Existing models of mental health (e.g., depression) statuses prediction from social media text are usually built from training data with no distinction between publications made before and after diagnosis. However, once individuals are made aware of a mental health disorder diagnosis, changes in discourse are likely to occur, which may, in turn, simplify the computational task and/or make it potentially less useful for practical purposes. Based on these observations, this article presents an analysis of word- and topic-level features for the prediction of mental health statuses from data produced before and after diagnosis, and compares the results of text classifiers built from both sources. Results confirm our initial assumption that diagnosis awareness may affect both the computational task and its practical application, and motivate the use of more curated training data in these settings.