1 Verbin, J. S. Kloppenborg, Excavating Q: The History and Setting of the Sayings Gospel (Minneapolis: Fortress, 2000)Google Scholar.

2 Goodacre, M. S., The Case Against Q: Studies in Markan Priority and the Synoptic Problem (Harrisburg, PA: Trinity, 2002)Google Scholar.

3 Robinson, J. M. et al. , eds., The Critical Edition of Q: Synopsis including the Gospels of Matthew and Luke, Mark and Thomas with English, German, and French Translations of Q and Thomas (Minneapolis: Fortress; Leuven: Peeters, 2000)Google Scholar.

4 Casey, P. M., An Aramaic Approach to Q (Cambridge: Cambridge University, 2002)CrossRefGoogle Scholar.

5 Foster, P., ‘Is it Possible to Dispense with Q?’, NovT 45 (2003) 313–37Google Scholar; Kloppenborg, J. S., ‘On Dispensing with Q? Goodacre on the Relation of Luke to Matthew’, NTS 49.2 (2003) 210–36CrossRefGoogle Scholar; see also Kloppenborg, J. S., ‘Variation in the Reproduction of the Double Tradition and an Oral Q?’ ETL 83 (2007) 53–80Google Scholar.

6 Watson, F., ‘Q as Hypothesis: A Study in Methodology’, NTS 55.4 (2009) 397–415CrossRefGoogle Scholar.

7 Burkett, D., Rethinking the Gospel Sources. Vol. 2. The Unity and Plurality of Q (Atlanta: SBL, 2009)Google Scholar.

8 Poirier, J. C., ‘Statistical Studies of the Verbal Agreements and their Impact on the Synoptic Problem’, CBR 7 (2008) 68–123Google Scholar.

9 Linmans, A. J. M., Onderschikking in de Synoptische Evangeliën (Leiden: FSW, 1995) esp. 96–9, 319Google Scholar. Linmans made careful and extensive use of Log-Linear Analysis and Correspondence Analysis. He divided the text of the Synoptics by classifying very short units as narrative, dialogue or sayings, and compared large blocks of these, but with less emphasis on partitioning the data to determine within-group as against between-group differences or similarities. He covered the whole of the three Synoptic gospels, and focused on text type (discourse type) preferences and gospel preferences. He found little evidence of source preferences. The present paper focuses more on possible source differences within Matthew, using more data partitioning, in order to try more specifically to pursue that particular aspect of the data.

10 Sewell, P., ‘The Synoptic Problem: A Stylometric Contribution Regarding Q’, Colloquium (The Australian and New Zealand Theological Review) 33 (2001) 59–74, 153–68, esp. 73Google Scholar. Sewell made careful use of a method which compared sections of Matthew against a set of passages made up from different parts of the NT; it also compared sections of Matthew directly against each other. It required a p value of 1%, which means that it may have rejected some differences which were significant but not highly significant. He concluded there was some variability within M, but that Q passages are not in general markedly different from the rest of Matthew. But it would seem that Q sayings and M sayings were not directly compared like for like.

11 D. Gentile, http://www.davegentile.com/synoptics/main.html (26.01.2010). The study reported on this website was done by the statistician D. Gentile with some collaboration by another statistician David Inglis. The source gives information about the origin and processing of the data. Large numbers of the more frequent words in 19 Synoptic categories were included in the counts. The statistics checked correlations based on comparisons adjusted for the varying size of the texts for each category. Though the original source data did not include some of the most frequent words, did use many content words and did not sub-classify by genre, or allow further partitioning, the careful analysis of correlations repays attention. Genre was considered at the interpretative stage. The main conclusions supported Markan priority and though initially the author was cautious about the merits of FGT, 2ST and 3ST (Three Source Theory), in recent discussion he has inclined more towards the latter.

12 See List A in the Appendix. Details of the numeric data can be obtained from the author by email, requesting file syndat1a.doc.

13 See Lists B, C and D in the Appendix.

14 It would, in theory, be equally possible to test for the use of Luke by Matthew, with similar methods to those reported here on Matthew, to see if there is, or is not, a clear stylistic difference in Luke between double- and single-tradition material. In practice in addition to the paucity of narrative in the double tradition, there might be difficulty in finding enough sayings samples in the Lukan single tradition, as parables and apophthegms are more prominent in that material.

15 Somers, H. H., ‘Statistical Methods in Literary Analysis’, The Computer and Literary Style (ed. Leed, J.; Kent, OH: Kent State, 1966) 128–40Google Scholar.

16 Greenwood, H. H., ‘St. Paul Revisited: A Computational Result’, Literary and Linguistic Computing 7 (1992) 43–7CrossRefGoogle Scholar.

17 Forsyth, R. S. and Holmes, D. I., ‘Feature Finding for Text Classification’, Literary and Linguistic Computing 11 (1996) 163–74, esp. 164, 170Google Scholar.

18 Holmes, D. I., ‘The Evolution of Stylometry in Humanities Scholarship’, Literary and Linguistic Computing 13 (1998) 111–17, esp. 113–14CrossRefGoogle Scholar.

19 Biber, D., Dimensions of Register Variation: A Cross-linguistic Comparison (Cambridge: Cambridge University, 1995) esp. 153, 165, 237CrossRefGoogle Scholar; Burrows, J. F., Computation into Criticism: A Study of Jane Austen's Novels and an Experiment in Method (Oxford: Clarendon, 1987)Google Scholar esp. 163–75.

20 Greenacre, M. J., Theory and Applications of Correspondence Analysis (London: Academic, 1984)Google Scholar.

21 Burrows, J. F., ‘Not Unless You Ask Nicely: The Interpretative Nexus Between Analysis and Information’, Literary and Linguistic Computing 7 (1992) 91–109 esp. 96, 101–2CrossRefGoogle Scholar.

