Statistical Methods In Epidemiology--general

点击翻译按钮获取中文摘要

This project develops new statistical methods for epidemiology with broad applications and also methods as needed for ongoing projects in epidemiology, particularly those related to reproductive studies. The work this year involved three projects. (1) One project concerned using the bivariate birth weight and gestational length data for US births to estimate and correct for measurement errors in use of last menstrual period for estimating gestational length. Our goal was ultimately to develop improved standards for birth weights and variances for each completed week of gestation and also to correct the estimated rates of preterm birth within ethnic and maternal age categories. We also hope to improve the assessment of the relationship between birth weight, gestational length and risk of perinatal mortality within those categories. (2) Another project concerns pooled assessment of expensive-to-assay biomarkers based on human samples. Earlier work had shown that in a case-control setting one can pool together specimens from sets of cases and sets of controls and carry out a set-based analysis. With a slightly modified logistic model that analysis can estimate the individual-level risk parameters and loses almost no power compared to analysis based on individual assays. This means that if an exposure is based on an expensive assay that uses human samples, one can markedly improve efficiency by pooling specimens prior to assay. In work with my new post-doc, Paramita Saha, we are now extending these methods to time-to-pregnancy data, where one pools specimens within strata defined by the time to conception. Again the power suffers almost not at all, compared to individual level assays, and the costs are greatly reduced. Another great benefit to pooling in general is in its conservative use of irreplaceable human specimens. (3) A third project developed a method for testing the homogeneity assumption that is required in conditional logistic regression applications. For example, one might carry out a sibship-based analysis involving a genetic and an environmental factor. If the measured genetic marker is not itself causative, but serves as a linked surrogate for an unmeasured SNP, then there could be heterogeneity among families and this test could detect that heterogeneity.