You are on page 1of 5

The Coming Paradigm Shift in Forensic Identification Science

Michael J. Saks, et al.

Science 309, 892 (2005);
DOI: 10.1126/science.1111565

This copy is for your personal, non-commercial use only.

If you wish to distribute this article to others, you can order high-quality copies for your
colleagues, clients, or customers by clicking here.

Downloaded from on February 15, 2012

Permission to republish or repurpose articles or portions of articles can be obtained by
following the guidelines here.

The following resources related to this article are available online at (this infomation is current as of February 15, 2012 ):

Updated information and services, including high-resolution figures, can be found in the online
version of this article at:
This article has been cited by 60 article(s) on the ISI Web of Science
This article has been cited by 13 articles hosted by HighWire Press; see:
This article appears in the following subject collections:

Science (print ISSN 0036-8075; online ISSN 1095-9203) is published weekly, except the last week in December, by the
American Association for the Advancement of Science, 1200 New York Avenue NW, Washington, DC 20005. Copyright
2005 by the American Association for the Advancement of Science; all rights reserved. The title Science is a
registered trademark of AAAS.
The Coming Paradigm Shift in
Forensic Identification Science
Michael J. Saks1 and Jonathan J. Koehler2

that different objects share a common set of

Converging legal and scientific forces are pushing the traditional forensic identification observable attributes. Without the discernible
sciences toward fundamental change. The assumption of discernible uniqueness that re- uniqueness assumption, far more scientific work
sides at the core of these fields is weakened by evidence of errors in proficiency testing and would be needed, and criminalists would need
in actual cases. Changes in the law pertaining to the admissibility of expert evidence in to offer more tempered opinions in court.
court, together with the emergence of DNA typing as a model for a scientifically de- Legal and scientific forces are converging
fensible approach to questions of shared identity, are driving the older forensic sciences to drive an emerging skepticism about the

Downloaded from on February 15, 2012

toward a new scientific paradigm. claims of the traditional forensic individual-
ization sciences. As a result, these sciences

ittle more than a decade ago, forensic in- different, criminalists conclude that the marks are moving toward a new scientific paradigm.
dividualization scientists compared pairs were made by the same person or object. EWe use the notion of paradigm shift not as
of marks (handwriting, fingerprints, tool Although lacking theoretical or empirical a literal application of Thomas Kuhn_s con-
marks, hair, tire marks, bite marks, etc.), in- foundations, the assumption of discernible cept (9), but as a metaphor highlighting the
tuited whether the marks matched, and testified uniqueness offers important practical benefits transformation involved in moving from a pre-
in court that whoever or whatever made one to the traditional forensic sciences. It enables science to an empirically grounded science.^
made the other. Courts almost never excluded forensic scientists to draw bold, definitive con- Two such forces are outgrowths of DNA typ-
the testimony. Cross-examination rarely ques- clusions that can make or break cases. It ex- ing: the discovery of erroneous convictions and
tioned the foundations of the asserted expertise cuses the forensic sciences from developing a model for a scientifically sound identification
or the basis of the analyst_s certainty. measures of object attributes, collecting popu- science. A third force is the momentous change
Today, that once-complacent corner of lation data on the frequencies of variations in in the legal admissibility standards for expert
the law and science interface has begun to those attributes, testing attribute independence, testimony. A final force grows from studies
unravelor at least to regroup. The news car- or calculating and explaining the probability of error rates across the forensic sciences.
ries reports of erroneous forensic identifica-
tions of hair, bullets, handwriting, footprints,
bite marks, and even venerated fingerprints.
Scientists have begun to question the core
assumptions of numerous forensic sciences
(16). Federal funding has materialized to sup-
port research that examines long-asserted but
unproven claims. Courts have started taking
challenges to asserted forensic science exper-
tise seriously (1). A dispassionate scientist or
judge reviewing the current state of the tra-
ditional forensic sciences would likely regard
their claims as plausible, underresearched, and
The traditional forensic individualiza-
tion sciences rest on a central assumption: that
two indistinguishable marks must have been
produced by a single object. Traditional foren-
sic scientists seek to link crime scene evi-
dence to a single person or object Bto the
exclusion of all others in the world[ (7, 8).
They do so by leaning on the assumption of
discernible uniqueness. According to this as-
sumption, markings produced by different
people or objects are observably different. Thus,
when a pair of markings is not observably

College of Law, Arizona State University, Tempe, AZ Fig. 1. Factors associated with wrongful conviction in 86 DNA exoneration cases, based on case
85287, USA. E-mail: 2McCombs School analysis data provided by the Innocence Project, Cardozo School of Law (New York, NY), and
of Business, University of Texas, Austin, TX 78712, computed by us. Percentages exceed 100% because more than one factor was found in many
USA. E-mail: cases. Red bars indicate factors related to forensic science.

