The analyses of efficacy data used the combined intention to treat (ITT) (Full Analysis) populations from three randomised parallel group double-blind trials comparing Sativex® with placebo. The results of each of these studies have been separately published, and we will not describe them further here [9
]. Each study included a similar population of patients, each study was of similar design, and each study used a patient-reported spasticity severity score as the primary assessment of the severity of spasticity. In one study [9
], the patients assessed their spasticity on a 0 to 100 mm visual analogue scale. These data were converted to a 0 to 10 scale by dividing each individual observation by 10. Both the other studies [10
] used a patient-reported 0 to 10 Numeric Rating Scale (NRS) on which 0 represented the complete absence of spasticity and 10 the worst ever spasticity.
The three studies were combined in a pre-planned meta-analysis [16
] which showed no statistically significant study-by-treatment interaction (P
>0.1) or heterogeneity of variance (P
>0.1). Overall, the meta-analysis showed Sativex® to be superior to placebo as assessed by the change in spasticity severity according to the NRS at the end of the study (treatment difference of 0.32 units, SE diff = 0.145; 95% CI −0.61 to −0.04; P
= 0.026). Secondary measures of anti-spasticity effect included a 30% responder analysis (OR 1.62; 95% CI 1.15 to 2.28; P
= 0.0073), a Patient Global Impression of Change (OR 1.66; 95% CI 1.16 to 2.9: P
= 0.0036) and an Ashworth Score (no significant difference between treatments).
The analyses of adverse events (AEs) and dose of test medication used study participants who received at least one dose of study medication (the combined safety populations). In all of the analyses, imputation for missing values was not done. Consequently, cases with missing values were removed from the affected analyses. In particular, this includes missing values for baseline covariates (12 participants) and for prior exposure to cannabis (1 participant).
The spasticity data were analysed using a general linear model. The dependent variable was the change from baseline in mean spasticity assessment. Fixed factors considered in the models were treatment group (Sativex®/placebo) (TG), prior use of cannabis (Yes/No) (PC), study (ST), and the interaction terms TG*PC, TG*ST, PC*ST, TG*PC*ST. Mean baseline spasticity (BS), measured on a continuous 0 to 10 scale, was included as a covariate in all models. Models containing the following terms were fitted sequentially; with each subsequent model being considered only if the term to be removed from the current model did not contribute significantly (P <0.05):
Model 1. BS, ST, PC, TG, TG*PC, PC*ST, TG*ST, TG*PC*ST
Model 2. BS, ST, PC, TG, TG*PC, PC*ST, TG*ST
Model 3. BS, ST, PC, TG, TG*PC, PC*ST
Model 4. BS, ST, PC, TG, TG*PC
Model 5. BS, ST, PC, TG
In the event of statistically significant interaction terms, TG*PC*ST, TG*ST or PC*ST, the interpretation of the analysis was explored further.
The absence of a statistically significant TG*PC interaction would suggest that prior use of cannabis has no association with the assessment of efficacy. Thus, there would be no evidence to suggest that the treatment difference differs between prior and non-prior cannabis users, that is, even if prior users are able to distinguish between the treatments, this is not likely to have led to bias in the assessment of efficacy.
At all stages, the interactions were to be carefully evaluated to determine whether they are of degree (quantitative) or direction (qualitative), and how they might influence the treatment contrast.
It is noted that these analyses could have been done on the three individual studies. However, the studies were only powered for their primary comparison and would have lacked power for examining the interaction terms that are so important to this methodology. In addition, the issue of blinding was raised in relation to a regulatory submission, by which time the studies had already concluded with no possibility of re-sizing them to provide reasonable power for the assessment of interaction. Consequently, it was decided to take this meta-analytic approach in order to look for systematic evidence of bias and to maximise the power of these comparisons.
The distribution of the residuals from the final model was examined for evidence of departures from the model’s assumptions, and, in particular, the Shapiro-Wilk test was used to assess the normality of the fitted residuals. In the event of gross-departures for the underlying assumptions, the data (response and baseline covariates) were ranked and these transformed data then analysed using the same linear model approach.
