Studies considered

There were a total of 29 trials comparing various antidepressants with placebo. Most trials were on venlafaxine (all studies used extended release [XL] preparations), duloxetine, escitalopram, sertraline and paroxetine. These trials were all large, high-quality studies funded almost exclusively by drug company sponsorship. There was no evidence of publication bias at the study level for any of the antidepressant comparisons as assessed by visual inspection of funnel plots and formally by the Egger’s test.

A summary of study characteristics can be found in Table 32 with full details in Appendix 15d, which also includes details of excluded studies.

Table 32

Study information table for trials comparing antidepressants with placebo.

Clinical evidence for antidepressants versus placebo

Evidence from the important outcomes and overall quality of evidence are presented in Table 33. The full GRADE profiles and associated forest plots can be found in Appendix 18c and Appendix 16c, respectively.

Table 33

Evidence summary table for trials of antidepressants versus placebo.

Clinical evidence summary

There was limited or no data for a number of interventions: there was only one study assessing imipramine; one study assessing citalopram; no data on mirtazapine, bupropion, trazodone, fluvoxamine, fluoxetine and amitriptyline; and no data on most TCAs (for example, clomipramine, doxepin, dosulepin, lofepramine, nortriptyline and trimipramine). A further limitation of the data was the lack of long-term studies (only two studies, one for venlafaxine and one for escitalopram, provided data on use beyond 6 months) and no available follow-up data beyond end of treatment.

The benefits in terms of reducing the risk of non-response, non-remission and the mean anxiety rating score was similar for most antidepressants suggesting a small-to-moderate improvement in anxiety relative to placebo.

The harms were also relatively consistent across drugs. Discontinuation due to adverse events was greater than placebo for most antidepressants but particularly high for paroxetine, duloxetine and venlafaxine. Specific side effects such as nausea and insomnia were more common in people receiving antidepressants compared with placebo. Sexual problems were relatively rare but there was an increased risk associated with antidepressants.

Studies considered

A total of eight trials compared pregabalin with placebo. A summary of study characteristics can be found in Table 34 with full details in Appendix 15d, which also includes details of excluded studies.

Table 34

Study information table for trials comparing pregabalin with placebo.

Clinical evidence for pregabalin versus placebo

Evidence from the important outcomes and overall quality of evidence are presented in Table 35. The full GRADE profiles and associated forest plots can be found in Appendix 18c and Appendix 16c, respectively.

Table 35

Evidence summary table for trials of pregabalin versus placebo.

Clinical evidence summary

Pregabalin was associated with a moderate benefit in terms of mean anxiety rating scores and non-response. However, although there was statistically significant evidence of benefit in relation to non-remission, the effect size was small.

In terms of harms, there was a small borderline statistically significant increase in the risk of discontinuation due to adverse events. For specific side effects, there was a different pattern from that found for antidepressants. There was no statistically significant increase in risk of experiencing nausea or insomnia. In addition, sexual problems were not reported as frequent side effects in any of the studies. However, there were large increases in risk of dizziness and fatigue (although for the latter this was not statistically significant).

Study characteristics

A total of four trials compared benzodiazepines with placebo. A summary of study characteristics can be found in Table 36 with full details in Appendix 15d, which also includes details of excluded studies.

Table 36

Study information table for trials comparing benzodiazepines with placebo.

Clinical evidence for benzodiazepines versus placebo

Evidence from the important outcomes and overall quality of evidence are presented in Table 37. The full GRADE profiles and associated forest plots can be found in Appendix 18c and Appendix 16c, respectively.

Table 37

Evidence summary table for trials of benzodiazepines versus placebo.

Clinical evidence summary

The evidence base for benzodiazepines was much smaller than for antidepressants and pregabalin reported above. There were inconsistent effects for most outcomes. On the mean anxiety rating score there were small-to-moderate benefits found but the effect for diazepam was not statistically significant. On non-response there was a moderate reduction for diazepam but no statistically significant effects were identified for lorazepam and alprazolam. For non-remission, no data was found for diazepam and there were no statistically significant effects for lorazepam or alprazolam.

There was inconsistent reporting of harms, therefore the data on side effects is relatively limited. There was no statistically significant increase in risk of discontinuation for diazepam and alprazolam but there was a higher risk in lorazepam. Increased risk of experiencing sexual problems was found for diazepam but this was not reported for the other drugs. There was an increased risk of dizziness for diazepam, lorazepam and alprazolam.

Studies considered

There were a total of five trials comparing buspirone with placebo. A summary of study characteristics can be found in Table 38 with full details in Appendix 15d which also includes details of excluded studies.

Table 38

Study information table for trials comparing buspirone with placebo.

Clinical evidence for buspirone versus placebo

Evidence from the important outcomes and overall quality of evidence are presented in Table 39. The full GRADE profiles and associated forest plots can be found in Appendix 18c and Appendix 16c, respectively.

Table 39

Evidence summary table for trials of buspirone versus placebo.

Clinical evidence summary

There was a small benefit associated with buspirone on both the mean anxiety rating score and non-response. However, no data was reported on non-remission therefore it is not possible to draw conclusions on this outcome.

There was greater risk of discontinuation due to adverse events associated with buspirone. There was a higher risk of experiencing nausea and dizziness compared with placebo.

Studies considered

A total of three trials compared hydroxyzine with placebo. A summary of study characteristics can be found in Table 40 with full details in Appendix 15d, which also includes details of excluded studies.

Table 40

Study information table for trials comparing hydroxyzine with placebo.

Clinical evidence for hydroxyzine versus placebo

Evidence from the important outcomes and overall quality of evidence are presented in Table 41. The full GRADE profiles and associated forest plots can be found in Appendix 18c and Appendix 16c, respectively.

Table 41

Evidence summary table for trials of hydroxyzine versus placebo.

Clinical evidence summary

There was inconsistent reporting of data on hydroxyzine therefore it is difficult to draw conclusions concerning the harms and benefits of this drug. The mean anxiety rating score suggested a moderate reduction in anxiety. However, most studies did not report data in sufficient detail on non-response and non-remission. There were also very little data on discontinuation or reporting of specific side effects.

Studies considered

A total of four trials compared quetiapine with placebo. A summary of study characteristics can be found in Table 42 with full details in Appendix 15d, which also includes details of excluded studies.

Table 42

Study information table for trials comparing quetiapine with placebo.

Clinical evidence for quetiapine versus placebo

Evidence from the important outcomes and overall quality of evidence are presented in Table 43. The full GRADE profiles and associated forest plots can be found in Appendix 18c and Appendix 16c respectively.

Table 43

Evidence summary table for trials of quetiapine versus placebo.

Clinical evidence summary

A review of the clinical efficacy of quetiapine is included in a forthcoming NICE Technology Appraisal and therefore is not assessed in this guideline. The data is to inform the network-analysis only (see section 8.8.3).

Studies considered

There were a total of six trials comparing antidepressants with other antidepressants. A summary of study characteristics can be found in Table 44 with full details in Appendix 15d, which also includes details of excluded studies.

Table 44

Study information table for trials comparing antidepressants with other antidepressants.

Clinical evidence for antidepressants versus other antidepressants

Evidence from the important outcomes and overall quality of evidence are presented in Table 45. The full GRADE profiles and associated forest plots can be found in Appendix 18c and Appendix 16c, respectively.

Table 45

Evidence summary table for trials of antidepressants versus other antidepressants.

Clinical evidence summary

There was a small statistically significant effect in favour of escitalopram compared with paroxetine based on a reduction in HAM-A scores. In addition, there was a 40% reduction in risk of non-response for escitalopram compared with paroxetine. Moreover, there was greater risk (although not statistically significant) of discontinuation of treatment due to adverse events associated with paroxetine.

There were no differences found on reduction of anxiety symptoms between escitalopram and venlafaxine. However, venlafaxine was associated with a greater risk of discontinuation (although this was not statistically significant).

No difference was found between duloxetine and venlafaxine for reduction in anxiety but there was a greater risk of discontinuation for venlafaxine (although again this was not statistically significant)

There were no statistically significant differences found between paroxetine and sertraline on any outcomes. However, this was based on a small trial that was unlikely to have sufficient power to identify any differences.

Studies considered

There were a total of six trials comparing antidepressants with other pharmacological interventions. A summary of study characteristics can be found in Table 46 with full details in Appendix 15d, which also includes details of excluded studies.

Table 46

Study information table for trials comparing antidepressants with other pharmacological interventions.

Clinical evidence for antidepressants versus other pharmacological interventions

Evidence from the important outcomes and overall quality of evidence are presented in Table 47. Data for quetiapine, which is considered in the network meta-analysis, are reported in the same table. The full GRADE profiles and associated forest plots can be found in Appendix 18c and Appendix 16c, respectively.

Table 47

Evidence summary table for trials of antidepressants versus other pharmacological interventions.

Clinical evidence summary

Similar to the data in Section 8.3.1, there was limited data concerning comparisons between active interventions. There were no statistically significant differences in reduction in anxiety for venlafaxine in comparison with pregabalin, buspirone or diazepam. However there was an increased risk of discontinuation due to adverse events for venlafaxine compared with these drugs.

