Need input here

[aguahombre]

The two methods I KNOW are prone to skewing are raw averaging without eliminating extremes at all, and whatever they came up with for on-line IMP pairs. Those dividing lines between scores and IMPs were put where they are for a reason. Fractional IMPs seem just plain wrong.

[gwnn]

The two methods I KNOW are prone to skewing are raw averaging without eliminating extremes at all, and whatever they came up with for on-line IMP pairs. Those dividing lines between scores and IMPs were put where they are for a reason. Fractional IMPs seem just plain wrong.

Yes, they were put where they are for a reason, namely to compare reasonable bridge scores with other reasonable bridge scores. They are not put there to compare +420 to +295 or -50 to +231.5. So to me it makes more sense to compare real bridge scores and then average the IMP's than to compare a real bridge score with a average (usually impossible) bridge score.

[helene_t]

When there is a large dataset, sometimes median is better than average for skew distributions.

Which is "better" begs the question: "better" for which purpose?

 

Anyway, raw bridge scores are generally not skewed. Using a robust statistic like the median can sometimes lead to absurd results. Suppose there are 19 tables. 10 NS pairs score +1430, 9 score -100. The median is +1430. So if you score +1430 you get 0 IMPs. Serves you right for making the slam on a randomly chosen two-way finesse, maybe. The problem is, however, as Rik shows, that NS can only lose on this board. At least in Rik's example, the difference between the average IMPs for NS and EW would not be more than a couple of IMPs. Here the difference will be more than 15 IMPs.

 

I think the notion that one should exclude extremes is flawed. If you are seriously concerned that there is a single very weak pair that produces nonsense results and which shouldn't be used for comparison, you might want to play Swiss, or do some Swiss-like postprocessing of the data like removing all boards involving pairs that scored less than -2 IMPs/board before recalculating the datum and butler scores.

 

I am not seriously suggesting this, though, since there is a better and simpler suggestion: X-IMPs:

- Produces zero average for both NS and EW

- Has similar tactical implications as a team match

- Evens out the discreteness of the IMP-scale

- Is reasonably robust because the single 7NTxx-13 gets IMPd before averaging so the impact is reduced.

 

Of course you can also just play matchpoints if you really want to reduce the impact of outliers.

[jogs]

Which is "better" begs the question: "better" for which purpose?

 

 

There is rarely a large dataset in bridge. The median is better for some economic measures. Median income paints a more accurate picture than average income.