METHODS FOR EPIDEMIOLOGIC DATA WITH MISSING VALUES

此次重新提交 FIRST 奖提案的研究的长期目标是为研究人员提供处理研究应用中缺失数据的新方法。该 FIRST 奖的目标是 1) 开发和扩展预测变量包含缺失数据时倾向得分估计的统计方法，2) 将此方法应用于各种应用数据集。 为了实现这些目标，提出了几个目标。 其中包括：当预测变量包含缺失数据时，开发和扩展用于估计倾向得分的方法；开发允许预测变量包含可能不是“随机丢失”的丢失数据的方法；开发诊断方法来评估竞争模型的有效性和适合性；开发用户友好的软件来实施该方法；并将这些方法应用到四个实际数据应用中。 这些应用包括：1) 使用弗雷明汉心脏研究的数据来估计风险评估函数，以预测缺少风险因素的个体的心血管疾病或中风等结果； 2) 使用来自北加州 Kaiser Permanente 研究部提供的由超过 120,000 名会员组成的糖尿病登记处的数据，开发技术，帮助识别在缺少危险因素信息的情况下，有可能发生糖尿病相关并发症的高风险的糖尿病患者； 3) 使用绝经后雌激素/孕激素干预 (PEPI) 临床试验的数据拟合模型来估计倾向评分，该倾向评分表示以可能包含缺失数据的预测因子为条件的药物依从性概率，然后使用这些倾向评分来找到激素治疗对心血管疾病危险因素、骨矿物质密度和其他症状的影响的调整估计； 4) 使用腺瘤性息肉遗传流行病学研究的数据来拟合模型，估计特定基因与结果之间的关系，考虑到缺失风险因素的存在。这项研究的结果将对医学、流行病学和统计研究做出重要贡献。随着有关缺失值的统计方法的发展和扩展，预计将出现方法论和应用出版物。 此外，实质性的医学和流行病学问题将使用这种新方法在所提供的应用数据集上得到回答。

The long term objective of the research in this resubmission of a FIRST Award proposal is to provide investigators with new methodology for handling missing data in research applications. The aims of this FIRST Award will be 1) to develop and extend statistical methodology for propensity score estimation when predictors contain missing data and 2) to apply this methodology to a variety of applied data sets. To address these aims several goals are proposed. These include: developing and extending methodology for estimating propensity scores when predictors contain missing data; developing methodology that allows predictors to contain missing data that may not be "missing at random"; developing diagnostics to assess the validity and fit for competing models; developing user friendly software for implementing this methodology; and applying these methods to four real data applications. These applications include: 1) using data from the Framingham Heart Study to estimate risk appraisal functions for predicting such outcomes as cardiovascular disease or stroke for individuals with missing risk factors; 2) using data from a diabetes registry consisting of over 120,000 members, provided by the Division of Research at Kaiser Permanente, Northern California, to develop techniques that aid in identifying diabetic persons who are at high risk for developing diabetic related complications in the presence of missing risk factor information; 3) using data from the Postmenopausal Estrogen/Progestin Intervention (PEPI) clinical trial to fit models to estimate propensity scores which represent the probability of medication adherence conditional on predictors that may contain missing data, and then use these propensity scores to find adjusted estimates of the effects of hormone therapy on cardiovascular disease risk factors, bone mineral density, and other symptoms and; 4) using data from the Genetic Epidemiology of Adenomatous Polyps study to fit models estimating the relationship of specific genes to outcomes considering the presence of missing risk factors. The results from this research will make important contributions to medical, epidemiological and statistical research. Methodological and applied publications are anticipated as statistical methodology concerning missing values will be developed and extended. In addition, substantive medical and epidemiological questions will be answered using this new methodology on the applied data sets provided.