user2016

Poster #25

Resistant and multi-resistant bacteria are seriously affecting modern health care. Therefore, it is critically important to understand and control the increased resistance. Binomial data with resistant or non-resistant values, where the resistance proportion is bounded between 0 and 1, can be modelled by logistic regression.  In its simplest form, the log odds of resistance equals a linear equation with one variable representing sales and two parameters for the intercept and slope, respectively. We used the glm-function with family = quasibinomial to account for larger variances (overdispersion), than expected from the binomial distribution. The R-packages visreg and ggplot2 were used, respectively, to calculate the regression curves with confidence limits, and to visualize the results. Resistance against 7 antibiotics for E. Coli isolates (n = 210) in the Faroe Islands, 2009-2012, was compared with the corresponding antibacterial mean-sales, 2008-2011. A prop.trend.test for trend is clearly significant (p-value < 2.2e-16), while the logistic regression is highly overdispersed (≈ 32) indicating low model fit (slope p-value = 0.05). Considering different biological resistance mechanisms, we exclude resistance outliers and extreme sales to minimize overdispersion (≈ 1), and find for 5 of 7 resistances the parameter 3.24 for slope (Std. Error = 0.226, p-value = 0.0007), OR [95% CI] = 25.5 [16.6, 40.4].Similarly, including data from Iceland and Denmark, we show how about 7 of 18 antibiotic resistances in the 3 countries closely follow a logistic prediction for increased resistance with sales, while we also detect different mechanisms for the remaining resistances.