- Email: [email protected]
DNA-Profiling of stains in criminal cases: analysis of measurement errors and band-shift. Discussion of match criteria
Forensic
Science International 61 (1993) 21-34
Forensic Science International
DNA-Profiling of stains in criminal cases: analysis of measurement errors and band-shift. Discussion of match criteria Birthe Eriksen*, Inslituteof Forensic
Ole Svensmark
Genetics, University of Copenhagen, Frederik V’s Vej Il. DK-2/&l Copenhagen, Denmark
(Received 4 January 1993; accepted 12 March 1993)
Abstract DNA-profiling was performed on approximately 600 stains (blood, semen/vaginal secretion, tissue samples and saliva) deriving from criminal investigations. The restriction enzyme was Hi&, and the VNTR (variable number of tandem repeats) probes were MSl, MS31, MS43a and YNH24. DNA-profiles were obtained from 60% of the stains, and matches were seen for 65% of the profiles. The measurement errors and the differences between corresponding fragment lengths of blood and stain profiles were analysed statistically. Distinct bandshifts were observed for approximately 65% of the profiles. For 50% of the profiles, the fragments derived from the stain migrated faster than those from the blood sample, and for 15% of the profiles the stain fragments migrated slower. The difference between the migration distance of the stain and the blood fragments of a given pair of profiles increased with increasing migration distance, i.e. with decreasing fragment length. After correction for this slope the measurement errors were independent of the fragment length, and of the same order of size as for duplicate determinations of fragments from blood samples. The differences between the fragment lengths of corresponding profiles were highly correlated (p = 0.8). Based on the statistical analysis, different match criteria are discussed and an ellipsoid accept-area is suggested. Key words: DNA-profiling; RFLP; VNTR; Single locus probes; Stains; Criminal cases; Statistical analysis; Measurement errors; Band-shift; Correlation of errors; Match criteria * Corresponding author. Abbreviations: kb, kilobase
pairs;
RFLP, restriction fragment length polymorphism; VNTR. variable
number of tandem repeats. 0379-0738/931$06.00 0 1993 Elsevier Scientific SSDI 0379-0738/93/01325-L
Publishers
Ireland
Ltd. All rights reserved
B. Eriksen, 0. Svensmark/Forensic
22
Sci. ht. 61 (1993) 21-34
1. Introduction
The aim of this study is the assessment of a robust match criterion for DNAprofiles from stain material and blood samples. To this purpose the measurement errors and the differences between stain and blood profiles were analysed statistically. DNA-profiling was carried out with 4 single locus VNTR (variable number of tandem repeats) probes, MSl, MS31, MS43a and YNH24. Fragment lengths were calculated in units of kilobase pairs (kb) and transformed into normalized migration distance in units of millimetres. After transformation, the measurement errors are independent of the fragment length and normally distributed [ 11.DNA-profiles from 289 stains (blood, semen, vaginal secretion, tissue samples and saliva) were compared with the profiles obtained from blood samples from the persons supposed to deliver the stains. Comparisons were performed between determinations on the same plate as well as between determinations on different plates. The study comprised approximately 3800 band comparisons. Many stains exhibited a more or less pronounced band-shift in the electrophoresis, i.e. all bands of the profile migrated either faster or slower than the fragments from the blood sample, and the difference between the migration distances of the blood and the stain fragments increased with decreasing fragment length. The differences between the fragment lengths of the two profiles were highly correlated (p = 0.8). Match criteria are discussed on the basis of the statistical analysis. 2. Materials and Methods Data related to 585 stains and 418 blood samples from 261 criminal cases investigated in the period 1989- 1991 were collected. The numbers of cases and stains in different types of cases are given in Table 1, and the numbers of different types of stain material in Table 2. DNA-RFLP (DNA-restriction fragment length polymorphism) analysis was performed as described previously [l]. The restriction enzyme was Hi& (Boehringer), and the probes were MSl, MS31, MS43a (Cellmark Diagnostics) and YNH24 (Promega Corporation). Ethidium bromide (0.5 &ml) was included in the TBE electrophoresis buffer and the loading buffer. The Amersham marker SJ5000 was used
Table 1 Case types investigated
in the period
1989-1991
Type of case
Number
Homicide Assault Rape/indecent assaults Robbery/burglary Miscellaneous Total
35 19 176 22 9 261
of cases
Number 112 51 369 35 18 585
of stains
B. Eriksen, 0. Svensmark /Forensic
Sci. ht. 61 (1993) 21-34
23
Table 2 Types of stains examined in the period 1989-1991 Type of stain
Number of stains
Blood Semen/vaginal secretion Tissue Saliva
195 358 13 19
Total
585
as size marker. DNA was isolated from stain material essentially according to Gill et al. [2]. Semen stains supposed to contain body fluids other than semen were extracted using preferential lysis of non-sperm cells before isolation of DNA. In many cases the non-sperm fraction of DNA as well as the sperm fraction were analysed. In the following, the differential extracts from the same stain were recorded as one stain with two different profiles. Throughout this study comparisons were made between profiles derived from stains and those from a corresponding or ‘matching’ blood sample from the victim or the suspect in the case. In this context it is important that the match criterion does not lead to false exclusions. Therefore, in the visual inspection of the autoradiograms, pairs with distances between corresponding fragment lengths of up to 6 mm were considered as potentially matching pairs. In most cases a match was supported by other evidence obtained from classical blood grouping. The statistical analysis of measurement errors requires that the standard deviation (S.D.) of the error is independent of the fragment length. To obtain data with measurement errors which are independent of the fragment length and normally distributed all fragment lengths were transformed into normalized migration distance [l]. The transformation was accomplished by the function L f(b): m = 796/(3.7 + b1.5) + 32.3 where b is the fragment length in units of kilobase pairs (kb) and m the normalized migration distance in millimetres. In the case of single-band patterns (apparent or true homozygotes) the length of the fragment was included twice, i.e. as band 1 and as band 2. 3. Results 3.1 Forensic aspects
From the total material of 585 stains, 407 profiles were obtained; 289 of these matched with a blood-sample profile from one of the persons in the case. With blood stains, DNA-profiles were detected for 119 of 195 stains (61%). Three stains contained two different profiles. Match with a blood sample from a person in the case was seen for 96 blood-stain profiles. With tissue (stomach, foetal tissues, muscle, dental pulp and bone marrow), profiles were obtained for 31% of the stains. The low rate is due to a relatively high proportion of aged or decomposed samples. With saliva.
B. Eriksen, 0. Svensmarkl Forensic Sci. ht. 61 (1993) 21-34
24
Table 3 Stains of semen and/or
vaginal
secretion
Number of stains No bands
358 125 (35%) 193 (54%) 39 (I I%)
1 profile 2 profiles 3 profiles
I
Number of profiles No matches Match with the victim Match with the suspect
214 90 (33Y”) 13 (26%) 111 (41%)
profiles were obtained from only 21% of the stains, the low rate being due to a high proportion of cigarette ends and masks used in robberies. For stains containing semen, vaginal secretion or mixtures of semen and vaginal secretion or other body fluids, profiles were seen for 65% of the stains. The sperm fraction obtained by differential extraction of stains with large amounts of other cells, usually vaginal epithelial cells, often contained detectable amounts of non-sperm DNA. Only in a few cases the non-sperm fraction contained detectable amounts of sperm DNA. In three cases the sperm fraction contained sperm DNA from 2 or more sources. For 184 profiles matches were seen with profiles from blood samples from a person in the case (Table 3).
