Choice of the Reference Set in a Randomization Test Based on Linear Ranks in the Presence of Missing Values

The connection between the randomization procedure and analysis is often ignored in the statistical analysis of clinical trials although regulatory guidelines explicitly call upon (ICH E8). Randomization tests comply with these requirements, as they use the applied random allocation to determine the distribution of the test statistic. However, the impact of missing values on randomization tests is not studied up to now. We explore the use of a randomization test when analyzing data of randomized clinical trials with missing values. We will consider two scenarios leading to a reduced or non-reduced reference set depending on handling missing data. Investigations of p-values of randomization tests under normal approximation and, for small sample sizes, exact randomization tests were conducted.