Software for conducting the simulations in "Using short-term response information to facilitate adaptive randomization for survival clinical trials" by Xuelin Huang, Jing Ning, Yisheng Li, Elihu Estey, Jean-Pierre Issa and Donald A. Berry in Statistics in Medicine 28(12): 1680-1689, 2009.

Download

The BMA-CRM Simulator is an easy-to-use implementation of the BMA-CRM dose-finding method Bayesian Model Averaging Continual Reassessment Method by Guosheng Yin and Ying Yuan.

Download

This is an R program that computes the effective sample size of a parametric prior, as described in the paper “Determining the Effective Sample Size of a Parametric Prior” by Morita, Thall, and Muller (Biometrics 64, 595-602, 2008).

Download

Software for monitoring clinical trials with either binary or time-to-event endpoints using Bayesian hypothesis testing based on Bayes factors. This software implements the clinical trial monitoring method proposed by Valen Johnson and John D. Cook in “Bayesian Design of Single-Arm Phase II Clinical Trials with Continuous Monitoring,” to appear in Clinical Trials.

Download

Estimation and inference under semi-parametric proportional density model

The software includes parameter estimation and inference on treatment effects under the semi-parametric proportional density model, and also includes a goodness-of-fit test for model checking purposes. Details of the methods can be found in the paper “Inference of tamosifen's effects on prevention of breast cancer from a randomized controlled trial” by Yu Shen, Jing Qin, and Joseph Costantino (2007 Journal of the American Statistical Association, Vol. 102: 1235-1244).

Phase I/II dose-pair-finding based on utilities of 3-level ordinal efficacy and toxicity

This is a menu-driven computer program for implementing the dose-finding method described in detail in the paper “Utility-Based Optimization of Combination Therapy Using Ordinal Toxicity and Efficacy in Phase I/II Trials” by Houede, Thall, Nguyen, Paoletti, and Kramar (2009).

This software is for designing and conducting Phase I clinical trials that simultaneously optimizes both dose and schedule. The goal is to determine a maximum-tolerated dose and schedule (MTDS) in terms of the overall risk of toxicity. The method is Bayesian adaptive and uses time-to-toxicity as the outcome.

The software is based on the paper Simultaneously optimizing dose and schedule of a new cytotoxic agent, Thomas M Braun, Peter F Thall, Hoang Nguyen, and Marcos de Lima, Clinical Trials 2007; 4: 113–124.

We've created a short video demonstrating some of the features of Inequality Calculator.

The video shows one feature in particular that many users do not know about: the ability to modify and save the output graph.

To obtain the latest and greatest version of OOMPA, use the following command in an R session:

install.packages(repos="http://bioinformatics.mdanderson.org/OOMPA")

New features:

1. Added "uncentered correlation" (as used in Eisen's cluster software) as another metric for distanceMatrix.
2. Improved the Mosaic class for producing red-green heatmaps.

There is now a Mananged C++ wrapper for DCDFLIB available from a Gene Harris.

A new version of Phase II PP is available for download. This new version, 1.2, corrects an error in the previous version.

This morning we posted a revised version of the CI of Interaction Index software correcting a minor mistake in the previous version.

The following diagram illustrates three ways of tuning adaptively randomized clinical trials.

In an equally randomized trial, each arm is assigned with exactly the same probability. With simple adaptive randomization, each arm is assigned with probability equal to the probability that that arm is the best arm, given the data seen so far. With myopic optimization, each patient is simply given what appears to be the best treatment; no randomization involved.

Our randomization software supports three methods for interpolating between the characteristics of equal randomization (ER) and simple adaptive randomization (SAR).

1. Beginning with burn-in period of equal randomization followed by SAR.
2. Using SAR but with a minimum randomization probability.
3. Using a value of the exponential tuning parameter between 0 and 1.

Adaptive randomization with a burn-in equal to the length of the trial is ER, and adaptive randomization with no burn-in is SAR. Setting the minimum randomization probability to 0.5 yields ER, but setting the minimum randomization probability to 0 yields SAR. Setting the exponential tuning parameter to 0 yields ER while setting it to 1 yields SAR. (As the tuning parameter goes past 1 to larger values, the design approaches myopic optimization.) So these three parameters give different ways of interpolating between ER and SAR. Each gives ER at one extreme and SAR at the other extreme. Between the extremes, the operating characteristics vary continuously in the design parameters.