Studies considered

There were a total of six head-to-head trials of pharmacological interventions other than antidepressants. A summary of study characteristics can be found in Table 48 with full details in Appendix 15d, which also includes details of excluded studies.

Table 48

Study information table for head-to-head comparisons of pharmacological interventions other than antidepressants.

Clinical evidence for head-to-head trials of pharmacological interventions other than antidepressants

Evidence from the important outcomes and overall quality of evidence are presented in Table 49. The full GRADE profiles and associated forest plots can be found in Appendix 18c and Appendix 16c, respectively.

Table 49

Evidence summary table for head-to-head comparisons of pharmacological interventions other than antidepressants.

Clinical evidence summary

As above, there was a lack of head-to-head comparisons. There were borderline statistically significant effects favouring pregabalin over lorazepam and alprazolam in reduction of anxiety. In addition, pregabalin was associated with a reduced risk of discontinuation due to adverse events compared with lorazepam. However, both lorazepam and alprazolam were less likely to be associated with reporting dizziness as a side effect.

There was a small but not statistically significant difference in favour of hydroxyzine compared with buspirone based on a reduction in HAM-A scores. In addition, no statistically significant differences were found between buspirone and lorazepam.

Studies considered

There were four trials on venlafaxine comparing different doses. A summary of study characteristics can be found in Table 50 with full details in Appendix 15d, which also includes details of excluded studies.

Table 50

Study information table for trials of venlafaxine comparing different doses.

Doses used in studies of venlafaxine ranged from a mean of 37.5 mg to 225 mg but there was limited data for most comparisons. The most common comparison was of 75 mg versus 150 mg.

Clinical evidence for venlafaxine comparing different doses

Evidence from the important outcomes and overall quality of evidence are presented in Table 51. The full GRADE profiles and associated forest plots can be found in Appendix 18c and Appendix 16c, respectively.

Table 51

Evidence summary table for trials of venlafaxine comparing different doses.

Clinical evidence summary

There were no statistically significant differences between 37.5 mg and 75 mg of venlafaxine for discontinuation due to adverse events and dizziness. However, with 37.5 mg compared with 75 mg, there was a 35% reduction in the risk of nausea. There was a borderline statistically significant difference on mean HAM-A score in favour of 75 mg in comparison with 150 mg of venlafaxine based on a reduction in HAM-A scores (SMD −0.27; CI −0.57 to 0.03) and a reduction in the risk of side effects such as nausea (RR = 0.82; CI 0.68 to 0.98) and insomnia (RR = 0.59; CI 0.34 to 1.01). There were no statistically significant differences in regards to a reduction in the risk of non-response, in discontinuation for any reason and side effects such as nervousness, dizziness and asthenia.

There were no statistically significant differences between 150 mg and 255 mg for risk of side effects such as insomnia, nervousness, asthenia and dizziness.

Studies considered

There were limited studies (only two trials) comparing doses for SSRIs. Comparisons could only be made for escitalopram and paroxetine, with just one study found for each drug. A summary of study characteristics can be found in Table 52 with full details in Appendix 15d, which also includes details of excluded studies.

Table 52

Study information table for trials of SSRIs comparing different doses.

Clinical evidence for selective serotonin reuptake inhibitors comparing different doses

Evidence from the important outcomes and overall quality of evidence are presented in Table 53. The full GRADE profiles and associated forest plots can be found in Appendix 18c and Appendix 16c, respectively.

Table 53

Evidence summary table for trials of SSRIs comparing different doses.

Clinical evidence summary

There were borderline statistically significant effects in the reduction of anxiety in favour of 10 mg of escitalopram compared with 5 mg based on mean HAM-A scores. There was no significant difference between the two groups regarding side effects with the exception of a reduction in the risk of reported headache with 5 mg compared with 10 mg of escitalopram. There was a reduced risk of reported headaches in the 20 mg group compared with the 10 mg escitalopram group.

There were no clear differences on outcomes between 20 mg and 40 mg of paroxetine.

Studies considered

There were two trials on duloxetine comparing different doses. A summary of study characteristics can be found in Table 54 with full details in Appendix 15d, which also includes details of excluded studies.

Table 54

Study information table for trials of duloxetine comparing different doses.

Doses used in the studies ranged from a mean of 20 mg to a mean of 120 mg. Results were similar as those reported above – there was limited evidence of differences between doses.

Clinical evidence for duloxetine comparing different doses

Evidence from the important outcomes and overall quality of evidence are presented in Table 55. The full GRADE profiles and associated forest plots can be found in Appendix 18c and Appendix 16c, respectively.

Table 55

Evidence summary table for trials of duloxetine comparing different doses.

Clinical evidence summary

There was a reduction in anxiety in favour of 60 to 120 mg of duloxetine compared with 20 mg based on HADS-A scores. However, this did not reach statistical significance. There were no clear differences between 60 mg and 120 mg found on any outcomes.

Studies considered

There were five trials on pregabalin comparing different doses. A summary of study characteristics can be found in Table 56 with full details in Appendix 15d, which also includes details of excluded studies. Dosages used in the studies ranged from a mean of 150 mg to a mean of 600 mg.

Table 56

Study information table for trials of pregabalin comparing different dosages.

Clinical evidence for pregabalin comparing different doses

Evidence from the important outcomes and overall quality of evidence are presented in Table 57. The full GRADE profiles and associated forest plots can be found in Appendix 18c and Appendix 16c, respectively.

Table 57

Evidence summary table for trials of pregabalin comparing different dosages.

Clinical evidence summary

There were few differences found between doses of pregabalin. However, there was some evidence that a mean of 600 mg was associated with greater reduction in anxiety compared with 150 mg. But 150 mg was associated with fewer reported side effects (particularly somnolence and dizziness). In addition, 400 mg was associated with greater benefits in reduction of anxiety compared with 600 mg.

Studies considered

There were four trials on augmentation strategies. A summary of study characteristics can be found in Table 62 with full details in Appendix 15d, which also includes details of excluded studies.

Table 62

Study information table for trials of augmentation strategies.

Clinical evidence for augmentation strategies

Evidence from the important outcomes and overall quality of evidence are presented in Table 63. The full GRADE profiles and associated forest plots can be found in Appendix 18c and Appendix 16c, respectively.

Table 63

Evidence summary table for trials of augmentation strategies.

Clinical evidence summary

There was limited evidence because three of the four trials were small and there was high heterogeneity in HAM-A scores (I² = 73%) for risperidone. There was no statistically significant evidence of benefit for any of the antipsychotic drugs assessed individually. When combining the antipsychotic data there was still limited evidence of benefit.

Cardiovascular symptoms

SSRIs do not appear to be associated with an increase risk in cardiovascular adverse events (for example, Swenson et al., 2006; Taylor, 2008) and are associated with a relatively low fatal toxicity index (number of poisoning deaths per million prescriptions). However, TCAs are associated with a higher risk of cardiovascular adverse events and have found to be cardiotoxic in overdose (Taylor, 2008).

Duloxetine was associated with small increases in diastolic blood pressure, tachycardia and cholesterol compared with placebo (Duggan & Fuller, 2004; Wernicke et al., 2007). In addition, there is evidence of moderate acute toxicity associated with venlafaxine (Taylor, 2008).

Bleeding

Several observational studies utilising data from national prescribing databases have found a relatively strong association (approximately a three-fold increase) between SSRIs and increased risk of gastrointestinal bleeding (Weinrieb et al., 2003; Yuan et al., 2006). However, it should be noted that the outcome was relatively rare with approximately four to five events per 1000 person years. This effect was particularly strong (approximately a 15-fold increase of bleeding) in people concurrently using non-steroidal anti-inflammatory drugs (NSAIDs) and SSRIs.

Gastrointestinal symptoms

There is consistent evidence both in depression and anxiety populations of the increased risk of gastrointestinal (GI) symptoms such as nausea, vomiting and diarrhoea associated with SSRI use (Brambilla et al., 2005; Beasley et al., 2000). This has been confirmed in the current systematic review of SSRIs for GAD (see section 8.2.3). TCAs also appear to be associated with higher risk of constipation when compared with fluoxetine (Beasley et al., 2000).

Sexual dysfunction

There was consistent evidence of sexual adverse events associated with SSRIs, duloxetine and venlafaxine in people with depression (Werneke et al., 2006; Gregorian et al., 2002; Beasley et al., 2000; Keller, 2000). These results have been replicated in people with GAD in the current systematic review (see section 8.2.3).

Weight change

Fluoxetine appears to be associated with greater loss in weight compared with placebo (Beasley et al., 2000), TCAs and other SSRIs (Brambilla et al., 2005). However, as noted by Demyttenaere and Jaspers (2008), these effects are reported early on in treatment. When assessing continuation studies there is a possibility that paroxetine and fluoxetine may actually be associated with weight gain but this needs further research to establish this finding.

In addition, there is some evidence that duloxetine was associated with weight loss with a mean reduction of 2.2 kg compared with 1 kg for placebo (Duggan & Fuller, 2004).