892 5 AUGUST 2005 VOL 309 SCIENCE

former forensic scientist noted, this pressure- ious alleles occur at different locations on the
packed environment can lead to data fudging DNA strand. The traditional forensic sciences
and fabrication: All [forensic science] experts could and should emulate this approach (23).
are tempted, many times in their careers, to Each subfield must construct databases of
report positive results when their inquiries sample characteristics and use these data-
come up inconclusive, or indeed to report a bases to support a probabilistic approach to
negative result as positive [(15), p. 17]. identification. Fingerprinting could be one of
the first areas to make the transition to this
DNA Typing as the New Model for approach because large fingerprint databases
Fig. 2. Bite mark evidence exhibit from trial of Scientific Forensic Identification already exist. The greatest challenge in this
Ray Krone, suggesting alignment of a cast of Much of the above criticism does not apply effort would be to develop measures of the
Krones dentition with bite wounds in victims to the science of DNA typing as practiced complex images presented by fingerprints,
flesh [State v. Krone, 182 Ariz. 319 (1995)]. A
forensic odontologist testified that this showed
today. Indeed, DNA typing can serve as a tool marks, bite marks, handwriting, etc.
Krone to be the biter. Krone was convicted of model for the traditional forensic sciences in (Figs. 2 and 3). Forensic scientists will need
murder and sentenced to death, but a decade three important respects. First, DNA typing to work with experts in differential geometry,
later he was exonerated by DNA analysis. [Source: technology was an application of knowledge topology, or other fields to develop workable
E. Thomas Barham (Los Alamitos, CA) and Alan derived from core scientific disciplines. This measures.
Simpson (Phoenix, AZ), attorneys for Krone] provided a stable structure for future empirical A second data collection effort that would
work on the technology. Second, the courts strengthen the scientific foundation of the fo-

