# Dog Cloning For Special Forces: Breed All You Can Breed

Decision analysis of whether cloning the most elite Special Forces dogs is a profitable improvement over standard selection procedures. Unless training is extremely cheap or heritability is extremely low, dog cloning is hypothetically profitable.
topics: genetics, decision theory, R, bibliography
created: 18 Sep 2018; modified: 6 June 2019; status: finished; confidence: possible;

Cloning is widely used in animal & plant breeding despite steep costs due to its advantages; more unusual recent applications include creating entire polo horse teams and reported trials of cloning in elite police/Special Forces war dogs. Given the cost of dog cloning, however, can this ever make more sense than standard screening methods for selecting from working dog breeds, or would the increase in successful dog training be too low under all reasonable models to turn a profit?

I model the question as one of expected cost per dog with the trait of successfully passing training, success in training being a dichotomous liability threshold with a polygenic genetic architecture; given the extreme level of selection possible in selecting the best among already-elite Special Forces dogs and a range of heritabilities, this predicts clones’ success probabilities. To approximate the relevant parameters, I look at some reported training costs and success rates for regular dog candidates, broad dog heritabilities, and the few current dog cloning case studies reported in the media.

Since none of the relevant parameters are known with confidence, I run the cost-benefit equation for many hypothetical scenarios, and find that in a large fraction of them covering most plausible values, dog cloning would improve training yields enough to be profitable (in addition to its other advantages).

As further illustration of the use-case of screening for an extreme outcome based on a partial predictor, I consider the question of whether height PGSes could be used to screen the US population for people of NBA height, which turns out to be reasonably doable with current & future PGSes.

Military and police dogs are specially trained for their jobs. Only some dogs are up to the task, but like training seeing-eye guide dogs, it’s difficult to know in advance and many dogs will wash out of training as expensive failures, with even fewer being able to handle the extreme life of a Special Forces dog; then they may get injured on the job, develop hip dysplasia or cancer, cutting short their career, and leading to perennial shortages. This is despite the best efforts of the (mostly European) breeders who raise the Malinois, German Shepherd, Belgian Shepherd, and Labradors preferred for war dogs.

In 2014, Bloomberg reported on an interesting aspect of Sooam Biotech, the famous South Korean dog cloning company: they were cloning a Special Forces dog. If it’s hard to be a K9, it’s even harder to be a SF dog, able to jump out of airplanes (they have special parachute harnesses), go on raids, carry cameras with them, even (reportedly) wear little doggie hoods with infrared camera goggles for night work; so valuable and specialized are such dogs that special $20,000 animatronic dog models like the “K9 Hero-Trauma” are sold to train medics how to treat injuries like gunshot wounds or amputations. If you have a successful SF dog… maybe the clone will be much more likely to succeed than a random puppy picked from one of the usual breeders, and you can make as many clones as necessary long after the original has gone to Dog Heaven. Clones of elite individuals are increasingly common in agriculture; plants, like the myriads of apple varieties, have always been propagated clonally, but cloning of cattle has made major commercial inroads1—not just cloning of cattle for beef or cows for milk, but also clones of rodeo bulls (the logical extension of the highly successful selective breeding for rodeo bulls). A striking example of this approach is the world polo champion Adolfo Cambiaso, who is so enthusiastic about the benefits of horse cloning that he has cloned his prized polo horse not once but >10 times, and has rode entire teams of clones to repeated victory. On the other hand, dog clones are still extremely expensive (~$100,000) and prices have not yet come down to the ~$10,000–$20,000 of cattle.

There may be cheaper alternatives to improving SF dog yield: training is probably well-refined and can’t be watered down without risking lives, but that leaves a place for improvement of what is trained, the selection into training—better prediction of SF potential means fewer dogs washing out means less total money spent to produce a successful SF dog. The predictions don’t work well, but the descriptions of screening suggest there’s a lot of room for improvement: the research literature supports the generalization that dog and cat behavioral measurements are not all that predictive. They may be badly designed or testing the wrong things, or there may be inherent noise which can be fixed by doing multiple measurements. (Even something as apparently mechanical as offering catnip to a cat can have different results from occasion to occasion and may have rater-specific effects, perhaps because—“set and setting”—the cat is fearful and distrusts the person offering the catnip that day, with the anxiety shutting down any response or play.) Many described measurements in the literature measure a dog once, on one day, by one person, for example, measuring aggressiveness by taking away food and seeing if the dog snaps at the person, and that’s the whole test. Such a test will be hindered by day-to-day variation (perhaps he is stressed that day), different levels of liking for that particular food, disliking of the person taking the food, sheer randomness in the particular split-second decision of whether the dog decides to express their aggression—likely would be much stabler and predictive if they were done multiple times in multiple ways by multiple people etc (although such extended testing would increase the cost of testing). Of course, that would take more time and would cost a lot more, and it’s unclear the increase in predictions is worth it.

However, ranking for selection is easier than prediction of all datapoints: only the ordering matters, and only the ordering in a particular region (near the threshold) matters. When considered in a real-world context, such predictive improvements do not need to be all that large (a point long made by psychometricians & industrial psychologists eg Taylor & Russell 1939/Lubinski & Humphreys 1996); counterintuitively, a score or test which correlates, say, r=0.10 with an outcome, which in many areas of science would be dismissed as a trivial correlation of no interest, can be quite useful in screening & should not be dismissed as ‘small’—and the rarer the outcome, the larger the benefit.2 In the case of dog cloning, our ‘score’ is the extent to which a donor’s performance predicts the performance of its clones, through their shared genes.

Both approaches could wind up being expensive and there’s no a priori answer about which one would be more cost-effective. To a certain extent, they are also mutually exclusive approaches: dog cloning is so expensive that unless it results in high probability of success, it probably won’t be cost-effective at all, and if the probability is sufficiently high, then testing is no longer useful (because you would save money by simply trying to train all clones), so better testing is unlikely to then pay for itself. Testing to gain information is only profitable in a certain intermediate region of probabilities & costs/benefits.

So it’s not absurd to think that dog cloning could work out well for training SF dogs, and I took a closer look.

# Benefits

The benefits of dog cloning are not limited purely to replicating an elite SF dog. The potential benefits of dog clones include:

1. lower total cost: the primary reason for cloning is that since dog clones are more likely to succeed in training given any reasonable heritability, they may reduce washout costs enough to compensate for the expense of cloning.

But the total lifetime cost of a dog goes beyond success or failure in becoming a useful dog. Successful dogs can still learn at different rates, and require more or less intensive intervention by trainers. Dogs can learn multiple roles, so a ‘success’ may only be a partial success, like a dog who is approved for odor detection of bombs or drugs, but can’t be used on patrol or raids. They can have longer or shorter careers, reflecting levels of competence and medical issues (hip dysplasia constantly comes up in war dog discussions as a disabling medical problem, and is highly heritable).

Hence, discussing only success/failure in training and the reduction in average training cost will seriously underestimate the benefits of cloning the best: clones of the best SF dogs will train faster, with less effort/time, excel at more roles (more likely to be acceptable for at least one role), be less likely to have crippling medical issues that kill them or end their careers prematurely, and have longer careers in general.

2. greater scalability in dogs: there are only a few dog breeders, and they have only a relative handful of bitches at any time

Even if demand spiked in a war and 1,000 more dogs were needed yesterday, they wouldn’t exist—dogs take a certain amount of time to reach sexual maturity, have only so big litters, mating in inbred/narrow pedigrees like German Shepherds/Malinois must be managed carefully to avoid exacerbating existing genetic issues (and eating the seed corn), training takes a while (Ritland notes that the US Navy takes delivery of 2-year-old dog candidates), and so on. In reading about US war dogs, a perennial theme noted by Hammerstrom 2005 is that a war happens (WWII, Vietnam, War on Terror), war dogs become incredibly useful to frontline troops, and dog supply simply cannot keep up.

Use of cloning can break part of the bottleneck by enabling surrogacy in female dogs of other breeds which are not scarce, and by enabling unlimited reproduction of a particular dog. (This doesn’t require cloning, since one could create the necessary embryos with standard IVF, but since the IVF/surrogacy is necessary, why not use cloning as well?)

This option is highly valuable and justifies dog cloning on its own; and because this option is available, militaries can more steeply reduce war dog numbers during peacetime as no ‘reserve’ is necessary.3

3. greater scalability in facilities: another bottleneck might be not the number of dogs, but the infrastructure for housing/training/testing the dogs.

There might be only so many dog kennels and experienced dog trainers at any point, and increasing the number could take a while. (You probably want the trainers and program management to have SF dog handler experience themselves, but it might take decades for a recruit to become an experienced trainer.) So given the inelastic throughput, here it would be valuable to improve the quality of inputs, which will increase the total yield, simply because it means less dilution or waste of scarce fixed housing/training/testing slots on dogs less likely to succeed.

4. greater predictability:

• response to training: yield might be increased simply by the inherent homogeneity of clones allowing improved training by greater experience, rather than any increased genetic merit.

One of the reasons Adolfo Cambiaso gives for investing so heavily in clones of a single polo horse is that he has learned from his long experience with the donor horse how best to train them: each new clone can be given personalized training which he knows it’ll respond best to, because he’s trained many clones before them. If there is some consistent weakness the clones are prone to, he can start addressing it before it even shows up. He also has gained long experience with their injury propensities, preferences, and other behavior, instead of starting from scratch with each new colt. Their similarity avoids the need for learning or wasted pedagogy.

Dogs presumably vary as much as horses do, and training of clones could benefit from this sort of homogeneity. (Since dog trainers will have never encountered clones before, and identical twin dogs are vanishingly rare, there’s no way to know how useful this would be in practice until large numbers of dog clones have been trained by individual trainers.)

• reduced variance for experimentation or analysis: statistical power scales poorly with increasing variance; relatively small increases in noise can require much larger n to overcome. The most efficient experiments are within-subject experiments, which avoid comparisons between individuals, but these are often impossible—one could not test improvements in puppy rearing, for example, or most training program changes. This is true of many things in humans as well; for this reason, experiments with identical twins are highly efficient (in the Lanarkshire milk experiment, a sample of n>10,000 children could have been replaced with n~300). Identical twins are remarkably powerful even in the absence of randomization for inferring causation (Turkheimer & Harden 2014) and by controlling for all genetics (which in human research, debunks a large fraction of all correlational research in psychology/sociology), make correlational analyses much more likely to deliver useful causal insights. As dog identical twins hardly exist, this has hitherto been entirely unavailable a research design for dog researchers, but clones change that.

• reduced variance for process control: given the choice between a small group of clones and a much larger group of regular dogs, such that they have ostensibly identical average costs & the same number of expected successes, which should a breeder or trainer prefer? The small group of clones, of course.

The large group will, by the law of small numbers, have larger absolute fluctuations due to randomness, especially with a base rate like 1%. It’ll be ‘feast or famine’. Sometimes there will be considerably more, sometimes considerably less in absolute numbers. This will complicate planning greatly, stress facilities/trainers, risk delivering too few (or too many) dogs each year, and so on. Switching to clones with a higher base rate will make the overall process more controllable and predictable, and this is worth something.

5. use in selective breeding:

The major use of cloning in cattle is for accelerating breeding programs, and not for their immediate marginal increase in meat or milk yield. While dog breeding is not nearly as sophisticated, the benefits of cloning may also be larger for the long-term improvement in the breed than for its immediate benefits in each cloned dog:

• clones can improve estimates of genetic merit by providing the most accurate possible heritability estimates (genetically identical individuals reared in different environments), and correcting individual estimates of traits, which is vital for planning any kind of breeding or selection program
• a clone can have a higher genetic potential than the average SF dog if intensive selection is done among SF dogs: if the best SF dog is selected for cloning, it’ll have a higher genetic potential than the default calculation of a truncated normal+regression toward the mean on a random SF dog would imply.
• elite clones can be useful in breeding programs in allowing particular individuals to keep contributing genetically long after the original has become infertile or died, or contribute far more (as mentioned before, female dogs are highly limited in reproductive fecundity compared to males, but they could be cloned & born via surrogacy). For example, the first cloned dog, Snuppy, died in 2015, but is survived by 3 clones of himself.2, and the record for number of clones appears to be the 49 clones of the world’s tiniest dog, Miracle Milly.4
6. Value of Information: dog cloning may or may not be worthwhile, but if it is, the total returns from cloning hundreds of dogs per year indefinitely (plus the additional benefits) could be large. It would be valuable to know if it would work.

Since I have not found any SF/military-specific heritabilities reported in the scientific literature (and the SF dog programs generally seem genetically unsophisticated so there may not be any private or classified ones either), the only way to know is to try it out experimentally. The clones’ realized performance would also provide additional valuable information as it would estimate heritability, which would be useful for the regular kinds of breeding & selection as well—because they give an idea of how much one can predict a dog’s performance based on known relatives, and how fast a breeding program can/should proceed.

And since, to be profitable, the success rate of clones need to be >=9% (which is highly likely, see later), this is reasonably easy to estimate: a sample of ~50 clones would give a reasonably precise estimate as to the success rate and enable better decision-making as to whether to keep pursuing cloning (in which case more information will come in and firm up the decision) or drop it as a dead end due to too high costs and/or low heritability. (By the same logic, one could treat choice of donor itself as a multi-armed bandit problem to optimize the selection, since with success rates likely >50%, the necessary sample sizes will be not unreasonable and will be reached as clone use ramps up—like in South Korea, which has at least 42 clone dogs deployed in 2019 and appear to be increasing clone use as they claim great increases in success rates, decreases in costs, and net savings, implying substantial heritability.)

# Modeling the SF selection problem

## South Korea

How could we estimate the benefit of cloning? Given an active dog cloning program like Sooam and sufficient experience, it can be estimated directly.

Choi et al 2014 reports that normally-bred detector dogs have a training success rate of 30% vs 86% for cloned dogs; the 30% appears to be based on the drug detection program, and the 86% is based on their sample of 7 of which 6 passed (ie )5 Oh et al 2018 notes as a followup

…the Toppy [clones] had the exact same genetic information as the elite drug sniffing dog, whereas the control dogs were the offspring of sniffer dogs. Surprisingly, all seven Toppy were selected with high scores, in contrast with the control group, of which three of the seven trained dogs were selected (Choi et al. 2014). In the 6 months after the seven Toppy clones were added to airport security, the drug detection rate increased sixfold, at the same time saving the budget for selecting elite dogs. Thus, outstanding abilities can be passed on to the next generation by cloning identical dogs that inherit identical genetic material.

A 2017 Korea Bizwire provides a partial cost-benefit analysis in a press release:

Cloning and deployment of special forces dogs began in 2012 as part of an initiative by the Rural Development Administration (RDA), in an effort to slash spending on police dog training. Special forces dogs come at a high price. For every dog, an estimated 1.3 billion won ($112,554) is spent on training them for multiple purposes such as human rescue, explosive detection and custom service. Despite the price, only 3 out of 10 dogs [30%] make it through the exhaustive training process to serve on police forces. Clone dogs, on the other hand, have a much higher pass rate of 80%, bringing down the training costs to 46 million won ($39,775). Compared to regular dogs, they offer savings of 65%.

“Sharing a competent and well-trained dog is no longer impossible, thanks to cloning”, said Im Gi-soon, a chief animal biotechnologist at the National Institute of Animal Science (NIAS).