Fragment _
22
54
10
length
Fragment
kb
length
kb
2
3
t
40
60
80
100
120 Fragment
140
160 length
160 mm
40
60
80
100
120 Fragment
140
160 length
180 mm
Fig. I. The difference between the transformed fragment length of the band derived from the blood sample and that derived from the stain as a function of the fragment length of the band from the blood sample DNA. The lines are the regression lines. (A) Results from 1822 determinations with blood and stain DNA on the same plate. (B) Results from 1959 determinations with blood and stain DNA on different plates.
B. Eriksen.
Table 4 Statistics
0. Svensmark /Forensic
for the differences
Fragment
Sri. ht.
61 (1993) 21-34
25
b- s
length
Mean
S.D.
Minimum
Maximum
(mm)
(mm)
(mm)
(mm)
Number
mm
kb
Same plate 40-180 40- 80 80-120 120-180
1.5-21.5 5.5-21.5 3.0-5.5 1.5-3.0
-0.27 -0.21 -0.30 -0.45
0.60 0.49 0.62 0.89
-2.77 -2.28 -2.07 -2.77
1.97 1.07 1.54 1.97
1822 959 628 235
Different plates 40-180 40-80 80-120 120-180
1.5-21.5 5.5-21.5 3.0-5.5 1.5-3.0
-0.27 -0.17 -0.35 -0.44
0.76 0.69 0.74 1.02
-3.00 -2.33 -2.88 -3.00
2.66 2.13 1.84 2.66
1959 1009 685 265
b and s are the transformed fragment lengths of corresponding bands from the blood sample and the stain. respectively. The upper part of the table gives the results of determinations on the same plate, and the lower part those from determinations on different plates.
40
60
80
100
120 Fragment
140
160
180
length mm
Fig. 2. The differences between the transformed fragment lengths of the corresponding bands from blood and stain DNA were plotted as a function of the fragment length of the bands of the blood sample for three different DNA-profiles. Blood and stain DNA were analysed on the same plate. Full lines are regression lines. (0) no band-shift; (V ), band-shift with negative slope; (a), band-shift with positive slope (‘reversed’ band-shift).
B. Eriksen, 0. Svensmark /Forensic
26
3.2 Difference
between profiles from
Sri. ht. 61 (1993) 21-34
blood sample and stain
The comparisons between the profiles of the stain and the corresponding blood sample were performed for profiles determined on the same plate as well as on different plates. The number of comparisons of bands from the same plate was 1822, and that from different plates was 1959. Before comparison, all fragment lengths were transformed into normalized migration distance (mm). For each pair of fragments the difference [b - s] (where b and s are the transformed fragment length of the fragment derived from the blood sample and the stain, respectively) was plotted against the fragment length of the blood sample DNA (Figs. lA,B). The band pairs in Fig. 1A derive from the same plate, and those in Fig. 1B from different plates. The scatter was most pronounced for the results derived from different plates. It is also seen that the scatter increased with increasing migration distance (decreasing fragment length). This also appears from Table 4. For both sets of data the S.D. increased with increasing migration distance. At large fragment lengths (40-80 mm, 5.5-21.5 kb) the S.D. of the difference was 0.5 mm on the same plate and 0.7 mm on different plates. The latter value is identical with that found for the difference between duplicate determinations of blood samples on different plates over the whole range of fragment lengths [ 11. 3.3 Band-shift To investigate the cause of the increase in measurement errors with increasing migration distance the dependence of the differences [b - s] of the fragment length was studied for each pair of corresponding profiles. Regression analysis was performed for all pairs of profiles, and approximately 65% of the pairs showed linear slopes significantly higher or lower than zero. The zero point of the regression line was at
Fragment 22 3
r
40
10
5
4
3
100
120
length
kb
Fragment
length
kb
2
A
60
80
Fragment
140
160 length
1
180 mm
-3
i 40
60
80
100
120 Fragment
140
160 length
180 mm
Fig. 3. Scatter diagrams, as Fig. 1. Data from profiles with positive slopes of band-shift >0.005. (A) Results from 198 determinations with blood and stain DNA on the same plate. (B) Results from 257 determinations with blood and stain DNA on different plates.
B. Eriksen, 0. Svensmark/Forensic
Sci. ht. 61 (1993) 21-34
21
approximately 40 mm. In Fig. 2 three examples are shown: one profile with a positive slope, one with a zero slope and one with a negative slope. For 35% of the pairs of profiles the slope was not clearly different from zero. We have, arbitrarily, divided the slopes into 3 classes: Positive slopes >0.005, negative slopes < -0.005 and insignificant slopes for which the numerical slope was less than 0.005, i.e. less than 0.5 mm/l00 mm. For each of the 3 classes the differences [b - s] were plotted against the fragment length for the blood sample DNA (Figs. 3A,B, 4A,B, and 5A,B). The scatter in relation to the regression lines was independent of the fragment length. For the results derived from determinations on the same plate the SD. of [b - S] was 0.53, 0.51 and 0.54 mm for the 3 classes, and for the results derived from different plates it was 0.84, 0.69 and 0.69 mm, respectively. The distribution of the slopes of the regression lines of the difference [h - s] as a function of fragment length is given in Fig. 6. Approximately 50% of the profile pairs exhibited significantly negative slopes and 15% significantly positive slopes. Most of the profiles obtained from stains with semen and vaginal secretion exhibited a negative slope (mean value: -0.006). For profiles from blood stains the mean value of the slope was approximately zero (-0.0005). The majority of profiles with positive slopes (‘reversed’ band-shift) derived from blood stains which were compared with degraded blood samples. 3.4 Correlation The differences in fragment length between corresponding bands of the band pairs of two matching profiles were highly correlated. The correlation was investigated for both data sets, i.e. data derived from the same plate and those from different plates. Furthermore, the correlation was calculated for the 3 classes of slopes in each set.
Fragment 22
10
54
length
Fragment
kb
2
3
'7
10
54
length
kb
2
3
a 40
60
80
100
120 Fragment
140
160 length
180 mm
40
60
80
100
120 Frogment
140
160 length
180 mm
Fig. 4. Scatter diagrams as Fig. 1. Data from profiles with negative slopes of band-shift < -0.005. (A) Results from 722 determinations with blood and stain DNA on the same plate. (B) Results from 892 determinations with blood and stain DNA on different plates.
B. Eriksen. 0. Svensmark / Forensic Sci. ht. 61 (1993) 21-34
28
Fragment 22 3
10
length
3
54
kb
Fragment
length
kb
2 7
A
a, -1
r ._ c!
-2
-3 40
60
80
120
100
Fragment
140
160 length
180
40
60
80
100
mm
120 Fragment
140
180
160 length
mm
Fig. 5. Scatter diagrams as Fig. I. Data from profiles with non-significant band-shift (absolute value of slope <0.005). (A) Results from 902 determinations with blood and stain DNA on the same plate. (B) Results from 810 determinations with blood and stain DNA on different plates.
