Impact of incomplete geographic locators on small-area predictive models of invasive pneumococcal disease (IPD) rates in New Mexico (NM), 2004-2010

Small-area spatial analysis of health events adds information about disease predictors when individual-level information is unavailable. The US Census Bureau's American Community Survey (ACS) provides small-area demographic, social, and economic data. Rural states often have difficulty geo-locating individuals with post office box addresses. We sought to determine the impact of non-geocoded cases (NGCs) on a regression model predicting block-group-level invasive pneumococcal disease (IPD). New Mexico (NM) IPD cases identified through Emerging Infections Program (NMEIP) active surveillance during 2004-2010 were geocoded (ArcGIS 9.3). NGCs were assigned to the postal code centroid block-group. Block-group-level IPD incidence rates were calculated. Educational attainment, English-speaking ability, vehicle availability, poverty status, average household size, median year house built, and household plumbing data were obtained from ACS 5-year estimates (2005-2009). We compared two regression models, one excluding NGCs and the second assigning them. NMEIP identified 2,391 IPD cases; 30% were unmatched at the street level. NGCs were more likely to be non-white, younger, and reside in rural counties. A regression model with NGCs removed identified low educational attainment and older housing as significant block-group IPD rate predictors (p<0.05), whereas the second model with assigned NGCs predicted increased rates with decreased English fluency and incomplete plumbing facilities (p<0.05). Our analysis indicated that excluding NGCs resulted in differing predictive models. Health researchers should recognize potential bias when excluding cases. More rigorous population-weighted geo-imputation might produce better predictive models; however, error might be reduced by postal code assignment in the absence of adequate resources for more complex analyses.