1. Long‐acting beta₂‐agonists (LABAs)

Thirty‐seven LABA treatment arms were identified across 30 studies (n = 15,266). Twelve studies included a formoterol arm with treatment given at 12 mcg twice daily, and Rossi 2002 also randomly assigned participants to a higher dose of twice‐daily 24 mcg. Formoterol was most commonly given via a dry powder inhaler but sometimes via a metered‐dose inhaler (to retain blinding in the newer multiarm mometasone studies). Salmeterol was given at 50 mcg twice daily in 17 studies (n = 9019), primarily via the Diskus device. Indacaterol, a once‐daily preparation, was given as 150 mcg in three studies (Bateman 2013 [SHINE]; Donohue 2010 [INHANCE]; Kornmann 2011) and as 300 mcg in three studies (Dahl 2010; Donohue 2010 [INHANCE]; To 2011), one of which also used a higher dose of 600 mcg. Donohue 2010 [INHANCE] was conducted in two stages, and the data derived are for the four randomly assigned groups from stage 2.

2. Long‐acting muscarinic antagonists (LAMAs)

Thirty‐four LAMA treatment arms were included across 28 studies (n = 21,805). Most studies used tiotropium 18 mcg daily, delivered most often via the HandiHaler. Six studies included two arms that met criteria for inclusion in the LAMA category: Bateman 2010b used two tiotropium doses (5 and 10 mcg daily via Respimat), ATTAIN 2011 and Gelb 2012 used two aclidinium doses (200 and 400 mcg daily) and Bateman 2013 [SHINE], GLOW2 2012 and SPARK 2013 included both glycopyrronium 50 mcg and open‐label tiotropium. Two additional glycopyrronium studies used 50 mcg via the Breezhaler, and two other aclidinium studies used the lower of the two doses via the Genuair device. Two other tiotropium studies used the Respimat device to deliver 5 mcg daily (Abrahams 2013 and Bateman 2010a).

3. Inhaled corticosteroids (ICSs)

Twenty‐seven studies included an ICS arm (n = 6788), with two including two doses of the same drug that were relevant to the review (Calverley 2008; FLTA3025). Twelve studies (n = 3491) reported a fluticasone arm: two studies used 250 mcg twice daily, and the rest 500 mcg, delivered most often via the Diskus device. Twelve studies with around half as many people used budesonide (n = 1900): two at the higher dose of 800 mcg twice daily, and the rest at 400 mcg, with both doses delivered via dry powder inhalers. Mometasone was used in four studies (n = 1397) and was given as 400 mcg twice daily or 800 mcg once daily.

4. Long‐acting beta₂‐agonist/inhaled corticosteroid combinations (LABA/ICS)

Thirty LABA/ICS treatment arms were identified in 25 studies (n = 9727). Three studies used two budesonide doses (160 mcg and 320 mcg; Rennard 2009; Sharafkhaneh 2012; Tashkin 2008 [SHINE]), and seven additional studies used only the higher dose (total N = 3221). Two identically designed studies included two formoterol/mometasone doses each, 200 and 400 mcg (n = 888). Sixteen studies had a salmeterol/fluticasone arm (n = 5618), six at the lower ICS dose of 50/250 mcg and 11 at the higher dose of 50/500 mcg. All LABA/ICS inhalers were prescribed for twice‐daily use and drugs were delivered by dry powder or pressurised metered‐dose inhalers.

5. Placebo

Most studies included a placebo control arm (n = 55), which consisted of a total of 19,476 people. Inhalers varied across studies depending on the device used to deliver the study drugs. Once‐daily and twice‐daily placebos were combined in the same node as the reference treatment to increase the number of comparisons that could be made.

St George's Respiratory Questionnaire (SGRQ)

The SGRQ is a well‐validated and widely used measure of health status in patients with chronic airflow limitation; the total score ranges from zero (perfect health) to 100 (most severe status) and includes the three components of symptoms, activity and impacts.

In the main network of effects at six months, 18 of 25 possible treatments were represented across 72 treatment arms in 25 studies (n = 27,024). The fixed‐effect model was preferred, as statistical heterogeneity was not significant. We extracted and analysed data as change from baseline per arm and estimated population variance on the basis of data from studies reporting variance for change per arm (see Measures of treatment effect). The mean of SDs, weighted by sample size, was imputed as 14 when variance for change could not be obtained.