For all data sets the coefficient of correlation was 0.8. In Fig. 7A the correlation is shown for all data derived from different plates. For comparison, data derived from duplicate determinations of DNA from blood samples determined on different plates are given (Fig. 7B). In both figures the 95, 99 and 99.9% contour ellipses are shown
‘,
O.O”, -4
-3
’
-2-l
/
1, ’
0
1
t-y--y2
3
4
SLOPE Fig. 6. Distribution of the slopes of the band-shift comprising more than 5 bands obtained
(see text) calculated from determinations
%
for 249 pairs of DNA-profiles on the same plate.
B. Eriksen. 0. Svensmark/ Forensic Sci. ht. 61 (1993) 21-34
-4v -4
' -3
I
I
I
I
I
I
I
-2
-1
0
1
2
3
4
Bond
1 (blood - stain)
29
mm
4 3 2 1 0 -1 -2 -3 -4-4
-3
-2
-1
0 Bond
1
2
3
4
1 (d, - d2) mm
Fig. 7. (A) Correlation of the differences b - s for each band in a band pair where b and s are the transformed lengths of the fragments derived from the blood sample and the stain, respectively. Results for 940 band pairs obtained in determinations on different plates. (B) Correlation of the transformed differences between duplicate determinations of each band in a band pair derived from blood samples. d, and d2 are the transformed fragment lengths obtained from the first and the second determination. The data derived from 1583 duplicate determinations of blood samples [l]. The 95, 99 and 99.9% contour ellipses are shown. The formula for the ellipse is x2 + y* - 2pxy = X’LT’(I - p*) where (r* for 2 degrees of freedom is used. p = 0.8.
(p = 0.8; u = 0.7). For the stain/blood pairs (Fig. 7A) several outliers were observed at the lower right part of the ellipse. These outliers all derive from profiles with pronounced band-shift and band pairs with large distances between the bands. Correspondingly, outliers can be expected
30
B. Eriksen, 0. Svensmark / Forensic Sri. ht. 61 (1993) 21-34
Fig. 8. A two-dimensional model for a match criterion (see the text). For a base of the square of 1 the area of the ellipse is 0.55, and that of the circle is 0.23. The area of C + U as well as that of C + L is 0.39. The formula for the ellipse is: x2 + y* - 2pxy = b*(l - p*)/4 where p = 0.7 and b is the base of the square.
at the upper left part of the ellipse due to profiles with reversed band-shift. For the duplicate determinations on blood samples (Fig. 7B) there was good agreement between the observed and the expected distribution within the elliptic areas. 3.5 Match criteria In the investigation of match criteria only the data set obtained from determinations with blood sample and stain on different plates was considered. For this data set, the extreme values of the difference between the fragment lengths of the blood fragment and the corresponding stain fragment were -3.0 mm and 2.66 mm. This might suggest the use of a window of 3 mm width. If, for a given profile, all bands fall within the corresponding windows there will be a formal match. However, false inclusions may occur because the correlation of the deviations is not taken into account. An objective match criterion may be constructed on the basis of the finding that the errors are correlated. The following notation will be used: d, is the difference b, - sl where b, is the transformed fragment length of band 1 derived from the blood sample, and sI is the transformed length of band 1 derived from the stain. Correspondingly d2 is the difference b2 - s2 valid for band 2 of the band pair obtained in the same lane with the same probe. The term accept area will be used for the area within which all coordinates (d,, d2) related to a pair of profiles have to fall if a match has to be declared. For the simple windowing method with a window width of 3 mm the accept area is identical with a square with a base of &3 mm, or 6 mm (Fig. 8). This square accepts obviously mismatching band pairs, e.g. a pair with the coordinates (-2, 2 mm) or (2, -2 mm). Such profiles do certainly not match according to a visual inspection. Some improvement can be obtained by the use of the correlation ellipse shown in Fig. 8 as accept area. This area will not accept extremes as those mentioned above. But, it will still accept false matches between pairs
31
B. Eriksen, 0. Svensmark / Forensic Sci. ht. 61 (1993) 21-34
Table 5 Percent match
for pairs of DNA-profiles
Base of the square
at different
base lengths
Percent
of the square
match
w-r)
f * f f f * ??
f f ??
3.25 3.00 2.15 2.50 2.25 2.00 1.75 1.50 1.25 1.00
Stain/blood
Blood/blood
100 99.3 98.3 96.5 93.8 90.7 81.7 70.9 53.6 37.0
100 99.6 99.2 97.3 92.4 85.1 71.4
(See Fig. 8.) Column 2: Percent match for 289 pairs of DNA-profiles derived from stains and corresponding blood samples. Column 3: Percent match for 262 duplicate determinations of blood samples. The determinations derived from different plates. The accept area is either the area C + U or the area C + L. The coefficient of correlation is 0.7.
with uncorrelated differences, e.g. pairs of profiles which exhibit one band pair with the coordinates (2,2 mm) and another with the coordinates (-2, -2 mm). In our material we have never seen profiles with large deviations of different signs. Therefore, ideally, the requirement will be that for all band pairs belonging to a pair of matching DNA-profiles the coordinates must have the same sign, all positive or all negative, or zero. However, to allow for measurement errors the circular area (C in Fig. 8) has to be included in the accept area. In this model the accept area is either the area C + L or the area C + U. The lower area (C + L) accepts pairs of profiles exhibiting normal band-shift and the upper (C + U) accepts pairs of profiles exhibiting ‘reversed’ bandshift (Fig. 8). This model was tested with the data derived from determinations obtained from different plates. Due to the outliers in the correlation diagram (Fig. 7A) deriving from profiles with a pronounced bandshift, the coefficient of correlation was decreased to 0.7 and the ellipse was inscripted in a square with a base of f 3.25 mm. This corresponds to a maximum difference (window width) of 3.25 mm which again corresponds to 4.6 S.D. All pairs of profiles (289) in the data set were accepted by this ellipse. In Table 5 the percent match is given for different lengths of the base. For comparison, the model was tested on a data set consisting of 262 pairs of determinations of blood samples, i.e. duplicate determinations. In this data set large band-shifts are not expected, and the limits for 100% match are consequently expected to be smaller than for stain/blood sample pairs. The limit was found to be ~2.5 mm, corresponding to &3.3 S.D.
32
B. Eriksen.
0. Svensmark/Forensic
Sci. hf.
