Types of interventions

We included interventions used to treat or prevent deterioration of symptoms of antipsychotic-induced TD of relevance for people in the NHS, indicated as priority interventions: ‘switch to SGA (including switch to amisulpride, clozapine, olanzapine, quetiapine, risperidone, ziprasidone)’, ‘antipsychotic (AP) reduction’, ‘antipsychotic maintenance/TAU (including AP)’, ‘antipsychotic withdrawal (with placebo)’, ‘FGA (any)’, ‘anticholinergic and AP continuation’, ‘anticholinergic withdrawal and AP continuation’, ‘benzodiazepines and AP continuation’, ‘buspirone and AP continuation’, ‘hypnosis or relaxation and AP continuation’, ‘vitamin E and AP continuation’ and ‘placebo (with AP continuation)’.

We assumed that any patient who met the inclusion criteria was, in principle, equally likely to be randomised to any of the interventions and, thus, the transitivity assumption was likely to hold on the onset.

Types of outcome measures

The following outcomes were measured:

• primary outcome – no clinical improvement of TD symptoms (< 50% improvement on scales)
• secondary outcome – total discontinuation rates.

We intended to analyse all planned outcomes described in the main paper but we were unable to do so because of the limited data available. We estimated the relative ranking of the competing interventions according to both of the above outcomes.

Relative treatment effects

Odds ratios were employed for dichotomous outcomes. When continuous outcomes were measured, we analysed them using the MD if all studies used the same measure to assess the same outcome. Standardised mean difference, Hedge’s adjusted g, was used when a different measure was used across studies to assess a common continuous outcome.¹⁷⁰

Relative treatment ranking

We estimated p-scores, which are the most frequent analogues of surface under the cumulative ranking curves (SUCRAs), to obtain a hierarchy of the competing interventions.^171,172

1. Assessment of clinical and methodological heterogeneity within treatment comparisons.
   We assessed the presence of clinical and methodological heterogeneity within each pairwise comparison by comparing trial and study population characteristics across all eligible trials. Considerable differentiation in synthesised studies in terms of patient, study and intervention characteristics might lead to a lack of usefulness of obtained results.¹⁷³
2. Assessment of transitivity across treatment comparisons
   The assumption underlying NMA implies that one can learn about the relative effectiveness of ‘A versus B’ via a common comparator, for instance C.^155,174 We were unable to compare the distribution of effect modifiers across comparisons because of the limited data, but we compared the particular study characteristics qualitatively. Moreover, we assessed if the indication of the included interventions varied according to the alternative it is compared against.

Assumptions when estimating the heterogeneity

In pairwise meta-analysis we assumed different heterogeneity variances for each comparison. In NMA, we assumed a common heterogeneity variance across all treatment comparisons in the network.

Assessment of statistical inconsistency

Network meta-analysis assumes consistency between various sources of evidence; that means that direct and indirect evidence is expected to be in agreement. However, it might be that the assumption of consistency is violated either in certain parts or in the entire network. We intended to evaluate statistical inconsistency using both local and global methods. In particular, we intended to evaluate the consistency assumption using the loop-specific approach.¹⁸³ Employing this method, we would estimate the disagreement between direct and indirect evidence in each closed loop (inconsistency factors).

Moreover, we intended to evaluate inconsistency in the entire network using the design-by-treatment interaction model.^156,184,185 However, there was only one closed loop in the network for the ‘total discontinuation rates’ outcome and, thus, we only judged on inconsistency for this loop using the loop-specific approach.

Investigation of heterogeneity and inconsistency

Several metaregression and subgroup analyses were planned in order to assess the impact of potential effect modifiers on the treatment effects. Our intention was to explore the impact of study and population characteristics fitting network metaregression models in a Bayesian environment using the WinBUGS software version 1.4.3 (MRC Biostatistics Unit, Cambridge, UK) and considering vague prior distributions for the covariates. As these analyses are known to have low power,^186,187 their presentation would be of questionable usefulness in the case of very few data.

