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.
genetics, decision-theory, R, bibliography, order-statistics
2018-09-182019-08-24 finished certainty: possible importance: 3


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 with cur­rent & future PGSes.

and 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 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 , , , and pre­ferred for war dogs.

The 7 Tomor­row Dogs/Toppies cloned in 2007.

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 , 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 to a cat can have 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//); 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 & —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: scales poorly with increas­ing vari­ance; rel­a­tively small increases in noise can require much larger n to over­come. The most effi­cient exper­i­ments are , 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 , 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 Greater 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 + on a ran­dom SF dog would imply.
    • elite clones can be Heav­ily Used 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 , and the record for num­ber of clones appears to be the 49 clones of the world’s tini­est dog, Mir­a­cle Milly.4
  6. : 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 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 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 (“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 where a sin­gle indi­vid­ual is used as the ‘par­ent’ of the ‘next gen­er­a­tion’. (This is not a , because the clone is differ­ent from the selected donor indi­vid­u­al, and is a draw from a new dis­tri­b­u­tion.)

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

The answers seem to be:

  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 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”.

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 $24,291US 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% suc­cess rates.

A trainer at the USDA National Detec­tor Dog Train­ing Cen­ter in 2019 described screen­ing the gen­eral pop­u­la­tion of dogs for can­di­dates: “We could look at 100 dogs and not come back with any…once they go through ini­tial test­ing, the per­cent­age of those dogs that make it is maybe 70%.”

Clone Success Rates

One of the first work­ing dogs cloned was a par­tic­u­larly famous Cana­dian police dog , 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 7⁄7 suc­cess rate for the clones vs 3⁄10 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 3⁄3 suc­cess rate (as opposed to a more typ­i­cal 4⁄8 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 0⁄3 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%).

The first (and as of August 2019, only?) Chi­nese cloned police dog, Kunx­un, was report­edly suc­cess­ful in train­ing & accepted for duty. Another 6 dog clones began police train­ing in Novem­ber 2019.

As of late 2019, K9 dog clone suc­cess rates are appar­ently high enough to allow one K9 dog breeder to offer “a bet­ter, a five-year war­ranty instead of a sin­gle-year war­ran­ty, which is offered by the other ken­nel”; and an ex-Navy Seal try­ing to launch a busi­ness for pro­vid­ing guard dogs in large vol­ume for schools to guard against mass shoot­ings, Joshua Mor­ton, trains only clones (a team of dog & han­dler is $125,000/year), due to the reli­a­bil­ity of cloning & train­ing a par­tic­u­lar dog: “Cloning allows me to be con­sis­tent. Now, I know that I can tell a client, ‘Hey, I’ll have this dog ready in nine months.’…It’s way more effec­tive, way more effi­cient.”

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: , MacLean et al 201911; , 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 trait anal­o­gous to the , 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: /).

Sooam has pub­lished behav­ioral research on cloned dogs: Kim et al 2018/Lee 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 s 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 (some­times by train­ers who used to train for Spe­cial Forces), can sell for $51,643$77,465 with Rit­land sell­ing his dogs at $50,000–$100,000, and the best award-win­ning “exec­u­tive pro­tec­tion dogs” sell­ing for up to $296,947. 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 $38,696 (Lem­ish 1996/1999, War Dogs: A His­tory of Loy­alty and Hero­ism) and $58,044 (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 $70,593 “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 ~$24,291, with a 21% total fail­ure rate imply­ing a 1.2x higher cost per suc­cess of ~$30,747. A 2011 NYT arti­cle on Marines notes it is “an expen­sively trained canine (the cost to the Amer­i­can mil­i­tary can be as high as $51,643 per dog)”. Rit­land & Brozek 2013 quotes a US Navy SEAL dog’s indi­vid­ual cost at >$62,414.16 South Korean police quote a drug sniffer dog at $51,643 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; the clone was report­edly suc­cess­ful in train­ing. 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 1⁄6. 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 >$62,414, 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

Dog cloning, on the other hand, cost $100,000 list price in 2015 from Sooam, down from the orig­i­nal $194,004 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. (Vi­a­gen’s price remained the same in 2019.) The Chi­nese firm Sino­gene, which cloned a Kun­ming wolf­dog in 2019, report­edly charges $53,000–$56,000 for a dog clone, and claims to have done >40 total as of Sep­tem­ber 2019. A Novem­ber 2019 anec­dote involv­ing a bil­lion­aire’s cloned dog stop­ping a drone men­tions “Mini Vader, the $144,000 (£111,563) pup that sourced from Dr Olof Olsson’s Kore­an-based cloning lab Swis­sX, earned his jaw-drop­ping price tag with a mid-air res­cue that unfor­tu­nately left his owner with a nas­ti­ly-gashed hand.” 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, or why Swis­sX—­bet­ter known for mar­i­jua­na-re­lated activ­i­ties—­costs 2–3x as much as any­one else, I do not know.)

For com­par­ison, the other kind of com­mer­cial­ly-offered pet cloning is for cats. The first com­mer­cial cat clone, report­edly cost $73,275 in 2004; more recent­ly, cat clones from Via­gen cost $35,000 in 2016 and $25,000 in 2019 (although the arti­cle notes that “Cloning a dog costs $50,000 while a cat is now $35,000—the com­pany recently increased the fee by $10,000 to cover ris­ing costs.”). Over­seas, the first Chi­nese cat clone from Sino­gene Biotech­nol­ogy Com­pany in July 2019 as the start of a com­mer­cial ser­vice quotes Sino­gene as say­ing it “is expected to cost 250,000 yuan ($35,400) each. Zhao told the Global Times that sev­eral cat own­ers had already booked the ser­vice, hint­ing that the future mar­ket could be huge. The com­pany also offers a dog cloning ser­vice, cost­ing 380,000 yuan [$54,400].”