61 (1993)
21-34
4. Discussion
Only few reports on the use of RFLP-DNA analysis in casework have been published [3-l 11. These studies do not comprise statistical analyses of the measurement error, though, in one study, the difference in size measurements between blood and vaginal epithelial cell fragments was reported [lo]. The transformation of fragment length into normalized migration distance makes it possible to study the band-shift as a function of the migration distance. For profiles exhibiting a distinct band-shift the difference between the migration of the blood sample and the stain fragments increased with increasing migration distance. The slope may be negative or positive (‘reversed’ band-shift). It is of interest to note that 65% of the stains exhibit a distinct band-shift, negative or positive. Positive band-shift was observed for 15% of the stains, most often found for blood stains which were to be compared with degraded blood samples. Large differences in fragment lengths of corresponding bands (up to 3 mm) were always associated with a pronounced band-shift. Large differences do not reflect a greater measurement error for stains than for blood samples. After correction for the band-shift the S.D. of the difference between measurements of stain and blood sample fragments was identical with that of the difference between duplicate determinations of blood samples. The correlation of the differences between the lengths of fragments from blood sample and stain (Fig. 7A) is less symmetrical than the correlation of the measurement errors derived from duplicate measurements of DNA from blood samples (Fig. 7B). In Fig. 7A all outliers derived from profiles exhibiting pronounced band-shift. Generally, the correlation does not vary much with the difference between corresponding fragment lengths [l], but for profiles with pronounced band-shift the correlation is low for band pairs with large distances between the bands. Different match criteria have been discussed. In its simplest form the requirement for a match is that all differences between the fragment lengths of corresponding bands in absolute values have to be less than a given treshold or guideline, e.g. 5% [lo] or 2.8% kb [12,13]. In one study [14] 4”/0kb was suggested for fragments less than 10 kb and 8% for fragments greater than 10 kb. Windowing with a constant width in units of percent kb suffers from two shortcomings. First, it does not take account of the fact that the measurement error in percent kb is highly dependent on the fragment length [ 11.Second, neither does it take account of the fact that the measurement errors are correlated, and not independent [ 1,15,16]. According to our investigations a window of 2.8% kb corresponds to 0.7 mm (1.4 S.D.) at 20 kb and to 4.3 mm (8.6 SD.) at 2 kb. Such a window leads to false exclusions at large fragment lengths and to false inclusions at small fragment lengths. On the other hand, it accepts extreme band-shifts. However, when windowing is recommended the importance of a visual comparison by an experienced investigator is always emphasized (e.g., Refs. 10,12,13). Windowing on transformed fragment length (mm) is a better approach because the window now represents a constant measure of the S.D. of the measurement error. However, it still takes no account of the fact that the errors are correlated, and false matches will occur. A match criterion based on the Bayesian likelihood ratio has been proposed [ 17-201. The method has been shortly described for a bivariate normal distribution
B. Eriksen, 0. Svensmark /Forensic
Sci. Ini. 61 (1993) 21-34
33
[17], and the model has been been effectively modelled by a a multivariate normal distribution with the appropriate correlation coefficient and S.D. built in [18,20]. However, for 4 or more probes the mathematics become rather intricate. A rule for matching which takes into account correlated errors for single loci has recently been suggested [16]. The simple two-dimensional model suggested by us seems conceptually simple, and it performs very quickly and effectively in search procedures (to be published). 5. Acknowledgements We are obliged to Dr Niels Morling for valuable discussions and to Mrs Jane Hellung Lauridsen and to Mrs Susanne Billesbdle for excellent technical assistance. 6. References I 2 3
4 5
6
7
8
9
10
11 12 13
14
B. Eriksen, A. Bertelsen and 0. Svensmark, Statistical analysis of the measurement errors in the determination of fragment length in DNA-RFLP Analysis. Forensic Sci. Inr., 52 (1992) 181-191. P. Gill, A.J. Jeffreys and D.J. Werrett, Forensic application of DNA ‘fingerprints’. Nature, 318 (1985) 577-579. D.J. Werrett, P.D. Gill, I.W. Evett, J.E. Lygo, K.M. Sullivan and J. Buckleton, DNA analysis in home office laboratories. Its introduction, immediate future and statistical assessment. Proc. In!. Symp. on the Forensic Aspects of DNA Analysis, FBI Academy, Quantico, VI, 1989, pp. 147-162. B.H. Parkin, DNA analysis in the Metropolitan Police Forensic Science Laboratory. Proc. Int. Symp. on the Forensic Aspects of DNA Analysis, FBI Academy, Quantico, VI, 1989, pp. 163-167. H. Schmitter, S. Herrmann and W. Pflug, The use of DNA polymorphisms in the police laboratories of the Federal Republic of Germany. Proc. Int. Symp. on the Forensic Aspects of DNA Analysis. FBI Academy, Quantico, VI, 1989, pp. 169-172. D.E. Adams, L. A. Presley, H. A. Deadman and A.G. Lynch, DNA analysis in the FBI laboratory. Proc. Int. Symp. on the Forensic Aspects of DNA Analysis, FBI Academy, Quantico, VI, 1989, pp. 173-177. F.J. Burridge, M.J. Greenhalgh and G.M. Willot, An evaluation of single locus probes in casework. In: C. Rittner and P.M. Schneider (eds.), Advances in Forensic Huemogenetics, Vol. 4, Springer Verlag, Berlin, 1992, pp. 210-212. W. Pflug, J. Teifel-Greding, S. Herrmann, M. Gerhard, R. Wenzl and H. Schmitter, Results of DNA analysis from six forensic science laboratories in Germany. In: C. Rittner and P.M. Schneider (eds.). Advances in Forensic Haemogenetics, Vol. 4, Springer Verlag, Berlin, 1992, pp. 237-239. B. Mevsg, S. Jacobsen, B. Eriksen and B. Olaisen, DNA typing in forensic casework in Norway. Strategies and experiences. In: C. Rittner and P.M. Schneider (eds.), Advances in Forensic Haemogenetics, Vol. 4, Springer Verlag, Berlin, 1992, pp. 260-262. B. Budowle, A.M. Giusti, J.S. Waye, F.S. Baechtel, R.M. Fourney. D.E. Adams, L.A. Presley, H.A. Deadman and K.L. Monson, Fixed-bin analysis for statistical evaluation ofcontinuous distributions of allelic data from VNTR loci for use in forensic comparisons. Am. .I Hum. Gener., 48 (1991) 841-855. M. Greenhalgh, F. Burridge and G. Willot, Experiences with single locus DNA probes in casework. Forensic Sci. Inr.. 57 (1992) 29-37. I.W. Evett and P. Gill, A discussion of the robustness of methods for assessing the evidential value of DNA single locus profiles in crime investigations. Electrophoresis, 12 (1991) 226-230. P. Gill, I.W. Evett, S. Woodroffe. J.E. Lygo, E. Millican and M. Webster, Databases, quality control and interpretation of DNA profiling in the home office forensic science service. Electrophoresis. 12 (1991) 204-209. T. Staples, C.K. Goff, J.G. Wegel and G. Herrin, RFLP match criteria. Determination from data
34
IS 16 17 18 19 20
B. Eriksen. 0. Svensmark /Forensic
Sci. hi. 61 (1993) 21-34
analysed on a Bioimage system. Proc. ht. Symp. Human Identijication, Promega Corporation, Madison, WI, 1991, p. 316. N.J. Risch and B. Devlin, On the probability of matching DNA fingerprints. Science, 255 (1992) 717-720. N. Risch and B. Devlin, Letter. Science, 256 (1992) 1744-1745. I.W. Evett and D.J. Werrett, Bayesian analysis of single locus DNA profiles. Proc. hi. Symp. Human Identification, Promega Corporation, Madison, WI, 1989, pp. 77-101. I.W. Evett, J. Scranage and R. Pinchin, Efficient retrieval from DNA databases: Based on the second European DNA Profiling Group collaborative experiment. Forensic Sci. hf., 53 (1991) 45-50. D.A. Berry, I.W. Evett and R. Pinchin, Statistical inference in crime investigations using deoxyribonucleic acid profiling. Appl. Star., 41 (1992) 499-531. I.W. Evett, J.K. Scranage and R. Pinchin. An efficient procedure for interpreting DNA single locus profiling data in crime cases, J. Forensic Sri. Sot., 32 (1992) 307-324.
