When dealing with non-linear data, since the reliability of chi-squared tests diminishes with the number of samples, is it reasonable to divide a large dataset into sample spaces of, say, 20 or 40, calculate chi-squared and the associated probabilities on each sample space and then average the resulting probabilities together? Having an averaged probability for each feature, can one then select the features appropriately?
My logic is that since the data is non-linear, the data is not related across samples, only within a sample. Is my assumption correct?
Thanks in Advance!