General study characteristics. Of the 24 studies, 21 (87.5%) made use of existing data, while three (12.5%) applied the methods to simulated datasets. Descriptive statistics are shown in Table 5.2. The median total sample size was 920 patients, the median number of predictors was 7 (low-dimensional data), and the median percentage of censoring was 70.8% (10 of 24 studies considered multiple outcomes). Medical applications were mainly in the field of oncology (73.5%, 25 datasets). The majority of these studies conducted research on breast cancer (10 datasets), and cervical cancer, gastric cancer or prostate cancer (2 datasets each). Other fields of application comprised cardiovascular disease, coronary artery disease, liver transplantation and post-partum amenorrhea (2, 1, 1, and 1 datasets, respectively). Clinical endpoints of interest included overall survival (analysed 16 times, 47.1%) and disease-free (or progression free, recurrence-free, relapse-free) survival (analysed 12 times, 35.3%). Remaining endpoints were breast cancer specific mortality (5.9%), death or hospitalization due to cardiovascular events (5.9%), menstruation-free survival (2.9%) and time to clinical artery events (2.9%). Min 1st Qu. Median 3rd Qu. Max Excel lines Total sample size 96 242 920 1616 361239 33 # of predictors 1 5 7 25.75 97 32 % of events 6.60 21.32 29.25 47.58 97.90 20 Table 5.2: General characteristics for the 24 studies. If multiple outcomes were predicted, multiple lines were used in the extraction sheet. Maximum number of lines was 34 (10 studies used multiple outcomes). For simulation studies, the number of predictors and percentage of events were not considered, unless they were fixed (e.g., not varied across simulations). The strategy used to address the missing data (if any) was unclear for 9/21 (42.9%) studies (disregarding the 3 simulation studies that did not contain any missing data). Single or multiple stochastic imputation was used for 6 studies (28.6%) and ad-hoc approaches (separate attribute or mean / median imputation) were used in 5 studies (23.8%). One dataset had no missing data. Ad-hoc approaches to missing data can be problematic, as they can alter the distribution of a variable (if there is a substantial number of missing values). Multiple stochastic imputation, which replaces each missing value with multiple plausible values, is the preferable option [52], as the variability in multiple predictions reflects the uncertainty of the imputation process. It is understandable that ...