Suicidal ideation and behaviour

One systematic review was identified on the association between antidepressant use and suicidal ideation and/behaviour (Stone et al., 2009). For those aged under 25 years there was an increased odds of suicidal behaviour (OR 2.30; 95% CI 1.04, 5.09) for people taking antidepressants compared with placebo. There was a borderline statistically significant increase in odds of suicidal ideation and suicidal behaviour (OR 1.62; 95% CI 0.97, 2.71).

Introduction – objective of economic modelling

The cost effectiveness of pharmacological interventions relative to other available treatments for people with GAD was considered by the GDG as an area with likely significant resource implications. The GDG was particularly interested in the cost effectiveness of pharmacological interventions compared with low- and high-intensity psychological interventions, as well as in the relative cost effectiveness between different pharmacological interventions, including no treatment (placebo).

The development of an economic model comparing pharmacological interventions with low- and high-intensity psychological interventions using clinical effectiveness data from the guideline systematic review was not possible: no RCTs directly comparing pharmacological with psychological interventions were identified in the systematic clinical literature review. Indirect (clinical) comparisons between pharmacological and psychological interventions using a common ‘baseline’ comparator were problematic because of important differences in study designs in terms of the following:

• Comparators: psychological studies used mainly waitlist or standard care as a comparator, while studies on pharmacological treatments used placebo as control (but never waitlist or standard care); therefore, it was not possible to make indirect comparisons between pharmacological and psychological interventions using a common ‘baseline’ comparator.
• Reported clinical outcomes: psychological studies tended to report mainly continuous outcomes. Few psychological studies reported rates of response or remission, which were often used as outcome measures in pharmacological studies; even then, the definitions of response and remission in psychological studies were not the same as the respective definitions in pharmacological studies.
• Study population: a number of studies on low-intensity psychological interventions were conducted on people with mixed anxiety rather than GAD only; in contrast, only studies on people with GAD were included in the systematic literature review of pharmacological interventions.

Due to the above limitations, which did not allow consideration of both psychological and pharmacological treatments in one economic analysis, an economic model was developed to assess the relative cost effectiveness between different pharmacological interventions for people with GAD in the UK. This analysis was considered as a priority by the GDG, because of the likely significant resource implications associated with the choice of drug in the treatment of people with GAD. Moreover, existing economic evidence in the area of pharmacological treatment for people with GAD is rather limited and not directly applicable to the UK setting, since only two of the five studies were conducted in the UK. The economic studies included in the systematic review were characterised by a number of limitations; besides, they did not assess the whole range of drugs available in the UK for the treatment of people with GAD.

Economic modelling methods

Model structure

A decision-analytic model in the form of a decision-tree was constructed using Microsoft Office Excel 2007. The structure of the model was determined by the availability of clinical data. According to the model structure, hypothetical cohorts of people with GAD were initiated on each of the six drugs assessed (first-line drug) or no pharmacological treatment. People initiated on the first-line drug could either continue treatment for 8 weeks, or discontinue due to intolerable side effects during this 8-week period. For modelling purposes, it was assumed that drug discontinuation because of intolerable side effects occurred at 2 weeks following initiation of treatment; this was based on the GDG’s estimate that the majority of people discontinuing treatment because of intolerable side effects do so within 2 weeks from starting treatment. People who continued on the first-line drug either responded to treatment or did not respond. Those who responded were given maintenance treatment (consisting of the same drug) for 6 months. During this period, they either experienced a relapse or did not relapse. In each cohort, people discontinuing the first-line drug due to intolerable side effects and those not responding to the first-line drug were switched to a second-line drug, which was a mixture of all drugs assessed in the economic analysis, except the first-line drug administered to this cohort. People taking the second-line drug were all assumed to continue treatment with this drug. From that point onwards they followed the same pathways as people who continued the first-line drug (that is, no response or response and maintenance treatment, during which they could relapse or not relapse). People receiving no pharmacological treatment were assumed to either discontinue treatment, in which case they did not clinically improve (‘no response’), or continue their treatment and follow a similar pathway to that experienced by people continuing pharmacological treatment (that is, no response or response followed by relapse or no relapse). The time horizon of the analysis was 42 weeks, based on the optimal duration of initial pharmacological treatment (8 weeks) and maintenance treatment (26 weeks), and in order to allow for switching to second-line treatment in case the 8-week first-line treatment did not lead to response. A schematic diagram of the decision-tree is presented in Figure 6.

Figure 6

Schematic diagram of the decision-tree constructed for the assessment of the relative cost effectiveness of pharmacological interventions for people with GAD.

Costs and outcomes considered in the analysis

The economic analysis adopted the perspective of the NHS and personal social services, as recommended by NICE (2009a). Costs consisted of intervention costs (drug acquisition and GP visit costs) and other health and social care costs incurred by people with GAD not responding to treatment or experiencing a relapse following response (including contacts with healthcare professionals such as GPs, psychiatrists, psychologists, mental health nurses and social workers, community care, inpatient and outpatient secondary care). The measure of outcome was the QALY.

Clinical input parameters and overview of methods employed for evidence synthesis

Clinical input parameters consisted of the probability of drug discontinuation due to intolerable side effects, the probability of response for those not discontinuing treatment due to side effects (conditional response), and the probability of relapse following response to treatment.

The guideline systematic review of the clinical literature on pharmacological treatments identified two dichotomous outcomes that could be utilised in economic modelling: response (defined in the vast majority of studies as 50% reduction in HAM-A scores) and remission (defined in the vast majority of studies as a HAM-A score below 7). Utilisation of both types of data was not possible because not all studies provided data on both outcomes; therefore, it was not possible to estimate the numbers of people with GAD who responded to treatment but did not meet criteria for remission, and of those who responded to treatment and remitted for the whole dataset. For the economic model, it was decided to utilise response (rather than remission) data for the following reasons:

• response data were available from a larger number of studies including a higher number of participants, compared with data on remission
• available relapse data referred to people who had responded to treatment; no relapse data following remission were available in the guideline systematic review
• utility data were available for the health state of ‘response’ but not for the health state of ‘remission’; in addition, there were utility data available for ‘relapse following response’ but not for ‘relapse following remission’.

The availability of clinical and utility data, and the subsequent selection of response data in order to populate the model, determined the economic model structure.

To take all trial information into consideration, network (mixed treatment comparison) meta-analytic techniques were employed to synthesise evidence on discontinuation due to intolerable side effects, as well as evidence on conditional response (the methods used can be found in Appendix 13). Network meta-analysis is a generalisation of standard pair-wise meta-analysis for A versus B trials to data structures that include, for example, A versus B, B versus C and A versus C trials (Lu & Ades, 2004). A basic assumption of network meta-analysis is that direct and indirect evidence estimate the same parameter; in other words, the relative effect between A and B measured directly from an A versus B trial is the same with the relative effect between A and B estimated indirectly from A versus C and B versus C trials. Network meta-analytic techniques strengthen inference concerning the relative effect of two treatments by including both direct and indirect comparisons between treatments and, at the same time, allow simultaneous inference on all treatments examined in the pair-wise trial comparisons while respecting randomisation (Lu & Ades, 2004; Caldwell et al., 2005). Simultaneous inference on the relative effect of a number of treatments is possible provided that treatments participate in a single ‘network of evidence’, that is, every treatment is linked to at least one of the other treatments under assessment through direct or indirect comparisons.

Details on the methods and clinical data utilised in the two network meta-analyses that were undertaken to estimate the probability of discontinuation due to intolerable side effects and the probability of conditional response for each treatment option considered in the economic analysis (that is, each first-line drug or no pharmacological treatment) are presented in Appendix 13. The findings of the two network meta-analyses are discussed in the next sub-section. The probability of response for the second-line drug in each decision node of the model was calculated as the average probability of conditional response of all drugs except the one that was used as a first-line treatment in this particular node of the model.

The probability of relapse following response to treatment was estimated based on relevant data included in the guideline systematic review. Four placebo-controlled trials assessed the efficacy of pharmacological treatments in preventing relapse in people with GAD: two of them assessed an SSRI (ALLGULANDER2006 – escitalopram; STOCCHI2003 – paroxetine), one assessed an SNRI (DAVIDSON2008 –duloxetine) and one assessed pregabalin (FELTNER2008). According to the expert opinion of the GDG, pregabalin is not a drug routinely used in the maintenance treatment of people with GAD. Moreover, the relative risk of relapse of pharmacological treatment versus placebo was higher in FELTNER2008 compared with the respective relative risk estimated for each of the other three studies, indicating that pregabalin may be potentially less effective than the other three drugs in preventing relapse in people with GAD that has responded to initial treatment. Inclusion of data from FELTNER2008 in a meta-analysis of the above four trials increased the heterogeneity of the analysis considerably (87% when data from all four studies were pooled together versus 0% when data from FELTNER2008 were excluded). For the above reasons, the estimation of the relative risk of relapse of pharmacological treatment versus placebo was based on meta-analysis of the remaining three studies (ALLGULANDER2006, DAVIDSON2008, STOCCHI2003). This estimate was utilised in all decision nodes of the model that involved pharmacological treatment, including the pregabalin node. Nevertheless, the probability of relapse following response in the ‘no treatment’ node of the model was estimated by pooling data from the placebo arms of all four studies.

