This is a research summary of a study that found one out of every five cancer trials that “did not finish” which actually means that they stopped early for futility, if I am reading between the lines properly. Of those studies, 40% stopped early because of poor accrual.

Cancer.net. One in Five Clinical Trials for Adults with Cancer Never Finish – New Study Examines the Reasons. Genitourinary Cancers Symposium. January 28, 2014. Available at http://www.cancer.net/one-five-clinical-trials-adults-cancer-never-finish-%E2%80%93-new-study-examines-reasons.

There was some recent discussion of issues with multi-center trials where one center dominates, contributing as much as 94% of all the patients. What does this do to the generalizability of the study. I wanted to summarize these comments here, because it relates to some of the issues I’m looking at right now in accrual models for multi-center trials.

One person shared the transcript of an FDA review of a drug that was tested in 41 patients, 38 of who came from Brigham & Women’s Hospital with the remaining 3 coming from a hospital associated with the University of Nebraska. As you might expect, the FDA panel was not thrilled with the ability to generalize from this study to the overall population, but they did acknowledge that the patients themselves were geographically dispersed and came to Brigham & Women’s Hospital because this was a rare condition that few places were qualified to offer treatment.

Another person suggested a 2.5 to 1 ratio maximum, meaning that if you had K centers, no more than 2.5 / K of them could come from a single center. For example, with 10 centers, no more that 2.5/10 = 25% should come from any one center. For pivotal studies you might want to consider a stricter limit like 2 / K. This is a very ad hoc approach of course, but you could run some simulations using an influence function and a score function to see how much the imbalance hurts. I can’t say that I completely follow this last comment.

Another person shared a guidance document CMPP/EWP/2330/99 which discusses meta-analyses and relying on a single pivotal trial. This guidance offer the general admonition that none of the centers in a multi-center trial should dominate the study either in the number of subjects or in the magnitude of the effect. The latter means, I presume that if the efficacy is a result of positive results mostly from just one of the centers, that is a serious concern.

Addendum: I did a quick google search and found a discussion on enrollment caps in a multicenter trial on the MEDSTATS forum on January 28, 2015. One reply had a link to an article in the Journal of the American College of Cardiology that argued that, at least in one clinical trial, high enrolling sites differed in important ways from low enrolling sites. This article has an accompanying editorial.

Another search found a Johns Hopkins report that discussed (among other things) the impact of enrollment caps on individual sites. And a 2007 blog entry suggests a cap of 3/K to 5/K.

There are lots of people that know a lot more about MCMC than I do. The only thing I can offer in my defense is that I know so little about MCMC that everything I do know qualifies as non-technical.

You should take a look at the examples in the BUGS manual. Most of these are simple illustrative examples that are very easy to follow.

I would also encourage you to look at some of the Bayesian MCMC approaches for missing data. Often we can make somewhat informed guesses about what a missing data point might be because, for example, males have a different pattern of responding than females. You need, however, to simulate missing data in a way that doesn’t overstate the certainty of this pattern. MCMC properly incorporates all of the uncertainties associated with predicting what a missing value might have been.

Another example that might be use of latent variables. Ordinal data, for example, can be thought of as cut-points of an underlying continuous variable. MCMC allows you simulate what this underlying continuous variable might be.

A paper that I co-authored in Statistics in Medicine (SIM) might be an interesting example. We were looking for predictive models for the duration of a clinical trial. Many clinical trials last a lot longer than expected. When you learn that the recruitment of subjects is a lot slower than expected, you need to revise your estimate of how long it will take for you to finish your research study.

Our solution, which has been worked on by others as well, involves a specification of a prior distribution (your estimate of the accrual rate prior to data collection) and combining that with actual accrual data as you progress with your study. If this study is one of many similar ones that you have conducted in the past, you will have provided a strong prior distribution, and you won’t overreact to a bit of early bad news. On the other hand, if this is a fairly new setting for you, and your prior is weak, you want to be sensitive to early warnings of problems.

The solution that we offer in the SIM paper does not really require MCMC because it has a closed form solution. But several extensions that we are looking at would require MCMC.

The key reference for these ideas is Gajewski BJ, Simon SD, Carlson SE. Predicting accrual in clinical trials with Bayesian posterior predictive distributions. Stat Med. 2008 Jun 15;27(13):2328-40 which is available on the web at  http://onlinelibrary.wiley.com/doi/10.1002/sim.3128/abstract

I have other material related to the Bayesian modelling of patient accrual at http://www.pmean.com/category/AccrualProblems.html.

Slow accrual of patients in clinical trials is the most critical factor hampering the completion of clinical trials and represents significant threat to the integrity of research. Delays in the completion of trials damage our ability to evaluate novel therapies and interventions in a timely fashion and undermine the Nuremberg Code’s mandate that research should “yield fruitful results for the good of society.” In collaboration with researchers at Kansas University Medical Center, Steve Simon has helped develop a Bayesian model for patient accrual that incorporates historical information from previous trials. As the trial progresses, the Bayesian model updates the predicted probability of successful completion of the trial on time with the recommended number of subjects. The model also allows researchers to assess the probability of meeting a “fallback” option which lengthens the trial duration, decreases the proposed sample size, or possibly both.

We’re currently seeking NIH funding to extend the Bayesian model and to write accessible and easy to use software for accrual that can run on a website, a stand-alone computer, or a smartphone/tablet.

http://blog.pmean.com/tag/accrual/

There are additional posting about accrual problems in clinical trials at my old website:

http://www.pmean.com/category/AccrualProblems.html

You can find other archive posts at

http://blog.pmean.com/category/archive/