It’s unclear if this 80% estimate is merely re-reporting Choi et al 2014, but if training each costs = ~$33,000, so going from 30% to 80% success rate means the clones have a training cost which is = 35% that of the regular dogs or$72,000. The costs here clearly exclude cloning, but as Viagen is able to offer consumer dog cloning at $50,000 and Sooam has the advantage of experience & much greater scale (in addition to any patriotic discounts), the SK police could be getting a substantially lower price. But if they pay the full$50,000 anyway, then they are still reducing the total cost to , saving $8,804. And at the$15k which may be the Viagen marginal cost, they would save $52,554. However, one might doubt these numbers or how applicable they are, and they appear to exclude the substantial cost of cloning, rendering the cost-benefit incomplete. A 2019 newspaper article states: According to the Animal and Plant Quarantine Agency, 42 of its 51 [82%] sniffer dogs were cloned from parent animals as of April, indicating such cloned detection dogs are already making significant contributions to the country’s quarantine activities. The number of cloned dogs first outpaced their naturally born counterparts in 2014, the agency said. Of the active cloned dogs, 39 are currently deployed at Incheon International Airport, the country’s main gateway…While the average cost of raising one detection dog is over 100 million won (US$85,600), it is less than half that when utilising cloned puppies, they said.

The lower price here may refer to lower levels of selectivity: “detection dog” vs “training them for multiple purposes”. But the wording implies this refers to total costs, since it states “raising” rather than just “training”, which usually means a total cost from the beginning. So if training each candidate dog costs the implied $25,680 and the success rates continue to be 30% vs 80%, and the clones have a per-success cost half that of normal dogs, then the implied amortized cloning cost would appear to be ~$8,560 ().

## Cost-benefit in selection problems

How would we approach this problem from first principles?

A SF dog is highly selected among candidate dogs, and it is either an acceptable SF dog or not. Being a SF dog requires a package of traits, ranging from physical health to courage to finely-controlled aggression (attacking if the handler orders, immediately stopping when counter-ordered), which sum up to an overall quality: somewhat poorer health can be made up by better smelling skills, say.

So a natural approach is to treat it as a logistic model, or more specifically, a liability threshold model (“Ch25, Threshold Characters”, Lynch & Walsh 1998): if a bunch of random variables all sum up to a certain high score, the dog becomes SF, otherwise, it is a normal dog. These random variables can be split into genetic variables, and everything else, ‘environmental’ variables.

Then the benefit of cloning can be estimated based on how much the genetic variables contribute to a high score, how high the genetic variables of a cloned SF dog might be (remembering that they are highly selected and thus imply regression to the mean), and this provides an estimate for increased probability that the clones will achieve a high score too. This is effectively an extreme case of truncation selection where a single individual is used as the ‘parent’ of the ‘next generation’.

Once the probability a clone will succeed versus a random candidate dog is calculated, then one can get the cost of screening candidate dogs for a SF dog versus cloning+screening clone dogs for a SF dog.

So we need to know:

1. how difficult it is for a regular SF dog candidate to succeed, and what the implied threshold for a ‘SF score’ is of a random SF dog, and of a elite SF dog

• if possible, how much less difficult it is for a cloned SF dog candidate to succeed, for the implied boost in their average scores
2. the cost of training a regular SF dog candidate

3. the cost of cloning an elite SF dog

4. the heritability of SF success, or failing that, dog traits in general as a prior

1. <1% of breeder puppies may eventually make it to successful SF deployment; most selection happens in the 2 years before handover from the breeder to the military, and failure rates are substantially lower during the military training. For more conventional military or police use, success rates are much higher, and from puppy to deployment, probably more something like 25%.

Of successful SF dogs, the SF cloning pilots appear to be choosing from dogs in the top 1% or higher of SF dogs.

2. the post-handover cost of training per SF dog is likely >$50,000, with total lifetime cost being higher; conventional military/police dogs are again much less stringently selected/trained and thus cost much less, perhaps as low as$20,000.

3. dog cloning costs have dropped steeply since the first cloned dog in 20056 (in large part thanks to consumer demand for pet cloning), with 2019 list prices at <$50,000 and marginal costs possibly as low as$16,000 (so cloning at scale could cost only >$16,000) # Base Rates ## Dog Success Rates Frost 1990, in a broad review of military working dog training: “However, this ‘European solution’ turned out to be only temporary, as rejection rates continued to remain high, and continue today in the range of 25 to 50% (Andersen, Burke, Craig, Hayter, McCathern, Parks, Thorton).”7 Hammerstrom 2005, discussing Vietnam-era war dogs, cites Lemish 1996 that there was “a high rejection rate of 30 to 50% of the potential canine recruits”. Sinn et al 2010 finds in their US Air Force sample, 21% of dogs failed both types of training, summarizes the overall failure rate as “In many selection and training programs for police and detection dogs, more than half of the candidate dogs are rejected for behavioral reasons (Wilsson and Sundgren, 1997b; Slabbert and Odendaal, 1999; Maejima et al., 2007)”8, and notes that given the costs of a failed candidate, “While the improvements in prediction observed here were small (2–7%), given the costs of purchasing, importing, housing, and training (approximately$18,500US per dog), this small percentage improvement results in a substantial potential savings.”

A 2011 article on the 341st estimates that “The suitability rate runs around 50%. In other words, to produce 100 serviceable dogs per year, the program will attempt to train about 200.”

The 2014 Bloomberg quotes Badertscher as saying “you’re lucky if one or two dogs out of a litter of eight might have the drive and focus to become the kind of dogs who can find bombs, take fire, and work independently on command—let alone jump out of airplanes at night.” A followup article quotes a trainer as estimating “maybe out of a litter of eight only four would be police service dogs or military dogs”.

Ritland 2013 describes dogs appropriate for Navy Seals as being “a one-in-a-thousand (or more) proposition…I call them 1 percenters…but they are more like one in ten thousand.”9

Stripes, reporting on Sooam, states in 2016:

But breeding and training programs are costly and often inefficient. For example, the school that trains K-9s for the Department of Defense has found that the suitability rate runs around 50%, so the program tries to train about 200 dogs per year to produce 100 that are serviceable.

The aforementioned South Korean newspaper article put regular dogs in the sniffer training program at 30%.

## Clone Success Rates

One of the first working dogs cloned was a particularly famous Canadian police dog Trakr, whose handler/owner James Symington, won a Sooam contest and received 5 clones of Trakr in 2009, which he began training in search-and-rescue under the auspices of his Team Trakr Foundation (TTF). TTF appears to have gone defunct sometime before 2014, with its last nonprofit filing in 2011, and I am unable to find any information about how the 5 clones worked out. (I have pinged TTF’s contacts.)

South Korean police in 2011 reported a success rate for the clones vs for normal dogs.

Australia was reported to be working on a 2011 deal to have up to 10 cloned sniffer dogs by 2013, but as there is no trace of these Australian dogs elsewhere, the deal must have fallen through.

The 2014 Bloomberg article on Sooam (“For $100,000, You Can Clone Your Dog: These two were made to order in a South Korean lab. They’re only the beginning”) reported that Sooam had a contract for 40 dogs for South Korean clones of which “several are already in service” (presumably the 7 reported before), and also on the birth of 2 clones of a particularly elite SF dog then serving in Afghanistan (name classified), “Ghost” and “Echo”, later joined by a third, “Specter”; the American trainer involved, Brannon, praises the results, reporting in 2016 a success rate (as opposed to a more typical estimate given in the article): Brannon says cloning seems to take the guess work out of normal breeding procedures. “Meaning, you have an excellent male an excellent female, and maybe out of a litter of eight only four would be police service dogs or military dogs,” according to Brannon. Specter is the third clone that the kennel has trained, and the other two are now working with federal SWAT units. “Right now were are three for three and they’re all successful,” said Brannon. A New Scientist report in 2016 visiting Sooam mentions 4 German Shepherds from 1 donor for SK police: “two 9-month-old German shepherds, cloned for the national police. Their original was a working dog deemed particularly capable and well-disposed…Further down is another pair of puppies cloned from the same donor; these ones are just 2 months old.”, commenting on how “incredibly eerie” it is to see dogs with the same “mannerisms” and “perky left ear…like looking at a living growth chart.” If formal police training began at 2 years of age, by 2019 all 4 should be known as successes or failures. Stripes reported in 2016 that 2 of Brannon’s clones finished training & were working for ATF and that Brannon was receiving another clone.10 3 Sooam-cloned Malinois were gifted to Russia in 2016; they reportedly badly failed initial testing in 2017, which was blamed on their thin fur coats being unsuited to the Yakutsk cold. (I don’t know if this should be considered a example or not, given that there were apparently extenuating circumstances.) The aforementioned South Korean newspaper article put clone dogs in the sniffer training program at 80% (vs 30%). # Heritability The connection between being a clone and success probability is mediated by the accuracy of prediction from a donor to the clone, and to what extent a high donor ‘score’ predicts a high clone score. A donor & clone are equivalent to a pair of identical twins raised apart (MZAs), and this prediction is simply the heritability of the trait (which is the square, so a r=0.10 is a h2=0.01). In humans, h2s for everything average ~0.50 or r=0.70; dogs seem to average lower heritabilities, with much of the genetic variance between dogs being the breed-level differences (which have already been exploited by dog breeders/SF trainers in their focus on Malinois etc), but even if the heritability is a fifth the size, a h2=0.10/r=0.31 is useful. SF-specific dog heritabilities should be calculable using existing pedigree records from breeders or the occasional government programs, but I didn’t find any mentioned in my reading. Dog heritabilities in general vary widely and are difficult to summarize because of equally widely varying methods, breeds, and analyses. Having been heavily selectively bred, there are large between-breed differences in behavior due to genetics (most recently: “Highly Heritable and Functionally Relevant Breed Differences in Dog Behavior”, MacLean et al 201911; “Absolute brain size predicts dog breed differences in executive function”, Horschler et al 2019); these group-level heritabilities are irrelevant to this analysis as candidate dogs are already drawn from the best-suited breeds, and it is the remaining within-breed individual genetic differences which matters. This narrower heritability is most frequently estimated <0.50 on measured traits, and the most recent meta-analysis,Hradecká 2015 finds global mean heritabilities like 0.15/0.10/0.15/0.09/0.12, which would seem to not bode well for cloning. However, the dog heritability literature is plagued with serious measurement error issues: the measured variables are unstable, unreliable, do not predict within the same dog over long periods of time, and are generally psychometrically inadequate. Measurement error biases heritability estimates towards zero: if a measurement of ‘temperament’ is not measuring temperament but something like how aggressive a particular trainer is, then regardless of how heritable temperament truly is, the measurement’s heritability will be near-zero; but one would be badly mistaken to then infer that temperament cannot be affected by breeding or that a clone will have a completely different temperament from the donor. In the papers which report relevant aspects of measurement error, the measurements are typically extremely bad, with r=0.1–0.2 being common (for comparison, a properly administered IQ test will be r>0.8). If one adjusts a measured heritability estimate like 0.09 for such noisy measurements, the true heritability could be easily be 0.66 or higher. MacLean et al 2019 reports a set of behavioral trait heritabilities within-breed averaging ~0.15 (see also table 4, Ilska et al 2017), using the C-BARQ inventory, developed with factor analysis and which has reasonable test-retest reliability r~0.5 and loading ~57% on the latent factors, suggesting a true mean heritability >0.24. Another issue is the interpretation of a low heritability on individual behavioral traits: should SF heritability be thought of as a single trait, perhaps the sum of a large number of more atomic behavioral traits, in which case low heritabilities mean that an elite SF dog still has only a somewhat higher total genetic advantage than a random SF candidate dog? Or should, given the need for long sequences of correct decisions & actions drawing on many traits without a single mistake, we see it as more of a log-normally-distributed trait analogous to the O-ring model, in which a small advantage on each atomic trait (due to stringent selection + low heritabilities) nevertheless multiplies out to a large difference in the final outcomes—and so a clone will outperform much more than one would expect from a low heritability on each atomic trait?12 Perhaps most relevantly, Goddard and Beilharz (1982) examine guide dogs, and estimate the heritability of a “success” trait rather than individual traits or subtests, which is much higher than the behavioral mean heritabilities: 0.44 (higher than 4 of the 5 behavioral traits they estimate). A “success” trait here is an “index score”, by weighting correlated variables according to their importance, tend to be optimal predictors with much greater heritability, and to outperform individual variables (Lynch & Walsh 2018: ch37 “Theory of Index Selection”/ch38 “Applications of Index Selection”). Sooam has published behavioral research on cloned dogs: Kim et al 2018/Lee et al 2018/Oh et al 2018 reviews. None of the Sooam papers take a behavioral genetics approach or attempt to estimate heritability/genetic correlations/liability threshold models despite those being necessary for a correct answer, so they have to be read closely. Choi et al 2014 has already been reviewed. Kim et al 2018 states that 4 clones of a cancer-sniffing dog were made but “ownership problems” prevented more than one from being evaluated, which Kim et al 2018 states had similar capabilities as the donor (cited to Kim et al 2015 which I am unable to download or read). Oh et al 2016 compared 2 clone puppies, finding them similar on the Puppy Aptitude Test. Shin et al 2016 tested learning/memory/exploration in 6 clones versus 4 controls, showing generally lower variance; no variance statistics are reported (just p-values), but counting dots on the plots, the implied variance all look >50% to me. Lee et al 2016 mated a cloned detector dog with a regular female dog and tested the 10 offspring; the offspring achieved above average scores with a pass rate of 60% (which is roughly intermediate regular dogs and cloned dogs, suggesting high heritability given a non-detector mother). Choi et al 2017/Choi 201813 found greater consistency of behavioral traits in clone than control puppies but did not estimate direct correlations/heritabilities, instead using calculating F-tests comparing the two groups’ variances; since these are unrelated control dogs and a single group of clone, the reduction in variance should be equivalent to heritability and can be read off from the F subscripts, in which case the various heritabilities are 0.20/0.35/0.40/0.23 etc. Since clones have already been deployed in practice, we can try to work backwards from observed success rates of clones vs normal dogs. The anecdotal instances imply high success rates, nearing 100%, vs standard success rates of <50%, but a tiny total sample size and unclear definitions of success. More specifically, the South Korean sniffer program reportedly has 30% vs 80% on a common outcome, and the sample size is unclear but potentially into the hundreds.14 Using the liability threshold model, one could work back from a threshold and difference in success rates to estimate an implied heritability. In this case, the cutpoints for 30% and 80% are -0.52SD and +0.84SD, implying the clones are +1.36SD above the normals, ignoring any selection before enrollment. That is the mean they regressed back to, based on the unknown heritability (how much back to regress) and a certain threshold (how high the donor/original started off). The lower the threshold, the greater heritability must be to avoid throwing away ability and still matching the observed success rate; the higher the threshold, the lower heritability can be while still providing enough ability-enrichment in the clones to have that higher success rate. In this case, we can assume a threshold like <1%, given that only elites are being cloned and this is consistent with everything else, in which case then the necessary heritability turns out to be… ~50%, which is plausible: qnorm(0.80) - qnorm(0.30) # [1] 1.36602175 0.513 * truncNormMean(qnorm(1-0.01)) # [1] 1.3672549 So based on the existing dog literature and extrapolating from the current observed 80–100% dog success rates, a heritability of ~50% seems most plausible to me. # Costs ## Training But that may be worthwhile depending on how expensive it is to train enough dogs to get a successful dog, and how expensive cloning is. For comparison, similar highly-trained civilian dogs, trained in Schutzhund (sometimes by trainers who used to train for Special Forces), can sell for$40,000$60,000 with Ritland selling his dogs at$50,000–$100,000, and the best award-winning “executive protection dogs” selling for up to$230,000. Bloomberg notes “Canines with finely trained noses now fetch $25,000 and up on the open market, where border patrol units, the State Department, and private security firms go for canine talent.” Hammerstrom 2005 quotes two cost estimates of a US military contractor BSI 1969–1970 at$10,000 (Lemish 1996, War Dogs: A History of Loyalty and Heroism) and $15,000 (cited to “Perry Money, a former Marines Corps handler of a BSI dog”), and mentions a program to breed dogs for better health & “superior ambush detection” (which failed for unknown reasons); he also quotes US Air Force LTC Bannister, commander of the 341st Training Squadron, as estimating the 120-day “DoD MWD Course” at$50,000 “per trained dog” (unclear if this refers to average over all dogs, or per successful ‘trained’ dog)15. Sinn et al 2010 quotes US Air Force training of patrol & detection dogs at ~$18,500, with a 21% total failure rate implying a 1.2x higher cost per success of ~$23,417. A 2011 NYT article on Marines notes it is “an expensively trained canine (the cost to the American military can be as high as $40,000 per dog)”. Ritland & Brozek 2013 quotes a US Navy SEAL dog’s individual cost at >$50,000.16 South Korean police quote a drug sniffer dog at $40,000 for training in 2011. Bloomberg 2017 reports “The U.S. military spends up to$283,000 to train a working war dog…Once it has a promising pup, the Pentagon spends an additional $42,000 to train a K9 unit…When all is said and done, a fully trained military dog costs about as much as a small missile.”17 A 2019 Chinese source on the first domestically-cloned police dog (cloned from a donor police dog who is “one in a thousand”) cites a standard police dog training cost of$75,000 over 5 years. A 2019 Wired notes “Highly trained bomb- and disease-sniffing dogs are in short supply and expensive, as much as $25,000 per pooch.” These estimates vary considerably, and seem to reflect heterogeneity in what cost is being calculated, how selective a role or facility is due to difficulty of role (police dogs cost less than single-role military dogs who cost less than multi-role dogs who cost less than Special Forces dogs). But it’s hard to see what is driving some of these differences, like the 6x difference between Ritland/Bannister and Bloomberg’s estimate, as a Navy SEAL dog is presumably a “working war dog”. It seems to me that the Ritland/Bannister figure is referring to the cost to train a single dog who happened to be successful (plausible since Bannister is providing a 2005 estimate from the dog training facilities he oversees), while the latter refers to the total cost to get one successful dog out of an unspecified number of candidates; if they are referring to the same dogs, then the implied success rate is . Ritland guesstimates that the failure rate among his SF dogs, after they have passed all of the thresholds which lead to formal acquisition by the US Navy (which Ritland emphasizes represents most of the selection done on the dogs), is “more like 3 or 4 in 10 instead of 7.5 out of 10”, which, combined with some additional failures or expenses elsewhere in the process, seems reasonably consistent: if each dog costs >$50,000, then even at face-value that failure estimate implies substantially higher price-tags. Alternately, the Bloomberg estimate might be a ‘total lifetime cost’ estimate of some sort, including all of the maintenance and equipment and costs during deployment and retirement, while Ritland/Bannister is referring only to the upfront training cost.

