Announcing the FightMetric Research Contest

Really looking forward to this!

Quick correction – unless I’ve got an incomplete spreadsheet, it “only” gives data for 1486 fights.

Wouldn’t it be more useful if we had fighters’ names? That way we could assess whether a short Fighter A was a far superior striker/fighter than a tall Fighter B (at least in fighters we recognized). I mean, reach doesn’t make up for a huge lack of skill.

I bet we see a few of those.

Awesome topic, will look into it more when I’m not mobile

Unfortunately I won’t have time to make a real thorough analysis :( But I’m confident others will, really looking forward to the outcome!

We’ll provide the paint, and the fence, and all you have to do is get to painting. We’ll select a couple of the best slats and feature them near the front of the house…

I keed. This is great.

It would have been nice to have dates on the fights though. I think it might be possible reach has more of an impact in more recent times.

Who is Semi Schilt.

it’s full of stars

I can’t read it now, but I’m curious how this wil turn out

why bloody elbow is the class of the MMA news community.

Outstanding project!

will study and post.

I’m not even gonna attempt because I suck at numbers/data, but damn man who has an 84.5 reach advantage? Struve is said to have 83 and Bones is what, 81.5 or something like that. I’ve seen this measurement pop up quite a few times and all on the winning side, who could it be? Alistair???

Here we go…

Of ring vs. cage. I’m at school so I can’t run the file, but is this data included? Maybe which orgs are included in these data?

how exactly is reach measured? is it fingertip-to-fingertip or fist-to-fist.Neither of these shouldn’t really be called reach, they should be called wingspan since it includes the chest in the measurements, as some chests (like Bader) make up a good portion of their ‘reach’.

I’ve seen a couple ‘Tale of the Tapes’ from HDNET that show reach being from the shoulder to the fist, indicating true reach and the advantage.

I didn’t read all the data, but if there is data available to do a graph of % of winning with respect to the number of lbs a fighter has above his opponent it would be interesting (in the HW division).

I started playing around with this, with the intent of performing some regressions. However, I ran into some issues. I’ll start off with some background.

To oversimplify a little, what we’ve seen so far are basically mean comparisons of winners’ and losers’ reach. Ok, this is a big oversimplification – several have provided some very intriguing and nuanced breakdowns. But ultimately, the structure of the data has encouraged all the posts I’ve seen to look at mean reach or mean reach differentials. This is useful, but it doesn’t quite map on, in a logical sense, to the process we think is occurring. I.e., what we think is happening is that reach advantage is influencing – i.e., causing a positive change in – the probability of winning. With mean comparisons, the implied causal direction is backwards. We are categorizing fighters into winners and losers and “predicting” reach differential. Regression frameworks map more closely on to the logic of the likely underlying causal model, and they’re also better suited for predicting the future. For example, ideally, we would take reach differential and use that to predict the probability of someone winning. Such a model would enable us, for example, to quantify the exact increase in the odds of winning that Jones’ reach advantage gives him over Rua. Note the outcomes in question are actually categorical (e.g., dichotomous win vs. loss; nominal method of winning). So this is a second sense in which particular categorical regression frameworks (logistic or multinomial) map more closely on to the likely underlying process. Having said all this, associations are associations, regardless of direction. My points thus var solely concern model conceptualization and aesthetics.

So we want to run a logistic regression model predicting fight outcome (win vs. loss) from reach differential. Let’s run through the early steps. The database provided has variables for winner’s reach and loser’s reach. We take the difference. Now, we need a variable corresponding to wins – e.g., coded 1 for wins and 0 for losses, to prepare for a logistic regression analysis. We can’t just assign a win any time a fighter in column A wins, because that column is defined as winner’s reach. I.e., they win 100% of the time, which means our outcome would not vary (zero variation implies zero covariation with other variables – i.e., regression models, which detect covariation, cannot be fitted). Likewise for loser’s reach – they win zero % of the time.

We could try taking another approach. E.g., we could restructure our data so that fighters with the longer reach were always in column A, and fighters with the shorter reach were always in column B. We could then 1) calculate the difference between these two variables as a third variable (which will always be positive), and 2) assign 1’s and 0’s to a fourth variable, corresponding to whether the fighter with the longer reach won. Now, we’ve got an outcome that shows some variation. But step back a moment and look more closely at step 2 above. What do we do when the fighters’ have equal absolute reaches? We have no basis upon which to assign an outcome.

Now, let me restate the penultimate sentence from an earlier paragraph above: “ideally, we would take reach differential and use that to predict the probability of someone winning.” With respect to the provided database and the issues discussed above, the question is: the probability of WHO winning? To calculate this variable, we have to have some basis on which to assign one person as “winner” and another person as “loser.” The only information in the provided database that provides this basis – this “anchor” – is reach differential. When there is no reach difference, there is no basis upon which to distingish winners and losers. The fact that the “winner” appears in “Column A” is arbitrary and meaningless.

Alternative strategies are unsatisfactory. E.g., models could be fitted if we throw out cases with equal reaches. Throwing out that much relevant data would be kinda silly. Alternatively, we could assign random numbers to each case, sort, and use the resulting order to determine which column takes priority in the regression model. This is unsatisfactory on the grounds that we are unecessarily introducing a source of random error into our model, one that will undoubtedly perturb results from one sample to the next. We’d like results that are as generalizable as possible to different samples.

In sum, the core issue with the structure of the data is that there isn’t enough information to estimate regression models in a principled way. Additional variables – that show some non-trivial amount of variability – would enable regression models to be fit for all cases, because it would establish a baseline, objective win probability. To take a somewhat absurd example to drive home the point: even if we had something completely unrelated – like color of shorts, this would provide an objective basis for assigning wins or losses to all cases. We can then say, ok, people with darker colored shorts win 50% of the time. How much does reach difference influence this baseline win probability?

My recommendation is that in the next round, include some additional variables. Incidently, the more the better: As noted by others, more variables will also enable more sophistated models or crosstabulations that tease apart many potential confounding effects, such as weight class.

For a discussion of a closely related issue, see:
http://www.uvm.edu/~dhowell/StatPages/More_Stuff/icc/icc.html

This author’s example concerns cases in which there is no sensible ordering between a predictor and an outcome. This is not the exact issue in our example, so unfortunately his solution (use of something calld the ICC) does not apply to our case (note also that he discusses but ultimately rejects a resampling procedure conceptually similar to one of the unsatisfactory approaches I discussed above). The best solution for our case is to include more variables.