See Comparing Methods of Tuning Adaptively Randomized Trials for examples of operating characteristics as each parameter varies.

Version 1.5 of CONFINT was released today. This program performs calculations of the probability that a specified one-group study (specification includes sample size) will produce a confidence interval (CI) of at most a particular length. It also calculates the sample size necessary to achieve a desired such probability. CONFINT15 will perform calculations for

• One-sample binomial
• One-sample Poisson
• Mean of the One-Sample Normal
• SD of the One-Sample Normal
• Hazard of One-Sample Exponential Survival Distribution
• Mean survival of the One-Sample Survival Distribution

We are pleased (and relieved) to announce the official release of Version 1.0 of the SuperCurve R package, part of the OOMPA suite of tools.

SuperCurve is a package for the analysis of reverse phase protein arrays. Among the interesting features:

1. Implements three different models for the intensity response to dilution curves
   • Joint logistic model, as developed here and independently by Tabus et al
   • Non-parametric model developed by Jianhua Hu
   • A loess-based model
2. Uses S4 classes to define an API to allow researchers to plug in other models
3. Allows for flexible descriptions of different array layouts or designs
4. Extensive graphical tools to diagnose potential problems
5. Includes a detailed vignette (thanks to Keith Baggerly) showing how to use the package
6. Includes the world's worst GUI. (We won't say who wrote it or designed it. It's only virtue is that it works and has actually been used by some of the biologists down at the Kleberg Center, who only complain that it doesn't have enough features.) We hope to get off the world's worst list in later releases.
7. Has a full regression testing suite to form a solid platform for future development.

IMPORTANT: This release only works with the latest version or R, which is 2.7.0. (As of this morning. That could change by this afternoon....)

To obtain the package, start R and then run the command

install.packages(repos="http://bioinformatics.mdanderson.org/OOMPA")

In the window that pops up, select both "SuperCurve" and "SuperCurveGUI"

NOTE: This release includes source versions (which can be compiled and installed on any LINUX or UNIX box running R 2.7) and Windows binary versions. It does not yet include Macintosh binary versions, but we are working on it.

Note that we already have some key feature requests near the top of the list, among which are

1. Add the topographical normalization and variable slope normalization methods developed in Shannon Neeley's thesis
2. Add more robust and/or more flexible methods for truncating "out-of-range" values
3. Incorporate more explicit QC reports

Thanks to Paul Roebuck, Shannon Neeley, Wenbin Liu, and Zhenlin Ju for their work in getting this release out the door.

The OOMPA (Object-Oriented Microarray and Proteomic Analysis) package has been upgraded to work with R version 2.7.0. Instructions for obtaining the package are available on the OOMPA web site.

The software package BP1CI was released today. This software calculates exact confidence limits on the probability of event for the one-sample binomial distribution or on the total number of events for the one-sample Poisson distribution. The software uses the Clopper-Pearson method. Includes source code, documentation, and executables for Windows and OS X.

A new version of WFMM (Wavelet-based Functional Mixed Models) is now available for download. This version minimizes the amount of RAM required to process large data sets which reduces the chance that the application will run out of memory. This version also fixes a problem in wfmm3 that caused an exception to be thrown.

A new version of the software for Predictive probability design for phase II studies was posted yesterday. The user's guide was also updated.

Adaptively randomized trials attempt to treat more patients effectively than equally randomized trials by tilting the randomization probabilities in favor of what appears to be the best treatment as each patient arrives. But what happens if the population changes during the course of the trial so that what used to be the best treatment is no longer the best? Will adaptive randomization learn that things have changed and still treat patients more effectively, or will it persist in assigning what used to be the best treatment? How will operating characteristics compare to adaptive randomization with fixed response probabilities? A new paper explores these questions. The Effect of Population Drift on Adaptively Randomized Trials.

This latest version of Adaptive Randomization primarily fixes a number of minor problems in the user interface.

