Measuring Accuracy Facial Forensics Comparisons - Journal-Paper

Accuracy and reliability of forensic
latent fingerprint decisions
Bradford T. Ulerya, R. Austin Hicklina, JoAnn Buscagliab,1, and Maria Antonia Robertsc
a
Noblis, 3150 Fairview Park Drive, Falls Church, VA 22042; bCounterterrorism and Forensic Science Research Unit, Federal Bureau of Investigation
Laboratory Division, 2501 Investigation Parkway, Quantico, VA 22135; and cLatent Print Support Unit, Federal Bureau of Investigation Laboratory
Division, 2501 Investigation Parkway, Quantico, VA 22135

The interpretation of forensic fingerprint evidence relies on the
expertise of latent print examiners. The National Research Council
of the National Academies and the legal and forensic sciences communities have called for research to measure the accuracy and
reliability of latent print examiners’ decisions, a challenging and
complex problem in need of systematic analysis. Our research is
focused on the development of empirical approaches to studying
this problem. Here, we report on the first large-scale study of the
accuracy and reliability of latent print examiners’ decisions, in
which 169 latent print examiners each compared approximately
100 pairs of latent and exemplar fingerprints from a pool of 744
pairs. The fingerprints were selected to include a range of attributes and quality encountered in forensic casework, and to be
comparable to searches of an automated fingerprint identification
system containing more than 58 million subjects. This study evaluated examiners on key decision points in the fingerprint examination process; procedures used operationally include additional
safeguards designed to minimize errors. Five examiners made false
positive errors for an overall false positive rate of 0.1%. Eighty-five
percent of examiners made at least one false negative error for an
overall false negative rate of 7.5%. Independent examination of
the same comparisons by different participants (analogous to blind
verification) was found to detect all false positive errors and the
majority of false negative errors in this study. Examiners frequently
differed on whether fingerprints were suitable for reaching a
conclusion.

T

he interpretation of forensic fingerprint evidence relies on the
expertise of latent print examiners. The accuracy of decisions
made by latent print examiners has not been ascertained in a
large-scale study, despite over one hundred years of the forensic
use of fingerprints. Previous studies (1–4) are surveyed in ref. 5.
Recently, there has been increased scrutiny of the discipline
resulting from publicized errors (6) and a series of court admissibility challenges to the scientific basis of fingerprint evidence
(e.g., 7–9). In response to the misidentification of a latent print
in the 2004 Madrid bombing (10), a Federal Bureau of Investigation (FBI) Laboratory review committee evaluated the scientific
basis of friction ridge examination. That committee recommended research, including the study described in this report:
a test of the performance of latent print examiners (11). The need
for evaluations of the accuracy of fingerprint examination decisions has also been underscored in critiques of the forensic
sciences by the National Research Council (NRC, ref. 12) and
others (e.g., refs. 13–16).
Background
Latent prints (“latents”) are friction ridge impressions (fingerprints, palmprints, or footprints) left unintentionally on items
such as those found at crime scenes (SI Appendix, Glossary).
Exemplar prints (“exemplars”), generally of higher quality, are
collected under controlled conditions from a known subject using
ink on paper or digitally with a livescan device (17). Latent print
examiners compare latents to exemplars, using their expertise
rather than a quantitative standard to determine if the informawww.pnas.org/cgi/doi/10.1073/pnas.1018707108

tion content is sufficient to make a decision. Latent print examination can be complex because latents are often small, unclear,
distorted, smudged, or contain few features; can overlap with
other prints or appear on complex backgrounds; and can contain
artifacts from the collection process. Because of this complexity,
experts must be trained in working with the various difficult
attributes of latents.
During examination, a latent is compared against one or more
exemplars. These are generally collected from persons of interest
in a particular case, persons with legitimate access to a crime
scene, or obtained by searching the latent against an Automated
Fingerprint Identification System (AFIS), which is designed to
select from a large database those exemplars that are most similar
to the latent being searched. For latent searches, an AFIS only
provides a list of candidate exemplars; comparison decisions must
be made by a latent print examiner. Exemplars selected by an
AFIS are far more likely to be similar to the latent than exemplars
selected by other means, potentially increasing the risk of examiner error (18).
The prevailing method for latent print examination is known
as analysis, comparison, evaluation, and verification (ACE-V)
(19, 20). The ACE portion of the process results in one of four
decisions: the analysis decision of no value (unsuitable for comparison); or the comparison/evaluation decisions of individualization (from the same source), exclusion (from different sources),
or inconclusive. The Scientific Working Group on Friction Ridge
Analysis, Study and Technology guidelines for operational procedures (21) require verification for individualization decisions,
but verification is optional for exclusion or inconclusive decisions.
Verification may be blind to the initial examiner’s decision, in
which case all types of decisions would need to be verified.
ACE-V has come under criticism by some as being a general
approach that is underspecified (e.g., refs. 14 and 15).
Latent-exemplar image pairs collected under controlled conditions for research are known to be mated (from the same
source) or nonmated (from different sources). An individualization decision based on mated prints is a true positive, but if based
on nonmated prints, it is a false positive (error); an exclusion
decision based on mated prints is a false negative (error), but
is a true negative if based on nonmated prints. The term “error”
is used in this paper only in reference to false positive and false
negative conclusions when they contradict known ground truth.
No such absolute criteria exist for judging whether the evidence is
sufficient to reach a conclusion as opposed to making an inconclusive or no-value decision. The best information we have to
Author contributions: B.T.U., R.A.H., J.B., and M.A.R. designed research; B.T.U., R.A.H., J.B.,
and M.A.R. performed research; B.T.U. and R.A.H. contributed new analytic tools; B.T.U.,
R.A.H., J.B., and M.A.R. analyzed data; and B.T.U., R.A.H., J.B., and M.A.R. wrote the paper.
The authors declare no conflict of interest.
This article is a PNAS Direct Submission.
Freely available online through the PNAS open access option.
1