Science International 61 (1993) 21-34
Forensic Science International
DNA-Profiling of stains in criminal cases: analysis of measurement errors and band-shift. Discussion of match criteria Birthe Eriksen*, Inslituteof Forensic
Ole Svensmark
Genetics, University of Copenhagen, Frederik V’s Vej Il. DK-2/&l Copenhagen, Denmark
(Received 4 January 1993; accepted 12 March 1993)
Abstract DNA-profiling was performed on approximately 600 stains (blood, semen/vaginal secretion, tissue samples and saliva) deriving from criminal investigations. The restriction enzyme was Hi&, and the VNTR (variable number of tandem repeats) probes were MSl, MS31, MS43a and YNH24. DNA-profiles were obtained from 60% of the stains, and matches were seen for 65% of the profiles. The measurement errors and the differences between corresponding fragment lengths of blood and stain profiles were analysed statistically. Distinct bandshifts were observed for approximately 65% of the profiles. For 50% of the profiles, the fragments derived from the stain migrated faster than those from the blood sample, and for 15% of the profiles the stain fragments migrated slower. The difference between the migration distance of the stain and the blood fragments of a given pair of profiles increased with increasing migration distance, i.e. with decreasing fragment length. After correction for this slope the measurement errors were independent of the fragment length, and of the same order of size as for duplicate determinations of fragments from blood samples. The differences between the fragment lengths of corresponding profiles were highly correlated (p = 0.8). Based on the statistical analysis, different match criteria are discussed and an ellipsoid accept-area is suggested. Key words: DNA-profiling; RFLP; VNTR; Single locus probes; Stains; Criminal cases; Statistical analysis; Measurement errors; Band-shift; Correlation of errors; Match criteria * Corresponding author. Abbreviations: kb, kilobase
pairs;
RFLP, restriction fragment length polymorphism; VNTR. variable
number of tandem repeats. 0379-0738/931$06.00 0 1993 Elsevier Scientific SSDI 0379-0738/93/01325-L
Publishers
Ireland
Ltd. All rights reserved
B. Eriksen, 0. Svensmark/Forensic
22
Sci. ht. 61 (1993) 21-34
1. Introduction
The aim of this study is the assessment of a robust match criterion for DNAprofiles from stain material and blood samples. To this purpose the measurement errors and the differences between stain and blood profiles were analysed statistically. DNA-profiling was carried out with 4 single locus VNTR (variable number of tandem repeats) probes, MSl, MS31, MS43a and YNH24. Fragment lengths were calculated in units of kilobase pairs (kb) and transformed into normalized migration distance in units of millimetres. After transformation, the measurement errors are independent of the fragment length and normally distributed [ 11.DNA-profiles from 289 stains (blood, semen, vaginal secretion, tissue samples and saliva) were compared with the profiles obtained from blood samples from the persons supposed to deliver the stains. Comparisons were performed between determinations on the same plate as well as between determinations on different plates. The study comprised approximately 3800 band comparisons. Many stains exhibited a more or less pronounced band-shift in the electrophoresis, i.e. all bands of the profile migrated either faster or slower than the fragments from the blood sample, and the difference between the migration distances of the blood and the stain fragments increased with decreasing fragment length. The differences between the fragment lengths of the two profiles were highly correlated (p = 0.8). Match criteria are discussed on the basis of the statistical analysis. 2. Materials and Methods Data related to 585 stains and 418 blood samples from 261 criminal cases investigated in the period 1989- 1991 were collected. The numbers of cases and stains in different types of cases are given in Table 1, and the numbers of different types of stain material in Table 2. DNA-RFLP (DNA-restriction fragment length polymorphism) analysis was performed as described previously [l]. The restriction enzyme was Hi& (Boehringer), and the probes were MSl, MS31, MS43a (Cellmark Diagnostics) and YNH24 (Promega Corporation). Ethidium bromide (0.5 &ml) was included in the TBE electrophoresis buffer and the loading buffer. The Amersham marker SJ5000 was used
Table 1 Case types investigated
in the period
1989-1991
Type of case
Number
Homicide Assault Rape/indecent assaults Robbery/burglary Miscellaneous Total
35 19 176 22 9 261
of cases
Number 112 51 369 35 18 585
of stains
B. Eriksen, 0. Svensmark /Forensic
Sci. ht. 61 (1993) 21-34
23
Table 2 Types of stains examined in the period 1989-1991 Type of stain
Number of stains
Blood Semen/vaginal secretion Tissue Saliva
195 358 13 19
Total
585
as size marker. DNA was isolated from stain material essentially according to Gill et al. [2]. Semen stains supposed to contain body fluids other than semen were extracted using preferential lysis of non-sperm cells before isolation of DNA. In many cases the non-sperm fraction of DNA as well as the sperm fraction were analysed. In the following, the differential extracts from the same stain were recorded as one stain with two different profiles. Throughout this study comparisons were made between profiles derived from stains and those from a corresponding or ‘matching’ blood sample from the victim or the suspect in the case. In this context it is important that the match criterion does not lead to false exclusions. Therefore, in the visual inspection of the autoradiograms, pairs with distances between corresponding fragment lengths of up to 6 mm were considered as potentially matching pairs. In most cases a match was supported by other evidence obtained from classical blood grouping. The statistical analysis of measurement errors requires that the standard deviation (S.D.) of the error is independent of the fragment length. To obtain data with measurement errors which are independent of the fragment length and normally distributed all fragment lengths were transformed into normalized migration distance [l]. The transformation was accomplished by the function L f(b): m = 796/(3.7 + b1.5) + 32.3 where b is the fragment length in units of kilobase pairs (kb) and m the normalized migration distance in millimetres. In the case of single-band patterns (apparent or true homozygotes) the length of the fragment was included twice, i.e. as band 1 and as band 2. 3. Results 3.1 Forensic aspects
From the total material of 585 stains, 407 profiles were obtained; 289 of these matched with a blood-sample profile from one of the persons in the case. With blood stains, DNA-profiles were detected for 119 of 195 stains (61%). Three stains contained two different profiles. Match with a blood sample from a person in the case was seen for 96 blood-stain profiles. With tissue (stomach, foetal tissues, muscle, dental pulp and bone marrow), profiles were obtained for 31% of the stains. The low rate is due to a relatively high proportion of aged or decomposed samples. With saliva.
B. Eriksen, 0. Svensmarkl Forensic Sci. ht. 61 (1993) 21-34
24
Table 3 Stains of semen and/or
vaginal
secretion
Number of stains No bands
358 125 (35%) 193 (54%) 39 (I I%)
1 profile 2 profiles 3 profiles
I
Number of profiles No matches Match with the victim Match with the suspect
214 90 (33Y”) 13 (26%) 111 (41%)
profiles were obtained from only 21% of the stains, the low rate being due to a high proportion of cigarette ends and masks used in robberies. For stains containing semen, vaginal secretion or mixtures of semen and vaginal secretion or other body fluids, profiles were seen for 65% of the stains. The sperm fraction obtained by differential extraction of stains with large amounts of other cells, usually vaginal epithelial cells, often contained detectable amounts of non-sperm DNA. Only in a few cases the non-sperm fraction contained detectable amounts of sperm DNA. In three cases the sperm fraction contained sperm DNA from 2 or more sources. For 184 profiles matches were seen with profiles from blood samples from a person in the case (Table 3).
