Sample size determination for cost-effectiveness trials Anne Whitehead Medical and Pharmaceutical Statistics Research Unit The University of Reading MPS Research Unit CHEBS Workshop - April 2003 1 Comparative study • Parallel group design • Control treatment (0) New treatment (1) • n0 subjects to receive control treatment n1 subjects to receive new treatment MPS Research Unit CHEBS Workshop - April 2003 2 Measure of treatment difference Let be the measure of the advantage of new over control > 0 new better than control = 0 no difference < 0 new worse than control Consider frequentist, Bayesian and decision-theoretic approaches MPS Research Unit CHEBS Workshop - April 2003 3 1. Frequentist approach Focus on hypothesis testing and error rates - what might happen in repetitions of the trial e.g. Test null hypothesis against alternative H0 : = 0 H1+: > 0 Obtain p-value, estimate and confidence interval Conclude that new is better than control if the one-sided p-value is less than or equal to Fix P(conclude new is better than control | = R) = 1– MPS Research Unit CHEBS Workshop - April 2003 4 Distribution of ̂ =0 = R k Fail to Reject H0 MPS Research Unit Reject H0 CHEBS Workshop - April 2003 5 A general parametric approach Assume ˆ ~ N , 1 w Reject H0 if ̂ > k P(ˆ k | 0) 1 k w k z w where is the standard normal distribution function and P(Z > z) = where Z ~ N(0, 1) MPS Research Unit CHEBS Workshop - April 2003 6 Require P(ˆ k | R ) 1 1 (k R ) w 1 1 R k w (R k) w z z z w R MPS Research Unit 2 CHEBS Workshop - April 2003 7 Application to cost-effectiveness trials Briggs and Tambour (1998) = k (E1 – E0) – (C1 – C0) is the net benefit, where E1, E0 are mean values for efficacy for new and control treatments C1, C0 are mean costs for new and control treatments k MPS Research Unit is the amount that can be paid for a unit improvement in efficacy for a single patient CHEBS Workshop - April 2003 8 ˆ k(x E1 x E0 ) (x C1 x C0 ) E1 E 0 var(x E1 x E0 ) n1 n0 2 2 C1 C 0 var(x C1 x C0 ) n1 n0 2 2 cov (x E1 x E0 ),(x C1 x C0 ) var(x E1 x E0 )var(x C1 x C0 ) Set z z 1 w ˆ var() R 2 and solve for n0 and n1 MPS Research Unit CHEBS Workshop - April 2003 9 2. Bayesian approach Treat parameters as random variables Incorporate prior information Inference via posterior distribution for parameters Obtain estimate and credibility interval Conclude that new is better than control if P( > 0|data) > 1 – Fix P0 (conclude new better than control) = 1 – MPS Research Unit CHEBS Workshop - April 2003 10 ˆ ~ N , 1/w Likelihood function f (ˆ | ) exp w(ˆ ) 2 / 2 N 0 , 1/w 0 Prior h0() is Posterior h(|data) exp{w(ˆ )2 / 2}exp{w 0 ( 0 ) 2 / 2} exp (ˆ w 0 w 0 ) 2 w w 0 / 2 i.e. h(|data) is MPS Research Unit 0 w 0 ˆ w N , w0 w 1 w0 w CHEBS Workshop - April 2003 11 P ( > 0|data) > 1 – if 0 w 0 ˆ w w0 w z w0 w i.e. 0 w 0 ˆ w z w 0 w i.e. ˆ w 0 w 0 z w 0 w MPS Research Unit CHEBS Workshop - April 2003 12 Prior to conducting the study, ˆ ~ N , 1 w 1 ~ N 0 , w 0 so ˆ ~ , 1 1 0 w w 0 2 w ˆw ~ w, w 0 w 0 MPS Research Unit CHEBS Workshop - April 2003 13 Require P0 ˆ w 0 w 0 z w 0 w 1 w z w w w 0 0 1 1 0 0 2 w w / w0 (w w) z w w 0 1 0 0 2 w w / w 0 0 (w 0 w) z w 0 w z w w 2 / w 0 0 (w 0 w) z w 0 w z w w 2 w 0 Express w in terms of n0 and n1, provide values for 0 and w0 and solve for n0 and n1 MPS Research Unit CHEBS Workshop - April 2003 14 Application to cost-effectiveness trials O’Hagan and Stevens (2001) = k (E1 – E0) – (C1 – C0) ˆ k(x E1 x E0 ) (x C1 x C0 ) Use multivariate normal distribution for x E1 , x E0 , x C1 , x C0 - separate correlations between efficacy and cost for each treatment Allow different prior distributions for the design stage (slide13) and the analysis stage (slide 