It is worth noting, given how much selection takes place before sale to SF trainers, that these costs embed the cost of failure beforehand: breeders are not charities, so however much it costs on net to raise, train, and test the dogs which don’t get bought for training, the sale price of a candidate SF dog must cover the washouts as well (after recouping whatever is possible by washouts’ alternate uses, perhaps in less selective roles), or else the breeder would go out of business.

## Cloning

Cloning, on the other hand, cost $100,000 list price in 2015 from Sooam, down from the original$150,000 in 2008, when Bernann McKinney received 5 cloned pit bulls from Sooam.18 Viagen in 2018 reportedly offers a $50,000 plan; the profiled pet owner, Amy Vangemert, received 3 cloned puppies (one of which was adopted out) and incidentally intends to do it again as necessary The Chinese firm Sinogene, which cloned a Kunming wolfdog in 2019, reportedly charges$56,000 for a dog clone. Sooam apparently offers a guarantee, and Viagen will refund if not successful, so that price should be firm. (Why Sooam is able to charge twice as much as Viagen, I do not know.)

The real price per dog may be lower. Their practice seems to be to engage in overkill by implanting multiple clone embryos to ensure the minimum specified number of healthy clones, and offer all resulting healthy live-births to the customer—for example, Vangemert is not mentioned as being charged $150k instead of$50k for the 3 puppies, as would be the case if Viagen charged $50k for each success and she requested 2 and got 3. So the per-clone cost appears to have become surprisingly reasonable:$50k–$100k for 1 dog, and potentially <$16k (if a purchase results in 3 puppies as in Vangemert’s case), suggesting that the price has a large fixed cost to it and the marginal costs might be quite small, which is not helpful in the pet-replacement scenario but would be important to large-scale cloning of specific animals like elite drug sniffers. For comparison, cattle and horse cloning have become industrialized at ~$10–15k; dog cloning is apparently more difficult (harder to control estrous, Viagen notes), so that may be a lower bound for the foreseeable future. # Liability threshold model This requires us to estimate two things: the threshold and the heritability on the liability scale. For common police dogs and other working dogs, training appears to be not that hard, and estimates of 30–50% are seen. This gives a threshold of 50%, or in standard deviations, 0SD. A SF dog is much more selective, and the only specific estimate given is <1% by Mike Ritland, which in standard deviations, means each dog would be >=2.33SD, and the actual mean created by this selection effect is +2.66SD. (If this is confusing imagine a threshold like 50%: is the mean of everyone over 50% equal to 50%? No, it has to be higher, and the mean of everyone >=0SD/>=50% is actually more like 0.8SD/75%—not 0SD/50%!—and we need to use the truncated normal distribution to get it right.) The clone of the SF dog shares only genetics with it, it doesn’t benefit from the unique luck and environment that the original did which helped it achieve it success, so it will regress to the mean. If genetics determined 100% of the outcome, then the clones would always be +2.66SD just like the donor, and hence make the 1%/+2.33SD cutoff 100% of the time, as they have the same genetic potential and zero environmental input (although that is extremely unlikely a scenario, due to measurement error in the testing if nothing else). While if genetics contributed 0% to the outcome and did not matter, then the clones will make the 1% cutoff just as often as if they were a random dog sampled from their breeders ie. 1%. And in between, in between. Under a more plausible case like genetics determining 50% of the variability (a common level of heritability for better-studied human traits), then that is equivalent to a perfect genetic predictor correlating r=0.7; the r, remember, is equivalent to ‘for each 1 SD increase in the independent variable, expect +r SDs in the dependent variable’, so since the clone donor is +2.66SD, the clones will only be SD above the mean. If the clones are distributed around a mean of +1.86SD thanks to their genes, what’s the probability they will then reach up to a total of +2.33SD (the threshold) with help from the environment & luck? Half the variance is used up, and the environment has to contribute another = +0.47SD, despite causing differences of only 0.7SD on average. In that case, the clones will have ~26% chance of being successful—which is a remarkable 26x greater than a random dog, but also far from guaranteed. But one can do better, since it is not necessary to select a random SF dog (with their implied average of +2.66SD) but one can select the best SF dog and clone this elite specimen instead. Multi-stage selection is always more efficient than single-stage selection, particularly when we are interested in extremes/tails, due to the ‘thin tails’ of the normal distribution. At any time there are thousands of SF dogs worldwide, and more in retirement (and perhaps more if tissue samples have been preserved from earlier generations), so the gain from an additional selection step is potentially large (especially when we consider tail effects), and since only 1 dog is necessary for cloning, why settle for anything less than the best? If one can select at least the best SF dog out of 100019, then the new ‘threshold’ is +4.26SD and the expectation for our elite dog is +4.47SD, and likewise, the clones at 50% heritability would be +3.16SD, which is considerably above the original SF threshold of 2.33SD, and now fully 88% of the clones would be expected to succeed at SF training. Source code defining the truncated normal distribution, the cloning process, and a Monte Carlo implementation20: ## exact mean for the truncated normal distribution: truncNormMean <- function(a, mu=0, sigma=1, b=Inf) { phi <- dnorm erf <- function(x) 2 * pnorm(x * sqrt(2)) - 1 Phi <- function(x) { 0.5 * (1 + erf(x/sqrt(2))) } Z <- function(beta, alpha) { Phi(beta) - Phi(alpha) } alpha = (a-mu)/sigma; beta = (b-mu)/sigma return( (phi(alpha) - phi(beta)) / Z(beta, alpha) ) } ## If we select the top percentile, the cutoff is +2.32SD, but the mean is higher, +2.66SD: qnorm(0.99) # [1] 2.32634787 truncNormMean(qnorm(0.99)) # [1] 2.66521422 truncNormMean(qnorm(1-0.01^2)) # [1] 3.95847967 cloningBoost <- function(successP=0.01, preThreshold=0.01, heritability=0.5, verbose=FALSE) { threshold <- qnorm(1-preThreshold) successThreshold <- qnorm(1-successP) originalMean <- truncNormMean(threshold) cloneMean <- 0 + (sqrt(heritability) * originalMean) ## regress to mean regression <- originalMean - cloneMean cloneP <- pnorm(cloneMean - successThreshold, sd=sqrt(1-heritability)) if (verbose) { print(round(digits=3, c(threshold, successThreshold, originalMean, cloneMean, regression, cloneP))) } return(cloneP) } ## Alternative Monte Carlo implementation to check: cloningBoostMC <- function(successP=0.01, preThreshold=0.01, heritability=0.5, verbose=FALSE, iters1=10000000, iters2=1000) { threshold <- qnorm(1-preThreshold) successThreshold <- qnorm(1-successP) r <- sqrt( heritability) r_env <- sqrt(1-heritability) population <- rnorm(iters1, mean=0, sd=1) eliteDonors <- population[population>=threshold] clones <- as.vector(sapply(eliteDonors, function(d) { rnorm(iters2, ## sample _n_ clones per donor ## regress back to mean for true genetic mean: mean=d*r, ## left-over non-genetic variance affecting clones: sd=r_env) })) successes <- clones>=successThreshold cloneP <- mean(successes) if (verbose) { library(skimr) print(skim(population)); print(skim(eliteDonors)); print(skim(successes)) } return(cloneP) } ## Varying heritabilities, 0-1: cloningBoost(successP=0.01, heritability=1.0, verbose=TRUE) # [1] 2.326 2.326 2.665 2.665 0.000 1.000 # [1] 1 cloningBoost(successP=0.01, heritability=0.8, verbose=TRUE) # [1] 2.326 2.326 2.665 2.384 0.281 0.551 # [1] 0.551145688 cloningBoost(successP=0.01, heritability=0.5, verbose=TRUE) # [1] 2.326 2.326 2.665 1.885 0.781 0.266 # [1] 0.266071352 cloningBoost(successP=0.01, heritability=0.2, verbose=TRUE) # [1] 2.326 2.326 2.665 1.192 1.473 0.102 # [1] 0.102340263 cloningBoost(successP=0.01, heritability=0.0, verbose=TRUE) # [1] 2.326 2.326 2.665 0.000 2.665 0.010 # [1] 0.01 ## Enriched selection by selecting elites rather than random: cloningBoost(successP=0.01, preThreshold=0.01 * (1/1000), heritability=0.5, verbose=TRUE) # [1] 4.265 2.326 4.479 3.167 1.312 0.883 # [1] 0.882736927 For insight, we can look at how final success probability increases with different heritabilities/_r_s, in the single-step selection scenario (corresponding to a random selection of SF dogs for cloning) and for the double-step selection (selecting a top 1% SF dog for cloning): ## Plotting the increase in subsequent probability given various correlations: df1 <- data.frame(PriorP=numeric(), R=numeric(), Success.Rate=numeric()) for (p in c(0.01, seq(0.05, 0.95, by=0.05), 0.99)) { for (r in seq(0,1, by=0.01)) { df1 <- rbind(df1, data.frame(PriorP=p, R=r, Success.Rate=cloningBoost(successP=p, heritability=r^2))) } } library(ggplot2); library(gridExtra) p1 <- qplot(R, Success.Rate, color=as.ordered(PriorP), data=df1) + geom_line() + theme(legend.title=element_blank()) p2 <- qplot(R, log(Success.Rate), color=as.ordered(PriorP), data=df1) + geom_line() + theme(legend.title=element_blank()) grid.arrange(p1, p2, ncol=1) ## Double-step selection: df2 <- data.frame(PriorP=numeric(), R=numeric(), Success.Rate=numeric()) for (p in c(0.01, seq(0.05, 0.95, by=0.05), 0.99)) { for (r in seq(0.01,1, by=0.01)) { df2 <- rbind(df2, data.frame(PriorP=p, preThreshold=p * 0.01, R=r, Success.Rate=cloningBoost(successP=p, preThreshold=p * 0.01, heritability=r^2))) } } library(ggplot2); library(gridExtra) p1 <- qplot(R, Success.Rate, color=as.ordered(PriorP), data=df2) + geom_line() + theme(legend.title=element_blank()) p2 <- qplot(R, log(Success.Rate), color=as.ordered(PriorP), data=df2) + geom_line() + theme(legend.title=element_blank()) grid.arrange(p1, p2, ncol=1) ## Cost-benefit Does cloning minimize loss? My cost-benefit below takes the cost per final dog without cloning, computes the implied per-dog-candidate cost, and then computes the increased success rate for a given threshold+heritability, and sees if the expected cloning+training cost is less than the original total cost. dogCloningCB <- function(successP, heritability, totalTrainingCost, marginalCloningCost, verbose=FALSE) { normalLoss <- totalTrainingCost marginalTrainingCost <- totalTrainingCost / (1/successP) cloningP <- cloningBoost(successP=successP, heritability=heritability) cloningLoss <- ((1/cloningP) * (marginalTrainingCost + marginalCloningCost)) if(verbose) {return(list(Boost=cloningP, Cost.normal=normalLoss, Cost.marginal=marginalTrainingCost, Cost.clone=cloningLoss, Profitable=normalLoss>cloningLoss, Profit=normalLoss-cloningLoss)) } return(normalLoss-cloningLoss) } ## Example: 30% success rate, 50% heritability,$85k per-dog training cost, $15k per-clone cost dogCloningCB(0.30, 0.5, 85600, 15000, verbose=TRUE) #$Boost
# [1] 0.972797623
#
# $Cost.normal # [1] 85600 # #$Cost.marginal
# [1] 25680
#
# $Cost.clone # [1] 41817.5364 # #$Profitable
# [1] TRUE
#
# $Profit # [1] 43782.4636 ### Scenarios As the key heritability trait is almost completely unknown and heritabilities of dog behavioral traits are all over the map and seem to suffer from severe measurement error issues, we might as well consider a wide range of scenarios to get an idea of what it would take. For success/threshold, we continue with 1%; for heritability, we’ll consider the most plausible range, 0–90%; for training cost, we’ll do the full$50k–$283k range since while it’s unclear what these numbers mean, treating them as a total per-dog cost is being conservative and makes it harder for cloning to be profitable, and for cloning costs we’ll consider the Vangemert case up to Viagen’s list price of$50k (since there doesn’t seem to be any good reason to pay twice as much to Sooam).

