# Who is the most complete athlete? – An insight with the Mahalanobis distance (sport & data analysis)

Want to share your content on R-bloggers? click here if you have a blog, or here if you don't.

The Olympic Games have finished a couple of days ago. Two entire weeks of complete devotion for sport. Unfortunately I hadn’t got any ticket but I didn’t fail to watch many games on TV and internet. I was looking at decathlon men competition and I was very impressed by the general quality of these athletes. They have to be able to do everything: sprinting (100 m), jumping high, jumping fast (110 m hurdles), long, throw heavy (put shot) and light (javelin) things, running longer (400 m) and even longer (1500 m)… It became obvious in my mind that it was the quintessence of the sport, every athlete has to find the perfect balance between those different performances to compete efficiently. This sport induces all the quality of a strong man: power, endurance, flexibility, sprint…

Is it really true? Is it really the most balanced athlete who win the decathlon competition?

I decided to test this assumption with the results of the previous Olympic Games (Beijing, 2008). I only kept the athletes who have completed all the disciplines so that I can do the study on a data set without any missing values. I used the observations of the scores for each discipline which are calculated according to the time of the distance done by the athlete. If you are interested in those details, you can have a look at the way it is calculated on: http://www.iaaf.org/mm/Document/Competitions/ … _Tables_of_Athletics_2011_23299.pdf.

I have been very surprised to see that the winner, Bryan Clay who has an average of 879 points per discipline, did very poorly in 400 meter (865 points), high jump (794 points) and in the 1500 meters race (522 points). On the contrary, he performed very well in 100 meters, 110 meters hurdle and long jump disciplines. Thus, I started wondering if the decathlon was not about power rather than about my so-called balance capacity in all the different areas.

Sir Prasanta Chandra Mahalanobis answered to this question some decades ago. In 1936 he decided to create a new function to measure the distance separating two observations. The most common distance is the Euclidian distance. However, this distance does not take into account two important elements. The first element is the variance of the different variables. Indeed, let’s consider the high jump discipline and the pole vault, a gap of 30 centimeters between two athletes is huge in high jumping whereas it is a reasonable difference in pole vault. The reason is easy to understand, the variance in pole vaulting discipline is higher than in high jumping. Fortunately, most of the robustness to the variance is taken into account by the international athletic association (the federation who sets the scores) – although we will see that this is not perfectly true. But there is another problem which is even more important. The correlation of the different disciplines. For example the following graphic shows a positive correlation between shot put and disc throw, which, if we think about it, makes sense! Thus, if we look for the most complete athlete, there should be no cumulative rewards – we don’t want to give athletes too many points when they have performed well in two very similar disciplines. On the contrary, if two disciplines are negatively correlated such as 1500 meters and 100 meters we want to give extra points to athletes who perform well in both of the disciplines. The Mahalanobis distance has been created in this purpose.

If S is the matrix of variance-covariance of the data set, we can formally write the Mahalanobis distance between the vectors x and y as:

Once the matrix S is computed, we can calculate the Mahalanobis score for every athlete – say the distance between zero and the scores of the athlete in the different disciplines. It was unexpected to see that the gold medal would be claimed by Oleksiy Kasyanov who has finished 7th during the Olympic Games. On the contrary, Bryan Clay the Olympic champion would now rank 5th. You can find below two tables, the first one is the ranking of the athletes according to the Mahalanobis distance, and the second one is the official decathlon ranking. As you can see they are many differences. Therefore, decathlon is not the ultimate sport of complete athlete.

 Mahalanobis Ranking Athlete Mahalanobis score 1 Oleksiy Kasyanov 790.60 2 Andrei Krauchanka 789.16 3 Maurice Smith 767.85 4 Leonel Suárez 754.27 5 Bryan Clay 742.40 6 Yordanis Garciá 737.40 7 Michael Shrade 723.31 8 Romain Barras 709.31 9 Aleksandr Pogorelov 701.18 10 Andres Raja 696.00 11 Roman Sebrle 693.79 12 Aleksey Drozdov 690.95 13 André Niklaus 687.12 14 Massimo Bertocchi 681.92 15 Jangy Addy 681.16 16 Mikk Pahapill 677.04 17 Mikalai Shubianok 667.82 18 Hadi Sepehrzad 653.71 19 Damjan Sitar 651.63 20 Eugene Martineau 637.66 21 Haifeng Qi 631.22 22 Aliaksandr Parkhomenka 630.64 23 Slaven Dizdarevic 607.92 24 Daniel Awde 607.78

 Decathlon Ranking Athlete Decathlon Score 1 Bryan Clay 8791 2 Andrei Krauchanka 8551 3 Leonel Suárez 8527 4 Aleksandr Pogorelov 8328 5 Romain Barras 8253 6 Roman Sebrle 8241 7 Oleksiy Kasyanov 8238 8 André Niklaus 8220 9 Maurice Smith 8205 10 Michael Shrade 8194 11 Mikk Pahapill 8178 12 Aleksey Drozdov 8154 13 Andres Raja 8118 14 Eugene Martineau 8055 15 Yordanis Garciá 7992 16 Mikalai Shubianok 7906 17 Aliaksandr Parkhomenka 7838 18 Haifeng Qi 7835 19 Massimo Bertocchi 7714 20 Jangy Addy 7665 21 Daniel Awde 7516 22 Hadi Sepehrzad 7483 23 Damjan Sitar 7336 24 Slaven Dizdarevic 7021

The code (R):

#data and data3 are randomly generated for the example

a = rnorm(24)
data=data.frame(shotPut=a, discusThrow=0.5*a + 0.5 * rnorm(24))
data3=data.frame(X1=a, X2=0.5*a + 0.5 * rnorm(24), X3 = rnorm(24), X4 = rnorm(24), , X5 = rnorm(24), X6 = rnorm(24))

lm.shotPut = lm(data$shotPut~data$discusThrow)

plot(data$discusThrow, data$shotPut, axes=TRUE, ann=FALSE)
abline(lm.shotPut)
title(ylab=”Score at shot put”, xlab = ‘Score at discus throw’, col.lab=rgb(0,0,0))

Sigma = cov(data3)
distance = mahalanobis(data3,0 , Sigma, inverted = FALSE)