Dose Response Modeling in Epidemiologic Cohort Studies

描述（由申请人提供）：本修订提案建立在我目前的工作基础上，应用半参数统计方法对职业队列研究中具有光滑暴露函数的非线性暴露-反应曲线进行建模（R01 CA81345）。在将惩罚样条用于生存分析的过程中，我们确定了两个需要进一步发展的问题：选择最佳平滑量的标准和诊断统计，以确定有影响的观察结果并测量它们对曲线拟合值的影响。这些问题将在典型的工作场所暴露的右偏斜分布的存在下进行检查。我们建议将惩罚样条的拟合优度和阻力与另外两种常见的平滑器进行比较：受限三次样条和称为黄土的局部加权回归平滑器。在这个修改后的应用程序中，我们建议通过一系列模拟研究来评估模型拟合来比较三种平滑方法。数据将在线性、三次多项式和阈值（断棒）暴露-响应模型下生成，假设暴露具有对数正态分布。在校准三种方法的平滑度后，我们将确定最适合真实数据的技术。我们现在还建议开发统计方法，以量化影响并确定半参数模型中具有高杠杆作用的点，特别注意偏斜暴露分布高端的值。我们将计算参数和非参数成分的库克距离，并制定评估统计显著性的规则，以及确定平滑参数对受试者删除的敏感性的诊断。新的诊断方法将用于惩罚样条，然后扩展到限制三次样条和黄土。这些模型将应用于两项癌症死亡率研究的再分析；接触金属加工液（MWF）的汽车工人罹患喉癌、肺癌、前列腺癌和直肠癌，加州硅藻土行业接触结晶二氧化硅的工人罹患肺癌。我们计划将方法学研究的重点放在MWF和二氧化硅上，因为我们相信，澄清它们的暴露-反应曲线的形状，可以对当前有关它们在工作场所控制的争议做出重大贡献。

DESCRIPTION (provided by applicant): This revised proposal builds on my current work applying semiparametric statistical methods to model nonlinear exposure-response curves with smooth functions of exposure in occupational cohort studies (R01 CA81345). In the course of adapting penalized splines for survival analysis, we have identified two issues that need further development: criteria for choosing the optimal amount of smoothness and diagnostic statistics to identify influential observations and measure their influence on the fitted values of the curve. These issues will be examined in the presence of right skewed distributions typical of workplace exposures. We propose to compare the goodness of fit and resistance of penalized splines to that of two other common smoothers: restricted cubic splines, and a locally weighted regression smoother called loess. In this revised application, we propose to compare the three smoothing methods by evaluating model fit in a series of simulation studies. Data will be generated under linear, cubic polynomial, and threshold (broken stick) exposure-response models, assuming exposure has a lognormal distribution. We will identify the technique with the best fit to the true data, after calibrating the amount of smoothness across the three methods. We also now propose to develop statistical methods to quantify influence and identify points with high leverage in semiparametric models, with particular attention to values at the high end of the skewed exposure distribution. We will compute Cook's distance for parametric and nonparametric components and develop rules for assessing statistical significance as well as diagnostics for determining the sensitivity of the smoothing parameter to subject deletion. The new diagnostics will be developed for penalized splines and then extended to restricted cubic splines and loess. The models will be applied to the re-analyses of two cancer mortality studies; larynx, lung, prostate and rectal cancer in autoworkers exposed to metalworking fluids (MWF) and lung cancer in workers in California's diatomaceous earth industry exposed to crystalline silica. We plan to focus our methodologic investigations on MWF and silica because we believe that clarifying the shapes of their exposure-response curves can make a substantial contribution to current controversies over their control in the workplace.