parametric approaches to quality adjusted survival analysis

2
Abstracts 383 A16 NONPARAMETRIC SURVIVAL REGRESSION TREES AND APPLICATION TO THREE RADIATION THERAPY ONCOLOGY GROUP (RTOG) MALIGNANT GLIOMA TRIALS Charles B. Scott, Thomas F. Pajak, and Walter J. Curran, Jr. Radiation Therapy Oncology Group Philadelphia, Pennsylvania A nonparametric tree-structured recursive partitioning technique for censored survival data is presented. Recursive partitioning is an exploratorymethod which provides information on the importance of and interaction be.tween prognostic factors not available with the semi-parametric Cox proportional hazards model. The product limit estimate of the survival function was calculated for all remaining covariates within each node of the regression. This algorithm utilizes the modified Wilcoxon test by Peto and Prentice for the splitting criteria. The partition for each non-terminal node was based upon the maximum chi-squaro value of all binary covariates examined. At every level of the tree multiple significance tests are performed, therefore a modified Bonferroni multiple comparison procedure (MCP) was used to control the multiple test e. Terminal nodes were defined as those with fewer than 25 patients, or when no possible partitions exceeded the minimum chi-squaro value determined by a preselected and the MCP. This method was applied to a combined data set of three RTOG malignant glioma trials, that constituted 1578 patients. There was interest in examining 32 covariates~24 binary, 7 ordered, and one nominal---for their prognostic importance in survival. This procedure resulted in 12 terminal nodes with median survival times of 4.7 to 58.6 months and displayed age as the most important prognostic variable. Previous recursive partitioning techniques were parametric based upon the exponential distribution. However, the exponential provided a poor fit to the data, thus a nonparametric technique was desired. Results of the comparison of the proposed algorithm to previously reported parametric techniques will be presented for this database. A17 OBSERVATIONAL ANALYSES WITH TIME DEPENDENCY AND MISSING VALUES IN THE THROMBOLYSIS IN MYOCARDIAL INFARCTION (TIMI) II CLINICAL TRIAL Michael L. Terrln, Bruce Thompson, Margaret Frederick, Genell L. Knatterud, and the TIMI Investigators Maryland Medical Research Institute Baltimore, Maryland In TIMI II, patients (pts.) were treated with intravenous tissue-type plasminogen activator, heparin as an intravenous bolus followed by continuous infusion for up to 5 days, and randomly assigned to either an invasive or conservative strategy for management. Heparin therapy was adjusted according to the activated partial thromboplastin time (APTT) measurements. In conjunction with analyses comparing the frequency of hemorrhagic events between patients assigned to the two treatment strategies, TIMI II investigators had an interest in assessing the effect of heparin anticoagulation with AP'I-r levels prolonged more than usually desired. Among the 3339 pts. enrolled in TIMI II, major hemorrhagic events occurred in 185 (5.5%) pts. in the first 5 days from study entry. AP'n" measurements were available for 3287 pts. at some time from study entry through day 5; 3270 on day 1; 3132 on day 2; 2993 on day 3; 2819 on day 4; and, for 2523 on day 5. Consequences of different approaches to take account of missing AP'I-I" values will be considered. A Cox proportional hazards model for time dependent covariates was applied to analyze changing AP'I-I" status in relationship to major hemorrhage. These results illustrate the importance of taking all patients into account and of allowing for time dependency in observational analyses performed within clinical trials. A18 PARAMETRIC APPROACHES TO QUALITY ADJUSTED SURVIVAL ANALYSIS Bernard Cole, Richard Gelber, and Keaven Anderson Dana-Farber Cancer Institute Boston, Massachusetts We present a parametric methodology for performing quality adjusted survival analysis using multivariate censored survival data. The event times correspond to transitions between states of health, and overall survival is adjusted according to the amount of time spent in each state, The data are modeled by repeatedly modeling the conditional cause-specific hazard functions given the previous transitions. Covariates are in-