Table 71 provides all the clinical input parameters utilised in the economic model.

Table 71

Input parameters utilised in the economic model of pharmacological treatments for people with GAD.

Findings of the network meta-analyses undertaken to inform the economic analysis

The summary statistics of a number of parameters of the two network meta-analyses undertaken to inform the economic analysis, including the log hazard ratios of all drugs considered in the economic analysis versus placebo and the between-trial variation, are reported in Appendix 13.

Table 68 provides the results of the network meta-analysis of data on drug discontinuation due to side effects, as well as the findings of classical pair-wise comparisons of each drug versus placebo on the same outcome. Results of the network meta-analysis are reported as mean values with 95% credible intervals, which are analogous to confidence intervals in frequentist statistics. Only results on options considered in the economic analysis are presented. The table shows the probability of each option leading to discontinuation side effects over 8 weeks of treatment, the probability of each option being the ‘best’ among available options in averting discontinuation due to side effects, as well as the hazard ratio of each drug versus placebo in this outcome. In addition to these three parameters examined in the network meta-analysis, the table shows the relative risk of each drug versus placebo for discontinuation due to side effects (mean and 95% confidence intervals), as estimated in the guideline classical, pair-wise meta-analysis. Treatment options have been ranked from ‘best’ to ‘worst’ in terms of their ability to minimise discontinuation due to side effects, according to the results of the network meta-analysis.

Table 68

Pharmacological treatment discontinuation due to side effects: findings of the network meta-analysis and of classical pair-wise comparisons versus placebo.

The results of the network meta-analysis indicated that placebo had the lowest probability of discontinuation due to side effects (mean 5.8% over 8 weeks). Among drugs, sertraline had the lowest probability of leading to discontinuation due to side effects (mean 7.2% over 8 weeks), followed by pregabalin, escitalopram, paroxetine, venlafaxine XL and, finally, duloxetine (mean 17.5% over 8 weeks). The probability of sertraline being the best drug in limiting discontinuation due to side effects reached 61%. All drugs showed a significantly higher hazard of discontinuation compared with placebo, except sertraline. The results of the guideline classical meta-analysis of placebo-controlled trials were consistent overall with the findings of the network meta-analysis: it can be seen that the ranking of relative risks of each drug versus placebo with respect to discontinuation due to side effects (classical, pair-wise meta-analysis) was the same with the ranking of the hazard ratios (network meta-analysis), with the exception of venlafaxine XL. The relative effect of sertraline versus placebo was found to be non-significant in the classical pair-wise meta-analysis, in accordance with the finding in the network meta-analysis. However, it must be noted that data on sertraline were taken from a small number of studies relative to other drugs (two placebo-controlled studies with 706 participants). On the other hand, in contrast to pair-wise comparisons, which failed to demonstrate a significant effect of pregabalin compared with placebo in terms of discontinuation due to side effects, data combined in the network meta-analysis had sufficient strength to demonstrate that pregabalin significantly increases the risk of discontinuation due to side effects compared with placebo.

Table 69 provides the results of the network meta-analysis of data on conditional response (that is, response in people who have not discontinued the drug due to side effects), as well as the findings of classical pair-wise comparisons of each drug versus placebo on non-response. It must be noted that the classical meta-analysis was based on an intention-to-treat (ITT) approach, and therefore considered all trial participants, without excluding those who discontinued due to side effects (exclusion of people discontinuing due to side effects in the network meta-analysis was dictated by the economic model structure). The table shows the probability of conditional response of each option considered in the economic analysis over 8 weeks of treatment, the probability of each option being the ‘best’ in leading to conditional response among available options, as well as the hazard ratio of each drug versus placebo with respect to conditional response. In addition to these three parameters examined in the network meta-analysis, the table shows the relative risk of each drug versus placebo regarding non-response (mean and 95% confidence intervals), as estimated in the guideline classical, pair-wise meta-analysis. Treatment options have been ranked from ‘best’ to ‘worst’ in terms of their ability to achieve conditional response, according to the results of the network meta-analysis.

Table 69

Response to pharmacological treatment: findings of the network meta-analysis and of classical pair-wise comparisons versus placebo.

The results of the network meta-analysis indicated that duloxetine had the highest probability of conditional response (mean 65.1% over 8 weeks), followed by sertraline, venlafaxine XL, pregabalin, escitalopram and paroxetine (mean 51.9% over 8 weeks). Placebo had the lowest probability of conditional response among options assessed (mean 42.8% over 8 weeks). The probability of duloxetine being the best drug in terms of response in people who have not discontinued their drug treatment was approximately 37%. Hazard ratios demonstrated that all drugs were significantly better than placebo in the outcome of conditional response. On the other hand, the relative risks derived from the guideline classical meta-analysis of placebo-controlled trials showed that all drugs significantly reduced the risk of non-response compared with placebo, with the exception of paroxetine; the latter demonstrated a positive effect which, nevertheless, did not reach statistical significance. However, it must be noted that since the data in the classical meta-analyses were based on ITT, cases of discontinuation due to side effects were counted as non-responders. This means that the classical meta-analysis considered both people who discontinued due to side effects, as well as people who did not discontinue due to side effects but did not respond to treatment either, as non-responders. This difference between network and classical meta-analysis may explain the discrepancies observed between the results: for example, duloxetine was shown to have the highest probability of conditional response (that is, the highest probability of response in those not discontinuing due to side effects) in the network meta-analysis, but not the lowest relative risk versus placebo in terms of non-response in the classical meta-analysis. The latter finding may be attributed to the fact that duloxetine is characterised by a high rate of discontinuation due to side effects, which reduces the response rate measured using an ITT approach. The lowest relative risk versus placebo in terms of non-response in the classical meta-analysis was that of sertraline. This is consistent with the fact that sertraline had the lowest probability of discontinuation due to side effects among the drugs considered, and, at the same time, the second highest probability of response in people who did not discontinue treatment due to side effects (that is, of conditional response).

The probability of discontinuation due to intolerable side effects and the probability of conditional response of each treatment option comprised the outcomes of the network meta-analyses that were utilised in the economic model. These data are also provided in Table 71.

Utility data and estimation of quality-adjusted life years

In order to express outcomes in the form of QALYs, the health states of the economic model needed to be linked to appropriate utility scores. Utility scores represent the HRQoL associated with specific health states on a scale from 0 (death) to 1 (perfect health); they are estimated using preference-based measures that capture people’s preferences on the HRQoL experienced in the health states under consideration. The systematic search of the literature identified two studies that reported utility scores for specific health states associated with GAD (Allgulander et al., 2007; Revicki et al., 2008). Details on the studies, their methods and reported utility data are provided in the respective section of the economic model described in Section 6.6.4.

According to NICE guidance regarding the selection of utility scores for use in cost-utility analysis, the measurement of changes in HRQoL should be reported directly from people with the condition examined, and the valuation of health states should be based on public preferences elicited using a choice-based method, such as TTO or SG, in a representative sample of the UK population. NICE recommends the EQ-5D (Brooks, 1996) as the preferred measure of HRQoL in adults for use in cost-utility analysis. When EQ-5D scores are not available or are inappropriate for the condition or effects of treatment, NICE recommends that the valuation methods be fully described and comparable to those used for the EQ-5D (NICE, 2008a).

Available utility data for people with GAD were not generated using EQ-5D. However, both studies included in the respective review used SF-6D for the estimation of utility scores in this population. SF-36 (and its shorter form SF-12) is a validated generic measure of HRQoL. The SF-6D algorithm can generate utility scores for all health states described from SF-36 (Brazier et al., 2002) and SF-12 (Brazier & Roberts, 2004), which have been elicited from a representative sample of the UK general population using SG; thus the valuation method meets NICE criteria.

The utility data reported in Allgulander and colleagues (2006) corresponded to the health states described in the economic model (that is, response, non-response, relapse following response, and no relapse following response); moreover, the definition of response in Allgulander and colleagues (2006) was the same as that used in all RCTs considered in the network meta-analysis that provided data on conditional response for the economic model. In contrast, the utility data reported in Revicki and colleagues (2008) corresponded to the health states of mild, moderate and severe anxiety, which could not be directly linked to the model health states of response, no response, relapse following response and no relapse following response. Therefore, it was decided to use the utility data reported in Allgulander and colleagues (2006) in the economic analysis.

It was assumed that the improvement in utility for people with GAD responding to treatment occurred linearly over the 8 weeks of treatment, starting from the utility value of non-response and reaching the utility value of response. People responding and not relapsing were assumed to experience a linear increase in their utility during the 6 months of maintenance treatment, starting from the utility value of response and reaching the utility value of response and no relapse. In contrast, people relapsing following response were assumed to experience a linear reduction in their utility during maintenance treatment, starting from the utility value of response and reaching the utility value of relapse following response.

