I am planning an fMRI experiment which looks at similarity of human brain activation when prompted with known words.
The goal of the experiment is to compare the similarity of activation with predictions made by a model. To get a good comparison, we would ideally want an even spread of similarity between stimulus words.
Is there a potential issue with using the same model to select stimuli of a range of predicted similarities?
Note that the analysis will not make presumptions about similarity of words other than what is shown from the data, so we should avoid circular analysis (Kriegeskorte et al, 2009). The model predictions don't influence data analysis until the final comparison between observed and predicted similarities in activation. It still feels intuitively like there might be an issue of false findings through the stimulus selection, but I cannot find/verbalize an actual problem.
An alternative would be to use existing literature to identify a word list with an even spread of similarity between words. But as our model is of course based on that literature, in practice this would result in a very similar list of words.