Letter to the Editor on “Clinical Characteristics and Mortality Risk Factors of Candida lusitaniae Candidemia: A Multicenter Study in Korea”

Dear Editor,

We read, with great interest, the recent multicenter study by Kim et al. [1], investigating risk factors for mortality due to Candida lusitaniae candidemia. The authors provided valuable epidemiological and clinical data on this emerging pathogen and identified severe sepsis, intensive care unit admission, total parenteral nutrition, moderate-to-severe kidney disease, and non-wild-type (non-WT) or multidrug-resistant isolates as independent risk factors for 30- and 60-day mortality. While these findings are clinically important, we wish to raise important methodological concerns regarding the multivariable analysis that may affect the interpretation of the results.

First, a key principle in developing reliable multivariable regression models, particularly when the number of outcome events is limited, is the events per variable (EPV) criterion. Statistical guidelines often recommend approximately 10 EPV to ensure model stability and prevent overfitting [2–4]. In the study by Kim et al., the 30-day mortality analysis was based on 21 events (39.6% of 53 patients). However, the final multivariate Cox regression model for 30-day mortality incorporated six independent variables (Table 2 in their paper), yielding an EPV of 3.5 (21 events/6 variables), which is substantially lower than the recommended threshold. The EPV for the 60-day mortality analysis (25 events; 47.2% of 53 patients) was likely similarly low. This overfitting risk can inflate hazard ratios and overly optimistic estimates of statistical significance, as seen in the extremely wide confidence intervals for factors such as fluconazole non-WT (adjusted hazard ratio 105.06, 95% confidence interval 4.47–2466.99).

Second, the methodology used for variable selection in the multivariate models could have been described more transparently. The authors stated that “variables with P<0.1” in univariate analysis were considered as candidate variables for multivariate analysis, added in a stepwise manner, and only those with relevant contributions were retained in the final model. Although stepwise selection is common, it is susceptible to type I errors and can result in models that are not generalizable [5]. Presenting the results of the complete univariate analysis for all the considered candidate variables would enhance the interpretability and allow readers to better assess the selection process.

We commend the authors on their important contributions to the understanding of C. lusitaniae candidemia. Addressing these statistical considerations in future studies, perhaps by limiting the number of variables in the model or employing alternative statistical techniques such as penalized regression for smaller datasets, would further strengthen the validity of the findings.