Steve Mooney, a doctoral student with the BEH group, won a best poster presentation award at the Society for Epidemiologic Research annual conference for his work on the effects on causal inference of error in measuring contextual variables.
Proportion of residents living in poverty, median household income and per capita income data from the Census are commonly used neighborhood-level contextual measures all of which are derived from personal income information provided to the Census bureau, information likely to be provided with some degree of error. Steve shows that this error in individual’s reporting of income, even if random and non-differential, causes bias in effect estimates for analyses of associations between neighborhood-level socioeconomic data and individual’s health outcomes. However, the direction of the bias depends on which of the neighborhood-level socioeconomic variables available from the Census the investigator chooses to use in the analyses. In the face of error in the underlying data collected by the Census, the use of neighborhood-level variables expressed as proportions (proportion living in poverty) results in bias away from the null. While the use of neighborhood-level variables expressed as a continuum (median household income) can cause bias towards or away from the null. The extent of bias in this later case is determined by the extent of measurement error, number of neighborhood units included in the study and the number of individuals per unit whose individual-level data were aggregated to create the neighborhood level variable.