Nearest-Neighbors Matching for Case–Control Study Analyses: Better Risk Factor Identification From a Study of Sporadic Campylobacteriosis in the United States

Abstract

Background: Case–control studies are commonly used to explore factors associated with enteric bacterial diseases. Control of confounding is challenging due to a large number of exposures of interest and the low frequencies of many of them. Methods: We evaluated nearest-neighbors matching in a case–control study (originally 1:1 matched, published in 2004) of sporadic Campylobacter infections that included information on 433 exposures in 2632 subjects during 1998–1999. We performed multiple imputations of missing data (m = 100) and calculated Gower distances between cases and controls using all possible confounders for each exposure in each dataset. We matched each case with ≤20 controls within a data-determined distance. We calculated odds ratios and population attributable fractions (PAFs). Results: Examination of pairwise correlation between exposures found very strong associations for 1046 pairs of exposures. More than 100 exposures were associated with campylobacteriosis, including nearly all risk factors identified using the previously published approach that included only 16 exposures and some less studied, rare exposures such as consumption of chicken liver and raw clams. Consumption of chicken and nonpoultry meat had the highest PAFs (62% and 59%, respectively). Conclusions: Nearest-neighbors matching appear to provide an improved ability to examine rare exposures and better control for numerous highly associated confounders.

Risk Factors for Non-O157 Shiga Toxin–Producing Escherichia coli Infections, United States

“…Nearest-neighbor matching approaches have a solid theoretical basis in epidemiologic research ( 37 – 39 ), but applying this method to matching in case–control studies of enteric diseases is recent ( 13 ). Although it is impossible to account for every possible confounder when selecting controls, this approach allows the most closely matched controls to be selected for each case.…”

“…Those reporting international travel were excluded from other risk factor analyses, which were conducted separately for infants <1 and persons > 1 years of age. To control for confounding in the main risk factor analysis, we rematched controls with cases using the nearest-neighbors approach ( 13 ). For a given exposure, we calculated Gower distance on the basis of age, sex, state, and all exposures except the one under consideration ( 14 ).…”