2016-Statistical Methods for Analysing Cancer Data


Statistical Methods for Analysing Cancer Data 
Chair: Gregory Pond (McMaster University)
Organizer: Amy Liu (Cancer Care Ontario) 
[PDF]

PATRICK BROWN, University of Toronto
Spatial Statistics with Cancer Registry Data [PDF]
 
Spatial data in cancer registries is messy; data quality and confidentiality mean only area-level data are available and low counts are suppressed. Area-level spatial models can give results which change with the spatial resolution, a phenomenon known as the 'modifiable area unit problem'. Treating aggregated spatial data as censored point data eliminates this problem, with aggregation reducing power without incurring bias. This approach will be shown with: a spatial point process model for cancer data with exact locations; a kernel-smoothing local-EM algorithm for aggregated case counts; and a spatially discrete model for case counts with censoring. 
 
KAREN KOPCIUK, University of Calgary
Risk Estimation in Family Data  [PDF]
 
Estimating risk in families who harbour a genetic mutation predisposing them to several types of cancer presents many statistical challenges: (a) families are identified and selected directly from population disease registries or high risk cancer clinics or from two-stage sampling designs, (b) missing genetic information is common or may be completely unknown for putative genes, (c) residual familial correlation exists when additional risk genes or environmental factors are shared, and (d) sequential cancers are frequent. Models need to take into account these features for age-at-onset outcomes, as well as direct interventions to the disease process. Genetic and disease risk models for family data we have developed will be described as well as our new R package, FamEvent. 
 
OLLI SAARELA, University of Toronto
A New Weighted Partial Likelihood Method for Estimating Marginal Structural Hazard Models [PDF]
 
Parameters in marginal structural Cox models can be estimated through an inverse probability of treatment and censoring (IPTC) weighted Cox partial likelihood. We propose an alternative weighted partial likelihood for estimating flexible parametric marginal structural hazard models, based on case-base sampling of person-moments, resulting in a weighted logistic regression form estimating function. This enables estimation of absolute hazards, and can accommodate continuous-time IPTC weights. In terms of computational convenience, the proposed method resembles the conventional discrete time pooled logistic regression method, but works in continuous time. We study the properties of the resulting estimator theoretically and through simulations, and illustrate the method in modeling the effects of repeated treatment procedures in cancer patients.