How Stylometric Evidence Is Weighed

Stylometric results are weighed on this site, though stylometrics is not a line on the convergence table. What follows explains what the method can establish, which kinds of study are circular and which are not, and how the results that are not circular are weighed, including results that count against candidates this site assesses.

What stylometrics can establish

Computer-assisted authorship analysis has contributed a good deal to early modern literary scholarship. Its best-established application is collaborative authorship: identifying where stylistic patterns shift within a text attributed to a single author, suggesting a second or third hand. The finding that Marlowe contributed to the Henry VI plays, that Fletcher co-wrote Henry VIII, that Middleton revised Macbeth; these are stylometric findings that have entered mainstream scholarship, and they are incorporated into this site’s collaborative authorship framework. In this application the method is not circular: it identifies a shift in stylistic signature and tests it against the signed works of known contemporaries.

The same collaborative-authorship methodology also supports one finding that this site accepts without qualification: the canon (minus the identified late collaborations) is overwhelmingly the work of a single mind. The stylistic consistency across the uncontested plays is a finding about the plays’ internal coherence, which a circular corpus argument could not produce. Fletcher and a small number of other late collaborators are identifiable precisely because they interrupt a signature that is otherwise remarkably stable across three decades and multiple genres. Whatever the authorship question ultimately resolves to, it resolves to a single primary author for the great bulk of the work.

The circular corpus problem

The most common stylometric claim in the authorship debate takes a different form: measuring the plays against a “Shakespeare corpus” to confirm that they sound like Shakespeare. This is circular. The canon is defined by the traditional attribution; a method trained on that canon and used to confirm it is measuring consistency with itself.

Ros Barber’s peer-reviewed analysis (Digital Scholarship in the Humanities, Oxford Academic, 2021) showed that the Zeta test methodology, the basis for the most widely cited stylometric claims in the debate, cannot correctly attribute Hamlet to Shakespeare, and a test that fails on the canon’s most certain case cannot carry weight on contested ones.

Not every stylometric study has this problem, because the circularity comes from what a study measures against. Mosteller and Wallace’s study of the disputed Federalist essays compared the undisputed writing of the two men the records named, tested its method on essays of known authorship before applying it, and reported its results as odds. Evidence of that kind is circumstantial, and What Would Settle It sets out five conditions any circumstantial case must meet before this site treats it as identifying an author: its facts come from primary sources, each fact comes with an estimate of how many other people it would also fit, facts with one cause count once, the method is first tested on people nobody proposes as the author, and the facts together are rare enough that no one else would be expected to fit them all. The Federalist study is the page’s example of all five being met.

Comparing the candidates’ own writing

A study that trains on a candidate’s own signed writing and compares it with the works presupposes nothing about who wrote them. It is circumstantial evidence like any other, and it has to meet the same five conditions. It can count for a candidate or against one. Craig and Kinney’s Shakespeare, Computers, and the Mystery of Authorship (Cambridge University Press, 2009) and MacDonald Jackson’s comparative analyses produce results more consistent with the traditional attribution than with the major alternative candidates.

The Claremont tests

The largest study of this kind is the Claremont Shakespeare Clinic, run by Ward Elliott and Robert Valenza. Their 1996 report applied 51 tests of play authorship and 14 of poem authorship to 37 proposed claimants and found that “no claimant … matched Shakespeare” (Computers and the Humanities 30, 1996). The claimants tested include Oxford, Bacon and Marlowe.

The usual objection is that most candidates left too little comparable writing for a reliable test. Elliott and Valenza addressed it directly for Oxford. His surviving poems amount to about 3,000 words of testable verse, so they cut Shakespeare’s own verse into blocks of the same size, and into 1,500-word blocks matched for metre, and ran the tests on both. Shakespeare’s blocks fell within his range; Oxford’s poems were rejected strongly on seven of the fourteen tests. They also found no sign of stylistic development in Oxford’s poems between his ages of 22 and 44, which answers the objection that the poems are juvenilia (“Can the Oxford Candidacy be Saved?”, 2000).

The results count against the candidates tested, under this site’s rule for negative evidence: they lower a case without ruling it out, and any explanation offered in reply has to be named as one. Elliott and Valenza list the factors their tests are sensitive to (genre, editing, prosody, subject matter and date of composition) and say they can seldom match for all of them at once. Those are the explanations open to an advocate, and a case that relies on one should say which. The Clinic’s methods have had published critics, and Elliott and Valenza have replied to them.

What the tests cannot reach

The tests compare a claimant’s writing with the works. William Shakspere of Stratford left no writing to compare, apart from six signatures, so the tests say nothing about him either way. A reader who takes the Claremont results as confirming the traditional attribution has added that step; the tests exclude claimants and cannot identify anyone.

Why it is not a convergence line

The convergence table records documented facts about each candidate, one mark per line: where they went, what they could read, whom they knew, and what kinds of work they are known to have written. Stylometric results measure how closely a candidate’s own writing matches the works, with their own margins of error, and a mark on a line cannot carry them. They appear on this page, and on the pages of the candidates tested, as evidence to be weighed under the five conditions.

No investigative or scholarly field treats stylometrics as a necessary condition for establishing authorship. Forensic linguistics, corpus linguistics and literary attribution present it as one tool among several, and its practitioners describe its results as probabilistic. In early modern drama, where collaboration, scribal copying and printing-house habits blur the signal, it needs particular care. That is why this site weighs each study on its own design.