Kenenisa in 26:17 Form Could Run 2:00:36? So Say MIT Guys

Let's no kid ourselves. There is no ethical dilemma here. If someone finds a math error during the peer review process, so what.. Who cares? Who does it harm? I think Franz is getting a little to caught up in the academic review process, overstating the distinction of being peer reviewed and also the end value of his work.

I want to point the fact I'm not critic with the authors, on the contrary I appreciate the attempt to find a scientific way for predicting the human possibilities.

What I say, is that there are too many casula and individual factors that can't be reduced to a mathematic formula.

Look for example at the speed in 400m for two different athletes, one with a morphology of a sprinter (for example, 185 tall and 62 kt heavy) before training with weights, another with the morphology of a small Marathon runner (165/49).

They can have the same percentage of Slow Fibers, but their max speed is very different and, on the other side, their cost of running changes on individual basis when the distance becomes longer. When you run using more "frequency", you can maintain for longer time the same cost of running, so this technical attitude is important for longer distances, and sometimes is a problem for shortest distances. In the middle, these two athletes can run the same time, but their correlation with longer or shorter distances doesn't follow the same trend. This is, for example, the limit in the calculator of Ventolin.

There are many myths to discuss. For example, some poster above continues to think a road race of 10 km is 30 seconds slower than the track, or than a Marathon can be faster if run on track, using spikes.

The reality is the opposite : athletes can run faster on a road race (of course, if the tarmac is perfect and there are not too many turns), for the following points : more free mentality (psychologically, runners use more "wild instinct" in road races than on track), more necessity of change of pace, that can help to maintain max level of coordination (it's not a case that the WR of 15 km of Leonard Komon is in a race called "seven Hills", so the idea that a course completely flat can be faster than a course with some hill, not too hard and not too long, is wrong), and, in many cases, bigger Group of competitors.

To compare the official lists all time of 10000m on track and 10 km on road it's not possible, because normally the runners on road are only interested to win the race, and not to attack some record. When we look at record (like, for example, we did in 2014 and 2015 with Florence Kiplagat in Barcelona), we can see athletes running faster than what they could do on track.

Another wrong idea is that spikes can give some advantage in long distances. Also in this case, it's true the opposite : the only effect spikes can have in long distances is to increase the soreness of calves, and to increase the risk of injuries. At the speed of 2'50" per km, or slower, athletes don't have any mechanical advantage to use spikes, also on track (of course, the best tracks of today). On a very hard track (for example, in Mondo Rubber), sometimes it's better to run with racing shoes than with spikes. This is what, for example, the Italian Record Holder of 10000m, Maura Viceconte, did in 2000m, improving of 1 minute (from 32:05 with spikes to 31:05.53 with racing shoes) her PB in Heusded, and is also what Florence Kiplagat did when ran the Kenyan record of 10000m (30'11") in Utrecht in 2009, and when almost defeated Tirunesh Dibaba (30'24") in Eugene 2012, after going there for running the limit for OG only.

About Marathon, I repeat that the results in shorter events are not correlated, if we want to have some indication starting from short events. This because the bioenergetical mechanisms we have as common denominator in 5000/10000/HM are not correlated with the main quality we need in a Marathon runner, which is the ability to reduce the consumption of fuel at the Marathon Speed. This is like to think that a Formula One. because is more powerful, can be the best also in the 24 hours of Le Mans, when instead we need a completely different car (of course, very fast, but with different characteristics in order to last : it's not a case that two years ago the winner was a car with diesel engine...).

Another observation regards the statistical validity of a research. When we want to have more precise infos about some phenomen, wider is the research (so, more data we have and more subjects we investigate), more connected with the reality are the results. But, when we want to have some research about the "top of the pyramid", we need to limit the research to omogeneus subjects, because the general average has the effect to increase the gap between every type of Elite and the "average Joe".

For example, about the distribution of richness in one Country, if we put together all the existing money dividing for the population, we can have an average data that, practically, include only a very small percentage of people, since the reality is not "I eat two chicken, you don't eat any chicken, the statistic says we eat one chicken each one", but "I have 0ne hundred million dollars, one million people have hundred dollars, our average is everybody has 200 dollars".

This is the case of every Elite : the research. for having some statistical validity, must be done with omogeneus subjects.

This means that one of the limits of these research is exactly the fact that this research can have, instead, great validity for the average. Collecting 150.000 data (I don't know about how many subjects) can contributes to increase the deviation from the data referred to some elite.

About the specific case of Kenenisa, I can say that, before Dubai this year, also limiting his training because we had to maintain under control the tendon problem (so, not enough mileage and not enough specific training), he showed in any case to have the CURRENT possibility to run in 2:04.

