Effects of Socioeconomic and Racial Residential Segregation on Preterm Birth: A Cautionary Tale of Structural Confounding
Effects of Socioeconomic and Racial Residential Segregation on Preterm Birth: A Cautionary Tale of Structural Confounding
复制标题
DOI：
10.1093/aje/kwp435
复制
发表时间：
2010-03-15
影响因子：
5
通讯作者：
Mason, Susan
中科院分区：
文献类型：
--
作者：
Messer, Lynne C.;Oakes, J. Michael;Mason, Susan
与社会分层或其他选择过程相关的混杂被称为结构性混杂。在存在结构性混杂的情况下，某些协变量分层将仅包含永远不会暴露的受试者，这违反了阳性或实验治疗效应假设。因此，结构性混杂可能会阻碍在不同暴露水平之间进行有意义的因果对比所必需的交换。作者通过估计北卡罗来纳州威克县和达勒姆县（1999-2001年）邻里剥夺和邻里种族组成（隔离）对早产率的独立影响，探讨了结构性混杂的存在和程度。表格分析和随机截距固定斜率多层次逻辑模型描绘了不同的结构现实，在这些县。多层次模型的结果表明，一些不显着的影响，居住在大片的社会经济贫困或种族居住隔离的调整后的早产几率白色和黑人妇女生活在这些县，和置信限比表明相当一致的精度水平的估计。然而，表格分析的结果表明，这些回归建模结果中有许多是不支持的，并且没有基于实际数据。统计和公共卫生推断的影响，在没有数据的情况下，被认为是。
Confounding associated with social stratification or other selection processes has been called structural confounding. In the presence of structural confounding, certain covariate strata will contain only subjects who could never be exposed, a violation of the positivity or experimental treatment effect assumption. Thus, structural confounding can prohibit the exchangeability necessary for meaningful causal contrasts across levels of exposure. The authors explored the presence and magnitude of structural confounding by estimating the independent effects of neighborhood deprivation and neighborhood racial composition (segregation) on rates of preterm birth in Wake and Durham counties, North Carolina (1999-2001). Tabular analyses and random-intercept fixed-slope multilevel logistic models portrayed different structural realities in these counties. The multilevel modeling results suggested some nonsignificant effect of residence in tracts with high levels of socioeconomic deprivation or racial residential segregation on adjusted odds of preterm birth for white and black women living in these counties, and the confidence limit ratios indicated fairly consistent levels of precision around the estimates. The results of the tabular analysis, however, suggested that many of these regression modeling findings were off-support and based on no actual data. The implications for statistical and public health inference, in the presence of no data, are considered.