Comparison of different methodological approaches to identify risk factors of nosocomial infection in intensive care units.

Abstract

Comparison of statistical methods and measurement scales to identify nosocomial infection risk factors in intensive care units (ICU). Prospective study in 558 patients admitted to the ICU of a referral hospital between February and November 1994. Analysis using three logistic regression models, three standard Cox regression models, and two Cox regression models with time-dependent extrinsic factors. Different scales were used to measure exposures to risk factors (dichotomous, ordinal, quantitative, and time-dependent variables). The most appropriate models were those that measured exposure using dichotomous variables. Models using ordinal or quantitative variables estimated biased coefficients and/or failed to comply with the statistical assumptions underlying the analyses. The Cox regression model with quantitative time-dependent variables met all the statistical assumptions, obtained a precise assessment of risk by exposure time, and estimated unbiased coefficients. The Cox regression analysis with quantitative time-dependent variables is the most valid alternative for assessing the risk of nosocomial infection per day of exposure to an extrinsic risk factor in the ICU.