Reinterpreting Clinical Trials in Children With Multiple Sclerosis Using a Bayesian Approach

Abstract

This systematic review and meta-analysis investigates the use of the bayesian approach to estimate the plausible effect of teriflunomide in children with multiple sclerosis.

According to standard statistical approaches and assuming a similar magnitude of effect, clinical trials in children with multiple sclerosis (MS) should be the same size as trials in adults. Given the rarity of the disease in children, such trials would require decades to complete, unless unrealistic treatment effects are postulated, thereby increasing the risk of missing moderate yet clinically worthwhile effects.¹ Bayesian approaches² may overcome this problem because they consider the fact that once the efficacy of a drug has been established in adults, the hypothesis that it is also efficacious in children becomes more plausible.

We applied a bayesian approach for estimating the plausible effect of teriflunomide in children with MS, where the results of the Safety and Efficacy of Teriflunomide vs Placebo in Paediatric Multiple Sclerosis (TERIKIDS) study, a negative trial assessing teriflunomide in children (57 placebo vs 109 teriflunomide), was integrated with the available knowledge on teriflunomide in adults.

Methods

We used published data from 2 randomized clinical trials testing teriflunomide (14 mg) in adult patients with MS (TEMSO³: 363 placebo vs 359 teriflunomide; TOWER⁴: 389 placebo vs 372 teriflunomide) to build the prior probability distribution of treatment effect. We pooled hazard ratios (HRs) and 95% CIs on time-to-first relapse (log scale) by inverse of variance weighting. Teriflunomide effect in children was assessed in the TERIKIDS trial.⁵ The log (HR) values were assumed to be normally distributed. We applied Bayes rule¹ to obtain the distribution of plausible effects of teriflunomide in children with MS (posterior distribution). To account for differences between the adult and the pediatric populations⁶ and between some details of the study designs,^3,4,5 we down-weighted prior distributions by 50% or 75%. This study followed the Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) reporting guidelines.

Results

The observed HRs of teriflunomide on time-to-first relapse in TEMSO, TOWER, and in TERIKIDS were 0.72 (95% CI, 0.58-0.90), 0.63 (95% CI, 0.50-0.79), and 0.66 (95% CI, 0.39-1.11), respectively (Table).^3,4,5 The prior distribution obtained by pooling the results of the 2 trials in adults was centered at HR 0.68 (95% CI, 0.58-0.79).

Table. Results of the Posterior Distribution for the Effect of Teriflunomide in Children^a .

Source	HR (95% CI)	log (HR)	SDb
TEMSO	0.72 (0.58-0.90)	−0.33	0.11
TOWER	0.63 (0.50-0.79)	−0.46	0.12
Prior distribution	0.68 (0.58-0.79)	−0.39	0.08
TERIKIDS (likelihood)	0.66 (0.39-1.11)	−0.41	0.27
Posterior distribution, HR (95% credible interval)
Weight for prior = 50%	0.67 (0.51-0.87)	−0.39	0.14
Weight for prior = 25%	0.67 (0.44-0.99)	−0.40	0.21

Abbreviations: HR, hazard ratio; TEMSO, Randomized Trial of Oral Teriflunomide for Relapsing Multiple Sclerosis; TERIKIDS, Safety and Efficacy of Teriflunomide vs Placebo in Paediatric Multiple Sclerosis; TOWER, Oral Teriflunomide for Patients With Relapsing Multiple Sclerosis.

^a Obtained by combining the prior probability distribution of teriflunomide effect on adults from the pooled estimate of the TOWER and TEMSO trials and the TERIKIDS results.

^b SD of the prior distribution (equal to SE of the estimated treatment effect, μ:log [HR]).

The posterior estimate for efficacy obtained by combining the prior distribution, weighted at 50%, with TERIKIDS results was HR, 0.67 (95% credible intervals, 0.51-0.87) with a probability of 0.0027 that teriflunomide has no effect in children. With a weight of 25%, the posterior HR estimate was 0.67 (95% credible intervals, 0.44-0.99), and the probability of no effect was 0.0287 (Table and Figure).

Figure. Probability Density Functions for the Prior Distribution, the Likelihood, and the Posterior Distribution for the Teriflunomide Effect in Children.

The dashed black lines represent the credible intervals bands and the dotted black line represents treatment inefficacy (hazard ratio = 1).

Discussion

The 34% reduction in the incidence of relapses observed in the teriflunomide treatment group of the TERIKIDS trial failed to achieve statistical significance; therefore, according to standard frequentist approaches, it did not provide evidence in support of efficacy in children. However, no compelling biological or clinical reasons indicate that evidence obtained in adults should be ignored when deciding treatment strategies for pediatric MS. Whenever trials of the same drug have been conducted in both adults and children, a bayesian approach may allow clinicians to integrate results from trials in adults in the interpretation of pediatric trials. In this case, results of our analysis suggest compelling evidence of the efficacy of the drug in children, with the most plausible association an approximately one-third reduction in relapses; the probability that teriflunomide has no treatment effect in children was very low (<3% or even <3/1000). Two possible criticisms of the use of the bayesian framework in this setting are (1) the subjectivity involved in the generation of the prior distributions (which in this case were straightforward and derived from a standard meta-analysis) and (2) extrapolation when translating results from adults to children (children are not just small adults). To account for the possibility that results observed in adults cannot be directly applied to children, 2 down-weighting factors⁶ were applied (50% and 25%) to limit its influence on the results of the analyses with little effect on conclusions.