To whom correspondence should be addressed. E-mail: joann.buscaglia@ic.fbi.gov.

This article contains supporting information online at www.pnas.org/lookup/suppl/
doi:10.1073/pnas.1018707108/-/DCSupplemental.

PNAS Early Edition ∣

1 of 6

APPLIED BIOLOGICAL
SCIENCES

Edited by Stephen E. Fienberg, Carnegie Mellon University, Pittsburgh, PA, and approved March 31, 2011 (received for review December 16, 2010)

2 of 6 ∣

www.pnas.org/cgi/doi/10.1073/pnas.1018707108

Presentation of next latent
Analysis
No

Value for
exclusion only
(VEO)

No

Yes

Value for
individualization
(VID)
Yes
Latent print

Comparison and Evaluation

Latent &
exemplar
prints

Inconclusive
reason

Exclusion

Comparison
decision
Inconclusive

Study Description
This study is part of a larger research effort to understand the
accuracy of examiner conclusions, the level of consensus among
examiners on decisions, and how the quantity and quality of
image features relate to these outcomes. Key objectives of this
study were to determine the frequency of false positive and false
negative errors, the extent of consensus among examiners, and
factors contributing to variability in results. We designed the
study to enable additional exploratory analyses and gain insight
in support of the larger research effort.
There is substantial variability in the attributes of latent prints,
in the capabilities of latent print examiners, in the types of
casework received by agencies, and the procedures used among
agencies. Average measures of performance across this heterogeneous population are of limited value (25)—but do provide insight necessary to understand the problem and scope future work.
Furthermore, there are currently no means by which all latent
print examiners in the United States could be enumerated or used
as the basis for sampling: A representative sample of latent print
examiners or casework is impracticable.
To reduce the problem of heterogeneity, we limited our scope
to a study of performance under a single, operationally common
scenario that would yield relevant results. This study evaluated
examiners at the key decision points during analysis and evaluation. Operational latent print examination processes may include
additional steps, such as examination of original evidence or
paper fingerprint cards, review of multiple exemplars from a
subject, consultation with other examiners, revisiting difficult
comparisons, verification by another examiner, and quality assurance review. These steps are implemented to reduce the possibility of error.
Ideally, a study would be conducted in which participants were
not aware that they were being tested. The practicality of such an
approach even within a single organization would depend on the
type of casework. Fully electronic casework could allow insertion
of test data into actual casework, but this may be complex to the
point of infeasibility for agencies in which most examinations
involve physical evidence, especially when chain-of-custody issues
are considered. Combining results among multiple agencies with
heterogeneous procedures and types of casework would be problematic.
In order to get a broad cross-section of the latent print examiner community, participation was open to practicing latent print
examiners from across the fingerprint community. A total of 169
latent print examiners participated; most were volunteers, while
the others were encouraged or required to participate by their
employers. Participants were diverse with respect to organization,
training history, and other factors. The latent print examiners
were generally highly experienced: Median experience was 10 y,
and 83% were certified as latent print examiners. More detailed
descriptions of participants, fingerprint data, and study procedures are included in SI Appendix, Materials and Methods.
The fingerprint data included 356 latents, from 165 distinct
fingers from 21 people, and 484 exemplars. These were combined
to form 744 distinct latent-exemplar image pairs. There were 520
mated and 224 nonmated pairs. The number of fingerprint pairs
used in the study, and the number of examiners assigned to each
pair, were selected as a balance between competing research
priorities: Measuring consensus and variability among examiners