The 12‐month analysis contained slightly fewer data and had a broadly similar structure (containing data for 19 treatment comparisons in 24 studies). The fixed‐effect model was preferred, as with the six‐month analysis, but Calverley 2003 [TRISTAN] showed some lack of fit. Data from four studies were removed from the six‐month network in the blinding sensitivity analysis (Bateman 2013 [SHINE]; Donohue 2010 [INHANCE]; GLOW2 2012; SPARK 2013), and data from two studies were removed from the 12‐month network (GLOW2 2012; SPARK 2013).

The network structure for SGRQ at six months is shown in Figure 2.

Figure 2

---

---

Network structures for both outcomes at six months.

Trough forced expiratory volume in one second (FEV₁)

In the main network of effects at six months, 20 of 25 possible treatments were represented across 82 treatment arms in 31 studies (n = 29,271). A random‐effects model was preferred, as the fixed‐effect model has very poor fit. Data for eight studies were available only as treatment contrasts, and the rest were entered as change from baseline per arm, along with the associated variance. For trials reporting treatment contrasts, we included co‐variance, which was calculated as the variance of the mean in the control arm (placebo for all trials). For the rest of the trials, missing variance for change from baseline was imputed as 283 (mL) from the mean of available SDs weighted by sample size.

One very small trial was excluded from the six‐month analysis (SLMF4010, n = 34) because of the unusually high dropout from already very small groups (seven of 17 dropped out in both groups). The study showed very poor fit, even in random‐effects models, because the results contradicted all other trials.

The 12‐month analysis contained fewer data (19 treatments compared in 21 studies) but again had a structure broadly similar to the main six‐month analysis. As with the six‐month analysis, the random‐effects model was preferred, as the fixed‐effect model had very poor fit. Data from the same four studies were removed from the six‐month and 12‐month networks, as for the SGRQ sensitivity analyses (see Table 5).

The network structure for trough FEV₁ at six months is shown in Figure 2.

Class effects

Table 6 shows each class effect relative to placebo (i.e. the combined treatment benefit over placebo), expressed as change from baseline. The classes are also ranked (mean and median) and are presented with estimates of the probability that each is the best class. Figure 4 displays class effects against the reference treatment (placebo), shown by the vertical dashed line on the right‐hand side of the graph.

Combination LABA/ICS was the highest ranked treatment, with a mean improvement over placebo of 3.89 units (95% credible interval (CrI) ‐4.70 to ‐2.97). LAMAs and LABAs were ranked second and third, with mean differences of ‐2.63 (95% CrI ‐3.53 to ‐1.97) and ‐2.29 (95% CrI ‐3.18 to ‐1.53), respectively, but a large degree of overlap was noted between their credible intervals. Inhaled corticosteroids were the fourth ranked class but still showed a mean two‐unit benefit over placebo at six months (95% CrI ‐3.06 to ‐0.87). Figure 6 shows graphs of the probabilities that each class ranks in any one of the five possible positions. The median within‐class variance was 0.44 units (95% CrI 0.03 to 1.09).

Figure 6

---

---

St George's Respiratory Questionnaire (SGRQ) probability ranking diagrams—six months and 12 months.

Individual treatment effects

Effects relative to placebo are presented separately for each treatment in Table 7, along with ranks and estimates with and without the class model for comparison. The credible intervals for each treatment were wide and largely overlapping (for both ranks and mean differences from placebo).

Individual treatment effects of particular note within a class were associated with indacaterol and aclidinium for LABA and LAMA, respectively. Effects for both doses of each of these drugs were dampened by the class model in the six‐month SGRQ network (indacaterol by 0.72 and 0.92 for 300 mcg and 150 mcg, respectively, and aclidinium by 1.13 and 1.78 for 200 mcg and 400 mcg, respectively). Treatments of note for which estimates were inflated by fitting the class model were formoterol 12 mcg (by 0.64), budesonide 400 mcg (by 0.79) and formoterol/mometasone combination 2/400 mcg (by 0.66).

12 months

Nineteen treatments were compared in 24 studies, but a dose comparison of aclidinium made in Gelb 2012 was disconnected from the rest of the network. These treatments can be used in class analysis, as they “borrow” the class effect. The fixed‐effect model was preferred.

In the class analysis (Appendix 4 and Figure 4), LABA/ICS combinations remained the highest ranked at 12 months, with only a slightly smaller mean effect over placebo than at six months (‐3.60, 95% CrI ‐4.63 to ‐2.34). Compared with the six‐month analysis, LABA, LAMA and ICS showed more similar benefits over placebo at 12 months, ranging between ‐2.34 and ‐2.55, with similar credible intervals. Within‐class variance was slightly larger than in the six‐month analysis (median 0.53, 95% CrI 0.04 to 1.27).