Downloaded from on February 15, 2012

Post-Conviction DNA Exonerations and scientists scrutinized applications of the rensic sciences involves estimating error rates.
During the past decade, scores of people who technology in individual cases. As a result, Although the theoretical promise of forensic
were convicted of serious crimesincluding early, unscientific practices were rooted out. technology is considerable, the practical value
at least 14 who had been sentenced to death Third, DNA typing offered data-based, prob- of any particular technology is limited by the
have been exonerated by DNA analyses of abilistic assessments of the meaning of evi- extent to which potentially important errors
crime scene evidence that had not been tested dentiary matches. This practice represented arise. The best way to identify the frequency
at the time of their trials (10). It was not an advance over potentially misleading match/ with which errors occur is to conduct blind,
surprising to learn that erroneous convictions no-match claims associated with other forensic external proficiency tests using realistic sam-
sometimes occur, and that new science and identification sciences. ples. A proficiency test requires analysts to
technology can help detect and correct those Immediately after DNAs first courtroom make judgments about samples whose proper-
mistakes. Nor was it surprising to learn, from appearance in the 1980s, scientists from dis- ties are known. External proficiency tests are
an analysis of 86 such cases (Fig. 1), that ciplines as varied as statistics, psychology, and conducted by an agency unaffiliated with the
erroneous eyewitness identifications are the evolutionary biology debated the strengths and forensic scientists laboratory. Externality is
most common contributing factor to wrongful limitations of forensic DNA evidence. Blue- important to the integrity of proficiency tests
convictions. What was unexpected is that ribbon panels were convened, conferences were because laboratories have strong incentives to
erroneous forensic science expert testimony held, unscientific practices were identi-
is the second most common contributing fied, data were collected, critical papers
factor to wrongful convictions, found in 63% were written, and standards were de-
of those cases. These data likely understate the veloped and implemented. The scientif-
relative contribution of forensic science expert ic debates focused on the adequacy of
testimony to erroneous convictions. Whereas DNA databases (16), the computation
lawyers, police, and lay witnesses participate of DNA match probabilities (17), the
in virtually every criminal case, forensic training of DNA analysts (18), the pre-
science experts participate in a smaller subset sentation of DNA matches in the court-
of casesabout 10 to 20% of criminal cases room (19), and the role of error rates
during the era when these DNA exonerations (20). In some cases, disputants worked
were originally tried (11). together to find common ground (21).
Figure 1 also indicates that forensic scien- These matters were not resolved by the
tists are the witnesses most likely to present forensic scientists themselves, by fiat, or
misleading or fraudulent testimony. Deceitful by neglect. Most exaggerated claims
forensic scientists are a minor sidelight to this and counterclaims about DNA evidence
paper, but a sidelight that underscores cultural have been replaced by scientifically
differences between normal science and foren- defensible propositions. Although some Fig. 3. Image of two bullets viewed through a com-
sic science (12, 13). In normal science, academ- disagreement remains (22), the scientif- parison microscope. The bullets were fired from two
ically gifted students receive four or more years ic process worked. consecutively manufactured Smith & Wesson 38 Spe-
of doctoral training where much of the social- One of the great strengths of DNA cial revolver barrels. Whether fired through the same or
ization into the culture of science takes place. typing is that it uses a statistical ap- different barrels, numerous matching and nonmatching
striations are engraved onto bullets. To reliably identify
This culture emphasizes methodological rigor, proach based on population genetics the barrel through which a questioned bullet was fired,
openness, and cautious interpretation of data. theory and empirical testing. Experts an examiner must distinguish among class, subclass, and
In forensic science, 96% of positions are held evaluate matches between suspects and individual characteristics. These two bullets illustrate sub-
by persons with bachelors degrees (or less), crime scene DNA evidence in terms class characteristic agreement of striated markings on a
3% masters degrees, and 1% Ph.D.s (14). of the probability of random matches groove impression that could be mistaken for individual
When individuals who are not steeped in the across different reference populations characteristics. Without investigating the potential for
subclass carryover, the examiner could mistake these as
culture of science work in an adversarial, crime- (e.g., different ethnicities). These prob- having been fired from the same gun. [Source: Bruce
fighting culture, there is a substantial risk that a abilities are derived from databases that Moran, firearms examiner with the Sacramento County
different set of norms will prevail. As one identify the frequency with which var- (CA) District Attorney, Laboratory of Forensic Services] SCIENCE VOL 309 5 AUGUST 2005 893

be perceived as error-free. An even better of science, cannot, after Daubert, be regarded been remarkably little research on the ac-
test would be a blind proficiency test, in as scientific I knowledge (p. 1038). How- curacy of traditional forensic sciences. Pro-
which the analyst believes the test materials ever, the court did not exclude this unscientific ficiency tests in some fields offer a step in
are part of ordinary case work. Blindness in- testimony. It reasoned that handwriting identi- the right direction, even though simple tasks
creases the validity of proficiency test results fication did not have to reach the Daubert and infrequent peer review limit their value.
because it ensures that analysts treat the test standard because Daubert applied only to sci- Nonetheless, the available data hint that
sample as they would other case samples. Al- entific evidence, and handwriting identification some forensic sciences are best interpreted in
though proficiency tests are used in many plainly was not scientific evidence. Thus, when tandem with error rates estimated from sound
forensic sciences, the tests are generally in- a forensic science was found to stand on a studies.
frequent, internal, and unrealistic; blind tests weak foundation, the threshold of admission Unfortunately, forensic scientists often re-
are practically nonexistent. was lowered to accommodate this weakness. ject error rate estimates in favor of arguments
In Kumho Tire v. Carmichael [526 U.S. that theirs is an error-free science. For exam-
Changes in the Law 137 (1999)], the Supreme Court directly con- ple, an FBI document section chief asserted
Until recently, courts assessed expertise by fronted the question of whether Daubert that all certified document examiners in the
looking for superficial indicia of validity. In applies to nonsciences. A consortium of law United States would agree with his conclu-
the 19th century, courts were impressed by enforcement organizations prepared an amicus sions in every case [(25), p. 196]. Likewise,
qualifications and success in the market- brief urging that Daubert scrutiny not be ex- fingerprint experts commonly claim that all
place. If the market valued an asserted ex- tended to the testimony of police agency ex- fingerprint experts would reach the same con-
pertise or expert, courts generally did, too. pert witnesses. The brief argued that the clusions about every print (2). Such hubris