required multiple examiners for each image pair, while incorporating a broad range of fingerprints for measuring image-specific
effects required a large number of images.
We sought diversity in fingerprint data, within a range typical
of casework. Subject matter experts selected the latents and
mated exemplars from a much larger pool of images to include
a broad range of attributes and quality. Latents of low quality
were included in the study to evaluate the consensus among
examiners in making value decisions about difficult latents. The
exemplar data included a larger proportion of poor-quality exemplars than would be representative of exemplars from the FBI’s
Integrated AFIS (IAFIS) (SI Appendix, Table S4). Image pairs
were selected to be challenging: Mated pairs were randomly
selected from the multiple latents and exemplars available for
each finger position; nonmated pairs were based on difficult
comparisons resulting from searches of IAFIS, which includes
exemplars from over 58 million persons with criminal records, or
580 million distinct fingers (SI Appendix, section 1.3). Participants
were surveyed, and a large majority of the respondents agreed
that the data were representative of casework (SI Appendix,
Table S3).
Noblis developed custom software for this study in consultation with latent print examiners, who also assessed the software
and test procedures in a pilot study. The software presented
latent and exemplar images to the participants, allowed a limited
amount of image processing, and recorded their decisions, as
indicated in Fig. 1 (SI Appendix, section 1.2). Each of the examiners was randomly assigned approximately 100 image pairs out
of the total pool of 744 image pairs (SI Appendix, section 1.3).
The image pairs were presented in a preassigned order; examiners could not revisit previous comparisons. They were given
several weeks to complete the test. Examiners were instructed
to use the same diligence that they would use in performing casework. Participants were assured that their results would remain
anonymous; a coding system was used to ensure anonymity during
analysis and in reporting.

Individualization

evaluate the appropriateness of reaching a conclusion is the collective judgments of the experts. Various approaches have been
proposed to define sufficiency in terms of objective minimum
criteria (e.g., ref. 22), and research is ongoing in this area (e.g.,
ref. 23). Our study is based on a black box approach, evaluating
the examiners’ accuracy and consensus in making decisions rather
than attempting to determine or dictate how those decisions are
made (11, 24).

Exclusion
reason

Comparison
Difficulty

Review of
responses
Fig. 1. Software workflow. Each examiner was assigned a distinct, randomized sequence of image pairs. For each pair, the latent was presented first
for a value decision; if it was determined to be no value, the test proceeded
directly to the latent from the next image pair; otherwise, an exemplar was
presented for comparison and evaluation (SI Appendix, section 1.5).

Ulery et al.

VEO VID

80%

Individualization
Inconclusive

60%

Exclusion

40%

No Value
20%

0%

Mated
pairs
(68%)

Non-mated
pairs
(32%)

Fig. 2. Distribution of 17,121 decisions. 23% of all decisions resulted in novalue decisions (no comparison was performed); comparison decisions were
based on latents of VID and of VEO; 7.5% of comparisons of mated pairs
resulted in exclusion decisions (false negatives); 0.1% of comparisons of
nonmated pairs resulted in individualization decisions (false positives—too
few to be visible) (SI Appendix, Table S5).

Results
A summary of examiner decisions is shown in Fig. 2. We emphasize that individual examiner decisions are only a part of an overall operational process, which may include verification, quality
assurance, and reporting. Our results do not necessarily reflect
the performance of this overall operational process.
The true negative rate was greater than the true positive rate.
Much of this difference may be explained by three factors: The
amount of information necessary for an exclusion decision is
typically less than for an individualization decision, examiners
operate within a culture where false positives are seen as more
serious errors than false negatives (5), and the mated pairs included a greater proportion of poor-quality prints than the nonmated pairs (SI Appendix, section 1.3). Whereas poor-quality
latents result in the no-value decisions in Fig. 2, the poor-quality
exemplars contribute to an increase in the proportion of inconclusive decisions.
Rates of comparison decisions can be calculated as a percentage of all presentations (PRES), including latents of no value; of
comparisons where the latent was of value for individualization
(VID); or of all comparisons (CMP), which includes comparisons
Latent

Latent

Ulery et al.