Individual treatment effects for the 12‐month analysis are presented in Appendix 5.

Sensitivity analysis—blinding

Table 6 shows how the six‐month effects were affected by removal of data at high risk of bias for blinding (see 'Effect vs placebo sensitivity analysis'). The difference between point estimates of LAMA over LABA was larger with less overlap in the credible intervals, and within‐class variance was smaller than in the main analysis (median 0.35, 95% CrI 0.02 to 0.98). Inconsistency associated with Donohue 2010 [INHANCE] was resolved by removing the open‐label tiotropium arm, but unexplained poor fit of three further trials remained (Kornmann 2011; Rennard 2009; Tashkin 2012). As all open‐label arms removed were given tiotropium, it is possible that the expected treatment benefit over placebo in these studies was reduced by lack of blinding for these comparisons, which dampened the overall treatment effect within the LAMA class.

Class effects

Table 8 shows each class effect relative to placebo. As for SGRQ, the classes are also ranked (mean and median) and are presented with estimates of the probability that each is the best. Figure 6 displays trough FEV₁ class effects (mL) compared with placebo.

Results at six months showed a similar pattern to the SGRQ analysis. Combination LABA/ICS was the highest ranked treatment, with a mean improvement over placebo of 133.3 mL (95% CrI 100.6 to 164.0) after six months. As was the case for quality of life, LAMAs and LABAs were ranked second and third, with mean differences of 103.5 (95% CrI 81.8 to 124.9) and 99.4 (95% CrI 72.0 to 127.8), respectively, but again a large degree of overlap was noted between the credible intervals. Inhaled corticosteroids were the fourth ranked class but still showed a mean 65.4 mL benefit over placebo at six months (95% CrI 33.1 to 96.9). The median within‐class variance was 9.04 mL (95% CrI 0.41 to 34.85). Figure 7 shows graphs of the probabilities that each class ranks in any one of the five possible positions.

Figure 7

---

---

Forced expiratory volume in one second (FEV₁) probability ranking diagrams—six months and 12 months.

Individual treatment effects

Effects relative to placebo are presented separately for each treatment in Table 9, along with ranks and estimates with and without the class model for comparison.

The class model affected individual treatment effects to a lesser extent in the FEV₁ network than in the SGRQ one, because less intra class variation was present. Several treatment effects that were higher than the class mean were dampened by the class model, including indacaterol 300 (by 34.6 mL) and salmeterol‐fluticasone 50/250 mcg combination therapy (by 33.3 mL). Effects for formoterol 12 mcg and both formoterol/budesonide combination doses were less favourable within their classes and were inflated by the class model (by 41.0 mL, and 50.3 and 54.5 mL, respectively). Although aclidinium was particularly effective in the SGRQ analysis, the relative effects for both lower doses on trough FEV₁ appeared less favourable than other members of the LAMA class (MD 59.3 and 74.8 mL without the class model compared with the overall class mean of 103.5). Within the ICS class, the two budesonide doses showed less benefit over placebo than the two fluticasone doses. However, credible intervals for the individual treatment groups were wide and overlapping.

12 months

Nineteen treatments were compared in 22 studies. As with the 12‐month SGRQ analysis, the dose comparison of aclidinium made in Gelb 2012 was disconnected from the rest of the network. The random‐effects model was preferred (DIC = 485.4), as the fixed‐effect model had very poor fit (DIC = 557.4).

In the class analysis (Appendix 6 and Figure 6), LABA/ICS remained the highest ranked treatment strategy, although the effect over placebo was lower than at six months (100 mL, 95% CrI 55.5 to 140.1). As with SGRQ, the benefit of LAMA over LABA was much less clear at 12 months, and their credible intervals showed a large degree of overlap with each other and with the LABA/ICS class. Within‐class variance was larger than in the six‐month analysis (median 18.02, 95% CrI 1.22 to 47.80).

Individual treatment effects for the 12‐month analysis are presented in Appendix 7.

Sensitivity analysis—blinding

Table 8 shows how the six‐month effects were influenced by removal of data at high risk of bias for blinding. Within‐class variance was similar to that in the main analysis (median 10.49, 95% CrI 0.54 to 37.66). The effects were very similar with and without the open‐label data, unlike in the SGRQ analysis. This may reflect the nature of the two measurements, with SGRQ perhaps more susceptible to performance or detection bias than measures of lung function.