Downloaded from on February 15, 2012

In Frye v. United States [293 F. 1013 (D.C. great bulk of expert testimony provided by was on display in spring 2004 when the FBI
Cir. 1923)], a federal appellate court law enforcement officers does not involve sci- declared that a fingerprint recovered from a
confronted the question of admissi- suspicious plastic bag near the scene
bility of an expertise that had no life in of a terrorist bombing in Madrid pro-
any commercial marketplace. The vided a 100 percent match to an
court solved the problem by substitut- Oregon attorney (Fig. 4). The FBI
ing an intellectual marketplace. The eventually conceded error when Spanish
court asked whether the proffered fingerprint experts linked the print to
expertise had gained general accept- someone else (26).
ance in the particular field in which it The FBI and other agencies often
belongs. Sixty years later, the Frye seek to preserve the illusion of perfec-
test had become the dominant expert tion after disclosure of such errors by
evidence filter in American courts. distinguishing between human errors
In 1993, the law began to catch up (possible) and errors of method
with the scientific method. In Daubert (impossible). A leading FBI scientist
v. Merrell Dow Pharmaceuticals [509 explained the distinction to the court in
U.S. 579 (1993)], the U.S. Supreme United States v. Llera-Plaza I [58 Fed.
Fig. 4. (A) A latent fingerprint believed to belong to a terrorist
Court introduced a new standard for involved in train bombings in Madrid, Spain, in March 2004. R. Evid. Serv. 1 (E.D. Pa. 2002)]: We
the admissibility of scientific evi- (B) A database print belonging to Brandon Mayfield of Port- have to understand that error rate is a
dence. Under Daubert, proffered sci- land, Oregon. On the basis of these prints (though not neces- difficult thing to calculate. I mean, peo-
entific testimony must be shown to sarily these very images), FBI fingerprint examiners erroneously ple are trying to do this, it shouldnt be
stand on a dependable foundation. The identified Mayfield as the bomber (26). [Source: Problem Idents, done, it cant be doneI. An error rate
court suggested that trial judges mak-] is a wispy thing like smoke, it changes
ing this determination consider wheth- over timeI. If you made a mistake in
er the proffered science has been tested, the entific theories, methodologies, techniques, or the past, certainly thats valid information I
methodological soundness of that testing, and data in any respectI. Instead, law enforce- but to say theres an error rate thats definable
the results of that testing. The Daubert test in ment officers testify about such things as would be a misrepresentationI. Now, error
effect lowers the threshold for admission of accident reconstruction, fingerprint, footprint rate deals with people, you should have a
sound cutting-edge science and raises the and handprint [identification], handwriting method that is defined and stays within its
threshold for long-asserted expertise that lacks analysis, firearms markings and toolmarks limits, so it doesnt have error at all. So the
a scientific foundation. Seriously applied, the and the unique characteristics of guns, bullets, method is one thing, people making mistakes
Daubert test subjects the forensic sciences to a and shell casings, and bloodstain pattern iden- is another issue.
first-principles scientific scrutiny that poses a tification (24). Ironically, then, fields that Such claims are problematic. First, the
profound challenge to fields that lack rigorous initially gained entry to the courts by declaring suggestion that humans err but forensic tech-
supporting data. themselves to be sciences now sought to niques do not is unfalsifiable. It is impossible
United States v. Starzecpyzel [880 F. Supp. remain in court by denying any connection to disentangle method errors from prac-
1027 (S.D.N.Y. 1995)] offered an early in- with scientific methods, data, or principles. titioner errors in fields where the method is
dication of how Daubert could change judicial Despite efforts to preserve the nonscience primarily the judgment of the examiner. Sec-
views. After an extensive hearing on the sound- loophole, the Supreme Court doctrinally sealed ond, even if such disentanglement were pos-
ness of asserted handwriting identification ex- it shut when Kumho Tire held that all expert sible, it is a red herring. When fact-finders
pertise, a federal district court concluded that testimony must pass appropriate tests of hear evidence of a forensic match, a proper
the field had no scientific basis: [T]he tes- validity to be admissible in court. assessment of the probative value of that match
timony at the Daubert hearing firmly es- requires awareness of the chance that a mistake
tablished that forensic document examination, Error Rates was made. The source of such a mistake is
despite the existence of a certification pro- Although Dauberts testing recommenda- irrelevant for this purpose. If method errors
gram, professional journals and other trappings tions are familiar to most scientists, there has could be distinguished from practitioner errors,