Non-mated image pair B

Non-mated image pair C

Exemplar

Exemplar

Latent

Latent

Mated image pair X

Mated image pair Y

where the latent was of value for exclusion only (VEO) as well as
VID. Because standard operating procedures typically include
only VID comparisons, this is our default basis for reporting these
rates.
False Positives
Six false positives occurred among 4,083 VID comparisons
of nonmated pairs (false positive rate, FPRVID ¼ 0.1%) (SI
Appendix, Tables S5 and S8; confidence intervals are discussed
in SI Appendix, section 2.1). The image pairs that resulted in
two of the false positives are shown in Fig. 3. Two of the false
positive errors involved a single latent, but with exemplars from
different subjects. Four of the five distinct latents on which false
positives occurred (vs. 18% of nonmated latents) were deposited
on a galvanized metal substrate, which was processed with
cyanoacrylate and light gray powder. These images were often
partially or fully tonally reversed (light ridges instead of dark),
on a complex background (Fig. 3, image pair C). It is not known
if other complex backgrounds or processing artifacts would have a
similar increased potential for error.
The six errors were committed by five examiners, three of
whom were certified (including one examiner who made two
errors); one was not certified; one did not respond to our background survey. These correspond to the overall proportions of
certifications among participants (SI Appendix, section 1.4). In
no case did two examiners make the same false positive error:
Five errors occurred on image pairs where a large majority of
examiners correctly excluded; one occurred on a pair where the
majority of examiners made inconclusive decisions. This suggests
that these erroneous individualizations would have been detected
if blind verification were routinely performed. For verification to
be truly blind, examiners must not know that they are verifying
individualizations; this can be ensured by performing verifications
on a mix of conclusion types, not merely individualizations. The
general consensus among examiners did not indicate that these
were difficult comparisons, and only for two of the six false
positives did the examiner making the error indicate that these
were difficult (SI Appendix, Table S8).
There has been discussion (24, 26, 27) regarding the appropriateness of using qualified conclusions in investigation or testimony. The effects of qualified conclusions could be assessed
in this study, as “inconclusive with corresponding features” (SI
Appendix, section 1.5). Qualified conclusions potentially yield
many additional “leads”: 36.5% of VID comparisons resulted
in individualization decisions, and an additional 6.2% resulted
in qualified conclusions. However, 99.8% of individualization
decisions were mated, as opposed to only 80.6% of qualified conclusions (SI Appendix, section 2). Only one of the six image pairs
Exemplar

Exemplar

Fig. 3. Examples of fingerprint pairs used in the study that
resulted in examiner errors. Pairs B and C resulted in false
positive errors: 1 of 30 examiners made an individualization
decision on B (24 exclusions); 1 of 26 examiners made an
individualization decision on C (22 exclusions). The processing of the latent in C (cyanoacrylate with light gray powder) tonally reversed the image so that portions of ridges
were light rather than dark. Pairs X and Y resulted in false
negative errors, with no true positives made by any examiner: X was excluded by 13 of 29 examiners, presumably because the latent was deposited with a twisting motion that
resulted in misleading ridge flow; Y was excluded by 15 of
18 examiners; the exemplar was particularly distorted. For
use in this figure, these images were cropped to reduce
background area.
PNAS Early Edition ∣

3 of 6

APPLIED BIOLOGICAL
SCIENCES

100%

Predictive Value (estimated)

PPV (CMP)
PPV (VID)
NPV (CMP)
NPV (VID)

20%

0%
0%

20%

40%

60%

80%

100%

Mated image pairs as a percentage of test mix
Fig. 4. PPV and NPV as a function of mate prevalence in workload. The observed predictive values (PPVVID;59% ¼ 99.8% and NPVVID;59% ¼ 88.9% for VID
comparisons) correspond to the actual test mix (indicated) where 59% of VID
comparisons were mated pairs; other predictive values are calculated as a
function of mate prevalence. Sixty-two percent of all comparisons (VEO
and VID) were performed on mated pairs, and PPVCMP;62% ¼ 99.8% and
NPVCMP;62% ¼ 86.6%.

Count

Consensus
Each image pair was examined by an average of 23 participants.
Their decisions can be regarded as votes in a decision space
(Fig. 5). Consensus was limited on both mated and nonmated
pairs: VID decisions were unanimous on 48% of mated pairs
and 33% of nonmated pairs. Votes by latent print examiners
also provide a basis for assessing sufficiency for value decisions,
as shown in Fig. 6; consensus on individualization and exclusion
decisions is shown in SI Appendix, Fig. S6.
Lack of consensus among examiners can be attributed to
several factors. For unanimous decisions, the images were clearly
the driving factor: Unusable or pristine prints resulted in unanimous decisions, and therefore different data selection would
have affected the extent of consensus. When there was a lack of
consensus, much of the variation could be explained by examiner
differences: Examiners showed varying tendencies toward noMating

200

224 Non-mates

100

520 Mates

100%

80%

60%

e
iv

40%

20%

0%
100%

Count

300

80%

200

60%

100

40%

60

20%

Individualization (percent of examiners)