No clinical improvement of tardive dyskinesia symptoms

Evidence for the outcome ‘no clinical improvement of TD symptoms’ formed two disconnected networks that were analysed separately using NMA. The two formed networks for the outcome ‘no clinical improvement of TD symptoms’ are illustrated in Figure 14 [included treatments: ‘benzodiazepine (clonazepam, diazepam) and AP continuation’, ‘buspirone and AP continuation’, ‘MAO inhibitor (isocarboxazid, selengiline) and AP continuation’, ‘vitamin E and AP continuation’, ‘anticholinergic (biperiden, procyclidine) and AP continuation’, ‘antipsychotic maintenance/TAU (including AP)’, ‘hypnosis or relaxation and AP continuation’, ‘antipsychotic reduction (reduced dose FGA)’] and Figure 15 (included treatments: ‘switch to haloperidol’, ‘switch to thiopropazate’, ‘switch to quetiapine’). Nodes represent available treatments and edges represent available comparisons. Nodes and edges are weighted according to the number of studies involved in each treatment. Two studies^{105–109,115,116} compared treatments that were connected to neither of the two networks and, thus, were excluded from the NMA. ‘MAO inhibitor (isocarboxazid, selengiline) and AP continuation’ is included in the first subnetwork of Figure 14 despite the fact that it is not in the list of priority interventions as it connects ‘placebo (with AP continuation)’ to ‘anticholinergic (biperiden, procyclidine) and AP continuation’, the relative effectiveness of which is of interest.

FIGURE 14

Network plot for the first subnetwork for the outcome ‘no clinical improvement of TD symptoms’.

FIGURE 15

Network plot for the second subnetwork for the outcome ‘no clinical improvement of TD symptoms’.

Table 8 shows the NMA results for the network illustrated in Figure 14 for the outcome ‘no clinical improvement of TD symptoms’. Studies in which all participants were classified as events or non-events in both groups were excluded. The forest plot in Figure 16 shows the ORs of all treatments versus ‘placebo (with AP continuation)’ derived from the NMA. According to Table 8 and Figure 16, the NMA suggests that ‘hypnosis or relaxation and AP continuation’ has the greatest benefit over ‘placebo (with AP continuation)’, whereas ‘buspirone and AP continuation’ and ‘antipsychotic reduction (reduced dose FGA)’ are also more effective than ‘placebo (with AP continuation)’. ‘Anticholinergic (biperiden and procyclidine) and AP continuation’ appears to be less effective than ‘placebo (with AP continuation)’. The results are consistent with the corresponding effect estimates derived from pairwise meta-analysis. It should be noted, however, that any judgements on the relative effectiveness of the treatments are mitigated by the high imprecision associated with most network estimates.

TABLE 8

Network meta-analysis results for the outcome ‘no clinical improvement of TD symptoms’

FIGURE 16

Network meta-analysis results for comparisons ‘placebo (with AP continuation) vs. active treatments for outcome ‘no clinical improvement of TD symptoms’ (random-effects model, common heterogeneity parameter across comparisons). (more...)

The subnetwork corresponding to Figure 15 is formed by two studies only; a third study that was connected to the network¹⁸⁸ was excluded as all participants were classified as events. Thus, we do not present indirect estimates for the particular network as the value of drawing inferences would be doubtful because of the substantially limited data availability. The only study that compared ‘switch to FGA’ with ‘switch to SGA’ for the outcome ‘no clinical improvement’ was Emsley et al.,^110,111 in which an OR of 1.96 (95% CI 0.56 to 6.92) in favour of ‘switch to SGA’ was calculated. This comparison does not benefit from the NMA as it is not connected with the largest subnetwork of Figure 14 and there is no indirect evidence that can be used to strengthen evidence on the relative effectiveness of the two interventions.

