A comparison of modelling options to assess annual variation in susceptibility of generic Escherichia coli isolates to ceftiofur, ampicillin and nalidixic acid from retail chicken meat in Canada

Abstract

Statistical modelling of antimicrobial resistance (AMR) data is an important aspect of AMR surveillance programs; however, minimum inhibitory concentration (MIC) data can be challenging to model. The conventional approach is to dichotomize data using established breakpoints, then use logistic regression modelling for analysis. A disadvantage of this approach is a loss of information created by dichotomizing the data. The objectives of the study were to compare the performance and results of different regression models for the analysis of annual variation in susceptibility of generic Escherichia coli (E. coli) isolates to ceftiofur, ampicillin and nalidixic acid from retail chicken meat surveillance samples. E. coli susceptibility data for the three antimicrobials from retail chicken samples from 2007 to 2014 were obtained from the Canadian Integrated Program for Antimicrobial Resistance Surveillance (CIPARS). Annual variation in susceptibility for each antimicrobial was evaluated using multivariable linear, tobit, logistic, multinomial, ordinal and complementary log-log regression models (clog-log). MIC (log2), censored MIC (log2), resistant/susceptible, and categorized MIC (3 or 4 categories) data were used as outcome variables for the appropriate statistical models. Year and region were modelled as categorical predictor variables. Random intercepts were included in the ceftiofur and ampicillin models to account for clustering by retail establishment. The model assumptions evaluated for the mixed models included homoscedasticity and normality of residuals (linear and tobit), homoscedasticity and normality of best linear unbiased predictions (all models), proportional odds (ordinal), and proportional hazards (clog-log). Fixed effects models were used for the nalidixic acid models. The model assumptions evaluated for the fixed effects models included homoscedasticity and normality of residuals (linear and tobit), goodness-of-fit test (logistic and multinomial), proportional odds (ordinal), and proportional hazards (clog-log).Only logistic and multinomial models met model assumptions. Significant annual variation in susceptibility to all three antimicrobials was identified by the multinomial regression models, whereas the logistic regression models only identified significant annual variation in susceptibility to ceftiofur. The multinomial regression model consistently identified additional significant annual variation in susceptibility compared to the logistic regression model. The multinomial modelling approach was able to identify differences between MIC categories within susceptible MIC values, which were below the breakpoint (R) detection level. Given the convention of dichotomizing susceptibility data, the logistic regression approach is likely to remain the standard method of analysis for AMR surveillance data; however, the results of this study demonstrate that multinomial regression should be considered for the analysis of AMR surveillance data.