40

0%

20

www.pnas.org/cgi/doi/10.1073/pnas.1018707108

40%

us
cl
on

4 of 6 ∣

60%

C

Posterior Probabilities
False positive and false negative rates are important accuracy
measures, but assume a priori knowledge of true mating relationships, which of course are not known in forensic casework. In
practice, knowledge of mating relationships is based solely on
examiners’ decisions: It is important to know the likelihood that
these decisions are correct. Positive predictive value (PPV) is the
percentage of individualization decisions that are true positives;
negative predictive value (NPV) is the percentage of exclusion
decisions that are true negatives. Fig. 4 depicts PPV and NPV
as functions of the prior prevalence of mated pairs among the
examinations performed: As the proportion of mated pairs
increases, PPV increases and NPV decreases (SI Appendix,
section 9). The prior prevalence of mated pair comparisons varies
substantially among organizations, by case type, and by how candidates are selected. Mated comparisons are far more prevalent
in cases where the candidates are suspects determined by nonfingerprint means than in cases where candidates were selected
by an AFIS.

80%

0%
10

False Negatives
False negatives were much more prevalent than false positives
(false negative rate: FNRVID ¼ 7.5%) (SI Appendix, Table S5). Including VEO comparisons had no substantial effect: FNRCMP ¼
7.5%. Eighty-five percent of examiners made at least one false
negative error, despite the fact that 65% of participants said that
they were unaware of ever having made an erroneous exclusion
after training (SI Appendix, section 1.4, no. 25); awareness of
previous errors was not correlated with false negative errors
on this test. False negatives were distributed across half of the
image pairs that were compared. The likelihood of false negatives
varied significantly by examiner (discussed further under Examiner Skill, below), and by image pair (SI Appendix, Figs. S3 and
S5 C and D). Of the image pairs that were most frequently associated with false negatives, most had distorted latents and/or exemplars that gave an appearance of a different ridge flow pattern.
Verification of exclusions (especially blind verification) is not
standard practice in many organizations, in part due to the large
number encountered in casework. To investigate the potential
benefits of blind verification, we posed the following question:
Given a mated image pair, what is the probability, pv , that two
examiners would both reach exclusion decisions? If exclusions
were equally likely for all image pairs (independence assumption), we would estimate that exclusions by two examiners would
occur at the rate pv ¼ FNR2PRES ¼ 5.3% × 5.3% ¼ 0.3% (SI
Appendix, Table S5). However, the data show that the independence assumption is not valid: Some mated pairs are more likely
to be excluded than others. Because the outcomes of blind
verifications are not statistically independent but depend on the
image pairs, we estimate pv ¼ 0.85% (SI Appendix, section 11).
This suggests that blind verification of exclusions could greatly
reduce false negative errors; agency policy would have to balance
this benefit with the impact on limited resources.
For exclusions where the latent was VID, examiner assessment of comparison difficulty was a good predictor of accuracy,
but even “Very Easy/Obvious” exclusions were sometimes incorrect: Among 450 false negatives where the latent was VID,
13 were rated “Very Easy/Obvious” by 11 distinct examiners
(SI Appendix, Fig. S8). Latent value (VEO vs. VID) had no
predictive value for false negative errors; however, exclusions
were more likely to be true negatives when the latent was VID
than when it was VEO. This counterintuitive result is due to the
fact that VEO determinations were more often inconclusive,
hence most exclusion decisions were associated with VID latents
(SI Appendix, Fig. S7).

100%

Exclusion (percent of examiners)

that resulted in false positives had a plurality of inconclusive decisions, and none had a plurality “with corresponding features.”

Count

Fig. 5. Decision rates on each image pair. Percentage of examiners making
an individualization decision (x axis) vs. exclusion decision (y axis) on each
image pair; mean 23 presentations per pair. VEO and no-value decisions
are treated as inconclusive. Marginal distributions are shown as histograms.
Of mated pair decisions, 10% were unanimous true positives, 38% unanimous inconclusives. Of nonmated pair decisions, 25% were unanimous true
negatives, 9% were unanimous inconclusives. Points along diagonal represent pairs on which all examiners reached conclusions. The prevalence of false
negatives is evident in the vertical spread of mated pairs; the few false positives are evident in the limited horizontal spread of the nonmated pairs.

Ulery et al.

