Alon Honig points to this news article by Peter Smith and writes:
Beyond the fact that this seems crazy (why would anyone convict based off the actions of a dog) I am wondering how you would think about this from a bayesian perspective. P (murderer|dog visited) = P (dog visited|murderer) * P(murderer) / P(dog visited)
What parameters do these jurors have that would make sense?
I have no idea! But I’m posting this, first because we do so many cat pictures here that it’s only fair for a dog to be the star of the story for once, and second because it’s refreshing to read an article mentioning the University of California law school that doesn’t bring up that notorious torture apologist and political hack. It’s good to know they have other faculty there.
(Skimmed the article.) No false positive rate was mentioned. Right?
I don’t think the content in the actual article is as bad as the headline makes it seem
Still bad, but it’s not building a case on its own, just moving the line from a preponderance of evidence to reasonable doubt.
Ignoring ethical issues, 75% false positive rate is actually kind of amazing when you consider the base rate. Which makes me think–it would also be relevant to jurors how often these dogs are asked to find residual human remains, and testify in the negative. If they always find human remains when asked, there’s probably no signal there. Not in the article though.
This is by far the most concerning to me. Dogs are people pleasers. In this case, their handlers and family are the prosecution, who are explicitly trying their hardest to prove guilt. There’s no way to incentivize truthfulness in a dog–what are you going to do, charge them with perjury?
> is bridging the gap between pretty plausible suspect and guilty beyond a reasonable doubt,” Delger says, emphasizing that she could not comment on any particular case
Isn’t there some calculus example like this? I remember it being phrased in terms of cars driving places. The car averaged above the speed limit, but we will not comment on if it actually broke the speed limit at any point (and the theorem says it has to).
I don’t know about murder detection, but dogs seem to have a good record for melanoma detection. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3822260/
I remember reading an article about thirty years ago that showed a higher rate of true melanoma findings by dogs than by pathologists. I don’t remember all the details, but I often told my pathology friends that they would soon be replaced by a creature that growled a lot less than they did.
Yes, it has been claimed that “Researchers have discovered that your four-legged friend cannot only smell cancer, but also be more accurate than the most advanced laboratories when trying to detect certain cancers.”
https://www.homeoanimal.com/blogs/blog-pet-health/dogs-can-detect-cancer
Sounds doubtful. The Wikipedia article has some skepticism.
https://en.wikipedia.org/wiki/Canine_cancer_detection
https://www.npr.org/sections/health-shots/2020/03/23/820274501/her-incredible-sense-of-smell-is-helping-scientists-find-new-ways-to-diagnose-di
As far as I can tell, this is real
I don’t know the fisher’s exact test likelihood off the top of my head, but that’s a small p-value if I ever saw one,
Paper, for those who like that sort of thing
https://pubs.acs.org/doi/10.1021/acscentsci.8b00879#
It’s not clear what relevance cancer detection has to the linked article. In cancer detection, the source of the odor is present and identifiable. To compare to the relevant question in the article: if a melanoma patient sat on a chair, could a dog detect the melanoma a month or three months later?
Which brings up a question: the article seems to suggest that the dog detects the scent of “cadaver”, not the scent of the individual victim. Is that correct?
It’s not surprising, since dogs, like bears and other large mammals, eat carrion. But can the dog distinguish between different types of rotting flesh? Can it tell a person’s corpse from a deer’s corpse? If it can, has it been trained to “indicate” on that distinction? And, in the case in the article, how do we know it’s the same corpse as the victim? I can’t find anything in the article indicating it detects the *victim* as opposed to just any remains.
The article has a graphic showing the locations of physical evidence and dog indications. I guess this was a test to show that the dog accurately indicates where there is physical evidence? If so it’s a failure. Looking at the distribution of physical evidence, it seems to have been carried down and along a single drainage. Yet the dog also detected several drainages away, where no physical evidence was found. The dog and handler need to be blinded to the known physical evidence.
As Andrew often says, it may be true, but there doesn’t seem to be strong evidence for it.
The article mentions a prof at UC Hastings, but John Yoo is at Berkeley Law, not at Hastings
Days since Andrew took a shot at the Hoover Institution (where Yoo is a visiting fellow), either directly or by proxy: 0
The:
I had no idea that Yoo had a Hoover affiliation so this doesn’t count! I thought Hoover was for Stanford faculty, not Berkeley.
It’s a fundamental right to be able to confront your accuser. I believe this sort of “but the animal totally said so, because I totally know what the dog said” nonsense is dangerous to the project of ordered liberty.
Until I my lawyer can cross examine the dog, the most would should ever allow (and even then I’m skeptical) is the establishment of probable cause — which is already highly abused by police, since “yeah the dog totally told me he smelled illegals and drugs” is subjective (and lots of our police don’t care about “civil rights”).
You definitely want rigorous controlled studies to determine the credibility of the handler’s methods and interpretation of the dog’s actions. But demanding to cross-examine the dog would be like demanding to cross-examine the flies because a pathologist used the life cycle of flies on the body to infer the time of death.
We’re talking about cases where there literally isn’t a body. That’s egregious.
There is a pattern in law enforcement using junk science to try to get convictions polygraphs — fake. “Bite mark analysis” — fake. Ballistics pattern-matching (ala, it was fired from THIS gun) — oversold. Partial fingerprint analysis — close to fake, definitely oversold.
That’s gonna be a hard pass from me, dog.
Replace the dog with an electronic instrument that detects scent: I can already hear the defense demanding the instrument’s test and calibration records. The dog should have similar records.