMA = network meta-analysis; NR = not reported; RR = rate ratio; T2DM = type 2 diabetes mellitus.
- a
Source: Freemantle 2016.63
Summary of the NMA by Dawoud et al
The NMA by Dawoud et al. included 29 RCTs in patients with T1DM. Of these, one was excluded from the NMA, as it did not report A1C and there were no hypoglycemia events in any treatment group. All trials were considered to have moderate or high risk of bias, mainly due to allocation concealment or lack of blinding. The trials ranged from four weeks to 52 weeks in duration and had a sample size ranging from 51 to 629 patients. No other details regarding the study or patient characteristics were reported.
Three of the included RCTs compared IDeg with IGlar (Studies 1835, 3583, and 3770).5,67,68
Data from 25 RCTs were included in the NMA for the change from baseline in A1C (%). The included studies were rated as low to high risk of bias. Three studies were excluded because they did not report A1C. The analysis of severe hypoglycemia included 16 RCTs that had serious or very serious risk of bias (). Twelve RCTs were excluded because they did not report severe hypoglycemia, or did not report it as a rate, or had zero events in all treatment groups.
The direct meta-analysis of three RCTs showed no statistically significant difference between IDeg and IGlar in the change from baseline in A1C (mean difference 0.07; 95% confidence interval [CI] −0.02 to 0.17) (). For the NMA, the random-effects model had a better fit with a lower deviance information criterion than the fixed-effects model. Based on the random-effects model, no statistically significant differences were detected between IDeg and NPH (once or twice daily), IGlar (daily), or IDet (once daily or twice daily), but A1C was statistically significantly lower for IDeg versus NPH administered four times daily (mean difference −0.34%; 95% credible interval, −0.59% to −0.11%). No statistically significant difference was detected between IDeg and either NPH or IDet administered in a mixed population that received either once-daily or twice-daily basal insulin (data not summarized in this report).
No statistically significant difference was detected between IDeg and IGlar on the rate of severe hypoglycemia based on a direct meta-analysis of two RCTs (rate ratio 1.03; 95% CI, 0.63 to 1.67), or between IDeg and IGlar, IDet, or NPH based on the NMA (). There was substantial uncertainty in some of the NMA estimates given the wide credible intervals reported; however, the direct and indirect estimates were consistent for IDeg versus IGlar.
Table 72Results From Network Meta-Analysis by Dawoud 2017 (T1DM)
View in own window
| Change From Baseline in A1C (%) | Severe Hypoglycemia |
|---|
| Meta-Analysis | MD (95% CI)a | RR (95% CI)a |
|---|
| IGlar (q.d.) versus IDeg | 0.07 (−0.02 to 0.17) | 1.03 (0.63 to 1.67) |
|---|
| NMA | MD (95% CrI)a | Median HR (95% CrI)a |
|---|
| Insulin NPH (q.d.) versus: | | |
|---|
| IDeg (q.d.) | 0.07 (−0.25 to 0.38) | 1.09 (0.27 to 4.93) |
|---|
| IDeg (q.d.) versus: | | |
|---|
| IGlar (q.d.) | 0.07 (−0.08 to 0.22) | 1.04 (0.39 to 2.72) |
|---|
| IDet (q.d.) | 0.04 (−0.19 to 0.28) | 0.82 (0.11 to 5.68) |
|---|
| IDet (b.i.d.) | −0.13 (−0.12 to 0.39) | 1.11 (0.02 to 69.82) |
|---|
| Insulin NPH (b.i.d.) | −0.03 (−0.31 to 0.26) | 1.07 (0.02 to 63.67) |
|---|
| Insulin NPH (q.i.d.) | −0.24 (−0.59 to −0.11) | NR |
|---|
A1C = glycated hemoglobin; b.i.d. = twice daily; CI = confidence interval; CrI = credible interval; HR = hazard ratio; IDeg = insulin degludec; IDet = insulin detemir; IGlar = insulin glargine MD = mean difference; NMA = network meta-analysis; NPH = neutral protamine Hagedorn; NR = not reported; q.d. = once daily; q.i.d. = four times daily; RR = rate ratio; T1DM = type 1 diabetes mellitus.