Total discontinuation rates

Evidence for the outcome ‘total discontinuation rates’ formed two disconnected networks that were analysed separately using NMA, and are illustrated in Figures 17 and 18. Nodes represent available treatments and edges represent available comparisons. Nodes and edges are weighted according to the number of studies involved in each treatment. ‘MAO inhibitor (isocarboxazid, selengiline) and AP continuation’ is included in the subnetwork of Figure 17 despite the fact that it is not in the list of priority interventions as it connects ‘placebo (with AP continuation)’ to ‘anticholinergic (biperiden, procyclidine) and AP continuation’.

FIGURE 17

Network plot for the first subnetwork for the outcome ‘total discontinuation rates’.

FIGURE 18

Network plot for the second subnetwork for the outcome ‘total discontinuation rates’.

Studies in which all participants were classified as events or non-events in both groups were excluded. The forest plot in Figure 19 shows the ORs of all treatments versus ‘placebo (with AP continuation)’ derived from the NMA corresponding to the network plot of Figure 17. Tables 9 and 10 summarise the network estimates corresponding to the networks of Figures 17 and 18, respectively. As is shown in Tables 9 and 10 and Figure 19, most network estimates are highly imprecise (with rather wide CIs), rendering any conclusions on relative treatment effectiveness impractical. No statistically significant differences occur for any treatment versus ‘placebo (with AP continuation)’ in terms of discontinuation rates.

FIGURE 19

Network meta-analysis results for the comparisons of active treatments vs. placebo for the outcome ‘total discontinuation rates’ (using a random-effects model and using a common heterogeneity parameter across comparisons) corresponding (more...)

TABLE 9

Network meta-analysis results for the outcome ‘total discontinuation rates’ corresponding to the subnetwork of Figure 17

TABLE 10

Network meta-analysis results for the outcome ‘total discontinuation rates’ corresponding to the subnetwork of Figure 18

Comparison of heterogeneity parameters with empirical distributions

For a binary mental health outcome and a ‘non-pharmacological versus any’ comparison type, a median value of 0.13 is suggested for τ.¹⁸¹ The specific value is greater than our estimation of heterogeneity (0) for both outcomes ‘no clinical improvement of TD symptoms’ and ‘total discontinuation rates’.

No clinical improvement of tardive dyskinesia symptoms

The p-score value of ‘hypnosis or relaxation and AP continuation’ is 89%, indicating that it is 89% as effective as a treatment that would be ranked always first without uncertainty. ‘Anticholinergic (biperiden, procyclidine) and AP continuation’ appears to be the worst treatment in terms of ‘no clinical improvement of TD symptoms’ as it has a p-score close to 0. These findings are in agreement with the network effect estimates presented in Table 8 and Figure 16.

Total discontinuation rates

‘Antipsychotic reduction (reduced dose FGA)’ has the greatest p-score (90%) in terms of total discontinuation rates. Uncertainty in treatment effects escalates in uncertainty in treatment ranking resulting in many p-scores around 50%.

Clustered ranking plot for the outcomes ‘no clinical improvement of tardive dyskinesia symptoms’ and ‘total discontinuation rates’

In Figure 21 we have ranked treatments according to the outcomes ‘no clinical improvement of TD symptoms’ and ‘total discontinuation rates’. Hierarchical cluster analysis is performed to group the competing treatments. Different colours represent different groups of treatments considering jointly their relative ranking for two outcomes. Treatments that belong to the same group may be considered as being of comparable performance with respect to both outcomes. According to Figure 21, ‘antipsychotic reduction (reduced dose FGA)’ has the highest performance on both outcomes in terms of ranking for the two considered outcomes. ‘Anticholinergic (biperiden, procyclidine) and AP continuation’ and ‘MAO inhibitor (isocarboxazid, selengiline) and AP continuation’ can be considered as the treatments having the worst joint performance for the outcomes ‘no clinical improvement of TD symptoms’ and ‘total discontinuation rates’.

FIGURE 21

Clustered ranking based on p-scores for the outcomes ‘no clinical improvement of TD symptoms’ and ‘total discontinuation rates’. Hierarchical cluster analysis is performed to group the competing treatments.