Defining an optimal cut-off for subscales without a gold standard: a novel method using inflection points of latent attribute probabilities and raw scores