Distribution of adverse events: prior v naïve cannabis users
The patients most likely to be able to distinguish whether they were in the Sativex® or placebo groups were those who had previous exposure to cannabis. This could manifest itself in a different adverse event profile between prior and naïve users in the Sativex® group compared with the placebo group. The following adverse events were listed as potential effects of cannabis on the Information Leaflet given to patients prior to them giving informed consent for participating in the studies:
○ dizziness, disturbance in attention, euphoric mood, disorientation, feeling disconnected, loss of balance, difficulty with speaking, confusion, worry, paranoia, fainting, hallucinations, disorientation, poor concentration and/or forgetfulness, losing touch with reality, a feeling of general happiness or sadness, feeling abnormal or drunk (nervous system disorders)
○ fatigue, tiredness or drowsiness, weakness, increased incidence of falling, lethargy (general disorders and administration site conditions)
○ nausea, dry mouth, diarrhoea, thirst, throat irritation, feeling sick, being sick, abdominal discomfort, increase or decrease in appetite, changes in sense of taste, stinging or discomfort in the mouth and on the tongue (gastrointestinal disorders)
Consequently, events classified using the Medical Dictionary for Regulatory Activities (MedDRA) into system organ classes “nervous system disorders”, “general disorders and administration site conditions”, “gastrointestinal disorders” and “psychiatric disorders” were examined individually where they were reported by at least 10 patients (using the MedDRA preferred terms) in the analysis; events that occurred in fewer than 10 patients were considered to be too rare to provide reliable results.
The number of patients reporting these adverse events, by preferred term, was summarised across the included studies and displayed by treatment group and prior/naïve cannabis use within the treatment group. The odds ratio for patients having received Sativex® (as opposed to placebo) for both prior cannabis users and cannabis naïve users was calculated for each AE together with 95% confidence intervals.
Also, for each of these events, a logistic model was used to assess the odds ratio (Sativex®/Placebo) adjusted for prior cannabis exposure. Fixed factors included in the model were treatment group (Sativex®/placebo) (TG), Prior use of cannabis (Yes/No) (PC), Study (ST), and the interaction terms TG*PC, TG*ST, PC*ST, TG*PC*ST. The TG*PC interaction and PC main effect were to be inspected for evidence of an association between prior cannabis exposure and the adverse event profile, either as an interaction with treatment group or independent of treatment. If the TG*PC interaction was statistically significant then that would provide evidence of possible unblinding. If it were not, but the main effect of PC in the model excluding the TG*PC interaction was statistically significant, this would suggest that more (or fewer) prior cannabis users experienced the event than cannabis naïve patients. But it would not provide evidence that they would be more likely to identify their study treatment. The TG term was to be inspected for evidence of a differential adverse event profile between the treatment groups. Again, if it was statistically significant it would not necessarily provide evidence of unblinding. Odds ratios and 95% CIs were to be presented, as appropriate.
Association between the experience of central nervous system events and assessment of efficacy
The association of the three most frequently occurring central nervous system adverse events and the assessment of efficacy was investigated using the assessment of spasticity over the primary period during the three studies.
As in the previous analysis, the data were analysed using a general linear model. The dependent variable was the change from baseline in spasticity assessment. Fixed factors included in the model were treatment group (Sativex®/placebo) (TG), experienced one or more of the three AEs (Yes/No) (AE), study (ST), and the interaction terms TG*AE, TG*ST, AE*ST, TG*AE*ST. Baseline spasticity (BS) was included as a covariate in all models. The same stepwise model selection process described above (with the term PC replaced by AE) was completed.
The interpretation of these analyses would be that if the occurrence of one of the three most frequent CNS AEs has no association with the assessment of efficacy, this would be evidenced by no statistically significant TG*AE interaction. Thus, there would be no evidence to suggest that the treatment difference is itself different for patients who had experienced at least one of these events, that is, even if patients experiencing these events realise which treatment they are taking, it is not likely to have led to bias in the assessment of efficacy.
Dose of test medication: prior v naïve cannabis users
In each of the studies, participants titrated to their own preferred daily dose (number of sprays) of study medication. This was done by titrating the doses until satisfactory symptom relief was obtained or until unwanted side-effects occurred. If prior use of cannabis allowed patients to determine which medication they have been given then it might be expected that their titrated dose would be different from cannabis naïve patients; for example, in the placebo treatment group patients with previous exposure to cannabis might take fewer doses than the cannabis naïve patients if they realised that they were on a placebo (futility), especially given that the alcohol in the excipients is known to cause oral mucosal stinging in some study participants.
A general linear model was used exactly as for prior cannabis use in relation to efficacy with the exceptions that the dependent variable was to be the patients’ mean sprays per day of test medication during the studies and there was no baseline covariate. The interpretation of these analyses would be that if prior cannabis exposure has no association with the sprays of test medication used, as evidenced by no statistically significant TG*PC interaction, then any unblinding as a result of prior exposure to cannabis is not likely to have led to a different dosing pattern between prior exposure to cannabis and naïve patients.
In all of these analyses no formal adjustment was made for the multiplicity involved in considering such a large number of analyses, with a 5% significance level to be used for each individual analysis. This multiplicity needs to be considered when interpreting the above results.