scenarios <- expand.grid(SuccessP=0.01, Heritability=seq(0, 0.9, by=0.10), trainingCost=seq(50000, 283000, by=10000), cloningCost=seq(15000, 50000, by=10000))
scenarios$Profit <- round(unlist(Map(dogCloningCB, scenarios[,1], scenarios[,2], scenarios[,3], scenarios[,4]))) ## Plot relationships among profitable scenarios: scenariosProfitable <- scenarios[scenarios$Profit>0,]
library(ggplot2); library(gridExtra)
p1 <- qplot(cloningCost,  Profit, color=Heritability, data=scenariosProfitable) + geom_hline(yintercept=0, color="red")
p2 <- qplot(trainingCost, Profit, color=Heritability, data=scenariosProfitable) + geom_hline(yintercept=0, color="red")
grid.arrange(p1, p2, ncol=1)

## All profitable scenarios:
scenariosProfitable
# ...
The subset of profitable scenarios for SF dog cloning (typically requiring high heritabilities, and higher training costs / lower cloning costs).
Success Probability Heritability Training cost Cloning Cost Profit
0.01 0.6 50000 15000 4333
0.01 0.7 50000 15000 13962
0.01 0.8 50000 15000 21877
0.01 0.9 50000 15000 29015
0.01 0.5 60000 15000 1369
0.01 0.6 60000 15000 14038
0.01 0.7 60000 15000 23729
0.01 0.8 60000 15000 31695
0.01 0.9 60000 15000 38879
0.01 0.5 70000 15000 10993
0.01 0.6 70000 15000 23743
0.01 0.7 70000 15000 33497
0.01 0.8 70000 15000 41514
0.01 0.9 70000 15000 48744
0.01 0.4 80000 15000 2576
0.01 0.5 80000 15000 20617
0.01 0.6 80000 15000 33449
0.01 0.7 80000 15000 43264
0.01 0.8 80000 15000 51332
0.01 0.9 80000 15000 58609
0.01 0.4 90000 15000 12086
0.01 0.5 90000 15000 30242
0.01 0.6 90000 15000 43154
0.01 0.7 90000 15000 53032
0.01 0.8 90000 15000 61151
0.01 0.9 90000 15000 68473
0.01 0.4 100000 15000 21596
0.01 0.5 100000 15000 39866
0.01 0.6 100000 15000 52859
0.01 0.7 100000 15000 62799
0.01 0.8 100000 15000 70970
0.01 0.9 100000 15000 78338
0.01 0.3 110000 15000 2785
0.01 0.4 110000 15000 31106
0.01 0.5 110000 15000 49490
0.01 0.6 110000 15000 62565
0.01 0.7 110000 15000 72567
0.01 0.8 110000 15000 80788
0.01 0.9 110000 15000 88202
0.01 0.3 120000 15000 12119
0.01 0.4 120000 15000 40616
0.01 0.5 120000 15000 59114
0.01 0.6 120000 15000 72270
0.01 0.7 120000 15000 82334
0.01 0.8 120000 15000 90607
0.01 0.9 120000 15000 98067
0.01 0.3 130000 15000 21453
0.01 0.4 130000 15000 50126
0.01 0.5 130000 15000 68738
0.01 0.6 130000 15000 81976
0.01 0.7 130000 15000 92102
0.01 0.8 130000 15000 100425
0.01 0.9 130000 15000 107932
0.01 0.3 140000 15000 30787
0.01 0.4 140000 15000 59636
0.01 0.5 140000 15000 78362
0.01 0.6 140000 15000 91681
0.01 0.7 140000 15000 101869
0.01 0.8 140000 15000 110244
0.01 0.9 140000 15000 117796
0.01 0.3 150000 15000 40121
0.01 0.4 150000 15000 69146
0.01 0.5 150000 15000 87987
0.01 0.6 150000 15000 101386
0.01 0.7 150000 15000 111637
0.01 0.8 150000 15000 120062
0.01 0.9 150000 15000 127661
0.01 0.3 160000 15000 49455
0.01 0.4 160000 15000 78656
0.01 0.5 160000 15000 97611
0.01 0.6 160000 15000 111092
0.01 0.7 160000 15000 121404
0.01 0.8 160000 15000 129881
0.01 0.9 160000 15000 137526
0.01 0.2 170000 15000 6819
0.01 0.3 170000 15000 58789
0.01 0.4 170000 15000 88166
0.01 0.5 170000 15000 107235
0.01 0.6 170000 15000 120797
0.01 0.7 170000 15000 131171
0.01 0.8 170000 15000 139699
0.01 0.9 170000 15000 147390
0.01 0.2 180000 15000 15842
0.01 0.3 180000 15000 68123
0.01 0.4 180000 15000 97676
0.01 0.5 180000 15000 116859
0.01 0.6 180000 15000 130502
0.01 0.7 180000 15000 140939
0.01 0.8 180000 15000 149518
0.01 0.9 180000 15000 157255
0.01 0.2 190000 15000 24865
0.01 0.3 190000 15000 77457
0.01 0.4 190000 15000 107186
0.01 0.5 190000 15000 126483
0.01 0.6 190000 15000 140208
0.01 0.7 190000 15000 150706
0.01 0.8 190000 15000 159337
0.01 0.9 190000 15000 167119
0.01 0.2 200000 15000 33887
0.01 0.3 200000 15000 86791
0.01 0.4 200000 15000 116696
0.01 0.5 200000 15000 136107
0.01 0.6 200000 15000 149913
0.01 0.7 200000 15000 160474
0.01 0.8 200000 15000 169155
0.01 0.9 200000 15000 176984
0.01 0.2 210000 15000 42910
0.01 0.3 210000 15000 96125
0.01 0.4 210000 15000 126206
0.01 0.5 210000 15000 145732
0.01 0.6 210000 15000 159619
0.01 0.7 210000 15000 170241
0.01 0.8 210000 15000 178974
0.01 0.9 210000 15000 186849
0.01 0.2 220000 15000 51933
0.01 0.3 220000 15000 105459
0.01 0.4 220000 15000 135716
0.01 0.5 220000 15000 155356
0.01 0.6 220000 15000 169324
0.01 0.7 220000 15000 180009
0.01 0.8 220000 15000 188792
0.01 0.9 220000 15000 196713
0.01 0.2 230000 15000 60956
0.01 0.3 230000 15000 114794
0.01 0.4 230000 15000 145226
0.01 0.5 230000 15000 164980
0.01 0.6 230000 15000 179029
0.01 0.7 230000 15000 189776
0.01 0.8 230000 15000 198611
0.01 0.9 230000 15000 206578
0.01 0.2 240000 15000 69979
0.01 0.3 240000 15000 124128
0.01 0.4 240000 15000 154736
0.01 0.5 240000 15000 174604
0.01 0.6 240000 15000 188735
0.01 0.7 240000 15000 199544
0.01 0.8 240000 15000 208429
0.01 0.9 240000 15000 216442
0.01 0.2 250000 15000 79002
0.01 0.3 250000 15000 133462
0.01 0.4 250000 15000 164246
0.01 0.5 250000 15000 184228
0.01 0.6 250000 15000 198440
0.01 0.7 250000 15000 209311
0.01 0.8 250000 15000 218248
0.01 0.9 250000 15000 226307
0.01 0.2 260000 15000 88025
0.01 0.3 260000 15000 142796
0.01 0.4 260000 15000 173756
0.01 0.5 260000 15000 193852
0.01 0.6 260000 15000 208145
0.01 0.7 260000 15000 219079
0.01 0.8 260000 15000 228067
0.01 0.9 260000 15000 236172
0.01 0.2 270000 15000 97048
0.01 0.3 270000 15000 152130
0.01 0.4 270000 15000 183266
0.01 0.5 270000 15000 203476
0.01 0.6 270000 15000 217851
0.01 0.7 270000 15000 228846
0.01 0.8 270000 15000 237885
0.01 0.9 270000 15000 246036
0.01 0.2 280000 15000 106070
0.01 0.3 280000 15000 161464
0.01 0.4 280000 15000 192776
0.01 0.5 280000 15000 213101
0.01 0.6 280000 15000 227556
0.01 0.7 280000 15000 238614
0.01 0.8 280000 15000 247704
0.01 0.9 280000 15000 255901
0.01 0.8 50000 25000 3733
0.01 0.9 50000 25000 15476
0.01 0.7 60000 25000 478
0.01 0.8 60000 25000 13551
0.01 0.9 60000 25000 25341
0.01 0.7 70000 25000 10246
0.01 0.8 70000 25000 23370
0.01 0.9 70000 25000 35205
0.01 0.6 80000 25000 3986
0.01 0.7 80000 25000 20013
0.01 0.8 80000 25000 33188
0.01 0.9 80000 25000 45070
0.01 0.6 90000 25000 13691
0.01 0.7 90000 25000 29781
0.01 0.8 90000 25000 43007
0.01 0.9 90000 25000 54934
0.01 0.5 100000 25000 2282
0.01 0.6 100000 25000 23397
0.01 0.7 100000 25000 39548
0.01 0.8 100000 25000 52826
0.01 0.9 100000 25000 64799
0.01 0.5 110000 25000 11906
0.01 0.6 110000 25000 33102
0.01 0.7 110000 25000 49316
0.01 0.8 110000 25000 62644
0.01 0.9 110000 25000 74664
0.01 0.5 120000 25000 21530
0.01 0.6 120000 25000 42807
0.01 0.7 120000 25000 59083
0.01 0.8 120000 25000 72463
0.01 0.9 120000 25000 84528
0.01 0.4 130000 25000 1124
0.01 0.5 130000 25000 31154
0.01 0.6 130000 25000 52513
0.01 0.7 130000 25000 68851
0.01 0.8 130000 25000 82281
0.01 0.9 130000 25000 94393
0.01 0.4 140000 25000 10634
0.01 0.5 140000 25000 40778
0.01 0.6 140000 25000 62218
0.01 0.7 140000 25000 78618
0.01 0.8 140000 25000 92100
0.01 0.9 140000 25000 104257
0.01 0.4 150000 25000 20144
0.01 0.5 150000 25000 50403
0.01 0.6 150000 25000 71924
0.01 0.7 150000 25000 88386
0.01 0.8 150000 25000 101918
0.01 0.9 150000 25000 114122
0.01 0.4 160000 25000 29654
0.01 0.5 160000 25000 60027
0.01 0.6 160000 25000 81629
0.01 0.7 160000 25000 98153
0.01 0.8 160000 25000 111737
0.01 0.9 160000 25000 123987
0.01 0.4 170000 25000 39164
0.01 0.5 170000 25000 69651
0.01 0.6 170000 25000 91334
0.01 0.7 170000 25000 107921
0.01 0.8 170000 25000 121555
0.01 0.9 170000 25000 133851
0.01 0.3 180000 25000 1530
0.01 0.4 180000 25000 48674
0.01 0.5 180000 25000 79275
0.01 0.6 180000 25000 101040
0.01 0.7 180000 25000 117688
0.01 0.8 180000 25000 131374
0.01 0.9 180000 25000 143716
0.01 0.3 190000 25000 10864
0.01 0.4 190000 25000 58184
0.01 0.5 190000 25000 88899
0.01 0.6 190000 25000 110745
0.01 0.7 190000 25000 127456
0.01 0.8 190000 25000 141193
0.01 0.9 190000 25000 153581
0.01 0.3 200000 25000 20198
0.01 0.4 200000 25000 67694
0.01 0.5 200000 25000 98523
0.01 0.6 200000 25000 120450
0.01 0.7 200000 25000 137223
0.01 0.8 200000 25000 151011
0.01 0.9 200000 25000 163445
0.01 0.3 210000 25000 29532
0.01 0.4 210000 25000 77204
0.01 0.5 210000 25000 108148
0.01 0.6 210000 25000 130156
0.01 0.7 210000 25000 146991
0.01 0.8 210000 25000 160830
0.01 0.9 210000 25000 173310
0.01 0.3 220000 25000 38866
0.01 0.4 220000 25000 86714
0.01 0.5 220000 25000 117772
0.01 0.6 220000 25000 139861
0.01 0.7 220000 25000 156758
0.01 0.8 220000 25000 170648
0.01 0.9 220000 25000 183174
0.01 0.3 230000 25000 48200
0.01 0.4 230000 25000 96224
0.01 0.5 230000 25000 127396
0.01 0.6 230000 25000 149567
0.01 0.7 230000 25000 166526
0.01 0.8 230000 25000 180467
0.01 0.9 230000 25000 193039
0.01 0.3 240000 25000 57534
0.01 0.4 240000 25000 105734
0.01 0.5 240000 25000 137020
0.01 0.6 240000 25000 159272
0.01 0.7 240000 25000 176293
0.01 0.8 240000 25000 190285
0.01 0.9 240000 25000 202904
0.01 0.3 250000 25000 66868
0.01 0.4 250000 25000 115244
0.01 0.5 250000 25000 146644
0.01 0.6 250000 25000 168977
0.01 0.7 250000 25000 186061
0.01 0.8 250000 25000 200104
0.01 0.9 250000 25000 212768
0.01 0.3 260000 25000 76202
0.01 0.4 260000 25000 124754
0.01 0.5 260000 25000 156268
0.01 0.6 260000 25000 178683
0.01 0.7 260000 25000 195828
0.01 0.8 260000 25000 209922
0.01 0.9 260000 25000 222633
0.01 0.3 270000 25000 85536
0.01 0.4 270000 25000 134264
0.01 0.5 270000 25000 165893
0.01 0.6 270000 25000 188388
0.01 0.7 270000 25000 205596
0.01 0.8 270000 25000 219741
0.01 0.9 270000 25000 232497
0.01 0.2 280000 25000 8357
0.01 0.3 280000 25000 94871
0.01 0.4 280000 25000 143774
0.01 0.5 280000 25000 175517
0.01 0.6 280000 25000 198093
0.01 0.7 280000 25000 215363
0.01 0.8 280000 25000 229560
0.01 0.9 280000 25000 242362
0.01 0.9 50000 35000 1937
0.01 0.9 60000 35000 11802
0.01 0.8 70000 35000 5226
0.01 0.9 70000 35000 21666
0.01 0.8 80000 35000 15044
0.01 0.9 80000 35000 31531
0.01 0.7 90000 35000 6530
0.01 0.8 90000 35000 24863
0.01 0.9 90000 35000 41396
0.01 0.7 100000 35000 16298
0.01 0.8 100000 35000 34682
0.01 0.9 100000 35000 51260
0.01 0.6 110000 35000 3639
0.01 0.7 110000 35000 26065
0.01 0.8 110000 35000 44500
0.01 0.9 110000 35000 61125
0.01 0.6 120000 35000 13345
0.01 0.7 120000 35000 35833
0.01 0.8 120000 35000 54319
0.01 0.9 120000 35000 70989
0.01 0.6 130000 35000 23050
0.01 0.7 130000 35000 45600
0.01 0.8 130000 35000 64137
0.01 0.9 130000 35000 80854
0.01 0.5 140000 35000 3195
0.01 0.6 140000 35000 32755
0.01 0.7 140000 35000 55368
0.01 0.8 140000 35000 73956
0.01 0.9 140000 35000 90719
0.01 0.5 150000 35000 12819
0.01 0.6 150000 35000 42461
0.01 0.7 150000 35000 65135
0.01 0.8 150000 35000 83774
0.01 0.9 150000 35000 100583
0.01 0.5 160000 35000 22443
0.01 0.6 160000 35000 52166
0.01 0.7 160000 35000 74903
0.01 0.8 160000 35000 93593
0.01 0.9 160000 35000 110448
0.01 0.5 170000 35000 32067
0.01 0.6 170000 35000 61871
0.01 0.7 170000 35000 84670
0.01 0.8 170000 35000 103411
0.01 0.9 170000 35000 120312
0.01 0.5 180000 35000 41691
0.01 0.6 180000 35000 71577
0.01 0.7 180000 35000 94438
0.01 0.8 180000 35000 113230
0.01 0.9 180000 35000 130177
0.01 0.4 190000 35000 9182
0.01 0.5 190000 35000 51315
0.01 0.6 190000 35000 81282
0.01 0.7 190000 35000 104205
0.01 0.8 190000 35000 123049
0.01 0.9 190000 35000 140042
0.01 0.4 200000 35000 18692
0.01 0.5 200000 35000 60940
0.01 0.6 200000 35000 90988
0.01 0.7 200000 35000 113973
0.01 0.8 200000 35000 132867
0.01 0.9 200000 35000 149906
0.01 0.4 210000 35000 28201
0.01 0.5 210000 35000 70564
0.01 0.6 210000 35000 100693
0.01 0.7 210000 35000 123740
0.01 0.8 210000 35000 142686
0.01 0.9 210000 35000 159771
0.01 0.4 220000 35000 37711
0.01 0.5 220000 35000 80188
0.01 0.6 220000 35000 110398
0.01 0.7 220000 35000 133508
0.01 0.8 220000 35000 152504
0.01 0.9 220000 35000 169636
0.01 0.4 230000 35000 47221
0.01 0.5 230000 35000 89812
0.01 0.6 230000 35000 120104
0.01 0.7 230000 35000 143275
0.01 0.8 230000 35000 162323
0.01 0.9 230000 35000 179500
0.01 0.4 240000 35000 56731
0.01 0.5 240000 35000 99436
0.01 0.6 240000 35000 129809
0.01 0.7 240000 35000 153043
0.01 0.8 240000 35000 172141
0.01 0.9 240000 35000 189365
0.01 0.3 250000 35000 275
0.01 0.4 250000 35000 66241
0.01 0.5 250000 35000 109060
0.01 0.6 250000 35000 139514
0.01 0.7 250000 35000 162810
0.01 0.8 250000 35000 181960
0.01 0.9 250000 35000 199229
0.01 0.3 260000 35000 9609
0.01 0.4 260000 35000 75751
0.01 0.5 260000 35000 118685
0.01 0.6 260000 35000 149220
0.01 0.7 260000 35000 172578
0.01 0.8 260000 35000 191778
0.01 0.9 260000 35000 209094
0.01 0.3 270000 35000 18943
0.01 0.4 270000 35000 85261
0.01 0.5 270000 35000 128309
0.01 0.6 270000 35000 158925
0.01 0.7 270000 35000 182345
0.01 0.8 270000 35000 201597
0.01 0.9 270000 35000 218959
0.01 0.3 280000 35000 28277
0.01 0.4 280000 35000 94771
0.01 0.5 280000 35000 137933
0.01 0.6 280000 35000 168631
0.01 0.7 280000 35000 192113
0.01 0.8 280000 35000 211416
0.01 0.9 280000 35000 228823
0.01 0.9 70000 45000 8128
0.01 0.9 80000 45000 17992
0.01 0.8 90000 45000 6719
0.01 0.9 90000 45000 27857
0.01 0.8 100000 45000 16537
0.01 0.9 100000 45000 37721
0.01 0.7 110000 45000 2815
0.01 0.8 110000 45000 26356
0.01 0.9 110000 45000 47586
0.01 0.7 120000 45000 12582
0.01 0.8 120000 45000 36175
0.01 0.9 120000 45000 57451
0.01 0.7 130000 45000 22350
0.01 0.8 130000 45000 45993
0.01 0.9 130000 45000 67315
0.01 0.6 140000 45000 3293
0.01 0.7 140000 45000 32117
0.01 0.8 140000 45000 55812
0.01 0.9 140000 45000 77180
0.01 0.6 150000 45000 12998
0.01 0.7 150000 45000 41885
0.01 0.8 150000 45000 65630
0.01 0.9 150000 45000 87044
0.01 0.6 160000 45000 22703
0.01 0.7 160000 45000 51652
0.01 0.8 160000 45000 75449
0.01 0.9 160000 45000 96909
0.01 0.6 170000 45000 32409
0.01 0.7 170000 45000 61420
0.01 0.8 170000 45000 85267
0.01 0.9 170000 45000 106774
0.01 0.5 180000 45000 4107
0.01 0.6 180000 45000 42114
0.01 0.7 180000 45000 71187
0.01 0.8 180000 45000 95086
0.01 0.9 180000 45000 116638
0.01 0.5 190000 45000 13731
0.01 0.6 190000 45000 51819
0.01 0.7 190000 45000 80955
0.01 0.8 190000 45000 104905
0.01 0.9 190000 45000 126503
0.01 0.5 200000 45000 23356
0.01 0.6 200000 45000 61525
0.01 0.7 200000 45000 90722
0.01 0.8 200000 45000 114723
0.01 0.9 200000 45000 136367
0.01 0.5 210000 45000 32980
0.01 0.6 210000 45000 71230
0.01 0.7 210000 45000 100490
0.01 0.8 210000 45000 124542
0.01 0.9 210000 45000 146232
0.01 0.5 220000 45000 42604
0.01 0.6 220000 45000 80936
0.01 0.7 220000 45000 110257
0.01 0.8 220000 45000 134360
0.01 0.9 220000 45000 156097
0.01 0.5 230000 45000 52228
0.01 0.6 230000 45000 90641
0.01 0.7 230000 45000 120025
0.01 0.8 230000 45000 144179
0.01 0.9 230000 45000 165961
0.01 0.4 240000 45000 7729
0.01 0.5 240000 45000 61852
0.01 0.6 240000 45000 100346
0.01 0.7 240000 45000 129792
0.01 0.8 240000 45000 153997
0.01 0.9 240000 45000 175826
0.01 0.4 250000 45000 17239
0.01 0.5 250000 45000 71476
0.01 0.6 250000 45000 110052
0.01 0.7 250000 45000 139560
0.01 0.8 250000 45000 163816
0.01 0.9 250000 45000 185691
0.01 0.4 260000 45000 26749
0.01 0.5 260000 45000 81101
0.01 0.6 260000 45000 119757
0.01 0.7 260000 45000 149327
0.01 0.8 260000 45000 173634
0.01 0.9 260000 45000 195555
0.01 0.4 270000 45000 36259
0.01 0.5 270000 45000 90725
0.01 0.6 270000 45000 129462
0.01 0.7 270000 45000 159095
0.01 0.8 270000 45000 183453
0.01 0.9 270000 45000 205420
0.01 0.4 280000 45000 45769
0.01 0.5 280000 45000 100349
0.01 0.6 280000 45000 139168
0.01 0.7 280000 45000 168862
0.01 0.8 280000 45000 193272
0.01 0.9 280000 45000 215284