- a
Appraisal of NMAs
Key limitations of the three analyses have been summarized in . The quality of the included studies was an issue of concern in all three reports. The evidence was based largely on open-label trials; thus, subjective outcomes, such as hypoglycemia, may be prone to bias. The scope of the manufacturer’s NMA was limited to three basal insulins and excluded IDet. In addition, there was no systematic search for relevant trials, so it is unclear if all potentially relevant studies were included. In Dawoud et al.’s and the manufacturer-submitted reports, the data describing the study and patient characteristics were limited; thus, it was not possible for the reader to assess if the included studies were sufficiently similar to pool. There was no information on insulin dosing or glycemic targets in any of the reports. Limited information was available in all three reports on concurrent diabetes medications. Use of sulfonylureas, which have an increased risk of hypoglycemia, was more common in the trials for NPH versus IGlar than for those comparing IDeg and IGlar in the manufacturer’s NMA. The clinical importance of these differences in unclear.
The methods used to conduct the Bayesian NMA by Dawoud et al. were clearly and completely reported and appear to meet accepted standards. Several sensitivity analyses were run to test the robustness of the findings. The network was sparse, particularly for severe hypoglycemia; therefore, there is uncertainty in the estimates.
Freemantle et al. also conducted a Bayesian NMA, but only a random-effects model was run, with no justification for selecting this model and no exploration of other models or testing model assumptions (e.g., impact of informative versus non-informative priors). No information was provided on how convergence, goodness of fit, or inconsistency were assessed. Complete results from all treatment comparisons were not reported; thus, many comparisons of interest in this review were not available.
In the manufacturer-submitted NMA, the log rate ratio of hypoglycemia events was analyzed as a continuous outcome. This is in contrast to the other two NMAs that used a Poisson model with follow-up time as an offset. In the manufacturer’s NMA, the trial duration was 26 to 52 weeks for studies comparing IDeg with IGlar, whereas the duration of trials comparing NPH with IGlar ranged from 12 weeks to five years. The authors stated that a fixed-effects model was chosen because of the limited number of trials that informed the network; thus, there were too few studies with which to estimate between-study heterogeneity. Although model fit was tested, the deviance information criterion or residual deviance values were not reported. The authors stated that there were issues with model fit, and based on residual deviance values calculated for individual trials, some studies with poor fit were excluded in a sensitivity analysis. It is unclear if these sensitivity analyses were pre-planned, as they were not described in the methods section; furthermore, purposeful exclusion of studies based on deviance values calculated for individual studies is unconventional, and exclusion of these studies may increase the risk of bias in the results. For some analyses, a random-effects model was run, and in two cases, the model fit was better with the random-effects estimates. Ideally, both fixed-effects and random-effects models would have been run, and informative and non-informative priors used for the between-study variance parameter in the random-effects model. Given the issues with model fit in the analysis of nocturnal hypoglycemia in patients with T1DM and severe hypoglycemia in those with T2DM on basal insulin, the results of the random-effects model should be preferred over those of the fixed-effects model. The authors stated that model fit was poor for both the fixed-effects and random-effects models for overall hypoglycemia in patients with T2DM on basal insulin; thus, the results of this analysis should be interpreted with caution.
Another limitation of the manufacturer’s report is that no direct meta-analyses were conducted. Because the network had no closed loops, it was not possible to test for inconsistency with the usual methods. Thus, having access to both a traditional meta-analysis and NMA estimates would be desirable. This is especially true given the concerns with model fit in the NMA. Had the meta-analysis results been similar to the NMA findings, it would have increased confidence in the NMA’s findings. The lack of sensitivity analyses beyond model fit is a further limitation. There was limited exploration of the effects of heterogeneity on study results.