60%

60%

40%

80%

20%

100%
0%

20%

40%

60%

80%

0%
100%

356 latent images

Fig. 6. Examiner consensus on VID decisions, showing the percentage of
examiners reaching consensus (y axis) on each latent (x axis). Areas of unanimous (100%), decile (10%, 90%), and quartile (25%, 75%) consensus are
marked. For example, at a 90% level of consensus (y axes), examiners agreed
that 40% of the latents were VID (interval from 60% to 100% indicated by
a horizontal line in upper right) (SI Appendix, Table S11). Such measures of
consensus may be useful in developing quantity and quality metrics.

value or inconclusive decisions, or toward conclusions (SI
Appendix, Fig. S4). Examiners differed significantly in conclusion
rates, and we see this effect as secondary to image characteristics
in explaining lack of consensus. Other factors accounting for lack
of consensus include intraexaminer inconsistency and (presumably) test environment (SI Appendix, Fig. S3).
It was not unusual for one examiner to render an inconclusive
decision while another made an individualization decision on the
same comparison. This result is consistent with previous observations (1, 5, 28). Among all decisions based on mated pairs, 23.0%
resulted in decisions other than individualization even though at
least one other examiner made a true positive on the same image
pair; 4.8% were not individualization decisions even though the
majority of other examiners made true positives. This has operational implications in that some potential individualizations are
not being made, and contradictory decisions are to be expected.
When examiners reached contradictory conclusions (exclusion
and individualization) on a single comparison, the exclusion decision was more frequently in error: 7.7% of independent examinations of conclusions on mates were contradictory, vs. 0.23% on
nonmates. Which of the contradictory decisions is more likely to
be erroneous depends on the prior prevalence of mated vs. nonmated pairs: Exclusion decisions are more likely to be erroneous
except in situations where the prior prevalence of nonmated pairs
is very high.
Examiner Skill
The criminal justice system relies on the skill of latent print
examiners as expert witnesses. Currently, there is no generally
accepted objective measure to assess the skill of latent print
examiners. Skill is multidimensional and is not limited to error
rates (FPR and FNR), but also includes TPR, true negative rate
(TNR), VID and VEO rates, and conclusion rate (CR—the percentage of individualization or exclusion conclusions as opposed
to no-value or inconclusive decisions). Any assessment of skill
must consider these dimensions. Although most discussions of
examiner skill focus on error rates (e.g., ref. 13), the other aspects
of examiner skill are important not just to the examiner’s organization, but to the criminal justice system as well; e.g., an examiner who is frequently inconclusive is ineffective and thereby fails
to serve justice. Both individual examiners and organizations
must strike a proper balance between the societal costs of errors
and inappropriate decisions, and the operational costs of detection. Contradictory verification decisions, whether involving erroneous conclusions or inappropriate inconclusive decisions, should
be internally documented and addressed through an organization’s continual improvement processes.
We found that examiners differed substantially along these
dimensions of skill, and that these dimensions were largely independent. Our study measured all of these dimensions with the
exception of FPRs for individual examiners, which were too low
to measure with precision (SI Appendix, section 3). Fig. 7 shows
that examiners’ conclusion rates (CRPRES ) varied from 15 to 64%
(mean 37%, SD 10%) on mated pairs, and from 7 to 96% (mean
Ulery et al.

100%

A — Examiner decisions on non-mated image pairs (Rank order by CR)
Excl.
Individ.

0%
100%

Inconc.
NV

0%
B — Examiner decisions on mated image pairs (Rank order by CR)

Fig. 7. Decision rates by examiner. Proportions of decisions for all 169
examiners on (A) nonmated and (B) mated image pairs. Examiners in each
chart are sorted on CR. Each examiner was randomly assigned 51 to 74 mated
image pairs (mean 69, SD 5) and 26 to 53 nonmated image pairs (mean 33,
SD 7). In both, errors are shown in red. Column width indicates the number
of image pairs. Examiners who made false positive errors are indicated
with black dots (SI Appendix, Table S7).

71%, SD 14%) on nonmated pairs. The observed range in CRs
may be explained by a higher level of skill (ability to reach more
conclusions at the same level of accuracy), or it may imply a higher risk tolerance (more conclusions reached at the expense of
making more errors).
Fig. 7 shows substantial variability in CR among examiners.
These measured rates were based on an average of 69 mated
presentations and 33 nonmated presentations. The limited number of presentations resulted in a wide margin of measurement
error when evaluating the performance of an individual examiner
(SI Appendix, Fig. S5). Although the estimates for each examiner
are statistically unbiased, the sampling error in these estimates
contributed substantially to the observed variability among examiners. The observed variability is a biased estimate that overstates
the true variability (SI Appendix, Figs. S3B and S4).
Fig. 8 shows the relations between three of the skill dimensions
measured for each examiner. Blue squares near the lower right
of the chart represent highly skilled examiners: accurate (making
few or no errors) and effective (high TNR and TPR, and therefore high CR). The red cross at the bottom left denotes an accurate (0% FNRVID ), but ineffective (5% TNRVID , 16% TPRPRES )
examiner. The examiner denoted by the red cross at the top right
is inaccurate (34% FNRVID ), and has mixed effectiveness (100%
TNRVID , 23% TPRPRES ). Attempting to compare the skill of
any two examiners is a multidimensional problem. A combination
of multiple dimensions into a single hypothetical measure of skill
would require a weighting function to trade off the relative value
of each dimension; such weighting might be driven by policy,
based on the relative cost/benefit of each dimension for operational needs.
Tests could be designed to measure examiner skill along the
multiple dimensions discussed here. Such tests could be valuable
TPR PRES
11-24%
25-37%
38-57%

