Random effects (RE) models (continuous and binary) will be developed, implemented, and tested to improve analysis of correlated data encountered in clinical oncology research. Correlated clinical data are common in meta-analyses, family studies, and risk assessments. For example, two summary statistics (e.g., means) from the same study are correlated data. Measurements of cancer risk from members within a family tend to be more alike than cancer risk measurements from members of different families. Assumption of a common random component shared by outcomes of the same study or members of the same family is proposed. Two types of random effects models based on whether outcomes are continuous or binary (yes/no) will be developed theoretically and numerically. These statistical methods will be compared with existing methods through analysis and simulation. The new methods will be applied to epidemiological and clinical oncology research data sets. Continuous and binary random effects models arise from many applications, such as, meta-analysis of cancer research data and analysis of family cancer risk data. However, the two types of random effects models have the same data structure (i.e., correlated data) and the same statistical framework (i.e., random effects model). In meta- analysis, an EM algorithm (Dempster, et al., 1977) [20] will be obtained under the additive model for the maximum likelihood estimate (MLE). This will be generalized to a mixed model which permits both study- and group-specific covariates. Analytic models which allow both covariate types are needed for meta-analysis of the relationship between estrogen replacement therapy and breast cancer risk. A multiplicative random effects model will then be formulated as an alternative for meta-analysis of odds ratios (rather than log odds ratios in the additive model) to allow for combining relative risk in epidemiological studies. A random effects logistic regression model for analysis of clustered binary measurements will be developed to evaluate family cancer risk data. This model will be applied to multivariate survival data obtained from family studies. Fully documented computer programs that implement the proposed random effects models will be written in house and tested at several collaborative sites. The programs will be made generally available via anonymous ftp on the Internet. Performance of the proposed analytic models and computer programs will be assessed by simulations. Extensive analyses of actual data will be done including tumor measurements comparison, estrogen replacement therapy and breast cancer risk, breast cancer in African-American women, bone marrow transplantation versus chemotherapy in the treatment of acute leukemia, and the Chinese family study of esophageal cancer, among others. Analyses of these data will identify specific methodologic areas that require refinement and extension.
| Li, Z; Rao, D C (1996) Random effects model for meta-analysis of multiple quantitative sibpair linkage studies. Genet Epidemiol 13:377-83 |
| Li, Z (1995) A multiplicative random effects model for meta-analysis with application to estimation of admixture component. Biometrics 51:864-73 |
