Confounding and the analysis of multiple variables in hospital epidemiology.

Abstract

Most information in hospital epidemiology comes from observational studies of hospitalized patients rather than planned experiments, and in such observational studies the characteristics of study patients may vary widely, even within a single hospital. Any comparison between hospital populations will usually contain additional, unintended contrasts among patients with varying degrees of health. Adult patients, for example, may have vastly different underlying diseases, and infants may be of substantially different birth weights. We used both underlying disease and birth weight as indices of the basic severity of illness in order to adjust for confounding by differences in underlying disease in reanalyses of several published studies. We give an example in which differing birth weights among groups of infants compared artifactually double the apparent effect of nosocomial infections as a cause of mortality, and another example in which differing degrees of severity of underlying illness artifactually halve the apparent effect of appropriate antibiotics in preventing death from bacteremia with gram-negative bacilli. We describe simple intuitive methods based on stratification, adapted from chronic disease epidemiology, to remove confounding effects during analyses.