Fragment _
22
54
10
length
Fragment
kb
length
kb
2
3
t
40
60
80
100
120 Fragment
140
160 length
160 mm
40
60
80
100
120 Fragment
140
160 length
180 mm
Fig. I. The difference between the transformed fragment length of the band derived from the blood sample and that derived from the stain as a function of the fragment length of the band from the blood sample DNA. The lines are the regression lines. (A) Results from 1822 determinations with blood and stain DNA on the same plate. (B) Results from 1959 determinations with blood and stain DNA on different plates.
B. Eriksen.
Table 4 Statistics
0. Svensmark /Forensic
for the differences
Fragment
Sri. ht.
61 (1993) 21-34
25
b- s
length
Mean
S.D.
Minimum
Maximum
(mm)
(mm)
(mm)
(mm)
Number
mm
kb
Same plate 40-180 40- 80 80-120 120-180
1.5-21.5 5.5-21.5 3.0-5.5 1.5-3.0
-0.27 -0.21 -0.30 -0.45
0.60 0.49 0.62 0.89
-2.77 -2.28 -2.07 -2.77
1.97 1.07 1.54 1.97
1822 959 628 235
Different plates 40-180 40-80 80-120 120-180
1.5-21.5 5.5-21.5 3.0-5.5 1.5-3.0
-0.27 -0.17 -0.35 -0.44
0.76 0.69 0.74 1.02
-3.00 -2.33 -2.88 -3.00
2.66 2.13 1.84 2.66
1959 1009 685 265
b and s are the transformed fragment lengths of corresponding bands from the blood sample and the stain. respectively. The upper part of the table gives the results of determinations on the same plate, and the lower part those from determinations on different plates.
40
60
80
100
120 Fragment
140
160
180
length mm
Fig. 2. The differences between the transformed fragment lengths of the corresponding bands from blood and stain DNA were plotted as a function of the fragment length of the bands of the blood sample for three different DNA-profiles. Blood and stain DNA were analysed on the same plate. Full lines are regression lines. (0) no band-shift; (V ), band-shift with negative slope; (a), band-shift with positive slope (‘reversed’ band-shift).
B. Eriksen, 0. Svensmark /Forensic
26
3.2 Difference
between profiles from
Sri. ht. 61 (1993) 21-34
blood sample and stain
The comparisons between the profiles of the stain and the corresponding blood sample were performed for profiles determined on the same plate as well as on different plates. The number of comparisons of bands from the same plate was 1822, and that from different plates was 1959. Before comparison, all fragment lengths were transformed into normalized migration distance (mm). For each pair of fragments the difference [b - s] (where b and s are the transformed fragment length of the fragment derived from the blood sample and the stain, respectively) was plotted against the fragment length of the blood sample DNA (Figs. lA,B). The band pairs in Fig. 1A derive from the same plate, and those in Fig. 1B from different plates. The scatter was most pronounced for the results derived from different plates. It is also seen that the scatter increased with increasing migration distance (decreasing fragment length). This also appears from Table 4. For both sets of data the S.D. increased with increasing migration distance. At large fragment lengths (40-80 mm, 5.5-21.5 kb) the S.D. of the difference was 0.5 mm on the same plate and 0.7 mm on different plates. The latter value is identical with that found for the difference between duplicate determinations of blood samples on different plates over the whole range of fragment lengths [ 11. 3.3 Band-shift To investigate the cause of the increase in measurement errors with increasing migration distance the dependence of the differences [b - s] of the fragment length was studied for each pair of corresponding profiles. Regression analysis was performed for all pairs of profiles, and approximately 65% of the pairs showed linear slopes significantly higher or lower than zero. The zero point of the regression line was at
Fragment 22 3
r
40
10
5
4
3
100
120
length
kb
Fragment
length
kb
2
A
60
80
Fragment
140
160 length
1
180 mm
-3
i 40
60
80
100
120 Fragment
140
160 length
180 mm
Fig. 3. Scatter diagrams, as Fig. 1. Data from profiles with positive slopes of band-shift >0.005. (A) Results from 198 determinations with blood and stain DNA on the same plate. (B) Results from 257 determinations with blood and stain DNA on different plates.
B. Eriksen, 0. Svensmark/Forensic
Sci. ht. 61 (1993) 21-34
21
approximately 40 mm. In Fig. 2 three examples are shown: one profile with a positive slope, one with a zero slope and one with a negative slope. For 35% of the pairs of profiles the slope was not clearly different from zero. We have, arbitrarily, divided the slopes into 3 classes: Positive slopes >0.005, negative slopes < -0.005 and insignificant slopes for which the numerical slope was less than 0.005, i.e. less than 0.5 mm/l00 mm. For each of the 3 classes the differences [b - s] were plotted against the fragment length for the blood sample DNA (Figs. 3A,B, 4A,B, and 5A,B). The scatter in relation to the regression lines was independent of the fragment length. For the results derived from determinations on the same plate the SD. of [b - S] was 0.53, 0.51 and 0.54 mm for the 3 classes, and for the results derived from different plates it was 0.84, 0.69 and 0.69 mm, respectively. The distribution of the slopes of the regression lines of the difference [h - s] as a function of fragment length is given in Fig. 6. Approximately 50% of the profile pairs exhibited significantly negative slopes and 15% significantly positive slopes. Most of the profiles obtained from stains with semen and vaginal secretion exhibited a negative slope (mean value: -0.006). For profiles from blood stains the mean value of the slope was approximately zero (-0.0005). The majority of profiles with positive slopes (‘reversed’ band-shift) derived from blood stains which were compared with degraded blood samples. 3.4 Correlation The differences in fragment length between corresponding bands of the band pairs of two matching profiles were highly correlated. The correlation was investigated for both data sets, i.e. data derived from the same plate and those from different plates. Furthermore, the correlation was calculated for the 3 classes of slopes in each set.
Fragment 22
10
54
length
Fragment
kb
2
3
'7
10
54
length
kb
2
3
a 40
60
80
100
120 Fragment
140
160 length
180 mm
40
60
80
100
120 Frogment
140
160 length
180 mm
Fig. 4. Scatter diagrams as Fig. 1. Data from profiles with negative slopes of band-shift < -0.005. (A) Results from 722 determinations with blood and stain DNA on the same plate. (B) Results from 892 determinations with blood and stain DNA on different plates.
B. Eriksen. 0. Svensmark / Forensic Sci. ht. 61 (1993) 21-34
28
Fragment 22 3
10
length
3
54
kb
Fragment
length
kb
2 7
A
a, -1
r ._ c!
-2
-3 40
60
80
120
100
Fragment
140
160 length
180
40
60
80
100
mm
120 Fragment
140
180
160 length
mm
Fig. 5. Scatter diagrams as Fig. I. Data from profiles with non-significant band-shift (absolute value of slope <0.005). (A) Results from 902 determinations with blood and stain DNA on the same plate. (B) Results from 810 determinations with blood and stain DNA on different plates.