894 5 AUGUST 2005 VOL 309 SCIENCE

a 1% method error affects the probative value ficial testing situations; erroneous fingerprint constructed proficiency testswould inform
of the match in exactly the same way as a 1% identifications have made their way out of fact-finders about the probative value of the
practitioner error. Identifying sources of error the crime lab and into prosecutions in at least evidentiary match.
is relevant for improving forensic science prac- 21 documented cases (32). Simply put, we envision a paradigm shift
tice, but it plays no role in identifying the Forensic science proficiency tests and ex- in the traditional forensic identification sci-
probative importance of a match. aminations are obviously imperfect indicators ences in which untested assumptions and semi-
Third, the suggestion that error rates do of the rate at which errors occur in practice. informed guesswork are replaced by a sound
not exist because they change over time and This fact does not justify ignoring the worri- scientific foundation and justifiable protocols.
are not specific to the case at hand is a base- some data these tests have yielded. Indeed, Although obstacles exist both inside and out-
rate fallacy. In this fallacy of reasoning, peo- these data are probably best regarded as lower- side forensic science, the time is ripe for the
ple underuse (or willfully ignore) general bound estimates of error rates. Because the traditional forensic sciences to replace anti-
background data in judgment tasks because tests are relatively easy (according to test par- quated assumptions of uniqueness and per-
they believe the data are irrelevant to the ticipants), and because participants know that fection with a more defensible empirical and
instant case. However, general background mistakes will be identified and punished, test probabilistic foundation.
data (or base rates) are relevant for specific error rates (particularly the false-positive error
predictions (27, 28). For example, although rate) probably are lower than those in every- References
risk estimates for a disease fluctuate and are day casework (33, 34). 1. D. L. Faigman, D. Kaye, M. J. Saks, J. Sanders, Modern
developed on patients other than the patient The studies mentioned above cry out for Scientific Evidence: The Law and Science of Expert
Testimony (Thompson-West, St. Paul, MN, ed. 2, 2002).
now seeking medical advice, these estimates attention and follow-up investigations. In light