# Conclusion

Under the most plausible scenarios corresponding to US & South Korea costs & success rates, and the high heritabilities indicated by available evidence, dog cloning is profitable in theory. This is consistent with the reports from South Korea that dog cloning is profitable in practice.

So dog cloning for police/military use is profitable in both theory & practice.

The benefits of cloning or using partial predictors of extreme outliers is general, and will be applicable to many other areas, especially where screening/training is lengthy & expensive.

# Appendix

## Dog heritabilities

On measurement error and heritability:

• Wilsson & Sundgren 1997, “The use of a behaviour test for selection of dogs for service and breeding. II. Heritability for tested parameters and effect of selection based on service dog characteristics” notes that their use of factor analysis on multiple tests to derive an index value yields higher heritability estimates than the raw tests (compare Table 1 with Table 3):

It is remarkable that the heritability for the calculated index value and for the four factors from the factor analysis is comparatively high (Tables 2 and 3). This is normally expected to hold true for single well-defined characteristics. This study, however, shows a higher heritability for complex behaviour systems. The more complex parameters, index values and the four factors from the factor analysis show a higher heritability than most of the single characteristics that they are based on. One possible explanation is that the evaluated characteristics overlap and a higher degree of confidence can be achieved if the information from the evaluated characteristics are pooled. The probability of this explanation is further enhanced by the relatively high positive phenotypic correlation maintained between the characteristics (Wilsson and Sundgren, 1996). Goddard and Beilharz (1982) show a heritability as high as 0.44 to predict a dog’s ability to become a guide dog for the blind. The characteristic used was defined as “success”. Mackenzie et al. (1985) calculated the heritability of “temperament” to be 0.51 in 575 military dogs. In both cases the high heritability figures were calculated on a characteristic that summarises complex behaviour systems. With regards to this, it should be pointed out that the characteristic “temperament” in the study of Mackenzie et al. (1985) is defined as a military dog’s suitability for protection and tracking and must not be considered synonymous with the definition of temperament used in this study.

• If temperament tests are to be of any value, they must be shown to be both reliable and valid. Reliability is a prerequisite for validity, and so we review the evidence for reliability first.The first thing to conclude about reliability is that with the few exceptions we will discuss in more detail, researchers have rarely reported reliability of any kind.

…Table 3 is divided into two types of reliability: inter-observer agreement and test–retest reliability. The studies using inter-observer agreement used the traditional method of analysis in which each variable is analyzed across subjects (instead of computing reliability within subjects). The correlations suggest that inter-judge agreement varies greatly across studies and traits. Although strong agreement is possible, it is by no means guaranteed; the sample-weighted mean agreement correlation was .60, but the agreement correlations ranged from .00 to .86…Two studies appear in the test–retest reliability category, listed in the lower section of Table 3, examining the correlation between scores when dogs were tested twice. One of these studies, by Goddard and Beilharz (1986), reveals Activity level is reliable from test to test, but that this reliability decreases as puppies age. The other study, by Netto and Planta (1997), shows a strong mean correlation, but also included many insignificant correlations. Closer examination reveals that many of the Kappa coefficients reported are zero,indicating no reliability. However, this is partially an artifact of the testing situation because the subtests were not intended to elicit Aggression, so it makes little sense to assess the reliability with which they elicited aggression. Of the subtests in this study which were intended to elicit aggression, the lowest Kappa coefficient is -.03 for reaction to an artificial hand taking away food, and reaction to a stranger being mildly threatening when meeting the dog’s handler. However, Netto and Planta’s study should be commended for fully reporting their reliability data; when interpreted against an understanding of the testing situations, these are data are very valuable. Table 4 summarizes all the internal consistency estimates reported in the studies reviewed. Internal consistency measures estimate the degree to which items on a scale assess the same construct. In human personality research, they are often used following factor analyses to determine the internal coherence of the derived factors. Of the 16 studies in our review to focus on factor analysis, only three reported internal consistency. Two of these studies (Hsu and Serpell, 2003; Serpell and Hsu, 2001) gathered data using questionnaires with 5-point frequency (Likert) scales; the third (Seksel et al., 1999) used a 100-point scale. One additional study that did not focus on factor analysis also reported internal consistency (Gosling et al., 2003a) and is included in Table 4. Internal consistency varied greatly across studies and factors, ranging from .42 for “Handling”, to .93 for “Stranger-directed Aggression”. Although high consistency is possible, it is by no means guaranteed. Nonetheless, the internal consistency measures had a weighted mean of .76, well within the limits acceptable in most human personality research (John and Benet-Martinez, 2000)

Overall, the evidence for convergent validity is reasonably promising, with the various estimates averaging about .40 across the nine dimensions examined here.

• Cauchoix et al 2018, “The repeatability of cognitive performance: a meta-analysis” finds across 25 animal species “mean R estimates ranging between 0.15 and 0.28.”

• “Personality and performance in military working dogs: Reliability and predictive validity of behavioral tests”, Sinn et al 2010: finds considerable inter-rater disagreement, and low long-term test-retest reliability

• Hradecká 2015, “Heritability of behavioural traits in domestic dogs: A meta-analysis”, meta-analyzes global heritabilities across various domains as 0.15/0.10/0.15/0.09/0.12; narrowing down to the “Psychical” domain of traits which seem to be most key to SF training, and the breeds which are most often employed: Belgian Shepherd Dog, 0.13; German Shepherd Dog: 0.12; Labrador Retriever: 0.07. These are low but Hradecká 2015 comments on the high unreliability of the measurements being used in most dog heritability studies, which will have the effect of extremely reducing heritability estimates:

Multifactorial analysis revealed that values of heritability of behavioural traits were affected not only by biotic factors such as age and sex, suggesting importance of experience, training, and learning (e.g., Karjalainen et al., 1996; Meyer et al., 2012), but also by abiotic factors such as testing month, weather during the testing, place of testing, judges, etc. This questioned the methods of evaluating heritability…evaluations of the behavioural traits are often difficult due to the lack of testing repeatability between and also within judges. Performance testing is usually subjective as significantly different scores are given by the judges as shown, for example, in Finnish Spitz (Karjalainen et al., 1996).

For perspective, if we assume a test-retest as much as 0.20, and we correct the Belgian Shepherd Dog mean psychical heritability of 0.13 for the test-retest alone (which is only one form of measurement error), the Spearman correction yields a true heritability of .

Reviews:

• van den Berg 2017, “Genetics of dog behavior”

The dog genetic studies reviewed in this chapter used more subjective phenotypic measures. Most heritability studies used phenotypes based on the behavior of dogs in test batteries. Jones and Gosling (2005) have reviewed studies of canine personality and noted that, “In theory, test batteries were the closest to achieving objectivity, but in practice the levels of objectivity actually attained varied substantially.” The molecular genetic studies mostly used even more subjective measures such as owner-report questionnaires and expert ratings (experts being veterinarians, trainers, or dog obedience judges). Owner and expert ratings may be influenced by a variety of factors other than the behavior of the dog, e.g. owner personality and expectations of typical dog behavior. Intuitively, the use of specific and objective metrics in genetic studies seems preferable. However, behavior of dogs in a test battery may not be representative of their behavior in everyday life and it is often unclear what exactly is being measured. Van den Berg and colleagues used three methods for measuring canine aggressive behavior: a behavioral test of the dog (van den Berg et al ., 2003), a questionnaire for the dog owner (van den Berg et al ., 2006), and a personal interview with the dog owner (van den Berg et al ., 2003 , 2006). The most promising heritability estimates (i.e. high heritability with low standard errors) were obtained for the owner impressions collected during the personal interview (Liinamo et al ., 2007). This is rather surprising because of the subjectivity of these phenotypes. Large coordinated projects, such as the European LUPA consortium, make an effort to clarify dog behavioral phenotypes by following standard procedures to describe dog behavior (Lequarré et al ., 2011). This is of great value for progress in canine behavioral genetics.

• “Canine Behavioral Genetics—A Review”, Mackenzie 1986

Variable Proportion
Posture in Pavlov stand 0.43
Investigative behavior in Pavlov stand 0.46
Escape attempts while in Pavlov stand 0.56
Human avoidance and vocalization at 5 weeks 0.59
Playful fighting at 13–15 weeks 0.42
Leash fighting 0.77
Docility during sit-training 0.48
Running time for long barrier 0.78
Vocalization on U-shaped barrier 0.47

Table 2: Proportion of total variance due to breed differences between Basenjis and Cocker Spaniels (after Scott and Fuller, 1965)

…G. Geiger investigated the breeding-book of Dachshunds in Germany in 1973 and found the scores better distributed than the data studied by Sacher, perhaps due to the 12-point system used as opposed to the 4-point system used in the pointer prize classes. He conducted a three-level nested 379 analysis of variance on 1463 full- and half-sib progeny of 21 sires. In contrast to the earlier findings of Humphrey and Warner (1934), King (1954) and Mahut (1958), his results showed maternal effects but no effect due to sex. The heritabilities are shown in Table III (Geiger, 1973, cited in Pfleiderer-Hogner, 1979).

