Equivalence and Bioequivalence: Frequentist and Bayesian views on sample size Mike Campbell ScHARR CHEBS FOCUS fortnight 1/04/03 1 Equivalence • Many trials are not designed to prove differences but equivalences Examples : generic drug vs established drug Video vs psychiatrist NHS Direct vs GP Costs of two treatments Alternatively – non-inferiority (one-sided) 2 Efficacy vs cost • For some trials (e.g. of generics) one would like to show similar efficacy at less cost • Thus can have an equivalence and a cost difference trial in one study 3 Motivating example • AHEAD (Health Economics And Depression) • Trial of trycyclics, SSRIs and lofepramine • Clinical outcome - depression free months • Economic outcome – cost • Powered to show equivalence to within 5% with 90% power and 5% significance (estimated effect size 0.3 and SD 1.0) 4 Bio-equivalence (diversion) • For bio-equivalence we are trying to show that two therapies have same action • Usually compare serum profiles by e.g. AUC • Often paired studies • FDA: 80:20 rule 80% power to detect 20% difference 5 Frequentist view • Impossible to prove null hypothesis • All we can do is show that differences are at most Δ • Choose Δ to be a difference within which treatments deemed equivalent • General approach – perform two one-sided significance tests of H0: μ1-μ2> Δ and μ1-μ2< -Δ If both are significant, then can conclude equivalence 6 Figure from Jones et al (BMJ 1996) showing relationship between equivalence and confidence intervals 7 Assume sd is known. CIs Then 100(1-)% CI to compare difference in means is d z / 2 , d z / 2 where (n11 n21 )1 / 2 Treatments deemed equivalent if interval falls in -Δ to +Δ 8 Let τ = μA - μB H 0 :| | vs H 1 : | | Let η = Δ - zα/2 σλ Pr( d ( , )) ( z / 2 ) ( z / 2 ) (1) Maximal Type I error rate is (1) when τ=Δ Power is defined when (1) has τ=0 1 2 ( z / 2 ) 1 or 1 / 2 z / 2 9 If treatment groups have same size, n, then required sample size is n2 2 2 (za / 2 z / 2 )2 (2) This is similar to testing for a difference Except i) Usually Δ is smaller than for a difference trial ii) We use β/2 rather than β 10 Problems with equivalence trials • Poor trials (e.g. poor compliance and larger measurement errors bias trial towards null) • Jones et al (1996) suggest using an ITT approach and ‘per-protocol’ and hope they give similar results! 11 Bayesian sample size (O’Hagan and Stevens 2001) • Analysis objective Outcome is positive if the data obtained are such that there is a posterior probability of at least ω that τ >0 • Design objective We require the sample size (n1,n2) be large enough so there is a probability of at least ψ of obtaining a positive result. The probability ψ is known as the assurance 12 Bayesian assumptions • Let prior expectation of (μ1,μ2)T be ma according to analysis prior and md according to the design prior • Let variances be Va and Vd for analysis and design priors respectively • Let x be ( x1 , x2 )T, the observed data • Let S be the sampling variance matrix (note this depends on n1 and n2) 13 Let Wa =Va-1 ,Ws=S-1 and V*=(Wa+Ws)-1 and a=(1,-1)T Under analysis prior Posterior mean of (μ1, μ2) is Normally distributed with expectation and variance a T V * {Wa m a Ws x) T * a V a 14 Under design prior Unconditional distribution of x is Normal with mean and variance E ( x) m d ,Var (x) Vd S From which can get sample size calculation (See O Hagan and Stevens) 15 Frequentist interpretation If Va1 Wa 0 and Vd 0 then the Bayesian methods for determining sample size agree with frequentist If If Va-1 =0 – weak analysis prior –’vague’ prior Vd=0 - strong design prior 16 Bayesian equivalence (after O’Hagan and Stevens(2001) • Analysis objective: Outcome of study is positive if the upper limit of the (1-ω)% prediction interval for τ is < Δ (one sided) or upper and lower limits of prediction interval for τ are within ± Δ (two sided). • Design objective: Sample size is such that there is a probability of at least ψ of obtaining a positive result. 17 A modification of O’Hagan and Stevens suggests that for equivalence trials, a positive outcome occurs when | a T V * {Wa m a Ws x) | z1 a T V * a Two-sided and a T V * {Wa m a Ws x) z1 a T V * a One-sided Sample size also a modification of O’Hagan and Stevens 18 Parameters for non-inferiority ma - the analysis prior mean could be 0 md - the design prior mean could be 0 19 What if md and Vd>0 ? A weak design prior Then we have some information about the possible differences, so ‘proving’ the null hypothesis is difficult E.g. if we were 50% sure that δ>0, before the trial then cannot be 80% sure that δ=0 after the trial 20 What if Va-1>0? A strong analysis prior CIs will be shifted towards ma If ma=0, then probability of a positive event increased 95% CIs will be narrower than for the frequentist approach 21 Conclusions • Bayesian approach more natural for equivalence (Can prove H0) • More work on getting pragmatic suggestions for Va and Vd needed 22