Biomelrika (1962), 49, 1 and 2, p. 133 Printed in Qreat Britain 133 Ranks and measures BY M. G. KENDALL G-E-I-R (U.K.) Ltd. 2. I consider a situation in which we can rank in order of magnitude not only the original objects but also their differences. For instance, given a class of students ranked on merit from 1 to 10, we may be able to say that the biggest gap occurs between the fifth and sixth, the next biggest between the ninth and tenth, and so on. Such a ranking of differences might be: Student Ranking of differences 1 2 6 3 3 4 9 5 6 7 8 9 8 1 7 4 5 1 10,] . Suppose further than than we rearrange these first differences in order of magnitude and can rank the differences between them, e.g. 1 2 6 3 1 7 4 5 5 6 2 7 8 8 9 4 3 (2) We can imagine the process continued until we arrive at a single pair of differences which can be ranked either 1, 2 or 2, 1. In this whole array of rankings of successive differences we obviously have a large amount of supplementary information. The question with which I am concerned is: how can this extra information be used to set up a measure for the original objects? 3. The procedure proposed may best be illustrated by a numerical example. Column (1) of Table 1 shows, in order or occurrence, the first eight normal deviates from Quenouille's (1959) tables of random numbers. In column (2) these are multiplied by 100 and increased by 100 to remove negative values. Column (3) rearranges them in order of magnitude, starting with the smallest. Column (4) gives the first differences and column (5) the rank order of those differences, again starting with the smallest. They are rearranged in order in column (6) and differences in column (7). Column (8) gives the rankings. In this case two values in column (7) are tied so the corresponding ranks are split. And so we proceed down to column (19). 4. The method I propose for reconstituting the original values consists of starting from column (19) and working backwards. To begin, we need some assumption concerning the Downloaded from http://biomet.oxfordjournals.org/ at Penn State University (Paterno Lib) on May 9, 2016 1. In many situations occurring in psychology we are able to rank a set of objects according to a criterion, but cannot measure the quality expressed by that criterion. The desirability of transforming our information in some way from ranks to measures has led to procedures of various kinds, some of which appear to work satisfactorily in practice but none of which really touches the fundamental problem. And this is not surprising when we recall that, even on the assumption that a measure exists, a set of ranks has irretrievably lost much of the information implicit in a set of variate-values from which the ranks were derived. To make progress we must supplement the original ranking information in some way or (less reliably) substitute assumption for additional information. 134 M. G. Table 1 Columns A (1) -0-84 (2) 16 (3) 1 1-37 237 16 -018 32 15 2 66 4 135 134 382 135 212 312 237 0-34 1 134 5 52 14 14 66 70 75 7 . 6 70 5 102 102 6 4 1 14 5 2 27 9 . 0 13 75 10 312 37 5 . 2 3 . o 1 9 . 5 10 . 4 13 37 1 1* 1 8 4 1 1 1 3 3 3 . 8 0 . 2 . 5 . 1 2 2 . 3 1 m 2 3 . t 382 relative magnitudes of the last two differences which are ranked 1, 2. I shall assume them to be proportional to 1 and 3 units. The reason for this is that if a magnitude is broken at random into two parts the expected value of the larger part is three times that of the smaller. Generally, if a magnitude is broken into n parts the expectation of the parts in descending \ / - /- —?— -\ (3) 1. 1 n'n' Thus our last pair of differences are assumed to have, in units which it is unnecessary to specify, the values . „ ... in that order. Now we require a second assumption to proceed to the previous differences. The total' length' in the two differences arrayed in (4) is 1 + 3 = 4 units, with an average of 2. I assume that a piece of this average length precedes the pair in (4), so that the lengths of the three differences are 2, 2+1, 2 + 1 + 3, namely 2, 3, 6 (5) and they are in this order since the ranks in column (17) of Table 1 are 1, 2, 3. We now proceed by similar assumptions. The total number of units in the three numbers of (5) is 11, with an average of 11/3. The four differences preceding are then or proportional to 11, 17, 26, 44. (6) Downloaded from http://biomet.oxfordjournals.org/ at Penn State University (Paterno Lib) on May 9, 2016 2-82 1 4 15 37 3 . 