Side effects of medication are expected to result in a reduction in utility scores corresponding to GAD-related health states. However, no studies on people with GAD reporting such ‘disutility’ due to side effects were identified in the literature. On the other hand, Revicki and Wood (1998) examined the effect of the presence of side effects associated with antidepressants in the HRQoL of people with depression. According to the study, people with a side effect reported lower utility scores compared with those not experiencing side effects. The observed mean disutility ranged from 0.01 for dry mouth and nausea to 0.12 for nervousness and light-headedness. However, except for light-headedness and dizziness, the reduction in utility caused by side effects did not reach statistical significance.

Clinical evidence on people with GAD suggests that side effects from drugs considered in the economic analysis consist mainly of nausea, insomnia and sexual problems (SSRIs and SNRIs), as well as dizziness, fatigue and headaches (pregabalin). Less common side effects include palpitations, tachycardia and orthostatic hypotension associated with duloxetine; SNRIs may increase blood pressure. Both SSRIs and SNRIs may result in suicidal thinking and self-harming behaviour in a minority of young people. Finally, SSRIs can cause gastrointestinal bleeding, especially if they are administered alongside NSAIDs.

Data on the risk for common, tolerable side effects have not been consistently collected and reported across RCTs included in the guideline systematic review, but available evidence indicates that all drugs are associated with such side effects in a similar degree. Data on less common but more severe side effects were sparser. On the other hand, discontinuation due to intolerable side effects was consistently reported in clinical trials. Development of intolerable side effects is expected to reduce more significantly the HRQoL of people with GAD compared with tolerable side effects.

Based on the above facts, data availability and limitations of available evidence, the economic analysis did consider the reduction in utility caused by intolerable side effects, given that data on discontinuation due to side effects were consistently reported for all drugs and analysed using network meta-analysis. The disutility caused by intolerable side effects was assumed to equal 0.12, which was the highest reduction in utility caused by the presence of side effects reported by people with depression taking antidepressants (Revicki & Wood, 1998). This reduction in utility due to intolerable side effects was assumed to last only 2 weeks, as discontinuation of drug treatment due to intolerable side effects was estimated to occur usually within 2 weeks from initiation of the particular drug. However, the reduction in utility caused by tolerable side effects was not considered in the economic analysis for two reasons: (i) inconsistent reporting of clinical data on tolerable side effects in RCTs might introduce bias in the economic analysis, should such data be included in the economic model, and (ii) available evidence on people with depression indicated that the majority of common side effects of antidepressants do not significantly reduce the HRQoL. Regarding less common and more severe side effects, such as gastrointestinal bleeding and suicidal thinking, these are likely to have a stronger negative impact on the HRQoL, but, given their low frequency, the implications of their omission (in terms of utility losses) are deemed to be less substantial at a study population level. Nevertheless, the lack of full consideration of the impact of side effects on the HRQoL of people with GAD treated with medication is acknowledged as a limitation of the economic analysis.

Cost data

Costs associated with the pharmacological treatment of people with GAD were calculated by combining resource-use estimates with respective national unit costs. Costs consisted of intervention costs and other health and social care costs incurred by people with GAD not responding to treatment or relapsing following response. Intervention costs of pharmacological treatment consisted of drug acquisition costs and GP visit costs. Intervention costs of no pharmacological treatment related to GP visit costs only. All costs were expressed in 2009 prices, uplifted, where necessary, using the HCHS Pay and Prices Index (Curtis, 2009). Discounting of costs was not necessary since the time horizon of the analysis was shorter than 1 year.

Drug acquisition costs were taken from the British National Formulary (BNF) 59 (British Medical Association and the Royal Pharmaceutical Society of Great Britain, March 2010). For each drug the lowest reported price was selected and used in the analysis; where available, costs of generic forms were considered. The average daily dosage of each drug was determined according to optimal clinical practice (the expert opinion of the GDG) and was consistent with the respective average daily dosage reported in the RCTs considered in the economic model. People discontinuing treatment due to intolerable side effects were assumed to have been already prescribed 1 month’s drug supply for their initiated drug, and therefore incurred the initiated drug cost over 4 weeks before switching to second-line treatment. The average daily dosages and acquisition costs as well as the total ingredient costs over 8 weeks of initial treatment and 6 months of maintenance treatment for all drugs are presented in Table 70. The ingredient cost of the second-line drug in each arm of the model was assumed to equal the average ingredient cost of all drugs except the one that was used as first-line treatment in this particular arm.

Table 70

Average daily dosage, acquisition costs and estimated 8-week and 6-month ingredient costs of drugs used in the treatment of people with GAD included in the economic model.

Regarding GP visits, these included one visit at initiation, two visits over the first 8 weeks of treatment, and another visit during maintenance treatment. People who discontinued their first-line treatment due to intolerable side effects were assumed to pay one extra visit to their GP, and then were initiated on second-line drug treatment following the same pattern of GP visits as that estimated for the first-line drug treatment. This pattern of GP visits was also assumed to apply to the cohort of people under no pharmacological treatment.

Costs of managing tolerable side effects were not considered separately in the analysis, partly due to inconsistent reporting of side-effect data in the RCTs included in the guideline systematic review of clinical evidence. Nevertheless, the GDG estimated that the majority of common tolerable side effects, such as nausea, insomnia, sexual problems (associated with SSRIs and SNRIs), dizziness, fatigue and headaches (associated with pregabalin), as well as the less commonly observed suicidal thinking (associated with antidepressants administered to younger people), palpitations and tachycardia (associated with duloxetine), would be discussed during monitoring GP visits which were considered at the estimation of intervention costs relating to initial and maintenance pharmacological treatment. It was the GDG’s view that even if the presence of these common side effects led to extra GP visits and incurred additional costs, these were unlikely to be considerable compared with total intervention costs. Regarding less common side effects, such as hypertension (associated with SNRIs) and gastrointestinal bleeding (associated with SSRIs), these were thought to result in higher management costs at an individual level, but given their low frequency they were deemed to entail smaller economic implications at a study population level. Therefore, although omission of costs associated with management of tolerable side effects is acknowledged as a limitation of the analysis, it is not considered to have substantially affected the economic modelling results.

The extra health and social care costs incurred by people with GAD not responding to treatment or relapsing following response to treatment were estimated based on data reported in the Adult Psychiatric Morbidity in England survey (McManus et al., 2009), supported by the expert opinion of the GDG. Data on resources used by people with GAD (including inpatient care, outpatient services, contacts with GPs, psychiatrists, psychologists, community psychiatric nurses, social workers and services provided by community day care centres) were combined with appropriate national unit costs (Curtis, 2009; DH, 2010) in order to estimate a total weekly cost incurred by people with GAD. The average length of stay for people with GAD receiving inpa-tient care was taken from national hospital episode statistics (NHS, The Information Centre, 2009). Based on the above data, the health and social care cost incurred by people with GAD not responding to treatment or relapse following response was approximately £804 per year or £15 per week. Details on the methods of estimation of this cost are provided in the economic analysis described in Section 6.6.4. People who did not respond to second-line pharmacological treatment and those who did not respond to no pharmacological treatment were assumed to incur this weekly health and social care GAD-related cost for the remaining time horizon of the analysis following no response. People who relapsed following response to treatment were assumed to incur maintenance treatment costs over 3 months and this health and social care GAD-related cost over the remaining 3 months of the 6-month maintenance treatment period that led to relapse.

Costs of treating tolerable side effects were not considered in the economic analysis due to lack of consistency in reporting appropriate side-effect data across all drugs.

Table 71 reports the mean (deterministic) values of all input parameters utilised in the economic model and provides information on the distributions assigned to specific parameters in probabilistic sensitivity analysis.

Data analysis and presentation of the results

Two methods were employed to analyse the input parameter data and present the results of the economic analysis.

First, a deterministic analysis was undertaken, where data are analysed as point estimates; results are presented as mean total costs and QALYs associated with each treatment option are assessed. Relative cost effectiveness between alternative treatment options is estimated using incremental analysis: all options are initially ranked from most to least effective; options that are dominated (they are more expensive and less effective than other options) are excluded from further analysis. Subsequently, ICERs are calculated for all pairs of consecutive options. ICERs express the additional cost per additional unit of benefit associated with one treatment option relative to its comparator. Estimation of such a ratio allows consideration of whether the additional benefit is worth the additional cost when choosing one treatment option over another.

After excluding cases of extended dominance (which occur when an intervention is less effective and more costly than a linear combination of two other options), ICERs are recalculated. The treatment option with the highest ICER below the cost-effectiveness threshold is the most cost-effective option.

One-way sensitivity analyses explored the following:

• The impact of the uncertainty characterising the monthly health and social care cost incurred by people with GAD not responding to treatment or relapsing following response on the results of the deterministic analysis. Since the estimation of this cost was based on a number of assumptions and data extrapolations, a scenario of a 70% change in this cost was tested to investigate whether the conclusions of the analysis would change.
• The impact of an increase in the extra GP visits following discontinuation of the first-line treatment due to intolerable side effects. The impact of three extra GP visits on the results was tested (in base-case analysis one extra visit was assumed).
• The uncertainty around the probability of response achieved by second-line drug treatment. In the base-case analysis, this probability was calculated as the average probability of conditional response of all drugs considered in the analysis except the one that was used as first-line treatment in each particular decision node of the model. However, it is possible that responsiveness to a drug used as second-line is lower than that observed when the drug is used as first-line. Therefore a scenario in which the responsiveness of the second-line drug was reduced by 15% was tested.