30%

20%

10%

0%
0%

20%

40%

60%

80%

100%

TNR VID
Fig. 8. Examiner skill. Each of the 169 examiners is plotted on three skill dimensions: TNRVID (mean 88%, SD 13.6%), FNRVID (mean 7.5%, SD 7.3%), and
TPRPRES (shown in color, with red crosses denoting the lowest quartile and
blue squares the highest quartile; mean 32%, SD 9.4%). The five examiners
who made false positive errors are indicated with bold filled circles.
PNAS Early Edition ∣

5 of 6

APPLIED BIOLOGICAL
SCIENCES

40%

Proportion of decisions

80%

FNR VID

100%

Consensus:
VID

Consensus:
not VID

0%
20%

not just as traditional proficiency tests with pass/fail thresholds,
but as a means for examiners or their organizations to understand
skills for specific training, or for tasking based on skills (such as
selecting examiners for verification based on complementary
skill sets).
Certified examiners had higher conclusion rates than noncertified examiners without a significant change in accuracy
(significantly higher TPRVID and TNRVID ; FNRVID did not vary
significantly) (SI Appendix, section 6). Length of experience as
a latent print examiner did not show a significant correlation
with TPRVID , TNRVID , or FNRVID (SI Appendix, Table S9 and
Fig. S2).
Examiners with a lower TPRVID tended also to have a lower
TNRVID . Examiners with a higher FNRVID tended to have a
lower TPRVID . Examiners with a higher TNRVID tended also to
have a higher FNRVID (SI Appendix, Table S9 and Fig. S2).

the financial and societal costs and benefits of additional verifications.
Most of the false positive errors involved latents on the most
complex combination of processing and substrate included in the
study. The likelihood of false negatives also varied by image.
Further research is necessary to identify the attributes of prints
associated with false positive or false negative errors, such as
quality, quantity of features, distortion, background, substrate,
and processing method.
Examiners reached varied levels of consensus on value and
comparison decisions. Although there is currently no objective
basis for determining the sufficiency of information necessary
to reach a fingerprint examination decision, further analysis of
the data from this study will assist in defining quality and quantity
metrics for sufficiency. This lack of consensus for comparison
decisions has a potential impact on verification: Two examiners
will sometimes reach different conclusions on a comparison.
Examiner skill is multidimensional and is not limited to error
rates. Examiner skill varied substantially. We measured various
dimensions of skill and found them to be largely independent.
This study is part of a larger ongoing research effort. To further
our understanding of the accuracy and reliability of latent print
examiner decisions, we are developing fingerprint quality and
quantity metrics and analyzing their relationship to value and
comparison decisions; extending our analyses to include detailed
examiner markup of feature correspondence; collecting fingerprints specifically to explore how complexity of background,
substrate and processing are related to comparison decisions; and
measuring intraexaminer repeatability over time.
This study addresses in part NRC Recommendation 3 (12),
developing and quantifying measures of accuracy and reliability
for forensic analyses, and will assist in supporting the scientific
basis of forensic fingerprint examination. The results of this study
will provide insight into developing operational procedures and
training of latent print examiners and will aid in the experimental
design of future proficiency tests of latent print examiners.

Conclusions
Assessing the accuracy and reliability of latent print examiners is
of great concern to the legal and forensic science communities.
We evaluated the accuracy of decisions made by latent print examiners on difficult fingerprint comparisons in a computer-based
test corresponding to one stage in AFIS casework. The rates measured in this study provide useful reference estimates that can
inform decision making and guide future research; the results
are not representative of all situations, and do not account for
operational context and safeguards. False positive errors (erroneous individualizations) were made at the rate of 0.1% and
never by two examiners on the same comparison. Five of the six
errors occurred on image pairs where a large majority of examiners made true negatives. These results indicate that blind
verification should be highly effective at detecting this type of
error. Five of the 169 examiners (3%) committed false positive
errors, out of an average of 33 nonmated pairs per examiner.
False negative errors (erroneous exclusions) were much more
frequent (7.5% of mated comparisons). The majority of examiners (85%) committed at least one false negative error, with
individual examiner error rates varying substantially, out of an
average of 69 mated pairs per examiner. Blind verification would
have detected the majority of the false negative errors; however,
verification of exclusion decisions is not generally practiced in
operational procedures, and blind verification is even less frequent. Policymakers will need to consider tradeoffs between

ACKNOWLEDGMENTS. We thank the latent print examiners who participated
in this study, as well as William Fellner, Jill McCracken, Keith Ward, Stephen
Meagher, Calvin Yeung, Ted Unnikumaran, Erik Stanford, and William
Chapman. This is publication number 10-19 of the FBI Laboratory Division.
This work was funded in part under a contract award to Noblis, Inc. from
the FBI Biometric Center of Excellence and in part by the FBI Laboratory
Division. The views expressed are those of the authors and do not necessarily
reflect the official policy or position of the FBI or the US government.

