This paper describes an experiment to evaluate the performance of three different types of semantic vectors or word embeddings-random indexing, GloVe, and ELMo-and two different classification architectures-linear regression and multi-layer perceptrons-for the specific task of identifying authors with eating disorders from writings they publish on a discussion forum. The task requires the classifier to process texts written by the authors in the sequence they were published, and to identify authors likely to be at risk of suffering from eating disorders as early as possible. The data are part of the eRISK evaluation task of CLEF 2019 and evaluated according to the eRISK metrics. Contrary to our expectations, we did not observe a clear-cut advantage using the recently popular contextualized ELMo vectors over the commonly used and much more light-weight GloVe vectors, or the more handily learnable random indexing vectors.