Methods for Comparative Effectiveness Research in Cancer

Harvard School Of Public Health, Boston MA

Investigators

Linked publications & trials

Abstract

This is a proposal to develop, refine, and implement statistical methods for comparative effectiveness research in cancer. Using these methods, we will emulate complex randomized trials of personalized and dynamic strategies for treatment and diagnosis of cancer. We will focus our efforts on three areas. First, the comparative effectiveness of personalized strategies for the diagnostic work-up of cancer patients. Because questions about the allocation of health care for diagnostic work-up involves the comparison of strategies that are clearly assigned in the data, conventional statistical methods cannot be used and alternative methods, like artificial censoring plus inverse probability weighting, are required. In Specific Aim 1, we will develop methods for the comparison of personalized health care delivery strategies in cancer diagnosis. As a motivating example, we will use SEER-Medicare to compare personalized strategies for the attendance to multiple medical centers by patients' and providers' characteristics. In Specific Aim 2, we will develop methods for the comparison of personalized strategies in the presence of unmeasured confounding. Subgroup analyses are important in identifying subpopulations for which treatment is most effective, and thus to personalize treatment. Because confounding may change the interpretation of subgroup analyses and lead to bias in effect modification and interaction parameters, we will develop methodology to assess and correct for the impact of unmeasured confounding in the design, analysis, and interpretation of subgroup analyses in comparative effectiveness research. In Specific Aim 3, we will develop methods for the comparison of dynamic health care delivery strategies in cancer treatment. Because many questions in cancer involve clinical decisions that depend on the evolving responses of patients or the characteristics of the local health care system, i.e., they are dynamic decisions, conventional statistical methods cannot be used and alternative methods, like dynamic marginal structural models and the parametric g-formula, are required. As a motivating example, we will compare dynamic strategies regarding androgen deprivation therapy for prostate cancer in the CaPSURE cohort. In Specific Aim 4, we will develop user-friendly, high quality and open access software to be distributed to cancer researchers. This project complements the descriptive and inferential aims in Projects 2 and 3, and relies heavily on the Statistical Computing Core, and the organizational infrastructure, team building strategies, provided through the Administrative Core.