During the race in Dubai (I was on the car of the jury), at 28th km Jos Hermens (who followed the race on a motorbike) came to me telling me "Pacers are not too good, Kenenisa is very fresh and doesn't do any fatigue, what I have to tell him ?", and I answers "tell him to go alone after 30 km" ; and after 200m we saw him limping, behind the Group...

The Kenenisa of Dubai (without considering his tendon problems) could run 2:04, but not faster than 27:30 in 10 km. This means he has ALSO the specific attitude for running a Marathon, so it's possible to suppose than the Kenenisa of 12'37" (who NEVER prepared 10000m, therefore his 26'17" is far from his real potentiality) could move, in 6 months specific training, to Marathon with a value, at the moment of the Marathon, of about 26'30", that could have a projection on Marathon of sure under 2:02.

But, I repeat, for running Marathon we need some characteristic that is not connected with other events, and there are too many athletes producing good well balanced performances in a Group of events (3000/5000/10000) still able to run a fast HM, and totally enable to run a fast Marathon also trying proper training. For this reason, while I think possible to individuate the theorical possibilities of an athlete (Always becoming reality with proper training only) analyzing their performances plus some variable in a "limited Group of events, correlated one with the other", I don't think possible to find any formula for predicting the possibility in Marathon, because this event requires some specific physiological quality NOT required by any other event in athletics.

take another look wrote:

RoadToTrack wrote:

This is a great point. I've collected both race results AND training data for athletes. And controlling for the terrain of the race (road, track, off-road) or limiting the sample to only track races, improves the fit of my models, and I would assume would do so for Franz's, if it has not been done already.

I doubt that the fact even crossed his mind. Even experienced observers fail to recognise that the fastest road 10ks are run 30 seconds slower than those guys run a 10000.

None of that matters. Aside from your premise being wrong - road is not slower than track (see Leonard Komon) - if there is a difference between road and track, the model would figure it out.

Franz Király wrote:dear all,

would it be possible for you to say a few words on what you and Renato call Ventolin's formula? That would be great.

It looks like

T1/D1 - T*/D* = (T2/D2 - T1/D1)/(log(D2/D1)*log(D*/D1))

where T1,D1 and T2,D2 are known time performance/distance pairs, and T* is the time to predict, at distance D*.

Is this correct?

i'm not sure if that's the correct 1-line formula

if it gives

1'43.00 / 13'00 ->

2'12.07

3'27.15

4'44.75

7'25.21

27'43.07

then it's correct

if not, work out algorithm yourself or subscribe here :

here's some "free" advice

your statistical approach is not going to work, even with thousands of inputs, as too many variables & too few elites

if you had few thousand guys running 1'43 or 3'30, then you may get somewhere

approach i used is that theoretically, which i believe is correct & best model, any doubling up of a track distance from 400m with an even, whole number value, will have a theoretical equivalent of a whole, even number for double distance - this is the use of logarithms as that is what running is :

400m of 44.00 =

800m of 1'42.00 =

1600m of 3'44.00 =

3200m of 7'54.00 =

6400m of 16'30.00 =

12800m of 34'06.00

you then use the above caculator to pick off intermediate distances of 1k/2k/3k/5k/10k

for longer track distances, all on track :

27'00.00 = 59'12 = 2"09'12

for sprint distances,

44.00 woud theoretically = 20.00

however, they are not "symmetrical" races as 400 is 2 bends/2 straights & 200 is 1 bend/1 straight

for "symmetry" you don't use outdoor 200, but it's equivalent indoor value of 200 as that is symmetrical

an outdoor 200 is ~ 0.25s quicker than an indoor :

20.00i = ~ 19.75

therefore 44.00 does not = 20.00 but 19.75

you can tack that on top of above list

as for 100m, just go with 1/2 the 200 value :

19.75 = 9.875

so final list of = is

9.875

19.75

44.00

1'42.00

2'11.52

3'28.25

4'45.23

7'22.11

12'41.75

26'20.07

i guarantee, these are just about best equivalents you will ever find

It matters not. wrote:Aside from your premise being wrong - road is not slower than track (see Leonard Komon)

komon didn't run a track 10k in his 26'44 year

he did run a 12'59 but blown away on last lap by merga in 12'53 in slow race to bell

komon in his 26'44 road shape wouda been lot better than that on the track

that was his career peak & never been close since

The actual flaw in this equation is that an athlete who is in shape to run 26:17 CAN'T run their best Marathon. Their specialty is running substantially over the LT for less than 30 minutes. If you ask the athlete to run 2 hours, their engine would finish fuel at 30k or so.

