Bayesian adaptive clinical trials: Promises and pitfalls

Bayesian adaptive clinical trials:Promise and pitfalls

John D. Cook

March 30, 2016

Why adaptive clinical trials?

I Human subjects not like light bulbs.

I Have to proceed cautiously.

I Want to treat subjects effectively.

Why adaptive clinical trials?

I Human subjects not like light bulbs.

I Have to proceed cautiously.

I Want to treat subjects effectively.

Why adaptive clinical trials?

I Human subjects not like light bulbs.

I Have to proceed cautiously.

I Want to treat subjects effectively.

Objectives

I Maximize the probability/expectation of a good outcome

I Mimimize the probability/expectation of a bad outcome

I Stop early because things are looking bad

I Stop early because things are looking good (less common)

Objectives

I Maximize the probability/expectation of a good outcome

I Mimimize the probability/expectation of a bad outcome

I Stop early because things are looking bad

I Stop early because things are looking good (less common)

Objectives

I Maximize the probability/expectation of a good outcome

I Mimimize the probability/expectation of a bad outcome

I Stop early because things are looking bad

I Stop early because things are looking good (less common)

Objectives

I Maximize the probability/expectation of a good outcome

I Mimimize the probability/expectation of a bad outcome

I Stop early because things are looking bad

I Stop early because things are looking good (less common)

Why Bayesian?

I Flexibility

I Sequential nature of Bayes rule

I Can incorporate prior information

Why Bayesian?

I Flexibility

I Sequential nature of Bayes rule

I Can incorporate prior information

Why Bayesian?

I Flexibility

I Sequential nature of Bayes rule

I Can incorporate prior information

Traditional designs

I VERY crude, e.g. 3+3

I Infrequent monitoring

I Uses very little data

Traditional designs

I VERY crude, e.g. 3+3

I Infrequent monitoring

I Uses very little data

Traditional designs

I VERY crude, e.g. 3+3

I Infrequent monitoring

I Uses very little data

Opportunities for improvement

I Use continuous outcomes

I Monitor frequently or continuously

I Use patient characteristics

I Use multiple outcome events

Opportunities for improvement

I Use continuous outcomes

I Monitor frequently or continuously

I Use patient characteristics

I Use multiple outcome events

Opportunities for improvement

I Use continuous outcomes

I Monitor frequently or continuously

I Use patient characteristics

I Use multiple outcome events

Opportunities for improvement

I Use continuous outcomes

I Monitor frequently or continuously

I Use patient characteristics

I Use multiple outcome events

Dangers

I Fitting complex models with little data

I A priori overfittingI Exploring high dimensional design/outcome space:

I Arbitrary simulation scenariosI May have blind spotI Temptation to cherry pick scenarios

Dangers

I Fitting complex models with little data

I A priori overfitting

I Exploring high dimensional design/outcome space:I Arbitrary simulation scenariosI May have blind spotI Temptation to cherry pick scenarios

Dangers

I Fitting complex models with little data

I A priori overfittingI Exploring high dimensional design/outcome space:

I Arbitrary simulation scenariosI May have blind spotI Temptation to cherry pick scenarios

Dangers

I Fitting complex models with little data

I A priori overfittingI Exploring high dimensional design/outcome space:

I Arbitrary simulation scenarios

I May have blind spotI Temptation to cherry pick scenarios

Dangers

I Fitting complex models with little data

I A priori overfittingI Exploring high dimensional design/outcome space:

I Arbitrary simulation scenariosI May have blind spot

I Temptation to cherry pick scenarios

Dangers

I Fitting complex models with little data

I A priori overfittingI Exploring high dimensional design/outcome space:

I Arbitrary simulation scenariosI May have blind spotI Temptation to cherry pick scenarios

Contact info