Response to the Letter: "Bias estimation of predictors and internal validity of the study 'Admission characteristics predictive of in-hospital death from hospital-acquired sepsis: A comparison to community-acquired sepsis'".
Response to the Letter: "Bias estimation of predictors and internal validity of the study 'Admission characteristics predictive of in-hospital death from hospital-acquired sepsis: A comparison to community-acquired sepsis'".
复制标题
对这封信的回应：“‘预测医院获得性脓毒症院内死亡的入院特征：与社区获得性脓毒症的比较’研究的预测因素和内部有效性的偏差估计”。
DOI：
10.1016/j.jcrc.2019.04.023
复制
发表时间：
2020
影响因子：
3.7
通讯作者：
Guirgis,FaheemW
中科院分区：
文献类型：
--
作者：
Gautam,Shiva;Smotherman,Carmen;Guirgis,FaheemW
We want to express our sincere thanks to Drs. Yarandi and Panahi for raising a few important points related to our recent paper [1]. We agree that a large odds ratio (and a small p-value) does not necessarily translate into a strong association and we did not emphasize the strengths of individual associations based on these numbers. In fact, our team discussed several strategies with which to analyze the data prior to settling on the approach presented in the paper.Because of the repeated encounters in sepsis admissions, we initially gravitated towards a generalized mixed model with patients as a random factor instead of simple logistic regression. When we started to explore the nature of the data, we suspected quasi separation and possible upward bias of the estimates. We had the option to ignore the repeated nature of the data and use the penalized likelihood approach (eg Firth correction) under logistic regression [2]. We also thought of dropping repeat patients, but it would require us to ignore or discard a portion of information available in both of these approaches that we thought was unacceptable. In this study, our primary focus was to identify possible baseline characteristics that help predict the outcome rather than evaluate the strength of the associations of individual variables in isolation. From the prediction viewpoint, even the plain maximum likelihood’s performance is expected to be similar to that of any penalized likelihood’s performance if the number of ‘events per variable (EPV)’is 10 or more [3]. In our study, we had 140 events and 10 observations. We would also like to point out that Firth
