Author response to LTE JCH‐17‐0110

Dear Editor,

We appreciate the careful reading of our article¹ by Ayubi and Safiri.² The statistical method used in our study was according to chapter 8 (“Logistic Regression”) in The Application of SPSS Version 17.0 in Medical Statistics published in 2010.1 Similar statistical methods can also be found in studies performed by Veltman‐Verhulst,2 Zhang,3 and Ashrafi.4 Therefore, we insist that the results of our current study are indisputable.

However, we would like to thank Ayubi and Safiri for their kind and professional proposals on our methodology.

There are a few points that need to be addressed.

• First, the authors constructed a multivariate model in which only independent variables with P<.05 in the univariate model were included in the multivariate model, which is questionable. In the standard model construction, relaxed univariate P values such as P<.2 are suggested to be used. In fact, the only independent variables with a large effect are included in the multivariate model when rough univariate P values are used in the variable selection. This means that variables with low effects are missed. Steyerberg named this phenomenon Testimation bias. Hence, Xia and colleagues constructed a multivariate model in which variables with P<.05 were only imported into the multivariate model and we wonder why fasting plasma insulin (P=.131), area of insulin under the curve (P=.137), homeostasis model of assessment—insulin resistance (P=.099), and triglycerides (P=.058) were not included in the model.

Response: In the studies mentioned above, independent variables with P<.05 in the univariate model were included in the multivariate model. The criteria of univariate P values were apparently optional. Compared with P<.2, a P value such as P<.05 was more appropriate for us to screen out the most valuable risk factors.

• Second, the sample size calculation is a critical step in epidemiological studies. The minimum required sample size and power for statistical tests is determined in this step. The authors compared risk factors between patients with (n=15) and without (n=77) preeclampsia; however, it is not clear whether the minimum power for statistical tests was provided.

Response: In the study, to detect a difference of 0.80 SD (Cohen's d) in sex hormone–binding globulin between women with and without preeclampsia, at least 86 (n=14 and n=72) patients were included (power of 90% at a 0.05 level). Cohen's d of 0.80 is at the upper limit of what may be considered a medium size effect.

• Third, at least 10 events per variable should be provided per each variable included in the multivariate model to avoid overparameterization and sparse data bias. In the study conducted by Xia and colleagues, only 15 events (preeclampsia) exist; whereas, more than five variables were included in the multivariate model. In fact, some significant associations may not be detected in the multivariate constructed by Xia and colleagues, as the overparameterization and sparse data bias attenuates the statistical power.

Response: In the current study, we performed a preliminary analysis on the risk for preeclampsia with an acquired sample size. Larger prospective cohort studies are needed to identify a clinically useful prediction model such as that by de Wilde5 and an accurate cutoff value for preeclampsia among patients with polycystic ovary syndrome.