22 Gaston, L., Horae Synopticae Electronicae: Word Statistics of the Synoptic Gospels (SBLSBS 3; Missoula: Scholars, 1973)Google Scholar.

23 Martin, R. A., Syntax Criticism of the Synoptic Gospels (SBEC 10; Lewiston: Mellen, 1987)Google Scholar.

24 Poirier, Statistical Studies, 77–8.

25 Neirynck, F. and van Segbroeck, F., New Testament Vocabulary: A Companion Volume to the Concordance (BETL 65; Leuven: Peeters, 1984) 290, 119, 221, 229, 246Google Scholar.

26 Some of the variations were minor and considered the following objection. Someone might assert that samples of 250 words should only be used with the 4 most frequent rather than the 5 most frequent words. To counter this objection, two further tests were made using only the 4 most frequent words. When mixed and sayings samples from Matthew attributed to Markan, Q and M material were used the p value came in at 1.34% and so it was more, not less, significant than the original 2.73%. In a straight comparison of Q and M material, allowing for genre differences between mixed and sayings samples, and using the 4 most frequent words, the p value relating to source effect came in at 1.85%, again more significant than the original 2.91%. The main results should therefore stand, and any objector note that to press the objection would actually strengthen, not weaken, the case made here. Using only the 4 most frequent words would also produce a slightly more significant p value for the source effect between Markan and Q material, but not, however, for the comparison of Markan and M material.

Some further tests were made omitting the sample which contains the parable of the sheep and the goats (sample s56). The reason for this is that the style of this parable is very distinctive and different from the other parables. In the comparison of Markan, Q and M material this made the p value for source effect more significant, both when 5- and 4-word variables were used. Similar results occurred with the comparison of Q against M material. This does not mean that these more significant p values should be pressed into becoming main results. It may, however, tend to confirm the suspicion that there is something unusual about the style of this parable.

27 The difference between a 2.73% and a 4.4% likelihood that the result is due to chance is not massive. Any surprise arises from a higher expectation for Marcan priority due to the additional possibility of making comparisons with Mark itself, whereas discussion of Q passages is largely restricted to comparison of Matthew with Luke. The tests reported here are almost all based just on samples from Matthew attributed to Q, M and Mark, in order to focus attention mainly on one crucial aspect of the problem.

28 Poirier, Statistical Studies, 77–78.

29 This is an adjusted measure for calculating the distance between two clusters of data, or between outlying samples and a cluster. Because clusters may well not be perfect spheres, but bulge in one direction or another, a method needs to be used which calibrates the distance from the centre of the cluster allowing for the presence or absence of a bulge in the direction in question: see Figure 5 at the end of the Appendix.

30 Forsyth and Holmes, ‘Feature Finding’, 169.

31 Grieve, J., ‘Quantitative Author Attribution: An Evaluation of Techniques’ Literary and Linguistic Computing 22.3 (2007) 251–70CrossRefGoogle Scholar, esp. 261.

32 Hoover, D. L., ‘Multivariate Analysis and the Study of Style Variation’, Literary and Linguistic Computing 18 (2003) 341–60CrossRefGoogle Scholar, esp. 343.

33 Again any objector insisting on using only the 4 most frequent variables would be confronted with p comfortably under the 1% significance level at 0.62%, and only two samples of 17 cross-classified (sample R to M and 8 to Q), giving 88.23% correctly assigned and an even stronger conclusion.

34 At this point it is worth considering the implications of the results of this study for the delimitation of Q. If the main results above are accepted, it is still possible that some aspects of the results would suggest that minor modifications should be made to the 2ST. Some of the subsidiary tests explored which samples were more often assigned to a source other than the one posited by the 2ST. Samples R, S and 8 strayed most often, but if either R and 8 were omitted, or given a changed attribution, then S did not stray. This suggests that the main problem lies with the attribution of R to Q and 8 to M, rather than with S.

Sample R contains a set of verses from Matt 23 (the woes against the scribes), and the fact that cross-validation assigns this sample to M does, in fact, match the view of some conventional literary scholars that these woes stem at least in part from M. Scholars have long-since noted low levels of agreement in the double tradition here, indications of the use of another source (so Burkett, Rethinking, 2.163–5, 236) and some evidence of divergent translation from Aramaic (Casey, Q, 82). Matthew is known to practice conflation, so it would not be unreasonable to suspect that at least some of these verses do not derive from Q (or not from Q alone).

Sample 8 contains Matt 10.41; 11.14-15, 28-30; 12.36-37; 16.17-19; 18.10, 16-17. Some of these 7 short passages are preceded or followed by verses from the double tradition which are attributed to Q. In two cases they are both preceded and followed by such verses. The context can be summarized as Q Context: a) no, b) both sides, c) before, d) both sides, e) no, f) after, g) before. Long ago Schürmann argued on stylistic grounds that a range of passages found only in Luke, but in a Q context, should be attributed to Q, and made the equivalent proposal for a few passages in Matthew. On this see Schürmann, H., ‘Sprachliche Reminiszenzen an abgeänderte oder ausgelassene Bestandteile der Spruchsammlung im Lukas- und Matthäusevangelium’, NTS 6 (1960) 193–210CrossRefGoogle Scholar (reprinted in Traditionsgeschichtliche Untersuchungen zu den synoptischen Evangelien [Düsselforf: Patmos, 1968]Google Scholar). Though Kloppenborg, J. S., The Formation of Q: Trajectories in Ancient Wisdom Collections (SAC; Philadelphia: Fortress, 1987)Google Scholar 83 and n. 147 is more cautious about this view, it might be worth reconsidering the attribution of some of these verses.

35 Forsyth and Holmes, Feature Finding, 169.