So if the 26:17 athlete trains for the Marathon, they no longer would be a 26:17 athlete, but a 2:03/04 athlete. Their 10k would be more along the lines of 27-27:20.

So if the athlete is 25"20 or better, then their "Marathon-Shape" 10k would be 26:17. Then the equation makes sense.

Maybe the author took that into the equation, but historically, Haile, Mo, Sammy, Kipchoge, Tergat, and Bekele have proved that 26:45 PBs or under doesn't come close to 2-Flat. If they trained for 10k, they'd blow up. If they trained for Marathon, they'd slow down. Maybe the title should be "A 25:20 WR holder in 10k should be able to run 2:00:36". I think that makes sense.

No, the flaw in your premise is that just because a runner can run a 26:17 then his specialty must be the 10k.

The other flaw is you saying that Haile, Mo, Sammy, etc have proven that 26:45 PBs doesn't come close to 2-flat. None of those runners even showed that they were in sub 26:45 shape when they ran the marathon.

Amerikano wrote:

The actual flaw in this equation is that an athlete who is in shape to run 26:17 CAN'T run their best Marathon.

No that's a flaw in the thread title. Apparently you didn't read the actual article.

I think you are utterly and completely right - I think so too after spending too much time over-thinking it and sleeping badly due to this.

The press coverage (which I need to say has in no way been promoted or endorsed by myself or Duncan, or our universities) so far has been about the "fair race" business anyway and not about medical/physiological implications, so I guess no harm. Does anybody really care about whether it is 492 or 508 m?

Also, I'm not giving an interview here but participating in scientific discussion with experts. Nor am I actively promoting or marketing our work since I did not start the discussion - instead the discussion is useful in improving the scientific quality of the work.

But I don't think I was overstating the value of proper peer review - the quality the discussion here has gained goes in a similar direction, thanks; and having a paper honestly reviewed and criticized by people who are able to judge every bit of it, is invaluable.

(though occasionally, you may get it reviewed and criticized by ... other kinds of people.)

Apologies, the press calling just triggered some flight reflexes. Fortunately this seems to be over now, at least for the moment, and there will be no interviews anyway before everything is scientifically watertight (maybe by then, nobody will ask anyway), so I'll get just back to our discussion.

Where I will need some time to go through all the posts...

Just want to point out that this is all preliminary work etc, you know, the IMPORTANT caveats when dealing with things unpublished.

Dear Renato,

Thanks for your comments!

Please let me answer the points you raise separately.

> It's like to study a dog Terranova or Bull dog, and after to transport the same statistical and physiological data on a levrier which competes for winning a race.

Indeed the very elite athletes are “at the boundary” of the 150.000 subjects – that is, they seem to be very special. We already remarked this in finding (IV.ii) of the pre-print. The point you raise is very valid, and also overlaps, partly, with what I said in my reaction to your previous post.

We do very well for terranovae and bull dogs, and part of our reasoning in the paper is that levriers behave similarly. Which, while plausible via inductive analogy, need not necessarily be, and it has to be checked in the future.

So we added some caveats which will appear in the final version, among them this one:

“N.B.: We would like to stress that such predictions need to be taken with much caution, as they are only insofar correct as our model extends, from the general population of British runners (who successfully participated in official events), to the very extremes of human performance.”

Though one thing to point out is that our framework is, in principle, applicable to levriers as well – we just need enough of them. They might be rare, but with a bit of effort one might get hold of a high enough number…

About mathematics, deduction and induction:

>Mathematics are Inductive Science :

>Physiology, instead, is a DEDUCTIVE SCIENCE :

I would disagree here, and not draw a separating line in either case.

Mathematics allows for deductive arguments (axioms -> logical reasoning -> necessary conclusions), often found in the pure fields, and inductive ones, on which large part of statistics relies (observation -> experiment -> generalization), including inference, the statistical syllogism, and predictions in the sense of our paper.

We have used both kinds of arguments in our work – the argument that the model is rank three is inductive, from the prediction experiment; while the statement that this implies three coefficients per athlete is deductive, from theoretical properties of “matrix rank” (determinant rank = column/row rank).

The predictions about the elite athletes are again inductive, thus not necessarily true in the sense of logical stringency, but only plausible under generalizability assumptions.

In physiology and medicine you also have both kinds of arguments. Deductive: knowledge about chemistry, physics -> necessarily implied knowledge about physiological processes, or inductive: lab experiment/clinical trial -> working hypothesis about physiology, therapy, training etc.

Reiterating, I would really not want to draw a boundary along lines that are, very often, rather social phenomena than scientific necessity.

