Two Thirds of Whom

What the record keeps

The Survey of Scottish Witchcraft is a database of the people its team found accused of witchcraft in Scotland, built by Julian Goodare, Lauren Martin, Joyce Miller and Louise Yeoman from surviving legal and church records and published by the University of Edinburgh. For each of 3,212 named people it keeps what the documents hold: a trial, a verdict, a sentence, a mark that an execution took place. Where the documents are silent, the database is silent too, and its documentation says how that silence is to be read: If no evidence was found, ‘no’ is entered, but this should be read as ‘unknown’.

The record gives a death for 249 named people. For 2,687 it records no outcome of any kind.

The number of those executed was probably much higher than those for whom we have definite records of their execution.

First, the published method, reproduced

The Survey's own FAQ gives the rate and doubts it in the same answer. After its list of known sentences it says This seems to suggest that 67%, two-thirds, were executed. Then: But this figure is probably not very accurate. It names two reasons that pull in opposite directions. The known sentences are unrepresentative because most of them come from trials in the central justiciary court, which probably acquitted more often than local courts; and our 3,212 people include a number whom we found being investigated by the church authorities. Probably some of these went on to receive a criminal trial, and may then have been executed; but others' cases were probably dropped before they came to trial.

Liv Helene Willumsen and Terje Vassdal took the question further in 2025. They divided the recorded sentences by type of trial, by panic and non-panic years and by three periods, estimated each part separately, and put a confidence interval around every total. Their method stands on an assumption they state plainly: We assume that the surviving information on verdict and sentence has not been deliberately disposed of. Therefore, it is appropriate to use the method of statistical random sampling to analyse these documents. Their paper is what makes a check like this one possible. The page runs their method before anything else, with the same engine it uses everywhere.

From the frozen export, the engine reproduces 112 of the 117 values the article prints in its tables and in its text exactly, 3 more to within 0.03, and 2 differ. The article's verdict-by-sentence table also leaves 34 cells blank. A blank is not a printed value, so none of them is counted above, and each is shown as blank in the article. Nothing the authors computed is fed to the engine.

By type of trial it finds the rates they print. Of central trial rows with a sentence, 60 of 102 are sentences of death (58.82 per cent); local, 101 of 133 (75.94); mixed central and local, 45 of 72 (62.50). Applied to the trial rows of each type, those rates give 1,758. The 803 remaining rows have no central, local or mixed trial process, and the authors give them the average rate of the rest (We assume here that missing cases have the same relative proportion of executions as known cases.), which gives 2,345.

Whom their count of named people counts

The article describes its population this way: When counting the number of persons in the Sentence table in the SSW, two persons cannot be accounted for. The total number of accused named individuals is thus 3,210. The export gives exactly that number by another route. 3,210 is the number of trial rows attached to named cases, and the article's verdict-by-sentence table and its type-of-trial table both add up to it over trial rows. Those rows belong to 2,835 people: 375 are a second or third trial row for the same person, and 377 named people have no trial row at all. Counted as people, the named total is 3,212. For the totals this barely matters (2,345 over rows, 2,348 over people). For the question of whom, it is the whole matter.

What did not come out exactly, and why

• Panic years. The article takes its panic years from Willumsen's 2008 thesis, p. 40, which this page has not read. The FAQ's own panic years give different counts; they are one of the strata below, labelled as the FAQ's. A declared search over 186,624 candidate sets of years found 1 that gives all six counts of the panic-year table: 1590 to 1591, 1597, 1628 to 1630, 1643 to 1644, 1649 to 1650, 1658 to 1659, 1661 to 1662. It is a reconstruction found by search, not the authors' definition, and it is labelled that way wherever it is used.
• Standard errors. The article prints 81.87, 74.36 and 73.25. Its own formula on the reproduced counts gives 81.86, 74.33 and 73.23. The type-of-trial value, 70.98, agrees exactly.
• One interval. The article prints 2,202 to 2,488 around 2,345. Plus or minus 1.96 times that standard error gives 2,206 to 2,484. The panic-year and period intervals reproduce.
• A label. The article describes the central-trial 102 as persons who received a verdict of Guilty. In the export, 102 is the number of central trial rows with any sentence, and central trial rows with a Guilty verdict number 63. The rates in the table use the first count.
• The FAQ's list. Its 305 known sentences (205 to be executed, 52 released, 27 banished, 11 declared fugitive, 6 excommunicated, 2 put to the horn, 1 prison, 1 public humiliation) are reproduced by counting trial rows and leaving out the export's one Hang row and one Branded row. Counted by case instead, banishment is 26, although the number of cases with any sentence is again 305. The FAQ calls the 52 acquitted; the record's word is Released, and 43 of those cases carry no verdict, 5 Not Guilty and 4 Not Proven. These rows show that the parser reads the Survey's own tallies. They are not independent of the data.

What the confidence interval measures

The sophisticated replyHistorians know two thirds is uncertain. The Survey says so, and Willumsen and Vassdal answered it with stratified sampling and a confidence interval. The published totals are about right.

Their interval measures sampling variation: how far a total would move if a different set of sentences, of the same size, had happened to survive, on their stated assumption that the surviving records are a random sample of all of them. The authors put it exactly: The sample of 307 is by nature arbitrary. It could have been some other 307 records that had survived. What it does not measure is whether the people the rates are applied to, but whose records hold no trial, share the rate of the people whose sentences survive.

Of the 3,212 named people, 839 have no central, local or mixed trial process on record. For 462, every trial row is of the kind the page reads as the documentation's record of someone mentioned in another person's trial; 377 have no trial row at all. None of the 839 has a trial date, a verdict, a sentence, an arrest, a flight or an execution mark on record. No division of the sentences can say anything about them. Every total built from a rate gives them one by assumption, and Willumsen and Vassdal, to their credit, state theirs.