This evening we released TTEDesigner, software for designing single-arm safety monitoring trials with time-to-event endpoints. This software is the design counterpart to TTEConduct. TTEDesigner aims to eliminate the typical simulate-and-tweak cycle. Instead of specifying scenarios and tweaking parameters until you get the desired operating characteristics, you can specify the desired operating characteristics and the software will attempt to satisfy these.

We released a new version of CONFINT this morning adding calculations for exponential survival. Includes Windows and Mac OSX (Intel and PowerPC) executables.

The distribution file for STPLAN 4.3 now includes Intel and PowerPC executables for Macintosh OSX. The source code has not changed.

We made a couple changes to Predictive Probabilities. We gave it a new icon more suggestive of what the software does: a crystal ball. Also, we swapped a couple Greek letters in the interface, changing lambda to mu where the former had caused some confusion.

We also updated the icon for Inequality Calculator and slightly changed the wording for part of CRMSimulator.

Lyle Broemeling's new book Bayesian Biostatistics and Diagnostic Medicine was published recently. A part of this book gives an introduction to some of the clinical trial software available on this site.

A new version of the Confidence Interval of Interaction Index software by Jack Lee and Maiying Kong was posted today.

A new paper has been posted online that explores some of the reasons why typical phase II oncology trials do such a poor job of predicting which treatments will succeed in phase III. See A Review of Phase 2-3 Clinical Trial Designs by Peter Thall.

We have released a new build of Predictive Probabilities version 1.4. The new build opens the most commonly used dialog, interim analysis for binary endpoints, automatically when the program opens. Also, the User's Guide has been updated; some errors were corrected and some unclear areas expanded. If you would like the new User's Guide without re-installing the software, the file is available online.

A new version of Multc Lean was released today that adds the option to specify a minimum number of patients to treat before evaluating the stopping rules. This option is not generally recommended though it is occasionally necessary.

Toxicity Probability Intervals, software for implementing the method described in "Dose-finding in oncology clinical trials based on toxicity probability intervals" by Yuan Ji, Yisheng Li, and Nebiyou Bekele, was posted today. This software uses Microsoft Excel to produce a table that can be used to conduct the trial. Simulation software, written in R, is also provided.

Today we released Phase II PP, Source code (R and S-PLUS) is provided to accompany the paper "A predictive probability design for phase II cancer clinical trials" by J. Jack Lee and Diane D. Liu. The method provides a flexible alternative to frequentist two- or three-state designs for phase II trials based on Bayesian predictive probability.

WFMM is a Windows command-line application that implements a Bayesian wavelet-based functional mixed model methodology for functional data analysis developed by Jeffrey Morris and Raymond Carroll.

The PRT software implements the statistical method described in the paper "Monitoring Late Onset Toxicities in Phase I Trials Using Predicted Risks" by Bekele, Ji, Shen, and Thall.

This software implements Bayesian inference for the ANOVA DDP model described in De Iorio, M., Mueller, P., Rosner, G., and MacEachern, S. An ANOVA Model for Dependent Random Measures, Journal of the American Statistical Association, 99(465), 205–215 (2004). The ANOVA DDP model is a model for repeated measurements data. The random effects distribution includes a regression on subject-specific covariates.

New Predictive Probabilities user's guide (16 January 2007)

There is a new version of the Predictive Probabilities user's guide posted on the product's download page. The new version corrects errors in two equations, one on page 4 and one on page 5.

The Adaptive Randomization software distributed by our department supports three ways to control how the randomization probabilities change in response to incoming data. One way is to begin the trial with a burn-in period of equal randomization. Another is to set a lower bound on the randomization probabilities. A third way is to use a power transformation of the inequality probability. This raises several interesting questions. How does each method impact the statistical power of the design? How does each impact the treatment of the patients in the trial? Which approach should one take? A technical report released today, Comparing methods of tuning adaptively randomized trials, addresses these questions in the context of a two-arm trial with a binary endpoint.

A new working paper, Practical Bayesian Adaptive Randomization in Clinical Trials by Peter F. Thall and J. Kyle Wathen was released recently. The abstract follows.