In addition to deterministic analysis, a probabilistic analysis was also conducted. In this case, all model input parameters were assigned probability distributions (rather than being expressed as point estimates), to reflect the uncertainty characterising the available clinical and cost data. Subsequently, 10,000 iterations were performed, each drawing random values out of the distributions fitted onto the model input parameters. This exercise provided more accurate estimates of mean costs and benefits for each intervention assessed (averaging results from the 10,000 iterations), by capturing the non-linearity characterising the economic model structure (Briggs et al., 2006).

The distributions of the probability of discontinuation due to intolerable side effects and the probability of conditional response for each drug, which were obtained using mixed treatment comparison techniques, were defined directly from values recorded in each of the 10,000 iterations performed in WinBUGS, as described in Appendix 13.

The probability of relapse for no pharmacological treatment was given a beta distribution. Beta distributions were also assigned to utility values, using the method of moments. The relative risk of relapse of drug treatment versus no treatment was assigned a log-normal distribution. The estimation of distribution ranges was based on available data in the guideline meta-analysis (relapse data) and the published sources of evidence (utility data). Costs (with the exception of drug acquisition costs) were assigned a gamma distribution; in order to define the distribution, a 30% standard error around the mean costs was assumed.

Table 71 provides details on the types of distributions assigned to each input parameter and the methods employed to define their range.

Results of probabilistic analysis are presented in the form of CEACs, which demonstrate the probability of each treatment option being the most cost effective among the strategies assessed at different levels of willingness-to-pay per unit of effectiveness (that is, at different cost-effectiveness thresholds the decision maker may set). In addition, the cost-effectiveness acceptability frontier is provided alongside CEACs, showing which treatment option among those examined offers the highest average net monetary benefit (NMB) at each level of willingness-to-pay (Fenwick et al., 2001). The NMB of a treatment option at different levels of willingness-to-pay is defined by the following formula:

Where E is effectiveness (number of QALYs), C is the costs associated with the treatment, and λ is the level of the willingness-to-pay per unit of effectiveness.

Results of deterministic analysis

According to deterministic analysis, sertraline was the most cost-effective option among those assessed because it produced the highest number of QALYs and was associated with the lowest costs (dominant option). ‘No pharmacological treatment’ was dominated by all drugs except pregabalin; the latter was more effective than placebo at an extra cost of £3,768 per QALY.

Table 72 provides mean costs and QALYs for every treatment option assessed in the economic analysis. The seven options have been ranked from the most to the least effective in terms of number of QALYs gained. It can be seen that sertraline is associated with lowest costs and highest benefits (QALYs) and consequently dominates all other drugs as well as no treatment. Figure 7 provides the cost-effectiveness plane showing the incremental costs and QALYs of all drugs versus paroxetine. It can be seen that sertraline is in the southeast quadrant and has the highest number of QALYs and the lowest costs relative to all other drugs assessed (no treatment is not shown in this graph).

Table 72

Mean costs and QALYs for each pharmacological treatment option for people with GAD assessed in the economic analysis - results per 1,000 people.

Figure 7

Cost-effectiveness plane of all drugs assessed in the economic analysis plotted against paroxetine – incremental costs and QALYs per 1,000 people with GAD.

Results were robust under all scenarios examined in one-way sensitivity analyses: sertraline remained dominant when the health and social care costs incurred by people with GAD not responding to treatment or relapsing following response increased by 70%, when three extra GP visits (instead of one) were assumed in the case of discontinuation of first-line treatment, and when conditional response for the second-line drug was reduced by 15%. Sertraline dominated all options except no treatment when the health and social costs incurred by people with GAD not responding to treatment or relapsing following response decreased by 70%. In this case, the ICER of sertraline versus no treatment was £946 per QALY gained, which is well below the lower cost-effectiveness threshold of £20,000 per QALY set by NICE (NICE, 2008b).

Results of probabilistic analysis

Results of probabilistic analysis were very similar to those of deterministic analysis: sertraline dominated all other treatment options when mean costs and QALYs derived from 10,000 iterations were estimated. Sertraline had also the highest probability of being the most cost-effective treatment option, at any level of willingness-to-pay per additional QALY gained. At the lower NICE cost-effectiveness threshold of £20,000/QALY (NICE, 2008b) the probability of sertraline being cost effective was 0.70, whereas venlafaxine XL, which was the second most cost-effective option, had a probability of only 0.13. The cost-effectiveness acceptability frontier coincided with the CEAC for sertraline, because sertraline produced the highest average net benefit at any level of willingness to pay.

Figure 8 shows the CEACs generated for each pharmacological treatment option assessed in the economic model. Table 73 shows the probability of each treatment option being cost effective at various cost-effectiveness thresholds, that is, at various levels of willingness-to-pay per QALY gained.

Figure 8

CEACs of all pharmacological treatment options for people with GAD assessed in the economic analysis.

Table 73

Probability of each pharmacological treatment option being cost effective at various levels of willingness-to-pay per QALY gained (WTP).

Discussion – limitations of the analysis

The results of the economic analysis suggest that sertraline is likely to be the most cost-effective pharmacological treatment for people with GAD. Sertraline dominated all other treatment options and had the highest probability of being the most cost-effective option at any level of willingness-to-pay per QALY gained, which reached 0.70 at the lower NICE cost-effectiveness threshold of £20,000 per QALY. The cost effectiveness of sertraline is attributed to a number of factors: sertraline had the lowest average probability of discontinuation due to intolerable side effects among all drugs assessed, and the second best probability of conditional response; in addition, sertra-line had the lowest acquisition cost among all drugs, as it is available in generic form. It must be noted that sertraline is currently not licensed for the treatment of people with GAD.

Clinical data on discontinuation due to intolerable side effects as well as response for those who did not discontinue due to intolerable side effects (conditional response) were synthesised using network meta-analytic techniques. Such methods enable evidence synthesis from both direct and indirect comparisons between treatments, and allow simultaneous inference on all treatments examined in pair-wise trial comparisons while respecting randomisation (Lu & Ades, 2004; Caldwell et al., 2005).

One limitation of the economic analysis was that data on conditional response for first-line drug treatment were also used to estimate the probability of response for second-line drug treatment, which, in every decision node of the model, was calculated as the average probability of conditional response of all drugs except the one that was used as first-line treatment in this particular node. This assumption was necessary in order to populate the model due to lack of response data on people with GAD switched to a second-line drug. However, it is possible that responsiveness to a drug used as second-line is lower than that observed when the drug is used as first-line. Nevertheless, one-way sensitivity analysis, in which the responsiveness of the second-line drug was assumed to be reduced by 15%, demonstrated that the results of the economic analysis were robust to this assumption.

Another limitation of the economic analysis is that it did not take into account the reduction in HRQoL and the costs associated with the management of tolerable side effects, which do not lead to treatment discontinuation. Consideration of these factors was not possible as there was no consistent reporting of side effects across trials included in the systematic review. Moreover, there is limited evidence on the reduction in HRQoL caused by the presence of side effects from drugs considered in the analysis and no such evidence in people with GAD. Regarding the reduction in HRQoL associated with the presence of side effects from antidepressants, available evidence has demonstrated that this is largely insignificant in people with depression (Revicki & Wood, 1998). Regarding costs associated with the management of tolerable side effects, these were considered to be non-substantial, as most side effects are expected to be managed during GP monitoring visits, which have already been considered at the estimation of intervention costs of pharmacological treatment. It should be noted that the economic analysis did consider the impact of the development of intolerable side effects, which lead to treatment discontinuation, on costs and HRQoL associated with pharmacological treatment of people with GAD.

The economic analysis revealed that drug acquisition costs may be an important factor in determining the relative cost effectiveness of pharmacological treatments for GAD: sertraline, which was found to be the most cost-effective option resulting also in lowest total costs, has currently the lowest acquisition cost, as it is available in generic form. Paroxetine, which is also available in generic form and has the second lowest acquisition cost among the drugs assessed, was ranked the second least costly drug and fourth most cost-effective option at a cost-effectiveness threshold of £20,000 per QALY (with probability of being cost effective about 0.05), despite the fact that it had one of the highest probabilities of discontinuation due to side effects and the lowest probability of conditional response among drugs. Venlafaxine XL, which has the lowest acquisition cost among patented drugs included in the analysis, was ranked the third least costly drug and second most cost-effective option at a cost-effectiveness threshold of £20,000 per QALY (with probability of being cost effective about 0.13). Based on these findings, it is expected that the relative cost effectiveness of drugs for the treatment of GAD is likely to change in the future, as eventually drugs will become available in generic form, resulting in a considerable reduction in their acquisition costs.

What is the relative effectiveness of sertraline compared with CBT in people with GAD that has not responded to guided self-help and psychoeducation in a stepped-care model?