Trait Sire Dam
Hare tracking 0.03 0.46
Nose 0.01 0.39
Seek 0.00 0.41
Obedience 0.01 0.19

Table 3: Heritability estimates in Dachshunds (after Geiger, 1973)

A second study of additive genetic variation in 1973 came from the Army Dog Training Center in Solleftea, Sweden. C. Reuterwall and N. Ryman reported on their study of 958 German Shepherds from 29 sires. The 8 behavioral traits studied were labeled A-H:

• Trait A was termed “Affability” (tested by having an unknown person con front the dog);
• Trait B was termed “Disposition for Self Defense” (tested by having an unknown person attack the dog);
• Trait C was termed “Disposition for Self Defense and Defense of Handler” (tested by having an unknown person attack the dog and handler);
• Trait D was termed “Disposition for Fighting in a Playful Manner” (tested by asking the dog to fight for a sleeve or stick);
• Trait E was termed “Courage” (tested by having a man-shaped figure approach the dog);
• Trait F was termed “Ability to Meet with Sudden Strong Auditory Disturbance” (tested by firing shots at some distance and making a noise with tin cans just behind the dog);
• Trait G was termed “Disposition for Forgetting Unpleasant Incidents” (tested by scaring the dog at a certain place and then asking the dog to pass the place again);
• Trait H was termed “Adaptiveness to Different Situations and Environments” (tested by observations during the other parts of the test).

In contrast to Geiger’s findings, Reuterwall and Ryman reported significant differences between the sexes, males handling noise (Trait F) better and exhibiting more controlled defense (part of Trait C) and playful fighting (Trait D). Sex differences had also been noted by Humphrey and Warner (1934), King (1954) and Mahut (1958). Reuterwall and Ryman noted that, in all 380 the traits studied, the additive genetic variation was small (Reuterwall and Ryman, 1973). The heritability estimates listed in Table IV were reported by Willis based on the information found in Reuterwall and Ryman (Willis, 1977). It should be noted that the scores used by Reuterwall and Ryman were transformed and extremely complex. Some workers in Sweden today, working on the genetics of the breeding program at the Statens Hundskola, feel that the findings of Reuterwall and Ryman’s study are based on scores too complex to have much meaning (L. Falt, personal communication, 1982).

Trait Males Females
A [Affability] 0.17 0.09
B [Disposition for self-defense] 0.11 0.26
C [Disposition for self-defense and defense of handler] 0.04 0.16
D [Disposition for fighting in a playful manner] 0.16 0.21
E [Courage] 0.05 0.13
F [Ability to meet with sudden strong auditory disturbance] -0.04 0.15
G [Disposition for forgetting unpleasant incidents] 0.10 0.17
H [Adaptiveness to different situations and environments ] 0.00 0.04

Table 4: Heritabilities in German Shepherds (after Reuterwall and Ryman, 1973)

The next year, M.E. Goddard and R.G. Beilharz stated their belief that fearfulness and dog distraction were heritable in Australian guide dogs (Goddard and Beilharz, 1974). In 1982, Goddard and Beilharz reported further on the genetics of Australian guide dogs…Fearfulness emerged as the most important and most highly heritable component of success. Estimates of heritabilities based on scores of 394 Labrador Retrievers computed from sire components, dam components and the two combined are listed in Table V (Goddard and Beilharz, 1982). In contrast to reports by Scott and Bielfelt (1976), Geiger (1973) and Scott and Fuller (1965), no strong maternal effects were evident (Goddard and Beilharz, 1982)

Trait Sire Dam Combined
Success 0.46 0.42 0.44
Fear 0.67 0.25 0.46
Dog distraction -0.04 0.23 0.09
Excitability 0.00 0.17 0.09
Health 0.40 0.10 0.25
Hip dysplasia 0.08 0.20 0.14

Table 5: Heritability estimates in Australian Labradors (after Goddard and Beilharz, 1982)

…Estimates of heritabilities based on scores of 249 Labrador Retrievers, calculated from combined sire and dam components, are listed in Table VI (Goddard and Beilharz, 1983). Nervousness had the highest heritability and was the only trait with a significant sire component. Estimates of genetic correlations between the traits are listed in Table VII (Goddard and Beilharz, 1983). In contrast to other workers (Castleberry et al., 1976; Bartlett, 1976; Rosberg and Olausson, 1976), Goddard and Beilharz (1983) found no negative correlations between important traits. However, they did not list correlations for hip dysplasia. They also noted the importance of sex; females being more fearful and distracted by scents but less aggressive and distracted by dogs than males. Sex differences were also noted by Humphrey and Warner (1934), King {1954), Mahut (1958), Reuterwall and Ryman (1973) and Pfleiderer-Hogner {1979). G. Queinnec, B. Queinnec and R. Darre reported on their work with French racing greyhounds (Queinnec et al., 1974). Breeding values for greyhounds were based 40% on the animal’s own performance and 60% on the performance of its progeny, both over 3 racing seasons to account for repeatability

Trait Heritability
Nervousness (N) 0.58
Suspicion (S) 0.10
Concentration (C) 0.28
Willingness (W) 0.22
Distraction (D) 0.08
Dog distraction (DD) 0.27
Nose distraction (ND) 0.00
Sound-shy (SS) 0.14
Hearing sensitivity (HS) 0.00
Body sensitivity (BS) 0.30

Table 6: Heritability estimates in Australian Labradors (after Goddard and Beilharz, 1983)

In 1975, the U.S. Army Biosensor Project reported a heritability estimate of 0.70 for their intermediate temperament evaluations. They also stated their intention to use heritability estimates of both hip dysplasia (previously estimated in their colony as 0.22) and temperament in the calculation of breeding values (Castleberry et al., 1975). The following year, they reported the first known estimate of the genetic correlation between temperament and hip dysplasia (considered by many to be the two major problems in breeding dogs for military or police work). Before listing the estimate, they noted that previous dysplasia-free litters had shown undesirable temperaments. Their estimate of the phenotypic correlation between the two traits was -0.25 and that of the genetic correlation was -0.35 (Castleberry et al., 1976). In 1976, C.R. Bartlett reported heritabilities and genetic correlations between traits studied in American guide dogs. The traits listed were hip dysplasia, body sensitivity (judged by how hard a jerk on the choke-chain leash the new dog could tolerate; low scores indicating a lack of sensitivity), ear sensitivity (judged by how loud a vocal correction the new dog required; low scores indicating lack of sensitivity), nose (olfactory acuity leading to distraction problems for all but the best trainers; low scores indicating greatest use of the nose), intelligence (the ability of the dog to understand things from its own viewpoint, not implying a willingness to obey; low scores indicating great intelligence, which may be a problem to all but the best trainers), willingness {willingness to do what the dog’s master asks of it, regardless of distractions; low scores indicating the most willing dogs), energy (activity versus laziness; low scores indicating active, energetic dogs), self right (the belief of the dog that it has a right to be where it is; negative scores indicating a tendency to give way to another), confidence (confidence shown with strange people or in strange environments; low scores indicating more confident dogs), fighting instinct (tendency to fight; low positive scores indicating the tendency to avoid fights, negative scores indicating even less tendency to fight, passing into submission) and protective instinct (a desire of the dog to protect its own; low positive scores indicating a dog which will speak if a stranger approaches its master with menace, but will not fight to protect the master). Heritability estimates of these traits, based on over 700 records for males and over 1000 records for females, both calculated by paternal half-sib analysis, are listed in Table VIII (Bartlett, 1976)

Trait Males Females Combined
Hips 0.72 0.46 0.54
Body sensitivity 0.26 0.05 0.10
Ear sensitivity 0.49 0.14 0.25
Nose 0.30 0.05 0.12
Intelligence 0.17 -0.07 -0.06
Willingness -0.14 -0.04 -0.03
Energy -0.03 0.06 0.05
Self-right 0.15 0.25 0.22
Confidence 0.04 0.26 0.16
Fighting instinct -0.05 -0.08 -0.04
Protective instinct -0.21 -0.13 -0.12

Table 8: Heritability estimates in American guide dogs (after Bartlett, 1976)

Rosberg and Olausson reported low heritability estimates for mental traits in the dogs at the Swedish Army Dog Center in Solleftea, Sweden. All dogs included in the study were German Shepherds. Phenotypic correlations between the mental traits they were studying and hip dysplasia were small, but negative. Genetic correlations were negative, ranging up to -0.55, but the authors felt they were unreliable due to problems with the material studied (Rosberg and Olausson, 1976). A study of the genetics of American guide dogs was completed in 1976 by C.J. Pfaffenberger, J.P. Scott, J.L. Fuller, B.E. Ginsburg and S.W. Bielfelt. They followed up Scott and Fuller’s (1965) work in behavior and obtained estimates of heritability for their puppy tests. The traits reported by Scott and Bielfelt (1976} in their chapter on analysis of the puppy-testing program included the following: sit (three repetitions of a forced sit with a vocal command}; come (five repetitions of the handler moving away, kneeling down, calling the puppy by name, followed by the command “come” while clapping the hands); fetch (three repetitions of playful retrieving with vocal command); trained response (a complex score, indicating if the puppy was afraid of the tester or not, was over-excited or cooperated calmly, did or did not pay attention to moving objects, adjusted slowly or readily to the new environment, showed no curiosity or was curious about new objects and people, did or did not remember previous experience, tried to do what the tester wanted or not, and showed persistence or not in performing a task); willing in training (also a complex score, indicating if the puppy was fearful or at ease, afraid to move or moved freely, was indifferent or friendly to the tester, was unresponsive or responsive to encouragement, urinated or was continent, was upset by the new situation or was confident, and was obstinate or willing in its responses); body sensitivity (another complex score, indicating if the puppy stood erect or cowered, turned head away or not, looked at or away from the tester, showed pain by action or not, came back after pain or attempted to escape, tucked in the tail or not, wagged tail or not after pain, and growled or not when in pain); ear sensitivity (similar to body sensitivity, except in relation to sound instead of pain); new-experience response (similar to trained response, but this time an emotional response to novel stimuli, not training); willing in new experience (similar to willing in training, except related to novel stimuli instead of training); traffic (indicates if puppy can avoid a moving and stationary cart without becoming fearful); footing-crossing (indicates if puppy noticed differences in footing between curbs and metal patches in the sidewalk); closeness {how close the puppy passed to obstructions); heel (how well the puppy accepted leash training). Eleven of the 13 traits, whose heritability estimates are listed in Table XI, had dam components much larger than the sire components, indicating strong maternal effects (Scott and Bielfelt, 1976). This agrees with the findings of Scott and Fuller {1965) and Geiger {1973). As part of the same study, J.L. Fuller examined the relationship between physical measurements and behavior. Once again, no substantial correlations were found (Fuller, 1976).

Trait Heritability
Sit 0.06
Come 0.14
Fetch 0.24
Trained response 0.08
Willing in training 0.12
Body sensitivity 0.16
Ear sensitivity 0.00
New-experience response 0.06
Willing new experience 0.24
Traffic 0.12
Footing-crossing 0.06
Closeness 0.04
Heel 0.10

Table 11: Heritability estimates for California guide dogs (after Scott and Bielfelt, 1976)

Comparing Scott and Fuller’s 1965 estimates with those of the U.S. Army Biosensor project (Castleberry et al., 1975), it seems possible that certain components of behavior may be highly heritable. The failure of other workers to find high estimates may indicate that such estimates are quite sensitive to the quality of the tests, size of the samples and statistical methodology.

In 1979, M. Pfleiderer-Hogner estimated heritabilities of Schutzhund scores in Germany. She analyzed 2046 test results in 1291 German Shepherds from 37 sires, all tested animals being born in 1973. The four criteria studied were tracking, obedience, man-work and character. She found sex and the number of dogs competing in a given trial to be significant, but not age or month of trial. Sex differences were previously noted by Humphrey and Warner (1934), King (1954), Mahut (1958) and Reuterwall and Ryman (1973). Estimates of heritabilities from sire components, dam components and their combination are listed in Table XII (Pfleiderer-Hogner, 1979).

Trait Sire Dam Combined
Tracking 0.01 0.20 0.10
Obedience 0.04 0.13 0.09
Man-work 0.04 0.07 0.06
Character 0.05 0.17 0.12

Table 12: Heritability estimates for German Schutzhund scores (after Pfleiderer-Hogner, 1979)

In 1982, L. Falt, L. Swenson and E. Wilsson reported their unpublished work on heritability estimates for behavioral traits studied at the National Dog School (Statens Hundskola) in Solleftea, Sweden. [Falt, L., Swenson, L. and Wilsson, E., 1982. “Mentalbeskrivning av valpar. Battre Tjanstehundar, Projektrapport II”. Statens Hundskola, Sveriges Lantbruksuniversitet and Stockholms Universitet. Unpublished.] The traits studied in 8-week-old German Shepherd puppies included: yelp (time from first separation from litter to first distress call); shriek (time from the same separation to the first serious, emphatic distress call); contact 1 (tendency to approach a strange person in a strange place after separation); fetch (pursue a ball and pick it up in the mouth); retrieve (bringing the ball back after picking it up); 389 reaction (to a strange object in a strange place); social competition (actually a form of tug-of-war); activity (number of squares entered when left in a marked arena); contact 2 (time spent near a strange person sitting passively in a chair in the middle of the marked arena); exploratory behavior (number of visits to strange objects placed in the corners of the marked arena). Estimates of heritabilities for the traits, calculated from sire components and dam components separately, are listed in Table XIII (Falt et al., 1982). Although some specific behaviors had low heritability estimates, others had quite high estimates.

Trait Sire Dam
Yelp 0.66 0.73
Shriek 0.22 0.71
Contact 1 0.77 1.01
Fetch 0.73 0.10
Retrieve 0.19 0.51
Reaction 0.09 1.06
Social competition 0.11 0.76
Activity 0.43 0.76
Contact 2 0.05 1.11
Exploratory behavior 0.31 0.83

Table 13: Heritability estimates for Swedish German Shepherds (after Felt et at., 1982)

…They felt that improved training and upbringing were as important as genetics in producing good behavior. Since the first-generation hybrids performed better than either of their pure-bred parents in problem-solving situations, Scott and Fuller recommended that cross-breds be considered as working dogs, provided that the pure-bred lines were properly maintained. Maintenance of the pure-bred lines seems important since they stated that the heterosis (hybrid vigor) lasted only for one generation. Consequently, inter-breeding of the hybrids should not result in any improvement in problem-solving ability. They also recommended against breeding one champion sire to many bitches, since they felt that good breeding programs need to consider multiple criteria to be effective (Scott and Fuller, 1965).

## NBA Screening Scenario

Analogous to the dog cloning scenario, I consider the case of selecting for extremes on PGSes, motivated by a scenario of scouting tall men for the NBA.

Setting up the NBA selection problem as a liability threshold model with current height PGSes as a noisy predictor, height selection can be modeled as selecting for extremes on a PGS which is regressed back to the mean to yield expected adult height, and probability of being tall enough to consider a NBA career.

Filling in reasonable values, nontrivial numbers of tall people can be found by genomic screening with a current PGS, and as PGSes approach their predictive upper bound (derived from whole-genome-based heritability estimates of height), selection is capable of selecting almost all tall people by taking the top PGS percentile.

