Abstract

Background: A challenge in the treatment of resistant bacteria has been the difficulty in identifying patients likely to be infected with such pathogens. Novel methods may be applied to surveillance data to determine patient- and institution-specific factors predictive of increased MIC. The censored nature of some MIC values (e.g. MIC ≤ 0.5 or MIC > 8) is a difficulty for SR analyses. Simulations were performed to compare CR versus SR in which MICs of the form MIC ≤ 2L or MIC > 2R (left- or right-censored) were replaced with specific values or excluded.

Methods: Using a model relating MIC of piperacillin-tazobactam for Enterobacter species to categories of patient age and hospital bed size, 200 simulations of 500 isolates were performed. Various MIC censoring patterns were imposed using 26 (L, R) pairs. Data were fit with CR, and with SR using 3 procedures: (1) censored MIC excluded, (2) censored MIC replaced by 2L or 2R, and (3) censored MIC replaced by 2L-1 or 2R+1.

Results: Censoring for the 26 pairs ranged from 7-86%. Using CR, average deviations from true parameter values were less than 0.10 log2 (mg/L) for all parameters and all (L,R) pairs, whereas for 7 of 8 parameters the average deviations were less than 0.10 log2 (mg/L) for less than 31% of (L,R) pairs. Coverage percentage of 2 standard error (SE) confidence intervals was as low as 0% for all SR methods, but not below 91.5% for CR.

Conclusions: When modeling censored outcomes such as MIC, CR is preferable to SR analyses to avoid biased parameter estimates.

American Society for Clinical Pharmacology and Therapeutics (ASCPT), Washington, DC, April 2003

By JP Hammel, SM Bhavnani, PG Ambrose, RN Jones, A Forrest, MR Piedmonte