While randomization is the established method for obtaining scientifically valid treatment comparisons in clinical trials, it sometimes is at odds with what physicians feel is good medical practice. If a physician favors one treatment over another based on personal experience or published data, it may be more appropriate ethically for that physician to use the favored treatment, rather than enrolling patients on a randomized trial. Still, the randomized trial may later show the physician's favored treatment to be inferior. This paper reviews a statistical method, Bayesian adaptive randomization, that provides a practical compromise between the scientific ideal of conventional randomization and choosing each patient's treatment based on a personal preference that may prove to be incorrect. The method will first be illustrated by a simple hypothetical example, then by a recent trial in which patients with unresectable soft tissue sarcoma were adaptively randomized between two chemotherapy regimens.

The Adaptive Randomization software distributed by our department allows the user to specify a tuning parameter λ. A recent technical report explains the motivation for this parameter and how it effects a trial design's operating characteristics. The report concludes with some suggestions for determining the value to use when designing a clinical trial.

The latest version of TTEConduct now has a Windows user interface for ease of use. The statistical functionality is the same as in the earlier commandline version.

A number of our users have had trouble extracting zip files. (See the first two questions on our FAQ page for details.) We released EffTox 2.10 as a self-extracting file a couple months ago, and the feedback has been positive. This week we converted a few more of our download products from .zip format to self-extracting .exe format. If the new format causes you difficulty, please let us know.

Several changes were made in this version of EffTox.

• Numeric inputs now work correctly with non-English versions of Windows. The software asks the operating system for the local convention for decimal separators.
• The tutorial has been expanded and converted to PDF format.
• Documentation for the new L^p contours has been added.
• Error handling has been improved.

Also, beginning with this release we are making products available as self-extracting zip files. We believe users will find this more convenient than customary zip files. If this causes a problem, please let us know.

Today we released BCSTTE, Bayesian Chi-square TTE fit software. This software uses the Bayesian chi square test of Valen Johnson to assess goodness-of-fit for several parametric models to right-censored time-to-event data. The following distribution families are supported: exponential, gamma, inverse gamma, Weibull, log normal, log logistic, and log odds rate.

Version 1.4 of Predictive Probabilities has several improvements over the previous version:

• The binary outcome model has been generalized. Where the model previous used P(X > Y) it now uses P(X > Y + delta). The delta term can represent a required level of improvement or an allowable slippage.
• The user's guide has been completely rewritten.
• A few changes in the user interface make the software easier to use.
• An incompatibility with non-English versions of Windows has been fixed. The software now queries the operating system for the correct local decimal separator character.

Multc Lean version 1.1 was released today. This version allows the user to specify the cohort size used for trial monitoring. Previous versions only supported continuous monitoring, i.e., cohorts of size 1. Now the cohort size can be any number that evenly divides the maximum number of patients.

This version also fixes a problem with running the software on non-English versions of Windows. (Earlier versions assumed that the decimal separator was always a period rather than asking the operating system for the local convention.) The input validation has also been improved.

This version of Inequality Calculator fixes a problem with running the software on non-English versions of Windows. (Earlier versions assumed that the decimal separator was always a period rather than asking the operating system for the local convention.) The input validation has also been improved.

This version also fixes an error in previous versions that occurred when an inverse gamma distribution had shape parameter equal to exactly 1 or 2.

This version of Parameter Solver fixes a problem with running the software on non-English versions of Windows. (Earlier versions assumed that the decimal separator was always a period rather than asking the operating system for the local convention.) The input validation has also been improved.

This version also fixes an error in previous versions that occurred when an inverse gamma distribution had shape parameter equal to exactly 1 or 2.

This version of CRMSimulator fixes a problem with running the software on non-English versions of Windows. (Earlier versions assumed that the decimal separator was always a period rather than asking the operating system for the local convention.) The input validation has also been improved.

Several of our software products require the Microsoft .NET framework. We have been including the framework as part of our zip files for distribution in case users do not have the framework installed. We continue to do this, but now we are adding the option of downloading a smaller file that does not contain the framework for the convenience of those who already have the necessary version of the framework installed.

The download site will suggest the smaller download file if it determines that the necessary version of the .NET framework is installed and will suggest the larger version otherwise.