For all data sets the coefficient of correlation was 0.8. In Fig. 7A the correlation is shown for all data derived from different plates. For comparison, data derived from duplicate determinations of DNA from blood samples determined on different plates are given (Fig. 7B). In both figures the 95, 99 and 99.9% contour ellipses are shown
‘,
O.O”, -4
-3
’
-2-l
/
1, ’
0
1
t-y--y2
3
4
SLOPE Fig. 6. Distribution of the slopes of the band-shift comprising more than 5 bands obtained
(see text) calculated from determinations
%
for 249 pairs of DNA-profiles on the same plate.
B. Eriksen. 0. Svensmark/ Forensic Sci. ht. 61 (1993) 21-34
-4v -4
' -3
I
I
I
I
I
I
I
-2
-1
0
1
2
3
4
Bond
1 (blood - stain)
29
mm
4 3 2 1 0 -1 -2 -3 -4-4
-3
-2
-1
0 Bond
1
2
3
4
1 (d, - d2) mm
Fig. 7. (A) Correlation of the differences b - s for each band in a band pair where b and s are the transformed lengths of the fragments derived from the blood sample and the stain, respectively. Results for 940 band pairs obtained in determinations on different plates. (B) Correlation of the transformed differences between duplicate determinations of each band in a band pair derived from blood samples. d, and d2 are the transformed fragment lengths obtained from the first and the second determination. The data derived from 1583 duplicate determinations of blood samples [l]. The 95, 99 and 99.9% contour ellipses are shown. The formula for the ellipse is x2 + y* - 2pxy = X’LT’(I - p*) where (r* for 2 degrees of freedom is used. p = 0.8.
(p = 0.8; u = 0.7). For the stain/blood pairs (Fig. 7A) several outliers were observed at the lower right part of the ellipse. These outliers all derive from profiles with pronounced band-shift and band pairs with large distances between the bands. Correspondingly, outliers can be expected
30
B. Eriksen, 0. Svensmark / Forensic Sri. ht. 61 (1993) 21-34
Fig. 8. A two-dimensional model for a match criterion (see the text). For a base of the square of 1 the area of the ellipse is 0.55, and that of the circle is 0.23. The area of C + U as well as that of C + L is 0.39. The formula for the ellipse is: x2 + y* - 2pxy = b*(l - p*)/4 where p = 0.7 and b is the base of the square.
at the upper left part of the ellipse due to profiles with reversed band-shift. For the duplicate determinations on blood samples (Fig. 7B) there was good agreement between the observed and the expected distribution within the elliptic areas. 3.5 Match criteria In the investigation of match criteria only the data set obtained from determinations with blood sample and stain on different plates was considered. For this data set, the extreme values of the difference between the fragment lengths of the blood fragment and the corresponding stain fragment were -3.0 mm and 2.66 mm. This might suggest the use of a window of 3 mm width. If, for a given profile, all bands fall within the corresponding windows there will be a formal match. However, false inclusions may occur because the correlation of the deviations is not taken into account. An objective match criterion may be constructed on the basis of the finding that the errors are correlated. The following notation will be used: d, is the difference b, - sl where b, is the transformed fragment length of band 1 derived from the blood sample, and sI is the transformed length of band 1 derived from the stain. Correspondingly d2 is the difference b2 - s2 valid for band 2 of the band pair obtained in the same lane with the same probe. The term accept area will be used for the area within which all coordinates (d,, d2) related to a pair of profiles have to fall if a match has to be declared. For the simple windowing method with a window width of 3 mm the accept area is identical with a square with a base of &3 mm, or 6 mm (Fig. 8). This square accepts obviously mismatching band pairs, e.g. a pair with the coordinates (-2, 2 mm) or (2, -2 mm). Such profiles do certainly not match according to a visual inspection. Some improvement can be obtained by the use of the correlation ellipse shown in Fig. 8 as accept area. This area will not accept extremes as those mentioned above. But, it will still accept false matches between pairs
31
B. Eriksen, 0. Svensmark / Forensic Sci. ht. 61 (1993) 21-34
Table 5 Percent match
for pairs of DNA-profiles
Base of the square
at different
base lengths
Percent
of the square
match
w-r)
f * f f f * ??
f f ??
3.25 3.00 2.15 2.50 2.25 2.00 1.75 1.50 1.25 1.00
Stain/blood
Blood/blood
100 99.3 98.3 96.5 93.8 90.7 81.7 70.9 53.6 37.0
100 99.6 99.2 97.3 92.4 85.1 71.4
(See Fig. 8.) Column 2: Percent match for 289 pairs of DNA-profiles derived from stains and corresponding blood samples. Column 3: Percent match for 262 duplicate determinations of blood samples. The determinations derived from different plates. The accept area is either the area C + U or the area C + L. The coefficient of correlation is 0.7.
with uncorrelated differences, e.g. pairs of profiles which exhibit one band pair with the coordinates (2,2 mm) and another with the coordinates (-2, -2 mm). In our material we have never seen profiles with large deviations of different signs. Therefore, ideally, the requirement will be that for all band pairs belonging to a pair of matching DNA-profiles the coordinates must have the same sign, all positive or all negative, or zero. However, to allow for measurement errors the circular area (C in Fig. 8) has to be included in the accept area. In this model the accept area is either the area C + L or the area C + U. The lower area (C + L) accepts pairs of profiles exhibiting normal band-shift and the upper (C + U) accepts pairs of profiles exhibiting ‘reversed’ bandshift (Fig. 8). This model was tested with the data derived from determinations obtained from different plates. Due to the outliers in the correlation diagram (Fig. 7A) deriving from profiles with a pronounced bandshift, the coefficient of correlation was decreased to 0.7 and the ellipse was inscripted in a square with a base of f 3.25 mm. This corresponds to a maximum difference (window width) of 3.25 mm which again corresponds to 4.6 S.D. All pairs of profiles (289) in the data set were accepted by this ellipse. In Table 5 the percent match is given for different lengths of the base. For comparison, the model was tested on a data set consisting of 262 pairs of determinations of blood samples, i.e. duplicate determinations. In this data set large band-shifts are not expected, and the limits for 100% match are consequently expected to be smaller than for stain/blood sample pairs. The limit was found to be ~2.5 mm, corresponding to &3.3 S.D.
32
B. Eriksen.
0. Svensmark/Forensic
Sci. hf.
