In this NYT video it is said that a 2002 paper found that the false positive rate is 11% when also tested via mtDNA. The paper referenced is probably "Correlation of microscopic and mitochondrial DNA hair comparisons" by Houck and Budowle.
Human hairs submitted to the FBI Laboratory for analysis between 1996 and 2000 were reviewed. Of 170 hair examinations, there were 80 microscopic associations; of these, only nine were excluded by mtDNA.
Looking at the Wikipedia page on the matter, that alas seems to be the only
quantitative study published... in decades.
The FBI (2009) claims the reliability of its method was based on studies from the 1940s... which reported 100% accuracy:
Paul Kirk conducted some of the first studies on the potential forensic application of microscopic comparison of hair in the United States (Gamble and Kirk 1941; Greenwell et al. 1941; Kirk 1940). In addition to his publications on the microscopic characteristics of human hairs, he conducted hair-comparison studies using his criminology students. All of the students were required to compare a single hair to 20 known samples, where all of the known samples were of a similar color and from individuals of a similar age. He reported that no student who completed the routine examination failed to report the association correctly (Kirk 1940).
Much later (after saying they send expert-matched hair to mtDNA analysis nowadays) they admit that
Some critics emphasize the fact that microscopic-hair examiners are unable to statistically quantify the significance of an association (see, for example, Robertson 1999). The development of a statistical model would involve frequency data across the entire population for all microscopic characteristics present in hair. Although this is an attractive idea, the difficulties associated with generating such a database have been, to date, practically insurmountable. In order to generate frequency data for hair characteristics, microscopic-hair examiners might be required to use a “checklist” or “archetype” approach rather than the pattern-recognition process normally used. [...]
Given that useful statistical data are not generated regarding the relative frequency of an evidentiary hair, one must accept that the answer to the question, what proportion of the population would have characteristics that are the same as the evidentiary hair? is we do not know. Similarly, the answer to the question, what is the probability of a coincidental match between the questioned hair and the known sample? is we do not know.
There were actually some kind of replications of those 100% experiments... using similarly well chosen samples, e.g.:
Strauss (1983) conducted a study using 100 individuals comprising 54 Caucasian, 19 Negroid, and 27 Mongoloid. From each of the 100 individuals, 7 hairs were chosen to represent the widest variation possible. These were mounted on glass microscope slides and were designated as the known samples. One hair was also chosen from each of the 100 samples, mounted on glass microscope slides, and designated as questioned hair samples. All 800 hairs (700 known hairs and 100 questioned hairs) were individually characterized using a checklist and punch cards.
A series of seven experiments was conducted. A neutral party selected a total of 10 single questioned hairs to be compared with 10 known samples. Comparison microscopy resulted in 100 percent accuracy in associating the correct questioned hair with its known source, showing that they could reliably associate a questioned hair with a known sample. In addition, the study showed that the examiners correctly identified each of the 100 individuals in the questioned hair pool to the correct known hair group, that is, 54 Caucasian, 19 Negroid, and 27 Mongoloid.
IMHO, unfortunately these [only] prove that if one person (who sets up the samples/experiments) can distinguish them, so can other people.
Perhaps more realistic, when using a sample not in the dataset:
Bisbing and Wolner (1984) conducted a series of studies using known head-hair samples recovered from 17 sets of twins and 1 set of identical triplets. [...]
This study involved removing 2 or 3 hairs from 7 randomly selected unmounted samples, which were then mounted on glass microscope slides. For each of these 7 “questioned” samples, between 5 and 10 known samples were randomly selected from the 74 mounted known samples for microscopic comparison. There were 52 comparisons made by each of the two examiners, for a total of 104 comparisons. Because of the random sampling, none of the true known samples for the questioned hairs was present in any of the comparison scenarios. The two examiners correctly excluded 96 of the known samples as being possible donors of the questioned hairs. Eight of the questioned hair samples were incorrectly associated to the known samples (5 by one examiner and 3 by the second examiner). In one of these cases, a sample of the fraternal twin’s hair was present in the known pool and was correctly eliminated. In the other simulated cases, the questioned hairs were incorrectly associated with control samples that were neither the true source nor the twin of the true source.
It is interesting to note that 7 of the 8 incorrectly associated hairs were classified by the authors as being blond, common, featureless hairs. These results serve to reinforce that human hair cannot be associated with one person to the exclusion of all others. In addition, this study served to show that caution is necessary when comparing common, featureless hairs.
For pubic hair even less data appears to exist. I could not easily access this study myself but according to one summary:
In Gaudette and Keeping’s study, 13 of 100 individuals had a strand of scalp hair that was indistinguishable from that of another individual. For pubic hair, the number was 25 of 60 individuals.
- Gaudette, B. D. and Keeping, E. S. (1974) Attempt at determining
probabilities in human scalp hair comparison. Journal of Forensic
Sciences, 19(3), 599–606.