Upload: bernard-cole

Post on 31-Aug-2016

216 views

Category:

Documents


2 download

TRANSCRIPT

Page 1: Parametric approaches to quality adjusted survival analysis

Abstracts 383

A16 NONPARAMETRIC SURVIVAL REGRESSION TREES AND APPLICATION TO THREE RADIATION

THERAPY ONCOLOGY GROUP (RTOG) MALIGNANT GLIOMA TRIALS

Charles B. Scott, Thomas F. Pajak, and Walter J. Curran, Jr. Radiation Therapy Oncology Group

Philadelphia, Pennsylvania

A nonparametric tree-structured recursive partitioning technique for censored survival data is presented. Recursive partitioning is an exploratory method which provides information on the importance of and interaction be.tween prognostic factors not available with the semi-parametric Cox proportional hazards model. The product limit estimate of the survival function was calculated for all remaining covariates within each node of the regression. This algorithm utilizes the modified Wilcoxon test by Peto and Prentice for the splitting criteria. The partition for each non-terminal node was based upon the maximum chi-squaro value of all binary covariates examined. At every level of the tree multiple significance tests are performed, therefore a modified Bonferroni multiple comparison procedure (MCP) was used to control the multiple test e. Terminal nodes were defined as those with fewer than 25 patients, or when no possible partitions exceeded the minimum chi-squaro value determined by a preselected and the MCP. This method was applied to a combined data set of three RTOG malignant glioma trials, that constituted 1578 patients. There was interest in examining 32 covariates~24 binary, 7 ordered, and one nominal---for their prognostic importance in survival. This procedure resulted in 12 terminal nodes with median survival times of 4.7 to 58.6 months and displayed age as the most important prognostic variable. Previous recursive partitioning techniques were parametric based upon the exponential distribution. However, the exponential provided a poor fit to the data, thus a nonparametric technique was desired. Results of the comparison of the proposed algorithm to previously reported parametric techniques will be presented for this database.

A17 OBSERVATIONAL ANALYSES WITH TIME DEPENDENCY AND MISSING VALUES IN THE

THROMBOLYSIS IN MYOCARDIAL INFARCTION (TIMI) II CLINICAL TRIAL

Michael L. Terrln, Bruce Thompson, Margaret Frederick, Genell L. Knatterud, and the TIMI Investigators

Maryland Medical Research Institute Baltimore, Maryland

In TIMI II, patients (pts.) were treated with intravenous tissue-type plasminogen activator, heparin as an intravenous bolus followed by continuous infusion for up to 5 days, and randomly assigned to either an invasive or conservative strategy for management. Heparin therapy was adjusted according to the activated partial thromboplastin time (APTT) measurements. In conjunction with analyses comparing the frequency of hemorrhagic events between patients assigned to the two treatment strategies, TIMI II investigators had an interest in assessing the effect of heparin anticoagulation with AP'I-r levels prolonged more than usually desired.

Among the 3339 pts. enrolled in TIMI II, major hemorrhagic events occurred in 185 (5.5%) pts. in the first 5 days from study entry. AP'n" measurements were available for 3287 pts. at some time from study entry through day 5; 3270 on day 1; 3132 on day 2; 2993 on day 3; 2819 on day 4; and, for 2523 on day 5. Consequences of different approaches to take account of missing AP'I-I" values will be considered. A Cox proportional hazards model for time dependent covariates was applied to analyze changing AP'I-I" status in relationship to major hemorrhage. These results illustrate the importance of taking all patients into account and of allowing for time dependency in observational analyses performed within clinical trials.

