1Carl-Fredrik Burman, 11 Nov 2008 RSS / MRC / NIHR HTA Futility Meeting Futility stopping Carl-Fredrik Burman, PhD Statistical Science Director AstraZeneca R&D
Carl-Fredrik Burman, 11 Nov 2008 RSS / MRC / NIHR HTA Futility Meeting 2 Stakeholder perspectives The patient A pharmaceutical company The public (MRC, NIHR)
Carl-Fredrik Burman, 11 Nov 2008 RSS / MRC / NIHR HTA Futility Meeting 3 The fundamental design requirement: Ethics ”My old mother – principle” The trial is ethical if (and only if) I would recommend my mother to take part in the trial, given that she would be eligible.
Carl-Fredrik Burman, 11 Nov 2008 RSS / MRC / NIHR HTA Futility Meeting 4 Interim stopping Stop the trial as soon as I would not include my mother, e.g. if One (publicly available) treatment is clearly better A “new” treatment fails to show sufficient effect, when it has known safety disadvantages No ethical obligation to stop If two treatments with similar safety have no clear difference in effect
Carl-Fredrik Burman, 11 Nov 2008 RSS / MRC / NIHR HTA Futility Meeting 5 (Genuine) informed consent The patient should get Full information regarding the trial treatments (and procedures), including previous data, potential risks, etc. Help to understand the information and Apply it to his/her specific situation (health status, preferences) When would a fully informed, fully competent patient give consent? If and only if it is better (not worse) for him/her to take part in the trial, as compared to receiving standard therapy. Cf. “my old mother” principle
Carl-Fredrik Burman, 11 Nov 2008 RSS / MRC / NIHR HTA Futility Meeting 6 Easy-going clinical equipose is not enough Clinical equipose If there is uncertainty about which treatment is better (Alternatively, compelling evidence of one treatment being better) (Alternatively, medical experts disagree) It’s far too easy to say that we are uncertain I expect my doctor to say what he believes is best
Carl-Fredrik Burman, 11 Nov 2008 RSS / MRC / NIHR HTA Futility Meeting 7 Our old Mother Scientific equipose Not every expert agree on CO 2 -induced global warming Do you suggest a randomised N-of-1 trial? Of course not — choose the treatment we believe is best Earth
Carl-Fredrik Burman, 11 Nov 2008 RSS / MRC / NIHR HTA Futility Meeting 8 What is ”best” for the patient? May depend on e.g. Effect (best guess + uncertainty) Safety Better care in the trial? Economic compensation (but beware of exploitation) Altruism Likely effect will differ between individuals (covariates) Preferences are different Decision theory may help decide (at least in theory …)
Carl-Fredrik Burman, 11 Nov 2008 RSS / MRC / NIHR HTA Futility Meeting 9 Decision analysis (DA) Patient perspective Utility function U(effect, safety, QoL, cost, …) Model for effect, safety, etc., based on best information (data, expert knowledge, …). Often Bayesian prior. Choose decision (volunteer to participate in trial, or not) to maximise expected utility The DA approach can also be used by a trial sponsor
Carl-Fredrik Burman, 11 Nov 2008 RSS / MRC / NIHR HTA Futility Meeting 10 A pharmaceutical company perspective (simplified) A new drug will be licensed if and only if the (next) phase III trial has a statistically significant effect (p<5%) If licensed, the company will make a profit of V (unit: £) The trial cost is k·N, where N is the sample size The assumed (believed) treatment effect is . Maximise V · Power(N) – k · N Of course, this model is wrong (as all models are). Should e.g. have V=V(T)=V(T(N)), where T is time.
Carl-Fredrik Burman, 11 Nov 2008 RSS / MRC / NIHR HTA Futility Meeting 11 Gain Net gain = Gain – Cost Cost Optimal sample size N opt = 1010
Carl-Fredrik Burman, 11 Nov 2008 RSS / MRC / NIHR HTA Futility Meeting 12 The interim decision (continue vs. stop for futility) Value V if significant Conditional power CP if trial is continued C additional trial cost if continued (compared to if stopped) Continue iffV · CP > C, that is, iff CP > C / V
Carl-Fredrik Burman, 11 Nov 2008 RSS / MRC / NIHR HTA Futility Meeting 13 DA vs. ”least clinically relevant” effect DA approach: Maximise expected utility based on ”best guess” effect (or prior) Traditional approach: 90% power at ”least clinically relevant” effect What is the least clinically relevant effect? If no adverse effects, no cost And the outcome is death One single saved life is clinically relevant … at least to the one saved What is a relevant effect depends on safety, cost etc.
