The theorem works, in part, by considering the background probability of whatever's being investigated.

For instance, at Los Alamos National Laboratory in Los Alamos, New Mexico, researchers use Bayesian methods to find out if someone has been exposed to radioactive plutonium on the job. Because these exposures almost never happen, even if a lab worker tests positive for exposure on a urine test, Bayesian analyses would suggest it's more likely due to a test error than actual contact with plutonium, unless the test results are unusually strong.


How does this help medical device trials? Elisabeth Crowe, co-author of the paper on Bayes, says it can help streamline clinical investigations. If the currently available data indicate an acceptable level of certainty of the device's efficacy and safety, the study can be ended sooner. This means recruiting fewer subjects and saving money.

And perhaps more important, it's more flexible.

"The limitations [of traditional studies] are once you planned your trial and started collecting data, you can't change the way you're analyzing depending on what you're learning," she tells DOTmed News. But with Bayes, companies can alter trial plans on the fly to adapt to new information.

And if the prior evidence was weak, that's OK, too. The Bayesian model "self-corrects."

"The model borrows strength from the prior information, and one gets to a conclusion much quicker, but if it looks like the current data are out of sync with the prior, these Bayesian hierarchical models figure that out, and in the process of figuring that out, end up requesting more patients -- basically, to try to figure out why the current data and the prior information are disagreeing," Dr. George Campbell, the head of the FDA's biostatistics division, tells DOTmed.

CAMBRIDGE CONSULTANTS' CONCERNS

Though the FDA believes Bayesian methods could be more cost-effective for some trials, they still require some traditional frequentist-type reporting, even for companies running Bayesian trials, a policy Sewell and Crowe want to see changed.

In essence, they want the FDA to reconsider requiring type 1 error reporting even in Bayesian models. Type 1 errors are "false positive" errors. That is, for medical device testing, a type 1 error would be a study suggesting the device works when it actually doesn't. (Or rather, it passes the statistical tests, even though it's all due to chance.) Typically, these error reports are not relevant for Bayesian models, Crowe says.

"The type 1 error rate is not the appropriate calculation for safeguarding the public; rather, it is the probability being high that the device is good given the data, a question which the Bayesian model answers directly," argue Crowe and Sewell.

More Industry Headlines