11) MPS Research Unit CHEBS Workshop - April 2003 15 3. Decision-theoretic approach Based on Bayesian paradigm Appropriate when outcome is a decision Explicitly model costs and benefits from possible actions Incorporate prior information Choose action which maximises expected gain MPS Research Unit CHEBS Workshop - April 2003 16 Actions Undertake study and collect w units of information on , then one of the following actions is taken: Action 0 : Abandon new treatment Action 1 : Use new treatment thereafter MPS Research Unit CHEBS Workshop - April 2003 17 Table of gains (relative to continuing with control treatment) 0 >0 Action 1 – cw – b – cw – b + r1 Action 0 – cw – cw c = cost of collecting 1 unit of information (w) b = further development cost of new treatment r1 = reward if new treatment is better G0,w() = – cw G1,w () cw b r1I( 0) MPS Research Unit CHEBS Workshop - April 2003 18 Following collection of w units of information, the expected gain from action a is Ga, w(x) = E {Ga,w()| x} Action will be taken to maximise E {Ga,w()|x}, that is a*, w* where Ga*, w*(x) = max {Ga, w(x)} (Note: Action 1 will be taken if P ( > 0|data) > b/r1) MPS Research Unit CHEBS Workshop - April 2003 19 At design stage consider frequentist expectation: E (Ga*, w(x)) G a*,w (x) f (x; , w) dx x and use this as the gain function Uw () MPS Research Unit CHEBS Workshop - April 2003 20 Expected gain from collecting information w is U w E0{U w ()} U w ()h 0 ()d So optimal choice of w is w*, where Uw* max {Uw } w MPS Research Unit CHEBS Workshop - April 2003 21 U w* max E0 E max E (G a,w () | x) w a max w max x a G a,w ()h( | x)d f (x; , w)dx h 0 ()d This is the prior expected utility or pre-posterior gain MPS Research Unit CHEBS Workshop - April 2003 22 Note: E max E (G a,w () | x) a = E{– cw + max(r1 P ( > 0|data) – b, 0)} MPS Research Unit CHEBS Workshop - April 2003 23 Application to cost-effectiveness trials Could apply the general decision-theoretic approach taking q to be the net benefit The decision-theoretic approach appears to be ideal for this setting, but does require the specification of an appropriate prior and gain function MPS Research Unit CHEBS Workshop - April 2003 24 Table of gains – ‘Simple Societal’ (relative to continuing with control treatment) 0 >0 Action 1 – cw – b – cw – b + r1 Action 0 – cw – cw c = cost of collecting 1 unit of information (w) b = further development cost of new treatment r1 = reward if new treatment is more cost-effective G0,w() = – cw G1,w () cw b r1I( 0) MPS Research Unit CHEBS Workshop - April 2003 25 Gains – ‘Proportional Societal’ (relative to continuing with control treatment) c = cost of collecting 1 unit of information (w) b = further development cost of new treatment r2 = reward if new treatment is more cost-effective G0,w() = – cw G1,w () cw b r2 MPS Research Unit CHEBS Workshop - April 2003 26 Gains – ‘Pharmaceutical Company’ c = cost of collecting 1 unit of information (w) b = further development cost of new treatment r3 = reward if new treatment is more cost-effective G 0,w (x) cw G1,w (x) cw b r3I(x A) where A is the set of outcomes which leads to Action 1, e.g. for which P ( > 0|data) > 1 – MPS Research Unit CHEBS Workshop - April 2003 27 References Briggs, A. and Tambour, M. (1998). The design and analysis of stochastic cost-effectiveness studies for the evaluation of health care interventions (Working Paper series in Economics and Finance No. 234). Stockholm, Sweden: Stockholm School of Economics. O’Hagan, A. and Stevens, J. W. (2001). Bayesian assessment of sample size for clinical trials of cost-effectiveness. Medical Decision Making, 21, 219-230. MPS Research Unit CHEBS Workshop - April 2003 28