61 (1993)
21-34
4. Discussion
Only few reports on the use of RFLP-DNA analysis in casework have been published [3-l 11. These studies do not comprise statistical analyses of the measurement error, though, in one study, the difference in size measurements between blood and vaginal epithelial cell fragments was reported [lo]. The transformation of fragment length into normalized migration distance makes it possible to study the band-shift as a function of the migration distance. For profiles exhibiting a distinct band-shift the difference between the migration of the blood sample and the stain fragments increased with increasing migration distance. The slope may be negative or positive (‘reversed’ band-shift). It is of interest to note that 65% of the stains exhibit a distinct band-shift, negative or positive. Positive band-shift was observed for 15% of the stains, most often found for blood stains which were to be compared with degraded blood samples. Large differences in fragment lengths of corresponding bands (up to 3 mm) were always associated with a pronounced band-shift. Large differences do not reflect a greater measurement error for stains than for blood samples. After correction for the band-shift the S.D. of the difference between measurements of stain and blood sample fragments was identical with that of the difference between duplicate determinations of blood samples. The correlation of the differences between the lengths of fragments from blood sample and stain (Fig. 7A) is less symmetrical than the correlation of the measurement errors derived from duplicate measurements of DNA from blood samples (Fig. 7B). In Fig. 7A all outliers derived from profiles exhibiting pronounced band-shift. Generally, the correlation does not vary much with the difference between corresponding fragment lengths [l], but for profiles with pronounced band-shift the correlation is low for band pairs with large distances between the bands. Different match criteria have been discussed. In its simplest form the requirement for a match is that all differences between the fragment lengths of corresponding bands in absolute values have to be less than a given treshold or guideline, e.g. 5% [lo] or 2.8% kb [12,13]. In one study [14] 4”/0kb was suggested for fragments less than 10 kb and 8% for fragments greater than 10 kb. Windowing with a constant width in units of percent kb suffers from two shortcomings. First, it does not take account of the fact that the measurement error in percent kb is highly dependent on the fragment length [ 11.Second, neither does it take account of the fact that the measurement errors are correlated, and not independent [ 1,15,16]. According to our investigations a window of 2.8% kb corresponds to 0.7 mm (1.4 S.D.) at 20 kb and to 4.3 mm (8.6 SD.) at 2 kb. Such a window leads to false exclusions at large fragment lengths and to false inclusions at small fragment lengths. On the other hand, it accepts extreme band-shifts. However, when windowing is recommended the importance of a visual comparison by an experienced investigator is always emphasized (e.g., Refs. 10,12,13). Windowing on transformed fragment length (mm) is a better approach because the window now represents a constant measure of the S.D. of the measurement error. However, it still takes no account of the fact that the errors are correlated, and false matches will occur. A match criterion based on the Bayesian likelihood ratio has been proposed [ 17-201. The method has been shortly described for a bivariate normal distribution
B. Eriksen, 0. Svensmark /Forensic
Sci. Ini. 61 (1993) 21-34
33
[17], and the model has been been effectively modelled by a a multivariate normal distribution with the appropriate correlation coefficient and S.D. built in [18,20]. However, for 4 or more probes the mathematics become rather intricate. A rule for matching which takes into account correlated errors for single loci has recently been suggested [16]. The simple two-dimensional model suggested by us seems conceptually simple, and it performs very quickly and effectively in search procedures (to be published). 5. Acknowledgements We are obliged to Dr Niels Morling for valuable discussions and to Mrs Jane Hellung Lauridsen and to Mrs Susanne Billesbdle for excellent technical assistance. 6. References I 2 3
4 5
6
7
8
9
10
11 12 13
14
B. Eriksen, A. Bertelsen and 0. Svensmark, Statistical analysis of the measurement errors in the determination of fragment length in DNA-RFLP Analysis. Forensic Sci. Inr., 52 (1992) 181-191. P. Gill, A.J. Jeffreys and D.J. Werrett, Forensic application of DNA ‘fingerprints’. Nature, 318 (1985) 577-579. D.J. Werrett, P.D. Gill, I.W. Evett, J.E. Lygo, K.M. Sullivan and J. Buckleton, DNA analysis in home office laboratories. Its introduction, immediate future and statistical assessment. Proc. In!. Symp. on the Forensic Aspects of DNA Analysis, FBI Academy, Quantico, VI, 1989, pp. 147-162. B.H. Parkin, DNA analysis in the Metropolitan Police Forensic Science Laboratory. Proc. Int. Symp. on the Forensic Aspects of DNA Analysis, FBI Academy, Quantico, VI, 1989, pp. 163-167. H. Schmitter, S. Herrmann and W. Pflug, The use of DNA polymorphisms in the police laboratories of the Federal Republic of Germany. Proc. Int. Symp. on the Forensic Aspects of DNA Analysis. FBI Academy, Quantico, VI, 1989, pp. 169-172. D.E. Adams, L. A. Presley, H. A. Deadman and A.G. Lynch, DNA analysis in the FBI laboratory. Proc. Int. Symp. on the Forensic Aspects of DNA Analysis, FBI Academy, Quantico, VI, 1989, pp. 173-177. F.J. Burridge, M.J. Greenhalgh and G.M. Willot, An evaluation of single locus probes in casework. In: C. Rittner and P.M. Schneider (eds.), Advances in Forensic Huemogenetics, Vol. 4, Springer Verlag, Berlin, 1992, pp. 210-212. W. Pflug, J. Teifel-Greding, S. Herrmann, M. Gerhard, R. Wenzl and H. Schmitter, Results of DNA analysis from six forensic science laboratories in Germany. In: C. Rittner and P.M. Schneider (eds.). Advances in Forensic Haemogenetics, Vol. 4, Springer Verlag, Berlin, 1992, pp. 237-239. B. Mevsg, S. Jacobsen, B. Eriksen and B. Olaisen, DNA typing in forensic casework in Norway. Strategies and experiences. In: C. Rittner and P.M. Schneider (eds.), Advances in Forensic Haemogenetics, Vol. 4, Springer Verlag, Berlin, 1992, pp. 260-262. B. Budowle, A.M. Giusti, J.S. Waye, F.S. Baechtel, R.M. Fourney. D.E. Adams, L.A. Presley, H.A. Deadman and K.L. Monson, Fixed-bin analysis for statistical evaluation ofcontinuous distributions of allelic data from VNTR loci for use in forensic comparisons. Am. .I Hum. Gener., 48 (1991) 841-855. M. Greenhalgh, F. Burridge and G. Willot, Experiences with single locus DNA probes in casework. Forensic Sci. Inr.. 57 (1992) 29-37. I.W. Evett and P. Gill, A discussion of the robustness of methods for assessing the evidential value of DNA single locus profiles in crime investigations. Electrophoresis, 12 (1991) 226-230. P. Gill, I.W. Evett, S. Woodroffe. J.E. Lygo, E. Millican and M. Webster, Databases, quality control and interpretation of DNA profiling in the home office forensic science service. Electrophoresis. 12 (1991) 204-209. T. Staples, C.K. Goff, J.G. Wegel and G. Herrin, RFLP match criteria. Determination from data
34
IS 16 17 18 19 20
B. Eriksen. 0. Svensmark /Forensic
Sci. hi. 61 (1993) 21-34
analysed on a Bioimage system. Proc. ht. Symp. Human Identijication, Promega Corporation, Madison, WI, 1991, p. 316. N.J. Risch and B. Devlin, On the probability of matching DNA fingerprints. Science, 255 (1992) 717-720. N. Risch and B. Devlin, Letter. Science, 256 (1992) 1744-1745. I.W. Evett and D.J. Werrett, Bayesian analysis of single locus DNA profiles. Proc. hi. Symp. Human Identification, Promega Corporation, Madison, WI, 1989, pp. 77-101. I.W. Evett, J. Scranage and R. Pinchin, Efficient retrieval from DNA databases: Based on the second European DNA Profiling Group collaborative experiment. Forensic Sci. hf., 53 (1991) 45-50. D.A. Berry, I.W. Evett and R. Pinchin, Statistical inference in crime investigations using deoxyribonucleic acid profiling. Appl. Star., 41 (1992) 499-531. I.W. Evett, J.K. Scranage and R. Pinchin. An efficient procedure for interpreting DNA single locus profiling data in crime cases, J. Forensic Sri. Sot., 32 (1992) 307-324.