This question should be addressed using a randomised controlled design in which people with GAD that has not responded to step 2 interventions are allocated openly to treatment with sertraline, CBT or waiting-list control for 12–16 weeks. The control group is important to demonstrate that the two active treatments produce effects greater than those of natural remission. The period of waiting-list control is the standard length of CBT treatment for GAD and is also commonly the length of time that it would take for specialist CBT to become available in routine practice. After 12–16 weeks all participants should receive further treatment chosen in collaboration with their treating clinicians.

The outcomes chosen at 12–16 weeks should include both observer- and participant-rated measures of clinical symptoms and functioning specific to GAD, and of quality of life. An economic analysis should also be carried out alongside the trial. The trial needs to be large enough to determine the presence or absence of clinically important effects and of any differences in costs between the treatment options using a non-inferiority design. Mediators and moderators of response should be investigated. Follow-up assessments should continue over the next 2 years to ascertain whether short-term benefits are maintained and, in particular, whether CBT produces a better long-term outcome.

Why is this important?

Both sertraline and CBT are efficacious in the treatment of GAD but their relative efficacy has not been compared. In a stepped-care model both CBT and sertraline are treatment options if step 2 interventions (guided self-help and/or psychoeducation) have not resulted in a satisfactory clinical response. At present, however, there are no randomised trial data to help prioritise next-step treatments and no information on how individuals with GAD may be matched to particular therapies. Clarification of the relative short- and longer-term benefits of sertraline and CBT would be helpful in guiding treatment.

Studies considered

There were a total of four trials comparing various herbal interventions with placebo. These were all small- to medium-sized trials, all of which were high quality. For two of the studies funding was provided by drug company sponsorship and one other from a national grant. One study failed to declare any funding. These trials could not be meta-analysed and therefore they are narratively reviewed below.

A summary of study characteristics can be found in Table 75 with full details in Appendix 15d, which also includes details of excluded studies. An evidence summary is provided in Table 76.

Table 75

Study information table for trials comparing herbal interventions with placebo.

Table 76

Evidence summary table for trials of herbal interventions versus placebo.

Narrative review of herbal interventions versus placebo

AMSTERDAM2009 conducted a randomised, double-blind efficacy trial in an outpatient clinic in the US comparing chamomile (n = 28) with placebo (n = 29) in participants with GAD. Participants met the DSM-IV diagnostic criteria for GAD and had a HAM-A score of greater than nine. Participants in the treatment group received one to five 220 mg capsules daily depending on tolerability levels. The placebo group received up to five capsules containing lactose monohydrate per day depending on their tolerability levels. Both treatment courses lasted for 8 weeks. Based on the evidence of this study, there is a moderate effect for chamomile over placebo in the reduction of clinician-rated anxiety scores. However, as this result has wide confidence intervals that just include the line of non-significance it should be interpreted with caution. This study also examined the difference in response rates as measured by a 50% reduction in HAM-A scores between the two groups. No statistically significant differences between the two groups were found. However, there was a 29% reduction in the level of non-response in favour of chamomile. With regards to discontinuation due to adverse events, there was no difference between the groups, suggesting that neither group was more likely to discontinue due to adverse events. Due to the limited evidence, it is difficult to draw any clear conclusions regarding the relative efficacy of chamomile to placebo.

WOELK2007 conducted a double-blind RCT in multiple outpatient centres in Germany, evaluating the therapeutic efficacy of gingko biloba (n = 70) versus placebo (n = 37) in participants with GAD. Some participants met the DSM-III-R diagnostic criteria for GAD (n = 82) and others met the diagnostic criteria for adjustment disorder with anxious mood by DSM-III-R (n = 25). Participants in the active treatment group received either a mean dose of 240 mg (n = 36) or a mean dose of 480 mg (n = 34) over 4 weeks. The placebo group took two film-coated drugs per day that were of the same appearance as the gingko biloba pills. Based on this limited evidence, there was a statistically significant moderate effect in favour of gingko biloba in the reduction of clinician-rated anxiety scores. For non-response, which was measured by a 50% reduction in HAM-A scores, there was a 25% reduction in non-response, which was statistically significant suggesting that those in the active treatment group were more likely to respond than those in the placebo group. In contrast, there were no significant differences in relation to non-remission between the two conditions. Finally, there were no significant differences in relation to dropout due to any reason. However, these results should be interpreted with caution as they are based on one medium-scale study, and given the wide confidence intervals it is difficult to make any firm conclusions from this evidence about the relative efficacy of ginkgo biloba to placebo.

HANUS2004 conducted a double-blind RCT in multiple outpatient centres in Paris, evaluating the therapeutic efficacy of combined plant extracts (n = 130) in comparison with placebo (n = 134) in participants with GAD. Participants met the DSM-IV diagnostic criteria for GAD and had a HAM-A score of between 16 and 28. Participants in the active treatment group received a mean dose of 375 mg (two tablets per day) of combined plant extracts (that is, Crataegus oxyacantha, Eschscholzia californica and magnesium) over a period of 3 months. The placebo group were given an indistinguishable tablet that was made from the same ingredients as the study drug except for the active ingredients. Firstly, in relation to HAM-A scores, there was a statistically significant small effect between treatments in favour of the combined plant extracts. Secondly, in relation to non-response (again measured as a 50% reduction in HAM-A scores), there was a 20% reduction in non-response for those taking the active treatment, which was statistically significant. Finally, there was no statistically significant difference between treatments in relation to dropout due to adverse events. Once more, firm conclusions are subject to cautious interpretation due to the limited evidence available and the small sample size.

ANDREATINI2002 conducted a double-blind RCT in Brazil, evaluating the therapeutic efficacy of valerian extract (n = 12) in comparison with placebo (n = 12) in participants with GAD. All participants met the DSM-III-R diagnostic criteria for GAD. Participants in the active treatment group received a mean dosage of 81.3 mg per day of valerian extract over 4 weeks. The placebo group were given identical capsules, which were administered three times per day. Firstly, in relation to HAM-A scores, there was no statistically significant differences between valerian extract versus placebo. There was no data reported for either non-response or non-remission. Finally, there was no statistically significant difference between treatments in relation to dropout due to any reason. Again, conclusions are subject to cautious interpretation because of the limited evidence available.

Studies considered

There were a total of four trials comparing various herbal interventions with benzodiazepines, including lorazepam, diazepam and oxazepam. These trials were all small to medium sized and of high quality. One study was funded by drug company sponsorship and the other three studies failed to declare any funding. These trials could not be meta-analysed therefore they are narratively reviewed below. A summary of study characteristics can be found in Table 77 with full details in Appendix 15d, which also includes details of excluded studies. An evidence summary is provided in Table 78.

Table 77

Study information table for trials comparing herbal interventions with benzodiazepines.

Table 78

Evidence summary table for trials of herbal interventions versus benzodiazepines.

Narrative review of herbal interventions versus benzodiazepines

Only one study (HERRERA-ARELLANO2007) examined the effectiveness of galphimia glauca (n = 72) in comparison with lorazepam (n = 80) for treating the symptoms of GAD. This was a medium-scale, high-quality RCT in an outpatient setting in Mexico. With regard to the comparative beneficial effects of these two treatments in reducing clinician-rated anxiety scores, there was no statistically significant difference at post-treatment. Also, there was no statistically significant difference in dropout for any reason between the two groups. However, there was a statistically significant difference in favour of the herbal intervention with regard to dropout due to adverse events, with only 7% dropping out due to adverse events in the herbal intervention compared with 20% in the lorazepam group. However, the results as a whole should be interpreted with caution due to the lack of placebo group, wide confidence intervals, and lack of statistical significance.

WOELK2010 conducted a double-blind RCT in multiple outpatient centres in Germany, comparing lavender capsules (n = 40) with lorazepam treatment (n = 37) in participants with GAD as diagnosed by DSM-IV criteria. Participants in the herbal treatment group received one capsule (80 mg) of silexan (an oil-produced form of lavender) and one capsule of lorazepam placebo. Participants in the drug condition received one capsule (0.5mg) of lorazepam and one capsule of silexan placebo. Both treatment courses lasted 6 weeks. In terms of reducing clinician-rated anxiety scores, there was a non-significant difference between the two treatments.

In addition, there was a 20% reduction in the risk of non-response in favour of lavender, however, this was not statistically significant. Moreover, there was an 18% reduction in the risk of non-remission in favour of lavender, which again was not statistically significant. Finally, there were no statistically significant differences between treatments in the risk of dropout for any reason. Due to the limited evidence, it is difficult to come to any firm conclusions about the relative efficacy of these two treatments.

ANDREATINI2002 is a double-blind RCT in Brazil, comparing diazepam (n = 12) with valerian extract (n = 12) in participants with GAD as diagnosed by DSM-III-R criteria. Participants in the active treatment group received a mean dosage of 81.3 mg per day of valerian extract over 4 weeks. Participants in the diazepam condition received a dosage of 6.5 mg per day in capsule form. The capsules were administered three times a day with the lowest dose consisting of two placebo pills and one active capsule based on response. In terms of reducing clinician-rated anxiety scores, there was a small but statistically insignificant effect in favour of diazepam treatment. There was no data available for either response or remission. Finally, there was no statistically significant difference between the two conditions on the outcome of dropout for any reason. It is difficult to come to any clear conclusions about the relative efficacy of these two treatments due to the small sample size, lack of statistical significance and large confidence intervals.