About correlation:

> At the max level, there is no correlation between 10000m and Marathon

I computed the (Spearman rank) correlation matrix of (male) athletes in their best year, here it is:

perf.1 perf.2 perf.3 perf.4 perf.5 perf.6 perf.7 perf.8 perf.9 perf.10

perf.1 1.000 0.935 0.785 0.541 0.208 0.628 0.248 0.323 0.241 0.122

perf.2 0.935 1.000 0.849 0.649 0.358 0.720 0.383 0.382 0.247 0.188

perf.3 0.785 0.849 1.000 0.777 0.559 0.806 0.559 0.495 0.415 0.373

perf.4 0.541 0.649 0.777 1.000 0.919 0.927 0.735 0.697 0.750 0.645

perf.5 0.208 0.358 0.559 0.919 1.000 0.966 0.855 0.824 0.846 0.808

perf.6 0.628 0.720 0.806 0.927 0.966 1.000 0.923 0.921 0.905 0.864

perf.7 0.248 0.383 0.559 0.735 0.855 0.923 1.000 0.960 0.949 0.889

perf.8 0.323 0.382 0.495 0.697 0.824 0.921 0.960 1.000 0.964 0.920

perf.9 0.241 0.247 0.415 0.750 0.846 0.905 0.949 0.964 1.000 0.931

perf.10 0.122 0.188 0.373 0.645 0.808 0.864 0.889 0.920 0.931 1.000

The events are ordered as in the paper. That is, perf.1, perf.2,… perf.10 are 100m, 200m, 400m, 800m, 1500m, the Mile, 5km, 10km, Half-Marathon, Marathon. So perf.8 = 10km, perf.10 = Marathon.

How this is to be read, for the non-statistician: the entries are between -1 and 1. The closer an entry in the row/column of two events A,B is to 1, the more it means: “if an athlete is better at event A, he will be better at event B”. This is why for A=B you have a 1. The closer to -1, the more it means “ … better at A, … worse at B”. (for more about the math, look on Wikipedia for “correlation”)

Note that all entries above are positive – so being better at one event makes you, on average, better at any other – if you are a random person among the 85,498 male athletes I looked at for this analysis, that is (females and athletes with only one attempted event are not in here, for technical reasons).

Also, the correlation between 10km and Marathon is 0.92, so quite high. Naively, a contradiction to your claims – but as said these are mostly “normal” athletes (bull dogs and terranovae).

Now here is the surprise (well, for me at least): I restricted the data to people who achieved top percentile performance in any event, i.e., the one-in-a-hundred performance elite – roughly as close to the very top as the data set allows to look without getting serious problems of estimating these numbers (2643 events of the 1064 very best athletes).

perf.1 perf.2 perf.3 perf.4 perf.5 perf.6 perf.7 perf.8 perf.9 perf.10

perf.1 1.000 0.296 -0.084 NA NA NA NA NA NA NA

perf.2 0.296 1.000 0.119 -0.855 NA NA NA NA NA NA

perf.3 -0.084 0.119 1.000 0.110 -0.537 -0.357 -0.484 -0.881 NA NA

perf.4 NA -0.855 0.110 1.000 0.337 0.527 -0.069 -0.021 0.117 -0.214

perf.5 NA NA -0.537 0.337 1.000 0.819 0.517 0.349 0.388 0.415

perf.6 NA NA -0.357 0.527 0.819 1.000 0.476 0.309 0.288 NA

perf.7 NA NA -0.484 -0.069 0.517 0.476 1.000 0.687 0.451 0.433

perf.8 NA NA -0.881 -0.021 0.349 0.309 0.687 1.000 0.575 0.517

perf.9 NA NA NA 0.117 0.388 0.288 0.451 0.575 1.000 0.710

perf.10 NA NA NA -0.214 0.415 NA 0.433 0.517 0.710 1.000

(NA here has no direct relation to Batman, it just means that there were not enough athletes who attempted both the row and column events for a reliable estimate).

What one sees is that the correlations drop a lot, some of them even become negative – even though from what we know, the data of the top percentile athletes are much “cleaner”. Correlation between 10 km and Marathon is around 0.5. So the general trend does seem to be in qualitative accordance with what you claim.

This is very interesting and indicates that something curious happens at the boundaries - and that we can pick it up. It also makes me hope we can perhaps get more levriers and/or data on terranovae training to become levriers…

^ Very interesting stuff

Dear rekrunner,

thank you very much!

Do I correctly understand that people pay a fee for evaluating this formula?

I really don't want to ruin the business model of somebody I don't know, so I would like to start by saying that it may be recommendable from a charitable standpoint to pay for evaluating the formula if that is the only, or main source of income for that person, even if you perhaps do not want to evaluate the formula.

