Forensic handwriting examiners are often tasked with identifying the writer of a particular document. Examples of handwriting evidence include ransom notes, forged documents and signatures, and threatening letters. At present, examiners rely on visual inspection of similarities and differences between the questioned document and reference writing samples. Here, we propose a principled modeling approach to compute the posterior predictive probability of writership when the author of the questioned document is part of a closed set of writers. Given a handwritten document, we extract document-level and character-level measurements which are the response variables in a multi-level model. We fit the model and test its posterior predictive performance using writing samples from the United States and from Europe. We find that as long as the questioned document is longer than a sentence or two, it is possible to correctly associate a writer with a document that he or she wrote with high probability. Earlier versions of this work have been well received by the community of forensic document examiners.