Authors: Henry Hui Luan*, University of Oregon
Topics: Spatial Analysis & Modeling, Medical and Health Geography, Geographic Information Science and Systems
Keywords: Censored data; spatiotemporal disease mapping; Bayesian
Session Type: Paper
Start / End Time: 2:35 PM / 4:15 PM
Room: Marshall South, Marriott, Mezzanine Level
Presentation File: No File Uploaded
Health data, especially those sensitive ones such as cancer and HIV/AIDS, are left-censored when they are released at small-area levels. For example, HIV surveillance data at the Zip Code level is suppressed if the number of new cases is smaller than five. This censored-data issue is increasingly common in spatiotemporal datasets. Conventionally, the censored observations are removed from the analysis, leaving a truncated dataset for the (potentially biased) inference. An alternative method is imputing the suppressed data with simple deterministic approaches, which fails to account for uncertainties associated with the imputation. This research develops a model-based approach to impute the censored health data at small-area levels. Specifically, a censored sampling distribution is used to fit the suppressed data. Spatial autocorrelation, temporal autocorrelation, and spatiotemporal interaction in the underlying process that generates the health data are explicitly modeled with random effects. The model is implemented in the Bayesian statistical framework. An HIV/AIDS dataset of Philadelphia at the Zip Code level between 2010 and 2015 is analyzed with the developed model. Over one-fourth of the observations (75 out of 265) are suppressed from data release.