Resource Labs
Biostatistics Unit
Overview
The collaborative activities of our group include assisting in the design and analysis of the Cancer Center's basic and clinical research, as well as serving as the Statistical Coordinating Center for the NCI-sponsored Cancer Genetics Network (CGN). The CGN is a multi-institutional initiative to support collaborative investigations of the genetic basis of cancer susceptibility. Following is an overview of our statistical research (funded by NCI R01 grants).
Our research on quantitative methods for clinical trials and epidemiology has been primarily motivated by the issues that we identified in our collaborative efforts of designing and analyzing cancer studies. A prime example how a Cancer Center collaboration resulted in a major statistical contribution to the design of randomized clinical trials is given by Finkelstein et al, (2003) describing comparison of survival of a sample to that of a population. This work was motivated by a collaboration with Cancer Center surgical oncologists interested in showing that patients with NMPD (Paget's disease) do not experience an increased mortality risk.
Many of the areas of statistical research have been drawn from our more than 25 years of experience working in AIDS and cancer research. One major area of our research has been in the analysis of missing progression time data. When patients are being followed to monitor for disease progression, they often miss visits, and the failure can be censored into the interval of missed visits (interval censored data). In this case, standard methods such as the Cox model cannot be directly applied. We have developed a methodology to analyze these data for several applications to AIDS and cancer, such as the case where the risk of progression and the visit compliance are correlated (Finkelstein et al 2002).
A second focus of our research in survival methods has been in the simultaneous analysis of longitudinal and event time data. We have applied Bayesian methods for modeling the risk of cancer as a function of a longitudinal biomarker such as CA125 for ovarian cancer (Skates et al. 2001) or PSA for prostate cancer (Pauler and Finkelstein, 2002). In response to the expanding field of genomics, we have developed methods for analysis of longitudinal microarrays (Rajicic et al, 2006).
Dr. Dianne Finkelstein is the Principal Investigator for an NCI $16.8M contract for the Cancer Genetics Network (CGN). This project is a resource for research on individuals with genetic susceptibility for cancer. The CGN has enrolled over 25,000 subjects willing to participate in these studies. Several studies have been completed in the first 7 years of the project, for example the 2000 subject ovarian cancer screening study (PI Skates). The Biostatistics Center has used the registry for epidemiology studies on cancer clusters in individuals and families (Matthews et al 2005, 2007), and is currently collecting DNA for a Harvard-based biorepository for basic research on multiple cancer families.
Dr. Rebecca Betensky collaborates with investigators in Neuro-Oncology program and the Molecular Neuro-Oncology Laboratory. Her current methodological research focus is in the areas of latent class modeling for genomic data. Latent class models are useful for both unsupervised clustering of moderate to high dimensional genomic data and for supervised clustering by clinical outcomes, such as survival. Dr. Betensky has made innovative contributions to this area through the introduction of penalization, both in frequentist and Bayesian settings, to enable model fitting with the high dimensional data (Houseman et al, 2006). She has developed these models for both unsupervised and supervised classification and for genomic data with spatial structure (e.g., loss of heterozygosity and array comparative genomic hybridization). She has applied these methods to brain tumor studies.
Dr. Hui Zheng, working with Dr. Alan Zaslavsky, is developing a hierarchical model for profiling colorectal cancer care-provider quality by combining the near- term post-surgical survival with long-term survival rate. The model includes provider level period-specific random effects and allows for within-provider correlations of the survival within 30 days of surgery and conditional survival (survival for those who are alive at the beginning of the period) in 30 day-1 year and 1-5 years periods. The method allows for improved statistical efficiency when short-term post-surgical mortality rates are low. It also makes it possible to assess the trend in the influence of patient and provider characteristics on survival over time.
Group Members
- Dianne Finkelstein, PhD
- Hang Lee, PhD
- Eric Macklin, PhD
- Steven J. Skates, PhD
- Beow Y. Yeap, ScD
Research at Mass General 
Discover the largest hospital-based research program in the U.S. and how clinicians and scientists chart new terrain in biomedical research to treat and prevent human disease and bring the latest advances to patient care