Carl-Fredrik Burman, 11 Nov 2008 RSS / MRC / NIHR HTA Futility Meeting 14 Conditional power at interim Final estimate is N( , 1/N). Stage i has sample size N i and estimate i. Then = (N 1 · 1 +N 2 · 2 )/N Statistical significance if > C / N (where C=1.96 say) CP = P( > C / N ) = ( · N 2 + 1 ·N 1 / N 2 -C (N/N 2 ) ) But which to use when calculating CP? Original alternative Alternative ? Interim estimate 1 ? Linear combination of 1 and Alternative ? Bayesian posterior based on interim data ?
Carl-Fredrik Burman, 11 Nov 2008 RSS / MRC / NIHR HTA Futility Meeting 15 Stop, continue, or something else? Run a new trial? Sample size reestimation, based on interim estimates Flexible design methodology (Bauer & Köhne –94) Predefined weights for the different stages (generally, weight not proportional to information) May change the sample size for stage 2 after viewing interim results Discussion on CP Somewhat controversial May be better than design with only futility stopping Group-sequential designs should often be preferred
Carl-Fredrik Burman, 11 Nov 2008 RSS / MRC / NIHR HTA Futility Meeting 16 Publicly funded trial: Treatments with similar safety Assume Whole patient population will receive one of these treatments Efficacy is the only unknown Same safety, cost, etc. The closer the interim effect is to zero, the more value in continuing Thus, no reason to stop for futility
Carl-Fredrik Burman, 11 Nov 2008 RSS / MRC / NIHR HTA Futility Meeting 17 Example 1: Value of information Compare 2 treatments with probabilities p A, p B for death. Assume total future population size is T (10,000 say) If we knew that , we would choose treatment A T· lives would be spared as compared to using B Similarily, choose B if <0 Net value T·Abs( ) or T·Abs( )/2 if compared to using random treatment
Carl-Fredrik Burman, 11 Nov 2008 RSS / MRC / NIHR HTA Futility Meeting 18 Maximal value of information Before trial, p 2 p 1 has approximately normal prior with mean=0, SD= (say 10%) What would the value be if we could learn the exact value of ? Take the Bayesian expectation of the value T·Abs( )/2, E prior [T·Abs( )/2] = T· / (2 ) With T=10,000 and =10%, about 400 lives would be spared Example cont’d
Carl-Fredrik Burman, 11 Nov 2008 RSS / MRC / NIHR HTA Futility Meeting 19
Carl-Fredrik Burman, 11 Nov 2008 RSS / MRC / NIHR HTA Futility Meeting 20 Publicly funded trial: Intervention vs. no treatment (placebo) Assume Intervention is associated with some cost, safety risks Not clear whether intervention has a positive effect If effect, then the size of the effect will determine the size of the patient population which will get a positive net benefit First objective: is there any effect? Reasonable to stop for futility if interim estimate is low Expected value by continuing study is then small
Carl-Fredrik Burman, 11 Nov 2008 RSS / MRC / NIHR HTA Futility Meeting 21 Information leakage In regulatory setting, large discussion on who should see interim data Does the DMC have to be independent from the sponsor What are the risks of potential information leakage? Problems may be over-emphasised? The ethical aspect
Carl-Fredrik Burman, 11 Nov 2008 RSS / MRC / NIHR HTA Futility Meeting 22 Summary Futility stopping may be an ethical requirement Industry funded trials: Tradeoff cost and expected value Publicly funded trials (examples) Don’t stop for futility if two active treatments differ only in effect May stop for futility if “active” treatment unlikely to have sufficient effect (tradeoff cost and value) (If basic science objective …)