The EffTox User's Guide posted on the EffTox download page has been extended to include more guidelines for selecting the program inputs. The new guide gives advice for choosing cohort size, starting dose, outcome model, etc.

The download page for Adaptive Randomization added a link to a short paper explaining the design tuning parameter in detail.

If you would be interested in volunteering as a beta tester for some of our software applications, please send us a note. Also, if you have any comments on released software (bug reports, suggestions for new features, suggestions for clarifying documentation, etc.) please let us know.

This S-PLUS software estimates interaction indices and their confidence intervals for assessing multiple drug interactions.

This commandline software calculates requisite sample size to achieve a specified probability of a confidence interval of at most a specified size. Windows and Macintosh OS X executables and source code included.

The Windows and Macintosh OS X executables have been recompiled and the documentation was converted to LaTeX.

Today we released an update to Predictive Probabilities. The previous version had a bug that effected time-to-event interim analysis if the exponential distribution were parameterized in terms of rate. Since most users parameterize the exponential in terms of its mean, we simply removed the rate parameterization option and now require the mean parameterization.

CRMSimulator version 5.04 was released today. This minor release fixes a bug in the user interface of the previous version.

EffTox 2.9 was released today. This version uses a new method for specifying efficacy-toxicity trade-offs, described detail in this technical report. The new trade-off functions are simpler and more flexible than their counterparts in previous versions. Color coding was added to the efficacy and toxicity input labels to reduce the chance that the user might accidentally reverse these inputs.

Wiley has just published Statistical Method for Dose-Finding experiments, edited by Sylvie Chevret, a book which features several software applications available on this download site. Chapter 13, "A two-stage design for dose-finding with two agents", is about the statistical method implemented in the ToxFinder software. Chapter 14, "Using both efficacy and toxicity for dose-finding", describes the method used in the EffTox software. Chapter 15 also mentions our CRM and CRMSimulator software.

A new version of STPLAN was released today. A summary of the changes follows. For details, see the "news" file included in the distribution.

1. In previous versions, there was an error in the significance calculations for the two-sided test for one-sample randomized clinical trials. This has been corrected.
2. The algorithm for computing the two-sample Poisson test has changed.
3. The information time method for randomized clinical trials with an exponential distribution of time to event has been added.
4. In the one-sample binomial test, the code now picks the smallest sample if several samples have the same power.
5. The documentation was rewritten as PDF (from LaTeX) with bookmarks.

We plan to release new versions of STPLAN, EffTox, and Multc Lean over the next several weeks.

The coming version of STPLAN changes the way the two-sample Poisson test is calculated. Also, the information time method for randomized clinical trials with an exponential distribution of time to event has been added. Finally, the documentation has been revised.

The coming version of EffTox will introduce the new L^p contours and will resolve a problem with dependent scenarios. The new release may also modify the algorithm used to solve for hyperparameters based on elicited probabilities.

The coming version of Multc Lean will add an option for monitoring in cohorts; the current version supports only continuous monitoring, that is, cohorts of size 1.

TTEConduct version 1.2 was released this afternoon. The new version has three changes:

1. The previous version had assumed an upper limit of 100 years for the total time on test. This limit was too restrictive for large trials. The new limit is trial size times 10 years.
2. The new user's guide includes an example at the end.
3. The new software runs several times faster than the previous version.

This report describes the a method of computing efficacy-toxicity trade-offs in the EffTox dose-finding method. This method will be used in a new version of the EffTox software to be released in a few weeks.

We recently discovered a sequence of events that could prevent Multc Lean version 1.0.0 from displaying help files. This bug is fixed in version 1.0.3. As part of the same release, we also revised the statistical tutorial.

TTEConduct was re-released today with an updated ReadMe file. The new documentation gives recommendations for how one may select the design parameters. The new ReadMe file is posted on the TTEConduct page on the software download site as well as being bundled with the TTEConduct download distribution.

Note that TTEConduct trial designs need to be simulated in order to understand their operating characteristics. We currently have internal software for simulating such trials but do not have anything ready to release to the public yet. We hope to release such software in the future, perhaps this summer.

