Comparison of mixed effects models of antimicrobial resistance metrics of livestock and poultry Salmonella isolates from a national monitoring system

Antimicrobial use in agriculture is considered a pathway for the selection and dissemination of resistance determinants among animal and human populations. From 1997 through 2003 the U.S. National Antimicrobial Resistance Monitoring System (NARMS) tested clinical Salmonella isolates from multiple animal and environmental sources throughout the United States for resistance to panels of 16-19 antimicrobials. In this study we applied two mixed effects models, the generalized linear mixed model (GLMM) and accelerated failure time frailty (AFT-frailty) model, to susceptible/resistant and interval-censored minimum inhibitory concentration (MIC) metrics, respectively, from Salmonella enterica subspecies enterica serovar Typhimurium isolates from livestock and poultry. Objectives were to compare characteristics of the two models and to examine the effects of time, species, and multidrug resistance (MDR) on the resistance of isolates to individual antimicrobials, as revealed by the models. Fixed effects were year of sample collection, isolate source species and MDR indicators; laboratory study site was included as a random effect. MDR indicators were significant for every antimicrobial and were dominant effects in multivariable models. Temporal trends and source species influences varied by antimicrobial. In GLMMs, the intra-class correlation coefficient ranged up to 0.8, indicating that the proportion of variance accounted for by laboratory study site could be high. AFT models tended to be more sensitive, detecting more curvilinear temporal trends and species differences; however, high levels of left- or right-censoring made some models unstable and results uninterpretable. Results from GLMMs may be biased by cutoff criteria used to collapse MIC data into binary categories, and may miss signaling important trends or shifts if the series of antibiotic dilutions tested does not span a resistance threshold. Our findings demonstrate the challenges of measuring the AMR ecosystem and the complexity of interacting factors, and have implications for future monitoring. We include suggestions for future data collection and analyses, including alternative modeling approaches.