A18 PARAMETRIC APPROACHES TO QUALITY ADJUSTED SURVIVAL ANALYSIS

Bernard Cole, Richard Gelber, and Keaven Anderson Dana-Farber Cancer Institute

Boston, Massachusetts

We present a parametric methodology for performing quality adjusted survival analysis using multivariate censored survival data. The event times correspond to transitions between states of health, and overall survival is adjusted according to the amount of time spent in each state, The data are modeled by repeatedly modeling the conditional cause-specific hazard functions given the previous transitions. Covariates are in-

Page 2: Parametric approaches to quality adjusted survival analysis

384 Abstracts

corporated by accelerated failure time regression, and the model parameters are estimated by maximum likelihood. We then define quality functions which assign a "score" to a life having given transitions, and the modeling results are used to estimate the expectation of these functions. Standard errors and confidence intervals are computed using the bootstrap method and the delta method. The results are useful for evaluating treatments in terms of both quantity and quality of life.

As an example, we apply the methodology to data from the International Breast Cancer Study Group Trial V to compare short duration chemotherapy versus long duration chemotherapy in the treatment of node- positive breast cancer. The events studied are: (1) the end of treatment toxicity, (2) disease recurrence, and (3) overall survival.

A19 LESSONS LEARNED AND PITFALLS TO BE AVOIDED IN THE DESIGN OF STUDY DRUG

TITRATION SYSTEMS

Norma Lynn Fox, Evelyn Mlrenzi, and Frances LoPrestl Maryland Medical Research Institute

Baltimore, Maryland

The Post-CABG Coordinating Center has had more than two years of experience in titrating lipid-lowering medications for patients enrolled in this clinical trial. Coordinating Center staff issue recommendations for titration of each patient's medication regimen, based on information from multiple sources. An automated event-driven titration system reviews the pertinent history of each patient at each follow-up visit. If all data required for a titration decision are available, the system prints out the appropriate notification. If data are missing, inconsistent, or meet other criteria for monitoring, the system prints the history for staff review.

The automated titration system has been efficient and reliable, but the titration histories of a substantial proportion of patients have required staff review on one or more occasions. Some of the staff review could have been avoided by establishing guidelines for decision-making when data are missing or inconsistent. Minimizing the sources and amount of data required to prompt a decision would also have helped to limit staff review. The application of this experience to the design of study protocols and automated titration systems will be discussed.

A20 ADVERSE MEDICAL EVENTS IN CLINICAL TRIALS: REPORTING AND EVALUATION

Philip Day, Mark Jones, Clair Haakenson, Cindy Coiling, Carol Fye, and Mike Sather VA Cooperative Studies Program

Albuquerque, New Mexico

AMEs (adverse medical events) represent a major concern in all clinical trials involving drugs or medical devices. To prevent exposure of patients to unnecessary risk and to aid in documenting new AMEs, a concise mechanism for the collection, categorization and analysis of AME data is required.

The VA CSP (Veterans Affairs Cooperative Studies Program) has developed a system to uniformly handle AMEs in its trials involving drugs or medical devices. One component of this system is a computerized mechanism for evaluation of all AME reports. The heart of this computerized system is an AME coding scheme based upon a modified CoSTART (Coding Symbols for Thesaurus of Adverse Reaction Terms) terminology. CoSTART is the terminology developed and used by the Food and Drug Administration (FDA) for coding, filing and retrieving postmarketing adverse reaction reports.

The system allows study management to quickly evaluate each AME report and determine if findings suggest a new development which requires notification of investigators, the FDA and Data Monitoring Board. Implementing a unique coding system for incorporation into all VA CSP studies will also allow for trend analysis of AME data across trials.

A21 BAR CODING IN VALIDATING INVESTIGATIONAL DRUG PACKAGING

Mark S. Jones and Philip L. Day VA Cooperative Studies Program

Albuquerque, New Mexico

Providing the "right drug" to the "right patient" is a formidable task when dealing with large patient pop- ulations involving 'Iens-of-thousands" of dosage units. This task becomes even more ominous when dealing with large multi-center, randomized, double-blind, controlled clinical trials. Insuring that all investigational drug/