Some Doubts on the Study of Clinical Prognoses of Patients with a Bloodstream Infection Caused by Ampicillin-Susceptible but Penicillin-Resistant Enterococcus faecalis

LETTER

We read with great interest the article in an issue of Antimicrobial Agents and Chemotherapy by Kim et al. (1). The article concluded that the mortality rates of patients stratified by penicillin susceptibility might be relevant to the treatment failures of ampicillin and/or piperacillin in patients with an ampicillin-susceptible but penicillin-resistant (ASPR) Enterococcus faecalis bloodstream infection (BSI). Though an interesting and valuable study has been conducted, some methodological points need to be taken into account.

First, the values of crude effect estimates should change ≥10% after adjusting for other covariates in the multivariable model (2, 3). However, as shown in Table 1 in the article (1), for some predictors, the differences between the odds ratio (OR) values in univariate and multivariable models is far lower than the 10% cutoff. For example, the crude OR of cardiovascular disease changed only by 0.35% ([(OR crude – OR adjusted)/OR adjusted × 100%] = [(2.97 – 2.98)/2.98 × 100%] = 0.34%); thus, biases might exist in the confounder adjustment.

Second, in the study by Kim et al. (1), variables with a P value of <0.05 in the univariate analysis were analyzed whether to be included in the multivariate model. Nevertheless, only variables with P < 0.05 were imported into the multivariate model, which is questionable, as this strategy could lead to a phenomenon called testimation bias, which infers that only variables with a large effect were included and variables with a small effect did not enter into the model. This type of bias can be decreased when independent variables are imported into the multivariate model based on a P value of ≥0.2 (4, 5).

Third, Kim and colleagues applied Pearson’s correlation coefficients to describe the relationship between penicillin MIC and 30-day mortality rate. However, Pearson’s correlation should be used under the circumstance that both variables being studied are normally distributed. Thus, we suggest the authors carry out the Shapiro-Wilk test to examine whether the normality of distribution of the two variables is qualified; otherwise, the Spearman correlation should be used to investigate the correlation between two variables.