Most accused people have no outcome recorded: 2,687 of 3,212, while 249 have a recorded death. In trial rows, giving 803 untried rows the rate of the rest raises their total from 1,758 to 2,345, beyond twice the published interval.

Those two totals, 2,345 and 1,758, are 587 apart, and the published interval, 2,202 to 2,488, spans 286. All four are counted in trial rows. At this page's defaults the same trial-type rates are applied to all 3,212 named people instead. The total then runs from 1,735, if the 839 people with no trial process on record are given none of the rate, to 2,348, if they are given the rate of the rest: a gap of 613 people. That gap is not set beside the published interval, because the two are counted in different units, and the instrument above draws no sampling interval for settings that mix them. Counted in named people for the rates as well as the population, the gap is 604 people (1,711 to 2,315), and the authors' formula, applied by this page in that unit, gives a sampling interval 278 wide.

The interval the record itself supports

For named people the record sets its own limits, with no sampling assumption. The lowest total consistent with it is the number of people for whom it gives a death: 249 by a sentence of execution or the mark that an execution took place, 186 if both are required. The highest is 3,070: everyone except those with a recorded sentence of another kind or a verdict of Not Guilty or Not Proven and no sentence, or 2,961 if flight and dropped action are read as the end of a case. Every published total on this page lies inside those limits. Where inside them is decided by what is assumed about people the record does not follow to the end.

Why dividing the sentences raises the totals

Choose decade above and the reason is visible. Sentences survive most densely where sentences of death were rarest. The record names 85 people in cases of the 1700s, with 33 known sentences, 2 of them sentences of death. It names 594 in the 1640s, with 17 known sentences, 14 of death, and 674 in the 1660s, with 38, 33 of death. One rate for everyone lets the well-documented later cases speak for the crowded middle years. The authors saw the same direction in their periods: The stratification according to time periods clearly shows that the expected execution rate is highest in the period with the most cases.

People, not rates

The rates are made of people, and the record keeps their names in its own spelling. Choose a group to read each person's record as the Survey holds it. Every stage is shown as recorded, or as not recorded. A zero in a yes-or-no field is shown as not recorded, because that is what the documentation says it means.

What this instrument will not say

Each button asks for something the record cannot support. The answer is the engine's own refusal.

Sources, licence and limits

• Goodare, Julian; Yeoman, Louise; Martin, Lauren; Miller, Joyce. (2010). Survey of Scottish Witchcraft, 1563 - 1736, [Dataset]. University of Edinburgh. School of History, Classics and Archaeology. https://doi.org/10.7488/ds/100.
• Julian Goodare, Lauren Martin, Joyce Miller and Louise Yeoman, 'The Survey of Scottish Witchcraft', www.arts.ed.ac.uk/witches/ (archived January 2003, accessed 14 September 2026), the form of citation the Survey's documentation asks for; the site now lives at the University of Edinburgh.
• Liv Helene Willumsen and Terje Vassdal, “Execution rates in Scottish witchcraft trials”, Cogent Arts & Humanities 12(1), 2025, article 2495428, doi:10.1080/23311983.2025.2495428. The article.
• The Survey of Scottish Witchcraft, Introduction to Scottish Witchcraft (FAQ). The FAQ.
• Survey of Scottish Witchcraft, Database Documents (Witchcraftsurvey_download.pdf), in the DataShare deposit.
• Scottish Government, FOI release 202200295186 (Pardoning of witches: FOI review), information released, published January 2024. Its note on the numbers carries its caveat in a footnote: the figure of 3,212 named individuals includes a number who were investigated by the church authorities and some of these cases were probably dropped before they came to trial suggesting the actual number who were executed may be lower.
• Jane Recker, “Scotland Issues Formal Apology to Thousands Accused of Witchcraft”, Smithsonian Magazine, 5 April 2022.
• History News Network, “Scotland to Pardon Thousands Executed as Witches”, reprinting the Wall Street Journal, January 2022.

Licence

The Survey's data are used under the Creative Commons Attribution 4.0 International licence of the University of Edinburgh's DataShare deposit, with the attribution above. The page ships a derived extract: three of the deposit's tables, some of their columns, joined on the accused reference, with every value left as exported. The documentation in the same deposit also carries the website's older terms, The website and database and all of their contents are the copyright of the University of Edinburgh and reproduction is only permitted in accordance with the following terms, which allow use for private reference, research or study. This page follows the deposit's licence, and records that discrepancy here and in LICENCE.txt. Quotations from the FAQ, the documentation, the article, the Scottish Government release and the press are short and cited.

What this page cannot do

• It cannot say how many people were put to death. No total on it is a count.
• It cannot test whether the surviving sentences are a fair sample. It shows what each total assumes if they are.
• It does not give the unnamed groups a rate, because the record gives them no sizes it can trust and no outcomes.
• It does not set occupation, age or ProcessType against outcomes; the record holds too little of each.
• It works from the DataShare export. The online database may since have changed.

We searched the Survey of Scottish Witchcraft FAQ and documentation, Willumsen and Vassdal 2025, the Scottish Government's FOI release on the pardon review, press coverage of the 2022 apology, web searches for the 2,500 and two-thirds figures, and the Artificial Wasteland corpus on 2026-09-14, and did not find a published account of which population each circulating execution figure is applied to, of the 839 named accused with no trial process on record inside those populations, or of the 43 named people whose record says an execution took place while recording no sentence.