The latest version of Adaptive Randomization, version 3.2.2, contains a bug fix, and optimization, and a more convenient installer. (You can tell what version of Adaptive Randomization you are running by looking at Help -> About ARand.)

We discovered recently that previous versions of Adaptive Randomization would occasionally stop a trial early even if the early stopping rules were effectively turned off by setting impossibly high stopping boundaries. This has been corrected.

In addition, the procedure used to calculate the randomization probabilities has been made more efficient, speeding up simulation.

The new installer will automatically remove the previous version before installing the latest version; there is no need to manually remove the previous version from the control panel.

There were a few changes to the User's Guide and the code to bring it up. A few typos in the user interface were corrected.

Today we released a new version of our software for predictive probability interim analysis of clinical trials. The changes in version 1.3 are as follows.

• Online documentation has been revised
• Broken dialogs have been fixed
• Menus have been simplified
• User interface text has been revised

The numerical output of the software has not changed.

A new technical report entitled Power and Bias in Adaptively Randomized Clinical Trials was published today. The abstract from the report follows.

"This report examines the operating characteristics of adaptively randomized trials relative to equally randomized trials in regard to power and bias. We also examine the number of patients in the trial assigned to the superior treatment. The effects of prior selection, sample size, and patient prognostic factors are investigated for both binary and time-to-event outcomes."

We've added a resources page, listing some of the software, books, and web sites we've found useful in development the software on this site. We intend to add more to this page over time.

A new release of Adaptive Randomization 3.1.1 has been posted to our department's software download site. This release adds the User's Guide to the help menu, but otherwise does not change the program's functionality. The new release is available here.

The User's Guide is also available as a stand-alone document from the same page.

Software for generating the tables described in Continuous safety monitoring in single-arm, time-to-event trials without software was added to the download site today. This is the software that makes it possible to run the trial without software! That is, this software can be run at design time to produce a table containing stopping boundaries, eliminating the need for software to evaluate the stopping rules repeatedly during the course of the trial.

A new technical report entitled Numerical Evaluation of Gamma Inequalities was posted today. This report describes the algorithm used in Adaptive Randomization 3.1.1 to calculate the randomization probabilities for two- and three-arm time-to-event trials.

A new technical report entitled Continuous safety monitoring in single-arm, time-to-event trials without software has recently been posted. At first it appears that one needs special software to monitor time-to-event trials using the simple exponential/inverse gamma Bayesian model, but in fact one can determine the stopping conditions before the trial begins and carry them around on an index card.

Adaptive Randomization 3.1.1 was released this morning. This version corrects a problem with the previous version which prevented the installation program from working on some computers. Also, version 3.1.1 uses a new numerical algorithm for computing randomization probabilities for 3-arm time-to-event trials which in testing has shown to be faster and more robust.

Adaptive Randomization 3.1.0 was released this morning. The new version fixes a bug in version 3.0.0 that prevented the user from being able to simulate a 3-arm time-to-event trial. In addition, minor typos in the output were corrected.

A new STPLAN user's guide is available. This new version has the same content as the April 2000 version, aside from updated contact information. However, this version has been typeset in LaTeX, making the 199-page document easier to read.

A technical report Exact Calculation of Beta Inequalities was published today which explains how to find P(X > Y) in closed form where X and Y are beta random variables with some restrictions on the parameters. This technical report may be of interest to users of Adaptive Randomization, Inequality Calculator, and Multc Lean.

This software simulates randomized trials in which the randomization probabilities adapt in response to the outcome data. More patients are treated with the better treatment while retaining the benefits of randomization. The software supports both binary and time-to-event outcomes. Numerous design options are supported.

Software for simulating a wide variety of adaptively randomized clinical trials is in its final stages of testing and should be available for download early in November.

The software supports up to 10 arms. Stopping rules may be based on a maximum patient accrual or maximum trial length. Optionally one may specify a minimum number of patients in the trial. Also, one may specify the number of patients to randomize fairly before adaptive randomization begins. Stopping rules are based on posterior probabilities.

A frequently asked questions list has been added to the software download site.

The Department of Biostatistics and Applied Mathematics has just deployed a new software download site. Content from two previous sites has been merged. More information is available for each product, and search functionality has been added.