In any case here is some R code for everybody to use (for the free statistics software R) in case you want to run some empirical tests.

ventolin = function(T1,D1,T2,D2,Dnew) {

return (T1/D1 + (T2/D2 - T1/D1)*log(Dnew/D1)/log(D2/D1) )*Dnew

}

Just copy this into an R script window and run the code once. After that, you can evaluate the formula just by typing, say,

ventolin(12,23,34,45,56)

which will give you a ventolin-estimated performance at Dnew.

I really hope the parentheses are fine now, otherwise please feel free to change them.

Regarding use of the formula: is there a rule-of-thumb which two performances to substitute if you want to predict a third?

(often, the kind of black magic going in here is the interesting bit)

Dear HardLoper,

you mentioned an "original article" about ventolin, would it be possible for you to provide me with a link?

Thanks

The correlation matrix is awesome stuff. Thanks for sharing. One interesting quick takeaway based on the complete sample of male athletes is that 10k performance is most highly correlated with half marathon performance, then 5k, then mile(!), and then finally the marathon performance.

This reading a lot into slight differences, but is in an interesting way to frame the correlations: Based on this sample, your 10k is a better predictor of your mile time than it is for your marathon time.

Franz, do you agree?

I think this somewhat lines up with Renato's instincts and arguments about predicting marathon performance.

Also Franz thank you for not being scared away. I think many of us appreciate the discussion and comments you've provided on this thread.

Dear What the What?,

just want to point out that even if all the assumptions for generalizability of the prediction model are fulfilled (which maybe they are not always), every prediction comes with a "standard error" which gives you a region in which the prediction would be plausibly expected to be.

The media, or the performance calculators do often forget about telling you these standard error regions, or that they are there, and thus make a prediction which may take the false appearance of being exact.

I would not assume intention there - even the MIT tech review arXiv blog which is, to my knowledge, written by statistics experts, forgot.

Anyway, for the methods out there, the standard error of the prediction ranges between around 4 to 7%, which for the Marathon means that, on average, you would expect the prediction to be off by around 5 to 10 minutes.

Or, you could say the runner will be on average off by 5 to 10 minutes, but that is perhaps a matter of viewpoint.

(for experts: I'm aware of the difference between standard error, standard deviation and expectation of the absolute error, but this explanation was qualitative anyway)

You're welcome. It's not you guys who scared me nor Renato, really.

Regarding your question, three caveats at the beginning:

- I have not computed standard errors of the matrix, so please do not interpret anything into the second or third digits after the comma.

- there are only 6,590 Mile performances, as opposed to around 100,000 on the other 9 events each, in the vanilla database. So it may be better to treat the row/column with the Mile with extra care and not trust it too far.

- correlation does not imply causation!

So if correlation is 0.99, and you get better at event A, it does not mean you get better at event B. What the matrix means is just that people who are better at event A are, usually, also better at event B.

But apart from that, you could now hypothesize that event A predicts well event B if the correlations are high. Though the matrix does not directly prove that, you would still need to check this by actually trying to predict (out-of-sample).

Makes sense. Thank you for the response.

Everybody, hold up!

1. I just stumbled upon this thread - why was I not invited to this discussion?

2. Franz, beszelsz magyarul??

No I stand by my point. If you look at Sammy Wanjiru, he ran 26:41 as junior. He also set records in the half-marathon. Although he did pass too early in his career, he was FAR from 2-flat. I think he is the best example of a hypothetical mid-26 runner training for the Marathon in his peak of career.

The flaw is that every athlete so far who is capable of running under 26:45 has a period of adjusting their engine to be more efficient to last at the Marathon. If the athlete prepares for 10k, sure they can have a best of 26:20/26:30 (for instance Mo at the moment). If they attempt a Marathon, they get smoked by Kipsang like in 2014. Kipsang got beat by a 40 year old man last week, but is a top 3 Marathoner in the world.

Are you saying if Mo trained to run 26:20, then ran a marathon a few weeks later, he'd run close to 2:00:36? Doesn't take a complicated equation to figure that its not likely.

The reality of the distance is that no athlete can be in top shape for the 10k and Marathon at the same time. The engine needs to do two seperate things. Either you slow down the 10k time, race the Marathon or be in top shape for 10k and blow up the Marathon.

I will say that 25:20 is way too fast, but I would say a 25:40/25:50 athlete is a possible athlete. Also, the 10k is not a great distance to compare PBs because there's no money or even competitions to judge from. Mo hasn't run a fast 10k since 2011. That's a 4 year old PB, maybe he will get a better one in Eugene.