Bayesian Implementations in Drug Development: Why ...

17 downloads 0 Views 366KB Size Report
«Bayesian statistics is a great intellectual achievement, a great theory. But it's very hard to apply in ... (see also: Roger Hoerl. The World is Calling: Should We Answer? Deming. Lecture, JSM Miami ... Isaac Newton (1642-1727). ▫ A word to the ...
Bayesian Implementations in Drug Development: Why Progress is Slow ASA Joint Statistical Meetings, San Diego, 2012 Beat Neuenschwander Novartis Pharmaceuticals Basel, Switzerland

Agenda  Prélude  Statistical Thinking  Statistical Science  Statistical Engineering  Coda

2

| JSM - San Diego 2012 | Neuenschwander | Bayesian Implementations in Drug Development: Why Progress is Slow

Prélude  Bayesian statistics • A 250-year success story of practical and theoretical work • “Bayes: the theory that would not die” by Bertsch McGrayne 2011 - Bayes used for many highly relevant applied problems, e.g. • to crack the enigma code in word war II • for 1st US presidential election polls • ... + many more

 Bayes in drug development • Industry: increasingly used, mostly for internal decision making • Health authorities - Some acceptance at CDRH/CBER - FDA/Johns Hopkins Workshop in 2004 (Clinical Trials 2005)

- Not much progress since 2004

• Academia: very active research area 3

| JSM - San Diego 2012 | Neuenschwander | Bayesian Implementations in Drug Development: Why Progress is Slow

Prélude  Personal background • Trained as a mathematical (classical) statistician • 20 years of experience in Bayesian statistics (mainly self-taught) - 2 yrs in academia - 7 yrs in government (Swiss Federal Office of Public Health) - 11 yrs in industry

 The following thoughts • reflect my experience with applying the Bayesian approach to healthcare problems • may not necessarily represent the views of Novartis Pharmaceuticals

 For a more objective/representative view, see DIA Bayes survey (slide 19) 4

| JSM - San Diego 2012 | Neuenschwander | Bayesian Implementations in Drug Development: Why Progress is Slow

Statistics in Drug Development: 3 Layers Statistical Thinking - Statistical Science – Statistical Engineering

The following discussion will have 3 main layers

 Statistical Thinking • considerable uncertainty in many areas of drug development • understand these uncertainties and act adequately • all parties should be aware of the role and potential of statistics

 Statistical Science provides a rich source of know-how  Statistical Engineering • how to make best use of statistical thinking and statistical science in order to successfully solve applied problems (for a definition see slide 18) • Statistical enginieering provides the link! 5

| JSM - San Diego 2012 | Neuenschwander | Bayesian Implementations in Drug Development: Why Progress is Slow

Statistical Thinking Inference / Decisions

 Statistics in Drug Development • Lots of decisions (under uncertainty): informed decision making should be based on all relevant information • This is easier said than done

 Inference: what do we know? («Science») • (Incomplete) knowledge represented probabilistically: p-values, estimates, standard errors, confidence intervals, posterior distributions, predictive distributions

 Decisions: what should we do? («Policy») • Decision making should obviously be based on current knowledge (formal statistical inference) • Inference first, then decisions! • Decision making can be more or less formal 6

| JSM - San Diego 2012 | Neuenschwander | Bayesian Implementations in Drug Development: Why Progress is Slow

Statistical Thinking: P4 Four Perspectives • Operating Characteristics: for given truth , how often are we right? P( right decision |  ) - Phase I: for given toxicity profile, how often will we find an acceptable dose? - Phase II/III: under the null and and alternative parameters, how often will we reject the null hypothesis? (p-value, type-I and II error)

• Context: what does the context tell us about  ? P(  | context ) - Phase I: pre-clinical data, historical human data - Phase II: historical data on control treatment

• Evidence: what do we know about  ? P(  | evidence + context ) - Phase I: how confident are we that true toxicity rate is in target interval? - Phase II: how confident are we that true hazard ratio is