1 3£ 1 82 m -0-99 1 14 52 0-35 (7) (8) (9) (10) (11) (12) (13) (14) (15) (16) (17) (18) 19) (4) (5) (6) Ranks and measures 135 We can work in proportions throughout until we impose a scaling factor at the final stage. Now the rankings corresponding to these four quantities, from column (14) of Table 1 are l£, 4, 1|, 3. We therefore average the first two of (6) to obtain 14 for each and rearrange (6)aS 14, 44, 14, 26. (7) The average is 49/2 and, on multiplication by 2 to remove the fraction, the preceding five values become 49, 77, 165, 193, 245 and on rearrangement according to the ranking of column (11), 165, 49, 245, 193. (8) The reader will wish to verify the next stages for himself. Corresponding to columns (8) and (5) we find 4123, 8748, 4123, 1458, 2228, 6818. (9) 870, 2157, 1745, 458, 3208, 2526, 2303, total 13,267. (10) These numbers, if we have been successful, should be proportional to the differences between the eight original values. To make a comparison I make the ends of the scale based on (10)fitthe original set of numbers by identifying end points. The total range of the original values in column 3 of Table 1 is 381. The range in expression (10) is 13,267. I therefore multiply the values in (10) by 381/13,267 = 0-028,718 and accumulate from the term unity, to obtain Reconstructed values 1 26 88 137 151 243 316 382,) (11) Original values 1 16 82 134 135 237 312 382.1 The agreement seems good. 5. The experiment was repeated 24 times, with a sample of 5 and two samples of 10 from each of the eight tables given by Quenouille, which comprise samples from sundry different populations. The results for the samples of 10 are summarized in Table 2. The letters xv ...,zB indicate the eight populations in Quenouille's notation. The agreement on the whole seems very fair. The most serious discrepancies occur when one end value lies well away from the others, as in the first set of x3: and this is to be expected. 6. It is not self-evident that the final scaling should be determined by pinning the end points. Some experiments were carried out to see whether pinning at the second and ninth or third and eighth points would result in an improvement. There were undoubtedly cases where this was so, but I did not perceive any general effect which would justify one method rather than another. Where outliers are suspected, however, I should be inclined to pin at points inside the range. For example, in the first set of x3 we get these results: Original values Reconstructed: By pinning end points By pinning at A, B By pinning at C, D 41 G 43 A 54 69 69 70 75 B 89 D 111 409 41 41 38 59 45 43 99 54 53 151 66 67 161 68 70 175 72 73 202 78 80 250 89 93 319 105 111 409 126 135 The distortion due to the last value is evident. Apart from this the fits are quite good. Downloaded from http://biomet.oxfordjournals.org/ at Penn State University (Paterno Lib) on May 9, 2016 77, 136 M. G. KENDALL II oo t*» C N c s © c o —<•—i w r ,—I ~H .—I F-H ©r5Ot>CO<NCSCO .H_H,-Hi—1_j (M CQ CO <N CO "•# i-H ^H ^H CO Downloaded from http://biomet.oxfordjournals.org/ at Penn State University (Paterno Lib) on May 9, 2016 If 1 41 I* 3 •S-i 5COI>0 Is i—i i-t —< cq ?l 1 ! | £ <N co ««0«OIN05«N «; o CD i> oo co co o 1? Nf5CO0(NIOC IO O) W l> CO ^H N W TJ H ^ , H cq « (N CM < HHH«»ncqn SS- If HHHC4 JOHWCOO ,—I ^H i—I .—I <N ^ W D N M C O O HMOOJC50HP5100 ( I , , (^ S3 I * CJ o J * _~ CM ^H i—I r-H •—I M CM CO O 03 (p oHncoN ,—I ^H i-H i—( CO -o 1 T3 — T^#OOTtlCO i—<»f5G0C5O«—tC OSTf^HCCSOO FH IOCOCOOH "81 1O <—itMCiCscocoicoooseM c? CMI» wnitcocnrt M I H >^ *H ^ f« CN| n C D I e-< s 1O )H00 C O 73 >e S CO c O O f O N N ^ ^H ^ H r^ tM CN CNCq CM Ranks and measures 137 7. In actual practice, of course, it will scarcely ever be possible to rank differences beyond the second; the effort required of the imagination in psychological experiments would be too great. But this does not destroy the practical utility of the method. Even if only first differences can be ranked, some estimate of the original measures can be obtained. In fact, instead of starting with the ranking of two in the proportion 1:3 we can start with the ranking of seven in the proportions given by expression (3) with n = 7. This may be rather rough but it is obviously a good deal better than merely operating on the original ranks as if they were equally spaced measures, or replacing them by normal equivalent deviates. 9. In the following paper Prof. Daniels examines some theoretical questions prompted by the forementioned empirical results. I may, in conclusion, mention a further method which I have not yet had an opportunity of testing in practice. The successive rankings of differences are equivalent to linear inequalities in the variable on which they are based. For example, in Table 1 the first ranking is equivalent to x1<xi<x3...<x8. (12) The second ranking (of first differences) is equivalent to x5-xt < x2-xl < Z4-Z3 < . . . < x6-x5 (13) and so on. We can now pose a problem in linear inequalities. Given the inequalities of types (12), (13), etc., between what limits must the original variable lie? I hope to examine this problem on a subsequent occasion. REFERENCE QUENOUHXB, M. H. (1959). Tables of 1000 standardized deviates from certain non-normal distributions. Biometrika, 46, 178-202. Downloaded from http://biomet.oxfordjournals.org/ at Penn State University (Paterno Lib) on May 9, 2016 8. The foregoing method, or some modification of it, is in my opinion worth a trial as a method of scaling. As it stands, it separates a set of objects by distances which approximate to a measure, but it will not locate those objects on a previously determined scale. For scaling purposes I should recommend mixing the objects with a set of ' markers' whose position on the scale has already been determined, in whose position defines the scale. The relative position of the new unmarked objects can then be determined by reference to the markers.
© Copyright 2026 Paperzz