Table 73Key Limitations
View in own window
| Manufacturer12 | Freemantle et al., 201663 | Dawoud et al., 201762 |
|---|
| SR | No information was provided on the methods used to identify relevant studies that informed two of the previously published meta-analyses, and limited data were provided for the update conducted by the NMA authors. It is unclear if each meta-analysis used the same criteria to select trials. No search for recently published trials comparing IDeg with IGlar. At least one potentially relevant study has been published since (Study 3587 in insulin-naive patients with T2DM). No evaluation of potential sources of bias in the individual trials. The NMA authors report that many of the trials included in the 2008 CADTH report were considered poor quality. Except for one single-blind study, all the included studies were open label and, therefore, at high risk of bias.
| | |
|---|
| NMA | The scope of the NMA was limited by excluding IDet. Only hypoglycemia outcomes were assessed; glycemic control was not. No definitions of hypoglycemia events were provided; thus, it is unclear if all trials measured events the same way. Incomplete patient and study characteristics data were reported for individual trials; therefore, it was not possible for the reader to assess if the transitivity assumption was met. The authors reported that it was not possible to conduct any subgroup or metaregression analyses due to limited number of trials. There was no consideration of study quality in the NMA. A fixed-effects model was selected due to the limited number of trials included in the analysis. Several of the analyses reportedly had poor model fit; however, no details were provided and residual deviance values were not reported. Alternate models were not run in all cases; thus, data to justify the selection of the fixed-effects model were missing. In at least two instances, the random-effects model had a better fit. Sensitivity analyses were run excluding some trials that were said to have poor model fit based on residual deviance values that were calculated for individual trials. Rate data were analyzed as a continuous outcome using logistic regression model based on the log rate ratio data from treatment comparisons (rather than rate information from individual treatment groups within a trial). No direct meta-analysis was conducted.
| Appear to have conducted the NMA using accepted methods Random-effects model selected with no justification provided and no exploration of alternate models No information describing how convergence, goodness of fit, or inconsistency were assessed No mention if over-dispersion was accounted for in the Poisson model Reporting of results was incomplete as the focus of the NMA was the efficacy of IGlar (300 U/mL) versus other treatments
| Methods used to conduct NMA appear to be robust, with clear and complete reporting. Treatment duration varied substantially (4 weeks to 52 weeks). There were few details regarding the characteristics of patients enrolled in the RCTs or insulin doses received; thus, it was not possible to evaluate if the transitivity assumption was met. However, the authors did measure between-study heterogeneity and conducted sensitivity analyses to test for some sources of heterogeneity. Sparse network, particularly for severe hypoglycemia, thus, there was considerable. uncertainty in the estimates. No mention if over-dispersion was accounted for in the Poisson model. Many of the trials included had a high risk of bias (particularly an issue for hypoglycemia analysis).
|
|---|
IDeg = insulin degludec; IDet = insulin detemir; IGlar = insulin glargine; NMA = network meta-analysis; RCT = randomized controlled trial; SR = systematic review; T2DM = type 2 diabetes mellitus.
Discussion
Type 1 Diabetes Mellitus
Both the manufacturer-submitted NMA and Dawoud et al.’s NMA found no statistically significant differences between IDeg and IGlar on the rate of severe hypoglycemia events in patients with T1DM, which was consistent with the direct evidence. Similarly, no statistically significant differences were found when comparing IDeg with NPH insulin in both analyses, or for IDeg versus IDet in Dawoud et al.’s report. The 95% credible intervals for severe hypoglycemia were wide in Dawoud’s analysis, showing uncertainty in the results, and the authors stated that the trials had a high risk of bias.
The rates of nocturnal hypoglycemia and overall hypoglycemia also showed no statistically significant difference between IDeg and IGlar or NPH in the manufacturer’s submission. Given the issues with model fit described above, the random-effects model is preferred over the fixed-effects model for nocturnal hypoglycemia.
