Baseline characteristics and outcomes are described as one large group and divided by era and guideline adherence. Continuous data are expressed as mean (± SD) for normal distributions or median (plus quartiles) and compared with t tests or Mann-Whitney U tests. Baseline categorical data were compared with X2 tests (Yates corrected for 2×2 tables). Differences between empirical antimicrobial regimens were analyzed using %2 tests for categorical data and analyses of variance or Kruskall-Wallis tests for continuous data. Multivariable analyses consisted of linear and logistical regression for continuous and binary outcomes, respectively. Non-normally distributed explanatory variables were transformed with the natural logarithm. All multivariable analyses were controlled for pneumonia severity index (PSI) score and study site, and used backward selection procedures. To avoid overfitting, fewer than 10 explanatory variables per number of subjects or number of outcomes were introduced into automated selection procedures. No formal power calculations were performed.
Given that the PSI was derived for the prediction of mortality, multivariable analyses for LOS were also controlled for age, sex, long-term care residence, diabetes, congestive heart failure, cancer, or chronic lung, liver or renal disease. Morality was also controlled for in the analysis of LOS, drug costs and duration of IV therapy (high early mortality could result in shorter LOS, lower costs and shorter duration of IV therapy). All statistical analyses were performed using SAS version 8.0 (SAS Institute, USA).