Downloaded from on February 15, 2012

2. R. Epstein, S. Cal. Law Rev. 75, 605 (2002).
provide information useful for predicting of the laws growing reluctance to accept 3. M. J. Saks, Hastings Law J. 49, 1069 (1998).
whether this patient will contract the disease. experts personal guarantees in lieu of sci- 4. National Research Council, Forensic Analysis: Weighing
A 20% base-rate risk of contracting the dis- entific data, these studies should increase Bullet Lead Evidence (National Academy Press, Wash-
ington, DC, 2004).
ease makes it more likely that the patient candor about performance and create pres- 5. I. Pretty, D. Sweet, Sci. Justice 41, 85 (2001).
will get the disease than would a 1% risk. sure for improvement. 6. C. M. Cooley, Stanford Law Policy Rev. 15, 381 (2004).
Likewise, an X% base-rate risk of error in a 7. D. A. Stoney, J. Forensic Sci. Soc. 31, 197 (1991).
The Future 8. J. W. Osterburg, J. Crim. Law Criminol. 60, 97 (1969).
given forensic science provides some indica- 9. T. S. Kuhn, The Structure of Scientific Revolutions
tion of the chance that a particular conclusion The traditional forensic sciences need look (Univ. of Chicago Press, Chicago, 1962).
is in error (22). no further than their newest sister discipline, 10. B. Scheck, P. Neufeld, J. Dwyer, Actual Innocence
(Doubleday, New York, 2000).
Data from proficiency tests and other ex- DNA typing, for guidance on how to put the 11. J. Peterson, Forensic Science and the Courts (National
aminations suggest that forensic errors are not science into forensic identification science. Institute of Justice, Washington, DC, January 1986).
minor imperfections. Spectrographic voice This effort should begin with adoption of the 12. P. C. Giannelli, Issues Sci. Tech. 20, 55 (fall 2003).
13. P. C. Giannelli, Va. J. Soc. Policy Law 4, 439 (1997).
identification error rates are as high as 63%, basic-research model. Just as DNA scientists 14. K. G. Furton, Y. L. Hsu, M. D. Cole, J. Forensic Sci. 44,
depending on the type of voice sample tested tested the genetic assumptions that undergirded 128 (1999).
[(1), chap. 31]. Handwriting error rates aver- DNA typing theory (e.g., Hardy-Weinberg equi- 15. A. A. Moenssens, J. Crim. Law Criminol. 84, 1 (1993).
16. Committee on DNA Technology in Forensic Science,
age around 40% and sometimes approach librium), traditional forensic scientists should DNA Technology in Forensic Science (National Re-
100% [(1), chap. 28]. False-positive error rates design experiments that test the core assump- search Council, Washington, DC, 1992).
for bite marks run as high as 64% [(1), chap. tions of their fields. As basic research knowl- 17. R. C. Lewontin, D. L. Hartl, Science 254, 1745 (1991).
30]. Those for microscopic hair comparisons edge grows, experts will be able to inform 18. Technical Working Group on DNA Analysis Methods,
Crime Lab. Digest 18, 44 (1991).
are about 12% (using results of mitochondrial courts about the relative strengths and weak- 19. K. Roeder, Stat. Sci. 9, 263 (1994).
DNA testing as the criterion) (29). Fingerprint nesses of their theories and methods, and 20. J. J. Koehler, A. Chia, J. S. Lindsey, Jurimetrics J. 35,
examiners generally fare better, although suggest how that knowledge applies to indi- 201 (1995).
21. B. Budowle, E. S. Lander, Nature 371, 735 (1994).
data from a well-known forensic testing pro- vidual cases. 22. J. J. Koehler, Jurimetrics J. 37, 425 (1997).
gram contradict industry boasts of perfect, or At the same time, data should be collected 23. M. J. Saks, J. J. Koehler, Cardozo Law Rev. 13, 361 (1991).
even near-perfect, agreement (30). Since on the frequency with which markings and at- 24. Brief Amicus Curiae of Americans for Effective Law En-
forcement et al., submitted in Kumho Tire v. Carmichael,
1995, about one-fourth of examiners failed tribute variations occur in different popula- No. 97-1709, in the Supreme Court of the United
to correctly identify all latent prints in this tions. In addition to their case-specific benefits, States (October Term, 1997).
test (which includes 9 to 12 latent prints and these data may also facilitate the development 25. D. Fisher, Hard Evidence (Simon & Schuster, New
York, 1995).
palmprints). About 4 to 5% of examiners of artificial intelligence or computer-aided pat- 26. D. Heath, H. Bernton, Portland lawyer released in probe
committed false-positive errors on at least tern recognition programs for the identification of Spain bombings, Seattle Times (21 May 2004).
one latent. In one test, 20% of examiners sciences. Forensic scientists might also adopt 27. D. Kahneman, A. Tversky, Psychol. Rev. 80, 237 (1973).
mistook one persons prints for those of his protocols, such as blind examinations in com- 28. J. J. Koehler, Behav. Brain Sci. 19, 1 (1996).
29. M. M. Houck, B. Budowle, J. Forensic Sci. 47, 964 (2002).
twin. The editor of the leading fingerprint bination with realistic samples, that minimize 30. Collaborative Testing Services Inc., Forensics Testing
journal called this performance unaccept- the risks that their success rates will be inflated Program: Latent Prints Examination, 19932004.
able [(31), p. 524]. It is noteworthy that and their conclusions biased by extraneous ev- 31. D. Grieve, J. Forensic Ident. 46, 524 (1996).
32. S. Cole, J. Crim. Law Criminol., in press (available at
these misidentifications are not confined to a idence and assumptions (34). When matches
single lab, circumstance, or marking. More- are identified, forensic scientists in all fields 33. J. L. Peterson et al., Forensic Evidence and the Police
over, the misidentification rates do not show would compute and report random-match prob- (National Institute of Justice, Washington, DC, 1984).
34. M. J. Saks, D. M. Risinger, R. Rosenthal, W. Thompson,
a clear pattern of improvement (the mis- abilities similar to those used in DNA typing. Sci. Justice 43, 77 (2003).
identification rates in 2004 were 4 to 6%). These estimatesin combination with error
Nor are these errors limited to arguably arti- rate estimates provided by mandatory, well- 10.1126/science.1111565 SCIENCE VOL 309 5 AUGUST 2005 895