Statistical Analysis

A number of patients in our cohort were missing data on LDH, clinical T classification, and albumin. Patients who were evaluated preoperatively only by a surgeon were often missing LDH data, as surgeons at our institution order this test less commonly than medical oncologists do. Patients who had surgery before 1996 were often missing albumin measurements. We used multiple imputation by chained equations to impute values for the following variables: elevated lactate dehydrogenase (LDH) (N=131), clinical T3 disease or higher (N=70), low albumin (N=42), pathologic N classification (N=28), undefined race (N=8), liver metastasis (N=5), retroperitoneal lymphadenopathy (N=3), supradiaphragmatic lymphadenopathy (N=3), metastatic symptoms at presentation (N=3), and pathologic classification T3 or higher (N=2). This method of multiple imputation fills in missing values in multiple variables iteratively by using chained equations and accommodates arbitrary missing-value patterns¹³. Statistical analyses were performed utilizing the measured and imputed values combined across 10 imputations using Rubin’s rules. As a sensitivity analysis, we repeated these analyses using an additional imputed dataset and the original, non-imputed data. All analyses were conducted using Stata 13 (Stata Corporation, College Station, TX).

We then evaluated the preoperative nomogram predicting the risk of death at 6 and 12 mo, respectively, after CN. In our study, 280 patients were included in the analysis of the preoperative model as some patients were lost to follow-up before the endpoint could be determined. We reported model discrimination using the area under the curve (AUC) and created calibration plots (Fig. 2). Since there have not been previous studies regarding an optimal risk cut-point for CN, after discussion among several surgical experts we chose a 20% threshold for risk of death after CN, meaning that patients with a greater than 20% risk of death would not be recommended for surgery. This cut-point was used to assess clinical utility of the preoperative model, as this model could be used for counseling about whether or not to perform surgery. We recorded the number of patients who had a risk of death at 6 mo after surgery of less than 20% but died during that time period. We also used decision curve analysis to assess the clinical utility of the model for risk thresholds up to 30%.

Calibration plot for preoperative nomogram using imputed data to predict overall survival at 6 mo after surgery (N=280).