The selection problem considered above is fairly generic. The topic of ranking & selection based on a noisy predictor can be illustrated by considering a similar scenario.

### Genomic Prediction of Height

Can height be predicted? Yes: one of the best-performing PGSes for a common human complex traits as of 2019 is height PGSes, which has leapt from ~20% in 2014 to predicting ~40–42% of variance or r=0.65 in 2018 (Qian et al 2019/Lello et al 2017/Marquez-Luna et al 2018). Nor have current height PGSes have not hit a ceiling yet. Whole genome data indicates (Wainschtein et al 2019), using GCTA-like heritability techniques, that the pedigree-based heritability estimates of height are substantially correct and that whole genome sequencing will enable GWASes to eventually predict up to 79% variance or r=0.89 for height.21

Can extreme height be predicted? Also yes—while the extremes of human height can be caused by diseases, and can be affected by rare mutations, particularly for shortness, for the most part, it and tallness in particular are caused by common genetic variants and thus extremely tall people have elevated height PGS scores (Chan et al 2011, Liu et al 2013, Sexton et al 2018; the last provides the specific case of Shawn Bradley, who is +4.2SD on the 2014 height PGS). So, one can predict with considerable probability individuals of extreme height by looking for sufficiently extreme height PGSes.

Can extreme height of everyone be predicted? Yes, because eventually, everyone will be genetically sequenced. 23andMe and UK BioBank and Japan Biobank and All Of Us are only the start. The cost of SNP genotyping as of 2019 is far too low to not, as it would cost perhaps $20 in bulk (about what the UKBB paid years ago), and it is profitable considering only pharmacogenomics (Verbelen et al 2016, Chanfreau-Coffinier et al 2019) and treatment of rare monogenic diseases (only partially solved by the universal use of heel prick tests22) and screening newborns (McRae et al 2017, Farnaes et al 2018, Ceyhan-Birsoy et al 2019, Sharp et al 2019), never mind the benefits to research (which will drive further progress & make it even more cost-effective through experience curve effects and better PGSes) or lifelong uses such as CAD/T2D/IBD/breast cancer/stroke (Torkamani et al 2019, Khera et al 2018, Inouye et al 2018, Mavaddat 2019, Abraham et al 2019) prediction & prevention. So height will be predictable for everyone (including the deceased). ### Height in NBA Basketball “I’ll check up on anyone over 7 feet that’s breathing.” Ryan Blake, NBA scout23 Height is a guinea of human genetics because it is so visible, easily measured, highly heritable, and yet not a simple Mendelian trait but highly polygenic. Height is also interesting because a major professional sport, basketball, depends on the specific trait of height to an unusual extent–no other major sport (like soccer, baseball, football, or cricket) depends on a single physical highly-heritable highly-predictable trait the way basketball does. NBA players as of 2018 average 6 foot 7 inch (2.01 m), after a historical growth spurt which saw player heights grow enormously (stagnating in the past 2 decades, perhaps because they ran out of tall people and have had to emphasize athleticism & speed more than big men, although taller players are still paid more). The shortest player in the entire NBA as of 2018 is Isaiah Thomas at 5 foot 9 inch (1.75 m), and the shortest player ever was Muggsy Bogues, 5 foot 3 inch (1980s–1990s). (Graphing the current distribution of NBA player heights, it is fairly normal looking, but appears truncated at 6 foot and a bit right-skewed.) The importance of height to NBA entrance is demonstrated by the extreme rarity of NBA-like height. The US adult male population has a mean height of 69.2 (2.98SD) inches24, while NBA players are almost a foot higher at a mean height of 79 inches, putting the average NBA player at fully +3.29SD in height25. There are perhaps <80,000 men in the USA >=3.29SD. (Sports journalist Pablo S. Torre has famously argued that “while the probability of, say, an American between 6’6” and 6’8" being an NBA player today stands at a mere 0.07%, it’s a staggering 17% for someone 7 feet or taller.") And there were 3,853,472 babies total born in the USA in 2018 or ~1926736 male babies, so similarly, there are perhaps ~1000 male babies each year who will grow up to be >=3.29SD or the top 0.05% in height26, and perhaps ~25 with a PGS as extreme as Shawn Bradley27. Somewhat more realistic would be to ask what threshold a mean of +3.29 is; by the truncated normal28, that’s a threshold of +3.01SD or the top 0.13%, implying more like 2500 male babies per year, still a fairly small number. ### Height as Screening Problem A NBA or college basketball recruiter might be quite interested in knowing who those 2500 are—height is certainly not the only determinant of success, one has to at least want to play, but such a height would be a huge help in becoming a NBA player. Getting to potential recruits as early as possible could help develop an interest in basketball, accelerate their career, deal with rough patches, or just make them more attached to a particular college or team. Regardless of how plausible this particular scenario is, is it at least statistically possible? Can extremes in adult height, given the base rates & the <100% heritability of height & r=0.65 PGS, be predicted accurately enough to be plausibly useful for screening? So conceptually the model is: a large sample of normal variables are generated (the PGS), then the top n% are selected; this creates a new distribution of truncated normal variables with a much higher mean, yielding a certain boost in SD (like +4SD), but then to predict the correlated variable (adult height), it must then be regressed back to the mean due to the r<1 correlation of the 2 variables (error in the PGS), yielding a smaller outcome boost on the correlated variable (like +2SD); with the correct boost estimated, the probability any of the top n% will pass an additional threshold (eg NBA height thresholds at +3.01SD) can then be calculated. (Trickily, the distribution of samples after selection is not merely a normal distribution shifted higher, but also has a different, smaller, SD, so that must be adjusted for as well as the higher mean.) With the probability of success conditional on selecting the top n% with a PGS estimated, the total number of successful selected candidates be inferred from the total population and compared with the estimated number of all successful candidates to give an idea of screening efficiency. The dog cloning approach can be partially reused here: the ‘heritability’ is 0.42 (variance of PGS with adult height), the global success rate is set by +3.01SD, and we want to know what pre-screening must be applied for a reasonable probability of a candidate succeeding. A reasonable value here might be 10%: a recruiter isn’t investing that much time in each possible recruit, but at 1% they’d be wasting a lot of their time, but 10% seems like a reasonable value to look at. ### Model Implementing the necessary truncated normal apparatus (exact implementation & a Monte Carlo implementation to check): ## can check with Monte Carlo and against etruncnorm & vtruncnorm in ## 'truncnorm' package: https://cran.r-project.org/web/packages/truncnorm/truncnorm.pdf truncNormMean <- function(a, mu=0, sigma=1, b=Inf) { phi <- dnorm erf <- function(x) 2 * pnorm(x * sqrt(2)) - 1 Phi <- function(x) { 0.5 * (1 + erf(x/sqrt(2))) } Z <- function(beta, alpha) { Phi(beta) - Phi(alpha) } alpha = (a-mu)/sigma; beta = (b-mu)/sigma return( (phi(alpha) - phi(beta)) / Z(beta, alpha) ) } truncNormMeanMC <- function(a, mu=0, sigma=1, b=Inf, iters=1000000) { mean(Filter(function(x){x>a && x<b}, rnorm(iters, mean=mu, sd=sigma))) } truncNormMean(1) # [1] 1.52513528 truncNormMeanMC(1) # [1] 1.52510301 library(truncnorm) etruncnorm(1) # [1] 1.52513528 truncNormSD <- function(a, mu=0, sigma=1, b=Inf) { phi <- dnorm erf <- function(x) 2 * pnorm(x * sqrt(2)) - 1 Phi <- function(x) { 0.5 * (1 + erf(x/sqrt(2))) } Z <- function(beta, alpha) { Phi(beta) - Phi(alpha) } alpha = (a-mu)/sigma; beta = (b-mu)/sigma sqrt(1 + ((alpha * phi(alpha)) - phi(beta))/Z(beta,alpha) - (((phi(alpha)-phi(beta))/Z(beta,alpha))^2)) } truncNormSDMC <- function(a, mu=0, sigma=1, b=Inf, iters=1000000) { sd(Filter(function(x){ x>=a & x <=b }, rnorm(iters, mean=mu, sd=sigma))) } truncNormSD(1) # [1] 0.446203614 truncNormSDMC(1) # [1] 0.445727302 sqrt(vtruncnorm(a=1)) # [1] 0.446203614 Model the selection problem: doubleSelection <- function(successP=0.01, preThreshold=0.01, heritability=0.42, verbose=FALSE) { r <- sqrt(heritability) threshold <- qnorm(1-preThreshold) meanPost <- truncNormMean(threshold) * r sdPost <- sqrt(truncNormSD(threshold)) successThreshold <- qnorm(1-successP) p <- pnorm(meanPost - successThreshold, sd=sdPost) if (verbose) { print(c(r, threshold, meanPost, sdPost, successThreshold, p)) } p } doubleSelectionMC <- function(successP=0.01, preThreshold=0.01, heritability=0.42, iters=100000000, verbose=FALSE) { threshold <- qnorm(1-preThreshold) successThreshold <- qnorm(1-successP) r <- sqrt(heritability) library(MASS) data = mvrnorm(n=iters, mu=c(0, 0), Sigma=matrix(c(1, r, r, 1), nrow=2)) screen <- data[data[,1] >= threshold,] successful <- screen[screen[,2] >= successThreshold,] p <- nrow(successful) / nrow(screen) ## Skim to visualize non-normal post-selection distributions: if (verbose) { library(skimr); print(skim(screen[,1])); print(skim(screen[,2])) print(skim(successful[,2])); } p } doubleSelection(successP=(1-pnorm(3.01)), preThreshold=0.001, heritability=0.79, verbose=TRUE) # [1] 0.888819442 3.090232306 2.992735123 0.510269172 3.010000000 0.486504425 # [1] 0.486504425 doubleSelectionMC(successP=(1-pnorm(3.01)), preThreshold=0.001, heritability=0.79, verbose=TRUE) # variable missing complete n mean sd p0 p25 p50 p75 p100 hist # screen[, 1] 0 100076 100076 3.37 0.26 3.09 3.17 3.29 3.48 5.57 ▇▃▁▁▁▁▁▁ # screen[, 2] 0 100076 100076 2.99 0.52 0.85 2.64 2.98 3.33 5.7 ▁▁▅▇▅▁▁▁ # successful[, 2] 0 47745 47745 3.42 0.32 3.01 3.17 3.35 3.6 5.7 ▇▅▂▁▁▁▁▁ # [1] 0.477087414 Also of interest is calculating how many successful tall adults will be found in the screen, by applying the pre-selection & post-selection probability to the total sample: net <- function(pre, h2) { p <- doubleSelection(successP=(1-pnorm(3.01)), preThreshold=pre, heritability=h2) population <- 1926736 selected <- population * pre successful <- selected * p print(round(digits=4, p)); print(round(c(selected, successful))) return(c(p, successful)) } ## Example: with the top 1%, we will get ~207 tall adults (out of the 2500): net(0.01, 0.42) # [1] 0.0107 # [1] 19267 207 ## The tradeoff between spreading a wide net and catching as many as possible: as we spread a wider net, ## we catch more total, but each becomes much less likely to succeed. eg AUC/ROC curves round(digits=3, sapply(1/seq(from=1, to=100000, by=1000), function(pre) { net(pre, 0.42)}[1])) # [1] NaN 0.052 0.079 0.098 0.114 0.127 0.139 0.150 0.159 0.168 0.176 0.184 0.191 0.197 0.203 0.209 0.215 0.220 0.226 0.231 0.235 0.240 0.244 0.248 0.252 0.256 # [27] 0.260 0.264 0.268 0.271 0.274 0.278 0.281 0.284 0.287 0.290 0.293 0.296 0.299 0.301 0.304 0.307 0.309 0.312 0.314 0.317 0.319 0.321 0.324 0.326 0.328 0.330 # [53] 0.332 0.334 0.336 0.338 0.340 0.342 0.344 0.346 0.348 0.350 0.352 0.353 0.355 0.357 0.359 0.360 0.362 0.364 0.365 0.367 0.368 0.370 0.372 0.373 0.375 0.376 # [79] 0.378 0.379 0.381 0.382 0.383 0.385 0.386 0.387 0.389 0.390 0.391 0.393 0.394 0.395 0.397 0.398 0.399 0.400 0.402 0.403 0.404 0.405 round(sapply(1/seq(from=1, to=100000, by=1000), function(pre) { net(pre, 0.42)}[2])) # [1] NaN 101 76 63 55 49 45 41 38 36 34 32 31 29 28 27 26 25 24 23 23 22 21 21 20 20 19 19 18 18 18 17 17 17 16 16 16 15 15 15 # [41] 15 14 14 14 14 14 13 13 13 13 13 12 12 12 12 12 12 12 11 11 11 11 11 11 11 11 10 10 10 10 10 10 10 10 10 10 9 9 9 9 # [81] 9 9 9 9 9 9 9 9 9 9 8 8 8 8 8 8 8 8 8 8 ### Scenarios It turns out that examining the top 0.3% (n=600) by PGS2018 is enough to enrich candidates to a 10% probability of being tall enough for the NBA; combined, that implies ~60 tall people per year, and broadening to n=10,000 will select ~3x more, 175 tall people: net(0.0003, 0.42) # [1] 0.1036 # [1] 578 60 net(10000 / 1926736, 0.42) # [1] 0.0175 # [1] 10000 175 Screening 10,000 people is not unrealistic, and a payoff of 175 tall people is a potentially worthwhile one. Considering potential further improvements, as the PGS approaches the WGS upper bound of 79%, with the top 0.3%, the yield boosts to ~408: net(0.0003, 0.79) # [1] 0.7064 # [1] 578 408 With such a predictor, one might want to cast a wider net; going back to a 10% success probability, with the optimal predictor, one would be able to recover essentially all tall people per year by taking the top 1.3% (n=25,000): net(0.013, 0.79) # [1] 0.1008 # [1] 25048 2525 (At that point, uncaught tall people would be the exceptional cases: those not included in the screen to begin with, those who are tall because of diseases/environmental factors, those with novel de novo mutations not previously identified, etc.) If a selection sample of n=25,000 is too large despite being comprehensive, a sample 10x smaller (n=2,500) will still recover about half the tall people by taking the top 0.13%: net(2500 / 1926736, 0.79) # [1] 0.4367 # [1] 2500 1092 1. In 2015, plans were announced for a Tianjin cloning center aiming for 100,000 clones a year—while its present status is unclear and as of 2017 it “seems to be well behind schedule”, it shows the ambitions.↩︎ 2. For example, in apple breeding, there are so many seedlings and so few apple tasters, and the goal is to select apples so superior that they can potentially compete with existing commercial varieties (which have been selected out of countless millions of apple trees, on net, over the past few centuries), that the initial screening is a single bite of a single apple from a single seedling for several years in a row with no second chances—and only then can the real testing begin.↩︎ 3. The new bottleneck presumably becomes the human trainers, but they can give more intensive training or participate in research programs in peacetime to keep their skills sharp & train the next generation of human trainers, and given the popularity of ‘Schutzhund’ and ‘executive protection dogs’ among civilians, there may be enough civilian demand for trained dogs to maintain a reserve of trainers. This is probably a moot point for the USA, as the global ‘War on Terror’ and demand for SF dogs shows little sign of slackening soon.↩︎ 4. "“We made 49 because we were curious about the smallness,” explains Jeong, the head researcher. “Would it transfer?” He shakes his head. “It didn’t—the clones turned out bigger.”" The owner sued H Bion/Sooam in March 2019, claiming that only 10 clones were supposed to be made and she has accused them of lying about the results to reuse them in micro-pigs and other projects without royalties (case #6:2019cv00425; complaint).↩︎ 5. Choi et al 2013: Six cloned dogs that finished the training course were evaluated by a final drug-detection dog selection test and all of them passed. The pass level was a score of 60. [The cloned dog] To-Tue was graded as Excellent (score 90) and the remaining five dogs were evaluated as Good. In age matched-controls, seven puppies finished the training course and one of them passed the test. One of the eight puppies died before the training course was over. The pass rate of cloned dogs was 86% since six puppies passed among seven cloned ones. That of controls was 13% in the aggregate since one passed among eight control ones. This value was lower than generally found as 30% (Maejima et al. 2007) or 50% (Weiss and Greenberg 1997). ↩︎ 6. Interestingly, there was a previous successful cloned dog, in 2000, by the Missyplicity project (funded by a donation to Texas A&M University of$2,300,000)—but it was stillborn and not widely reported.↩︎

