Individuals with reduced speech and motor ability due to injury or motor neuron diseases have difficulties in communication. Such events at the terminal stage lead to the loss of all muscular activity which is referred as locked-in state. In this condition, communication can only be performed with electroencephalogram (EEG) signals. Brain computer interfaces (BCI) typing systems provide people without muscular control a communication baseline. In BCI typing systems, user is presented stimuli and corresponding EEG evidence is used to detect the user intent among a pre-defined alphabet. Due to low signal-to-noise ratio (SNR) of EEG evidence, multiple stimuli sequences are required. Thus, to limit the time spent on typing the stimuli presented to the user should be optimized. BCI typing systems are designed to operate including different modalities where modalities surpass each other either in SNR of the respective signal or time spent to acquire the response. Therefore, it is a fundamental problem when to choose cheap in time - weak in response questions and when to choose expensive in time - strong in response questions. In this study we propose a modality selection mechanism for systems that rely on recursive evidence collections under Gaussian evidence model assumption. Specifically, we focus on BCI typing systems that operate with error related potentials (ERPs) and feedback related potentials (FRPs). We analytically derive a decision threshold to select each of these modalities. We also demonstrate the performance of the proposed method using a BCI typing system.