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 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)

  ## NOTE: this is a brute-force approach chosen for simplicity. If runtime is
  ## a concern, one can sample from the extremes directly using the beta-transform trick:
  ## https://www.gwern.net/Order-statistics#sampling-gompertz-distribution-extremes
  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)
How the prob­a­bil­ity of post-s­e­lec­tion suc­cess increases given a prior base rate and a pre­dic­tor of r power for sin­gle-step selec­tion; absolute prob­a­bil­i­ties, and log-trans­formed.
Like­wise, but with an addi­tional selec­tion step prior to cloning to fur­ther select the best one.

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 SF dog cloning sce­nar­ios show­ing profit vs pos­si­ble cloning & train­ing costs, col­ored by her­i­tabil­i­ties.
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.

See Also

Appendix

Dog heritabilities

Notes on read­ing reviews & meta-analy­ses on the psy­cho­me­t­ric prop­er­ties & her­i­tabil­i­ties of dog behav­ioral traits, par­tic­u­larly for work­ing dogs. Dog her­i­tabil­i­ties might be expected to be low in the con­text of con­sid­er­ing dogs of the same breed (as would be rel­e­vant to a breed­ing or train­ing con­tex­t): heavy selec­tive breed­ing would tend to reduce with­in-breed her­i­tabil­i­ties (while increas­ing group her­i­tabil­i­ty).

Over­all, her­i­tabil­i­ties appear to differ by breed and be quite low (say, closer to 25% than to the ) but the psy­cho­me­t­ric prop­er­ties of dog behav­ioral tests also appear to be poor, with low item counts, reli­a­bil­i­ties, test-retests, and pre­dic­tive pow­er, rater/judge effects, and lit­tle use of latent fac­tors to extract more reli­able mea­sures, sug­gest­ing con­sid­er­able total mea­sure­ment error and thus con­sid­er­able under­es­ti­ma­tion of prediction/heritabilities. Pos­si­bly dog her­i­tabil­i­ties are much closer to human her­i­tabil­i­ties than they seem.

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.

  • Jones & Gosling 2005, “Tem­pera­ment and per­son­al­ity in dogs (Canis famil­iaris): a review and eval­u­a­tion of past research”

    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.”

  • , 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).

Fur­ther read­ing:

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 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 to pre­dict­ing ~40–42% of vari­ance or r = 0.65 in 2018 (Qian et al 2019//). Nor have cur­rent height PGSes have not hit a ceil­ing yet. Whole genome data indi­cates (), 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 (, Liu et al 2013, Sex­ton et al 2018; the last pro­vides the spe­cific case of , 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 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 (, Chan­f­reau-Coffinier et al 2019) and treat­ment of rare mono­genic dis­eases (only par­tially solved by the uni­ver­sal use of 22) and screen­ing new­borns (, 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 effects and bet­ter PGSes) or life­long uses such as CAD/T2D/IBD/breast cancer/stroke (Torka­mani et al 2019, Khera et al 2018, , Mavad­dat 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 at 5 foot 9 inch (1.75 m), and the short­est player ever was , 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.) Amus­ing­ly, many NBA play­ers are related.

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 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 aim­ing for 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 is a sin­gle bite of a sin­gle apple (which ) from a sin­gle seedling for sev­eral years in a row with no sec­ond chances—and only then can .↩︎

  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% () or 50% (Weiss and Green­berg 1997).

    ↩︎
  6. Inter­est­ing­ly, there was a pre­vi­ous suc­cess­ful cloned dog, in 2000, by the project (funded by a dona­tion to Texas A&M Uni­ver­sity of $4,263,096)—but it was still­born and not widely reported.↩︎

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

    • Ander­sen, Gary, LTC. Tele­phone inter­view, 1969-11-28. 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, 1989-12-05. 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, 1989-12-01. 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:

    ↩︎
  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).

    “Fig 1. Her­i­tabil­ity esti­mates, breed-level behav­ioral data, and clus­ter­ing based on behav­ioral and genetic data. A) Her­i­tabil­ity (h2) esti­mates (pro­por­tion of vari­ance attrib­ut­able to genetic fac­tors) for 14 behav­ioral traits. Geno­typic vari­a­tion accounts for five times more vari­ance in analy­ses across vs. within breeds (with­in-breed esti­mates com­piled from Ilska et al., 2017). Points for Hay­ward et al. and Parker et al. reflect the results of analy­ses with inde­pen­dent genetic datasets. Error bars reflect the 95% con­fi­dence inter­vals.”

    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. 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, 3⁄4 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. Ham­mer­strom 2005:

    “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 ) 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 $62,414 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 Dogs of War Are in High Demand: After send­ing hun­dreds of canines to post Sept. 11 bat­tle­fields, the Pen­ta­gon is buy­ing robot pooches to help train medics.” (2017):

    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 frame­work, for exam­ple, is widely used in agri­cul­ture for much bet­ter pre­dic­tions than pos­si­ble with merely indi­vid­ual SNPs (; in GWAS, eg , ). 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↩︎