1. Evett IW, Williams RL (1995) A review of the 16 point fingerprint standard in England
and Wales. Fingerprint Whorld 21.
2. Wertheim K, Langenburg G, Moenssens A (2006) A report of latent print examiner
accuracy during comparison training exercises. J Forensic Identification 56:55–93.
3. Gutowski S (2006) Error rates in fingerprint examination: The view in 2006. Forensic
Bulletin 2006:18–19 Autumn.
4. Langenburg G, Champod P, Wertheim P (2009) Testing for potential contextual
bias effects during the verification stage of the ACE-V methodology when conducting
fingerprint comparisons. J Forensic Sci 54:571–582.
5. Langenburg G (2009) A performance study of the ACE-V process. J Forensic Identification 59:219–257.
6. Cole SA (2005) More than zero: Accounting for error in latent fingerprint identification. J Crim Law Criminol 95:985–1078.
7. United States v Mitchell No. 96-407 (ED PA 1999).
8. United States v Llera Plaza Cr. No. 98-362-10, 11, 12 (ED PA 2002).
9. Maryland v Rose No. K06-0545 (MD Cir 2007).
10. Office of the Inspector General (2006) A Review of the FBI’s Handling of the Brandon
Mayfield Case (US Department of Justice, Washington, DC).
11. Budowle B, Buscaglia J, Perlman RS (2006) Review of the scientific basis for friction
ridge comparisons as a means of identification: Committee findings and recommendations. Forensic Sci Commun 8:1.
12. National Research Council (2009) Strengthening Forensic Science in the United States:
A Path Forward (National Academies Press, Washington, DC).
13. Koehler JJ (2008) Fingerprint error rates and proficiency tests: What they are and
why they matter. Hastings Law J 59:1077–1110.
14. Mnookin JL (2008) The validity of latent fingerprint identification: Confessions of a
fingerprinting moderate. Law Probability and Risk 7:127–141.
15. Haber L, Haber RN (2008) Scientific validation of fingerprint evidence under Daubert.
Law Probability and Risk 7:87–109.

16. Cole S (2006) Is fingerprint identification valid? Rhetorics of reliability in fingerprint
proponents’ discourse. Law Policy 28:109–135.
17. Scientific Working Group on Friction Ridge Analysis, Study and Technology (2011)
Standard terminology of friction ridge examination, Version 3., Available at http://
www.swgfast.org/documents/terminology/110323_Standard-Terminology_3.0.pdf.
18. Dror I, Mnookin J (2010) The use of technology in human expert domains: Challenges
and risks arising from the use of automated fingerprint identification systems in
forensic science. Law Probability and Risk 9:47–67.
19. Huber RA (1959) Expert witness. Criminal Law Quarterly 2:276–296.
20. Ashbaugh D (1999) Quantitative-Qualitative Friction Ridge Analysis: An Introduction
to Basic and Advanced Ridgeology (CRC Press, New York).
21. Scientific Working Group on Friction Ridge Analysis, Study and Technology (2002)
Friction ridge examination methodology for latent print examiners, Version 1.01.,
Available at http://www.swgfast.org/documents/methodology/100506-MethodologyReformatted-1.01.pdf.
22. Champod C (1995) Edmond Locard—Numerical standards and “probable” identifications. J Forensic Identification 45:136–155.
23. Neumann C, et al. (2007) Computation of likelihood ratios in fingerprint identification
for configurations of any number of minutiae. J Forensic Sci 52:54–64.
24. Mnookin JL (2008) Of black boxes, instruments, and experts: Testing the validity of
forensic science. Episteme 5:343–358.
25. Budowle B, et al. (2009) A perspective on errors, bias, and interpretation in the forensic
sciences and direction for continuing advancement. J Forensic Sci 54:798–809.
26. Saks M, Koehler J (2005) The coming paradigm shift in forensic identification science.
Science 309:892–895.
27. Stoney DA (1991) What made us ever think we could individualize using statistics?
J Forensic Sci Soc 31:197–199.
28. Grieve DL (1996) Possession of truth. J Forensic Identification 46:521–528.

6 of 6 ∣

www.pnas.org/cgi/doi/10.1073/pnas.1018707108

Ulery et al.