Contrasting methods to operationalize antibiotic exposure in clinical research: a real-world application on healthcare-associated Clostridioides difficile infection.

The goal of this article is to summarize common methods of antibiotic operationalization used in clinical research and demonstrate methods for exposure variable selection. We demonstrate three methods for modeling exposure, using data from a case-control study on Clostridioides difficile infection in hospitalized patients: 1) factor analysis, 2) logistic regression models, and 3) Least Absolute Shrinkage and Selection Operator (LASSO) regression. The factor analysis identified 8 variables contributing the most variation in the dataset: any antibiotic exposure; number of antibiotic classes; number of antibiotic courses; dose; and specific classes monobactam, 𝛽-lactam 𝛽-lactamase inhibitors, rifamycin, and cephalosporin. The logistic regression models resulting in the best model fit used predictors representing any antibiotic exposure and the proportion of a patient's hospitalization on antibiotics. The LASSO model selected 22 variables for inclusion in the predictive model, of which 10 were antibiotic exposure variables, including: any antibiotic exposure; classes 𝛽-lactam 𝛽-lactamase inhibitors, carbapenem, cephalosporin, fluoroquinolone, monobactam, rifamycin, sulfonamides, and miscellaneous; and proportion of hospitalization on antibiotics. Investigators studying antibiotic use should consider multiple characteristics of exposure informed by their research question and the theory on how antibiotics may impact the distribution of the outcome in their target population.