Only one study examined the effectiveness of passion flower extract (n = 18) versus oxazepam (n = 18) for the treatment of GAD (AKHONDZADEH 2001A). The study consisted of a double-blind RCT conducted in an outpatient setting in Iran. Both passion flower extract and oxazepam were found to be effective in reducing clinician-rated anxiety scores from baseline severity. In both groups, post-hoc comparisons of the baseline HAM-A scores at post-treatment revealed a significant reduction from baseline (p < .001). The differences between the two treatments were significant at day four (t = 2.84, df = 30, p = .008), however, after the fourth day the differences were no longer significant. Moreover, significantly more problems relating to impairment of job performance were encountered with people taking oxazepam (p = .049). However, there was no significant differences between the two treatments in terms of total side-effect profile (p = .83). These results are based on one small-scale study and thus it is difficult to make any firm conclusions from this evidence.

Narrative review of acupuncture

There were no studies concerning people with GAD who had received a diagnosis that met the eligibility criteria of the GDG. However, this partly reflected the fact that all identified studies were conducted in China and used the Chinese Classification of Mental Disorders (CCMD) criteria.

Zhiling and colleagues (2006) conducted a RCT in China, comparing acupuncture treatment (n = 35) with a medication control group (n = 30) in participants with GAD. Participants met the CCMD-3 criteria and had Self-rating Anxiety Scale (SAS) scores of 50 or more. Acupuncture treatment (which consisted of six acupuncture points) was given once daily. The control group were administered 0.5 to 2 mg of lorazepam (two or three times a day) with 20 mg of oryzanol, which is a mixture of plant chemicals (three times a day), or 10 to 20 mg of propanolol. Both treatment courses lasted for 30 days. The therapeutic effects between the groups looked at a measure of remission (disappearance of symptoms with stable emotions). No statistically significant difference between groups was found (RR = 0.90; CI, 0.65 to 1.24). Of participants in the treatment group, 65.7% did not achieve remission compared with 73.3% in the control group. For response (apparent improvement of symptoms with occasional anxious state) no statistical difference was found (RR = 0.90; CI, 0.59 to 1.38). Of participants in the treatment group, 54.3% did not respond compared to 60% in the control group.

Yuan and colleagues (2007) conducted a quasi-randomised trial, also in China, comparing the therapeutic efficacy of needling therapy with Western medication and a combination treatment. Participants were diagnosed with GAD using the CCMD-3-R criteria and had a HAM-A score of 15 or more. Participants in the Western medication group (n = 29) were treated with 20 mg of fluoxetine or paroxetine. In addition, 0.4 to 1.6 mg of alprazolam was given according to the participant’s condition. All drugs were administered once daily for 6 weeks. There were nine to ten acupuncture points selected in the needling therapy group (n = 29) and the treatment was given once daily, 6 times a week for 6 weeks. The same method for both Western medication and needling therapy groups was used for participants in the combination treatment group (n = 28). Clinical efficacy was scored using the CGI which includes a general index subscale. A high general index score indicates an inferior therapeutic effect. There was no statistically significant difference between the Western medication and needling therapy groups (SMD = 0.09; 95% CI, −0.44 to 0.63) or between the needling therapy and combination treatment groups (SMD = −0.16; 95% CI, −0.70 to 0.38).

Ruan (2003) conducted an RCT in China, comparing combined treatment of Chinese medicine with acupuncture (COM) with Western medication. Participants were diagnosed with ‘anxiety neurosis’ using CCMD-2 and the self-rated Anxiety Neurosis Scale. Those scoring more than 50 were eligible to participate. They were randomised into the COM group (n = 86) or the Western medication group (n = 83). The COM group were treated with Chinese medicine, taken twice each day, and received acupuncture daily for 30 to 60 minutes. The Western medication group were given the TCA doxepin (an average of 150 mg per day). Treatment lasted for 30 days. Thirty-nine out of 86 in the COM group and 30 out of 83 in the Western medication group remitted. There was no statistically significant difference between the COM and Western medication group. Clinical efficacy was scored using SAS-CR and there was no statistically significant difference between the COM and Western medication group (SMD = −0.14; 95% CI, −0.45 to 0.16).

Zhou and colleagues (2003) conducted a randomised trial comparing the combined effect (n = 50) of acupuncture and flupentixol (an antipsychotic drug) with flupentixol only (n = 50). Participants were diagnosed with ‘anxiety neurosis’ using CCMD-2-R. Participants were given acupuncture once per day for 10 days. They took 5 days’ rest before the second wave of treatment. There were three waves of treatment in total. They also took 20 mg of flupentixol three times daily continuously for 40 days. According to remission rates, combined treatment was statistically significantly better than treatment with a single drug (RR = 0.71; CI, 0.57 to 0.89). Of participants in the treatment group, 64% did not achieve remission compared with 90% in the control group. Participants also reported side effects. One person experienced dry mouth in the combined treatment group, and one person experienced insomnia in the drug-only group. One from each group had dizziness.

Zhang and colleagues (2003) conducted a randomised trial comparing the clinical efficacy of acupuncture (n = 157) and doxepin (n = 139). They selected people with ‘anxiety neurosis’ according to CCMD-2 and an SAS score of more than 50. The acupuncture group had treatment once a day, with a 1-day interval after six consecutive treatments; there were 30 sessions in total. They used any two of four methods (varying in methodology), one of which included giving an injection at an acupuncture point. Participants in the comparison group were given 25 mg of doxepin three times a day for 4 weeks, which was modified according to therapeutic or adverse events. The therapeutic effects between the groups looked at a measure of remission (disappearance of symptoms with stable emotions). No statistical difference was found (RR = 0.86; CI, 0.71 to 1.03). Of participants in the treatment group, 56.1% did not achieve remission compared with 65.5% in the control group. For response (clinical symptoms relieved with occasional emotional fluctuation) no statistical difference was found (RR = 1.04; CI, 0.90 to 1.21). Of participants in the treatment group, 72.1% did not respond compared with 69.1% in the control group.

Guizhen and colleagues (1998) conducted a randomised trial comparing the clinical efficacy of acupuncture (n = 80) with behavioural desensitisation (n = 80) and a combination of both treatments (n = 80) on people with ‘anxiety neurosis’ (with SAS scores of 50 or more). The acupuncture-only group were treated once every other day for ten sessions (this comprised one course). Each participant received between one and three courses. Behavioural desensitisation involved self-relaxation techniques (twice daily) and psychotherapy that incorporated desensitisation therapy (twice weekly for ten sessions). For acupuncture combined with behavioural desensitisation, each participant received both treatments in the same day and between one and four courses of treatment with 3 to 7 day intervals between courses. Physical examination and SAS evaluation measured remission (disappearance of symptoms – SAS score of less than 45). The results for the combined treatment group were significantly better than the acupuncture-only (RR = 0.59; CI, 0.46 to 0.77) or the behavioural desensitisation-only (RR = 0.64; CI, 0.49 to 0.84) groups. Of participants in the acupuncture-only group, 80% did not remit; in the behavioural desensitisation-only group, 73.8% did not remit; 47.5% in the combined treatment group did not remit. For response (marked improvement in symptoms and significant decrease in SAS scores, that is, more than 20 points) combined treatment was significantly higher than the acupuncture-only (RR = 1.30; CI, 1.02 to 1.65) and behavioural desensitisation-only (RR = 1.24; CI, 0.98 to 1.57) groups. Of participants in the combined treatment group, 72.3% did not respond compared with 55% and 57.5% in the acupuncture-only and behavioural desensitisation-only groups, respectively.

Is chamomile/ginkgo biloba more effective than placebo in increasing response and remission rates and decreasing anxiety ratings for people with GAD?

This question should be addressed using a placebo-controlled, double-blind randomised design to compare the effects of a standardised dose of chamomile (220–1100 mg) or ginkgo biloba (30–500 mg) in a readily available form, for example a capsule, with placebo. This should assess outcomes at the end of the trial and at 12-month post-trial follow-up. The outcomes chosen should include both observer-and participant-rated measures of clinical symptoms and functioning specific to GAD, and of side effects. There should be a health economic evaluation included and an assessment of quality of life. The trial should be large enough to determine the presence or absence of clinically important effects using a non-inferiority design. Mediators and moderators of response should be investigated.

Why this is important

GAD is a common mental health disorder and the results of this study will be generalisable to a large number of people. There is evidence for the efficacy of chamomile and ginkgo biloba in reducing anxiety in people with GAD but the evidence base is small (one study). However, the scarce literature on the effectiveness of other herbal interventions for treating GAD points to chamomile and ginkgo biloba as two of the more effective herbal interventions. Moreover, both these herbal remedies are widely available and relatively inexpensive. Furthermore, at present there is no scientific evidence of side effects or drug–herbal interactions in relation to chamomile or ginkgo biloba. As both these herbal interventions are readily available and have no known side effects, they could be used at an early stage as a means of preventing progression to drug treatments, which are associated with a number of undesirable side effects and dependency.