7. The citations are to interviews Frost conducted & summarized:

• Andersen, Gary, LTC. Telephone interview, 28 November 1969. LTC Andersen stated that the rejection rate at procurement of MWDs is approximately 50%, due to medical problems (primarily hip dysplasia). The washout rate for a patrol dog is about 16%. He also mentioned that this rate used to be 5% in the 70’s. He did not know why the rejection rates were higher today, since the quality of the dogs is better. He said that it could be that not enough time is spent trying to get the ‘slower’ dogs to pass the training course (Frost, Parks). He said he felt the high rejection rate was one reason for the large backlog of requisitions. He also said there are adequate numbers of trainers, but there is no apparent formalized certification process. When asked about command and control of the MWD Program, LTC Andersen said he understood it, but it was very confusing to people outside the program, and that it is largely ineffective in function since there is no one central manager for all facets of the MWD Program (Burwell, Parks, Stamp).
• Craig, Dan. Telephone interview, 5 December 1989. Dr. Craig said that all dogs procured for the MWD Program are bought for detector dogs, and those that wash out are trained as patrol dogs…He said that the rejection rate for Patrol/Explosive Dogs is 43%, while that for Patrol/Drug Dogs is 29% (Andersen; Burwell; McCathern, Telephone interview; Taylor, E.).
• McCathern, Marge. Telephone interview, 1 December 1989. While discussing rejection rates in the MWD Program, Ms. McCathern said the rejection rate for the new explosive course was 83%, and that it differs for each course. She also said the rejection rate at procurement was around 50%. She advised the author to talk to Dr. Craig for the average costs involved in training each specialty of MWD.
• Thorton, William H., LTC. “The Role of Military Working Dogs in Low Intensity Conflict”. Army-Air Force Center for Low Intensity Conflict. This paper discusses the historical and current roles of the MWD, and presents reasoning for the need to expand the role of the MWD. Such reasoning centers around economy of force, low technology, high capability, operational flexibility of the MWD, and the need for wider use of the MWDs capabilities other than as a law enforcement asset. Problems with the current MWD system are presented. The paper states that 98% of all dogs procured by DODDC come from Europe, that 45% are rejected after training, and that there is a backlog of 430 requisitions of MWDs (Taylor, E.).
↩︎
8. The citations from Sinn et al 2010:

• All dogs that did not fulfil the demands of a service dog were disqualified and were sold as companion animals or donated to the puppy walkers who had previously had cared for them. About 50% of the dogs selected for training were disqualified during the training period. After completion of the training program all dogs were finally run through a battery of working tests performed by the buyer of the dogs (the police force, guide dog association). In this work only the dogs that passed this final test are referred to as service dogs.

• Slabbert & Odendaal 1999, “Early prediction of adult police dog efficiency—a longitudinal study”:

Up to 70% of dogs that were bred at the South African Police Service Dog Breeding Centre (SAPSDBC) were not suitable for use.

• In Japan, approximately 30% of dogs that enter training programs to become drug detection dogs successfully complete training.

↩︎
9. Trident K9 Warriors, Ritland & Brozek 2013:

…As rare as it is to find a dog with the kind of prey drive that we seek, it is equally difficult to find a dog with the kind of nose that will help it succeed as a working dog with the SEAL Teams. Finding a dog with both those qualities is truly a one-in-a-thousand (or more) proposition. That’s where good breeding comes in, of course, and selecting for both those traits will invariably produce dogs that are stronger in one area over another…The last quality that I look for is difficult to describe in delicate terms. A dog has to have a big set of nuts on him—metaphorically speaking. Most dogs, even among those selected from the elite breeders from around the world, don’t have the kind of dominance and true forward aggression that is needed. Dogs have been domesticated and bred for so long that the type of dog that is willing to stand up to and fight a human—a human that is not frightened by that dog and physically capable of disabling that dog—is a very, very rare animal. I call them the 1 percenters (this was before the term had a political connotation), but they are more like one in ten thousand.

…It is also important to understand that when I acquire a dog from a breeder of Malinois, I’m not getting a very young puppy who hasn’t been trained at all. The two-to-three-year-old dogs have already gone through rigorous training; some even have become what is referred to as a “titled dog.” That means that they’ve been trained and have earned a certification in one of several different European dog sports. One of the more common types of those is Schutzhund, a dog sport popular in Germany. When this sport was first organized and the competitions formalized, a dog that had completed Schutzhund training and became certified in the sport was also essentially qualified to be a German police dog. That was the original intent of the program, but between politics and hurt feelings, the dogs that earn the “title” don’t necessarily have the competency to become actual working police dogs. The sport is so popular that other breeds of dogs now can enter into the competitions.

…Again, that comparison between the human members of the SEAL Teams and their canine co-workers applies. No one goes into the SEAL Teams without first completing basic training and then one additional level before starting BUD/S training. While there is a 75% attrition rate among those entering BUD/S, we don’t have that great a failure rate among the dogs. I haven’t kept statistics to track that rate among the dogs we acquire, but it is more like three or four in ten instead of seven and a half out of ten.

Part of the reason for that is that the early weeding-out process among the dogs is more vigorous than it is among the sailors. As I stated earlier, I felt the first test I had to pass to qualify as a SEAL Team candidate wasn’t very hard at all. When I’m evaluating prospective team dogs, my standards are much higher. In addition, when we select sires and bitches for breeding, we already have in mind the kinds of work that these dogs will be asked to do. As a consequence, we breed for those qualities, and from the moment those dogs are born—to be more precise, in the first several days of their lives—I’m already beginning their training.

↩︎
10. "“They’ve been performing excellent. They’re exactly like the original one,” he said in a telephone interview. “I can say it absolutely does work, and we have been able to create the same dog with the same qualities.”…Brannon, who also trains dogs for police departments around the U.S. as well as the military, said he was skeptical about cloning in the beginning but is now convinced it is more efficient than natural-breeding programs. He’s expecting another clone next year—this one the twin of a dog that has helped agents find millions of dollars in narcotics and apprehend many suspects."↩︎

11. MacLean et al 2019:

We assessed the heritability(h2) of 14 behavioral traits (Fig 1) measured by the Canine Behavioral Assessment and Research Questionnaire (C-BARQ), a well-validated instrument for quantifying diverse aspects of dog behavior (17, 18, 19, 20), including aggression, fear, trainability, attachment, and predatory chasing behaviors. We combined behavioral data from 14,020 individual dogs with breed-level genetic identity-by-state (IBS) estimates…Using a mixed-effects modeling approach (Efficient Mixed-Model Association; EMMA) to control for relatedness between breeds, we found that a large proportion of variance in dog behavior is attributable to genetic factors (Fig 1). The mean heritability was 0.51 ± 0.12 (SD) across all 14 traits (range: h2 0.27–0.77), and significantly higher than the null expectation in all cases (permutation tests, p < 0.001).

These estimates are also significantly higher than those identified in previous studies assessing heritability of these traits in large within-breed samples (t13 = -12.25, p < 0.001; 22, but see 23). Estimating between-breed variance thus yields h2 estimates that are on average, five times higher (range= 1.3–25.5 times higher), which is likely due to more variance among, compared to within breeds. Interestingly, the traits with the highest heritability were trainability (h2 = 0.73), stranger-directed aggression (h2 = 0.68), chasing (h2 = 0.62) and attachment and attention seeking (h2 = 0.56), which is consistent with the hypothesis that these behaviors have been important targets of selection during the cultivation of modern breeds.

↩︎
12. Shockley 1957 provides an analogous example: scientific productivity.

Some researchers are orders of magnitude more prolific and successful than others. Under a normal distribution conceptualization of scientific talent, this would be odd & require them to be many standard deviations beyond the norm on some ‘output’ variable. Shockley suggests that this isn’t so surprising if we imagine scientific research as more of a ‘pipeline’: a scientist has ideas, which feeds into background research, which feeds into a series of experiments, which feeds into writing up papers, then getting them published, then influencing other scientists, then back to getting ideas.

Each step is a different skill, which is plausibly normally-distributed, but each step relies on the output of a previous step: you can’t experiment on non-existent ideas, and you can only publish on that which you experimented on, etc. Few people have an impact by simply having a fabulous idea if they can’t be bothered to write it down. (Consider how much more impact Claude Shannon, Euler, Ramanujan, or Gauss would have had if they had published more than they did.) So if one researcher is merely somewhat better than average at each step, they may wind up having a far larger output of important work than a researcher who is exactly average at each step.

If SF dogs are similar, then there could be dogs which are orders of magnitude better than others, and this could stem from small advantages over competitors at each step; so small heritabilities producing small gains could still produce large output gains as long as many steps are being improved simultaneously.↩︎

13. Choi 2018 doesn’t seem to be using the same dataset as Choi et al 2014, because in Choi et al 2014, there were 8 clones and in Choi 2018 there were 6 clones; and in the latter, control puppies passed quarantine training.↩︎

14. Numbers >500 often come up in Sooam/police articles, and the sniffer article specifies that it has 42 currently and that 50% were clones by 2014 from the initial police clones somewhere ~2008; given substantial turnover in dogs with careers <10 years and an 80% success rate, that implies at least hundreds have entered sniffer training total.↩︎

15. “The DoD MWD Trainer/Supervisor Course provides kennel masters and trainers with the skills to enhance their MWD program. The course includes instruction in kennel management, administration, dog team training, and contemporary employment concepts” (Briefing by LTC Bannister the commander of the 341st Training Squadron, on September 7, 2005).

The DoD MWD Course [which produces the trained dogs] provides both patrol and dual certified patrol/detector dogs [Cost is about $50,000 per trained dog.] The course is 120-days long. The dogs are trained in either drug or explosive detection. The dogs are trained to detect marijuana, hashish, heroin, and cocaine and must meet a 90% accuracy standard to certify. Explosive detector dogs are trained to detect seven explosive substances (smokeless powder, nitro dynamite, ammonia dynamite, TNT, C-4, water gel, and det cord) and two chemical compounds (sodium and potassium chlorate) and must meet a strict 95% standard (briefing by LTC Bannister, commander of the 341st Training Squadron, on September 7, 2005). The 341st is a major training location as of 2011: With a second kennel facility located on Medina Annex about a mile away, Lackland AFB has approximately 900 dogs at any given time. The squadron’s school trains about 270 multipurpose dogs a year, according to school officials. Not only does the school train new dogs, but it trains handlers and supervisors as well… To keep up with the demand for trained dogs, the school uses a variety of procurement methods, including its own breeding program. The suitability rate runs around 50 percent. In other words, to produce 100 serviceable dogs per year, the program will attempt to train about 200. ↩︎ 16. Trident K9 Warriors, Ritland & Brozek 2013: …No matter that the navy had invested more than$50,000 in the acquisition, training, and care of Duco before Seth spent that year in our program pre-deployment, Duco was still “his.” That was as it should be; unfortunately, it isn’t always. I’ve trained hundreds of dogs for a variety of purposes, and it’s not always easy to let them go to another home, especially a quality dog like Duco. Training dogs to be of service to us is my job, and it’s also my passion. Seeing how a pair like Seth and Duco continue to operate does my heart good.

↩︎
17. The U.S. military spends up to $283,000 to train a working war dog. Once it has a promising pup, the Pentagon spends an additional$42,000 to train a K9 unit, a process that starts with obedience and drug and/or bomb detection at Lackland Air Force Base in San Antonio, Texas. Some of the dogs get a second round of training in how to patrol, detain an enemy and attack. A “dual-purpose” dog spends about 120 days completing both training cycles.

When all is said and done, a fully trained military dog costs about as much as a small missile. Keeping them in the field as long as possible is increasingly good business. (The Air Force declined to discuss canine casualty rates.)

↩︎
18. pg37, Dog, Inc.: The Uncanny Inside Story of Cloning Man’s Best Friend, Woestendiek 2010.↩︎

19. Perhaps by combining noisy ratings like training records & completed service records & handler ratings & general surveys, and selecting the best out of 10,000+ candidates; since there’s no available data for even speculating about what are plausible r values for such a procedure or what the total number of candidates might be if a government like the US federal government made a serious effort to screen all of its & available allies’ SF dogs, let’s consider ‘1 of 1000’ as something of a lower bound—there are surely tens of thousands of SF dogs available now & in decades to come, and a SF cloning program should be able to at least select from the top few hundred of those SF dogs. (Bloomberg 2017, of the US alone: “At the moment, roughly 1,600 Military War Dogs (MWDs) are either in the field or helping recuperating veterans.”; NYT 2011, 2700.)↩︎

20. The MC indicates that the exact implementation is slightly wrong, off by a relative -1%; I have not been able to figure out why the exact implementation is conservative but it probably has something to do with the variance of selected individuals being too small. So it is slightly biased against cloning efficacy.↩︎

21. Reaching such a predictor will be extremely difficult in the near-term, but it’s worth remembering that pedigree & GWAS are not mutually exclusive approaches, and predicting solely from SNPs is not the only nor the best possible way to do genomic prediction. GWASes can be extended to the mixed model approach, where genetic relatedness to other individuals of known phenotype is used for prediction. Explicitly modeling relatedness of individuals is a powerful method of prediction (if someone is more genetically similar to their tall paternal grandfather than their tall maternal grandmother, you can predict they will be taller without knowing any specific SNPs), and can be combined with PGS-based approaches based on individual SNPs. The BLUP framework, for example, is widely used in agriculture for much better predictions than possible with merely individual SNPs (Gianola & Rosa 2015; in GWAS, eg Loh et al 2018). Height is typically measured by biobanks, so as they get larger, use of SNP or WGS data for inferring genetic relatedness will become more feasible.↩︎

22. Which, incidentally, are often kept in storage and would allow overnight sequencing of a large fraction of the population. Every baby in California since 1983 has blood spots stored, which could be used for genome sequencing.↩︎

23. As quoted by Pablo S. Torre, 2011.↩︎

24. According to Table 12 of “Anthropometric Reference Data for Children and Adults: United States, 2011—2014”, so:

• all races: 69.2 inches
• 95th percentile: 74.1 inches
• 95th percentile = +1.64 SD
• SD = 2.98
↩︎
25. NBA mean of 6 foot 7 inches = 79 inches = = +3.29SD.↩︎

26. (1-pnorm(3.29))*1926736↩︎

27. 1926736 * (1 - pnorm(4.2))↩︎

28. truncNormMean(3.01) ~> 3.923↩︎