No statistically or clinically important differences in A1C were detected for IDeg versus insulin NPH, IGlar, or IDet.
Type 2 Diabetes Mellitus
Due to the focus on IGlar 300 U/mL in Freemantle’s report and differences in outcomes assessed in the manufacturer’s NMA, it is difficult to compare the findings. The direct meta-analysis found a statistically significantly lower risk of nocturnal hypoglycemia (rate ratio 0.79; 95% CI, 0.67 to 0.9) but a statistically higher risk of symptomatic hypoglycemia (rate ratio 1.35; 95% CI, 1.27 to 1.44) for IDeg versus IGlar 100 U/mL in Freemantle et al. 2016.63 The point estimates from the NMA were similar for these outcomes, but the credible intervals were wider and non-significant. There were no clinically important differences in A1C and weight found between IDeg and IGlar, although the direct meta-analysis was statistically significant. All studies included in the NMA were of relatively low quality; thus, the results are at high risk of bias.
Among patients with T2DM on basal insulin, the manufacturer’s NMA showed a statistically significantly lower rate of nocturnal hypoglycemia for IDeg versus IGlar and NPH; however, no statistically significant difference was found for the rate of severe hypoglycemia based on the random-effects model, which had better fit. Similar findings were reported for the analysis that included all T2DM patients. Data from the analysis of overall hypoglycemia could not be interpreted because the authors stated that both the fixed-effects and random-effects models showed poor fit with high residual deviance values. It appears that the model selection has an important impact on the findings, with fixed-effects models suggesting statistically significant differences, while random-effects models have wider credible intervals (as expected) and non-significant differences. Since the manufacturer’s NMA did not consistently run both models and did not report deviance information criterion and residual deviance values, it is difficult to interpret its findings.
Conclusion
Three indirect treatment comparisons were identified that compared IDeg with other basal insulins in patients with T1DM or T2DM. All three conducted Bayesian NMAs and analyzed hypoglycemia rates; two reports analyzed A1C; and one analyzed body weight. The manufacturer’s NMA12 and the NMA by Freemantle et al63 were limited in scope and did not include or report data for all basal insulins of interest to this review.
In patients with T1DM, the direct evidence for IDeg versus IGlar showed no statistically significant differences in the rate of severe hypoglycemia or change from baseline in A1C. The indirect evidence also suggested no statistically or clinically important differences between IDeg and IGlar, IDet, or NPH on the rate of hypoglycemia or change in A1C.
In patients with T2DM receiving basal insulin therapy, the direct meta-analysis found a statistically significantly lower risk of nocturnal hypoglycemia, but a statistically higher risk of symptomatic hypoglycemia for IDeg versus IGlar 100 U/mL in Freemantle et al. 2016.63 The point estimates from the NMA were similar for these outcomes, but the credible intervals were wider and non-significant. No clinically important differences in A1C and weight were found between IDeg and IGlar, although the direct meta-analysis was statistically significant.
Among patients with T2DM on basal insulin, the NMA suggested that the rate of nocturnal hypoglycemia was statistically significantly lower for IDeg versus IGlar and NPH in the manufacturer’s NMA. No statistically significant difference was found for the rate of severe hypoglycemia based on the random-effects model, which had better fit. Similar findings were reported for the analysis that included all patients with T2DM. Data from the analysis of overall hypoglycemia could not be interpreted because the authors stated that both the fixed-effects and random-effects models showed poor fit, with high residual deviance values.
All analyses were limited by the quality of the included studies. Also, the analyses were missing more recently completed head-to-head studies comparing IDeg with IGlar (e.g., SWITCH trials). The manufacturer’s NMA did not justify the model or analysis methods selected and indicated that there were issues with model fit. It appears that the model selection has an important impact on the findings. Given these issues, an exploration of alternate models and model assumptions was warranted, as was a comparison with the results of direct meta-analysis.
