New statistical technique for analyzing MIC-based susceptibility data

Abstract

Seventeen laboratories participated in a cooperative study to validate the regional susceptibility testing of Neisseria gonorrhoeae in The Netherlands. International reference strains were distributed. Each laboratory determined the MICs of ciprofloxacin, penicillin, and tetracycline, for each strain by Etest. To explore a more transparent assessment of quality and comparability, a statistical regression model was fitted to the data that accounted for the censoring of the MICs. The mean MICs found by all of the laboratories except three were closer than one 2-fold dilution step to the overall mean, and the mean MICs of each antimicrobial agent were close to the MICs for the international reference strains. This approach provided an efficient tool to analyze the performance of the Dutch decentralized gonococcal resistance monitoring system and confirmed good and comparable standards.

Fig 1

Illustration of censored data. Variable y has a normal distribution with a mean of 1 and a standard deviation of 2. Thirty values are shown at the bottom. Values of <−2 are left censored, values of >4 are right censored, and values in between are interval censored.

Fig 2

One thousand runs of 30 samples where the mean (left panels) and standard error of the mean (right panels) were estimated for each run (black dots). The “actual” values are based on the actual values, y_act; “naive” values are based on the observed values, y_obs (ignoring censoring); and “proper” values are based on the observed values, y_obs (accounting for censoring). The true mean (= 1) and standard error (= 2/√30) are indicated by diamonds. The means of the 1,000 estimates are indicated by circles.

Fig 3

Laboratory-specific fold differences (and 95% confidence intervals) from the expected mean MIC. A 1-fold difference is no difference [log₂(1) = 0].

Fig 4

Predicted mean MICs and 95% confidence intervals (solid squares), mode MICs (open squares), and reference MICs (solid circles), with their corresponding lower boundary accounting for interval censoring (open circles), for each combination of strain and antimicrobial agent.