Looking at last season's results, which amount to a data set of 7272 debates, the Affirmative won 52.5% of the time.  Whether one considers this a large split is somewhat subjective, but using a binomial distribution we can say that the odds of this happening randomly are functionally zero.  In terms of the Glicko ratings, a team gains something like a 17 point advantage by virtue of being affirmative.

To be clear, these numbers do not speak to the cause of the bias.  Some might like to say that it's the proliferation of non-topical cases or "new debate."  Or perhaps it could be due to the death of case debate, a general favoring of "Aff choice" on framework, or a biased resolution.  However, the cause is a question for another day - a question that would require a lot more work than I'm willing to give it at the moment.  What I can say is that it cannot be explained by the possibility that stronger teams somehow managed to be Aff more often.  I don't know how this would even happen in the first place, but it's not borne out by the data anyway.  The average Aff rating is only a fraction of a point higher than the average Neg rating.

While it is possible that the bias was due to the nature of last year's resolution, so far the trend has continued this year as well.  Aff win percentage is 52.2% over 2248 counted rounds, which would only happen randomly about 2% of the time.  Perhaps the advantage will level out as the season goes on.  Conventional wisdom sometimes says that the Aff's advantage is greatest at the beginning of the season when teams have not yet had the chance to fully prepare negative strategies.  However, if last season is any indication, we cannot expect to see any evening out.  Aff win percentage fluctuated a bit over the course of the year, but it always stayed in the black.  In fact, records showed the greatest advantage for the Aff during the second semester period in between the swing tournaments and district qualifiers.

**UPDATE**

The effect may also apply more narrowly to elimination rounds, perhaps even being magnified depending on how one interprets the data.  Keep in mind that the size of the sample is much smaller.  Whereas the total data set for 2014-15 was 7272 debates, there are only 563 elimination rounds counted.  Of these, the affirmative won 54.4% - a definite step up from the prelim data.  However, it's also possible to interpret the results in a way that suggests that the teams that were on the Aff were overall slightly better than the negatives.  While the mean rating for negatives was about 5 points higher than for the affirmative, it happens to be the case that in 52.1% of debates the Aff was favored to win.  In the end, the bias in elims appears to have been probably about the same as prelims.  However, the smaller sample size means that randomness is more likely to have played a role.  We could expect to see numbers like these or worse around 14% of the time even if sides were equally balanced.

The data for this year is even more limited.  I have 184 elim rounds, of which the affirmative has won 57.6%, a number that is quite high.  However, once again, the Aff was also the favored team in 53.9% of these debates, suggesting that the advantage might not be that different than it has been for prelims.

**UPDATE 2**

​​The graph below breaks down the 2014-15 season to more narrowly show how each side performed as varying degrees of favorites.  The breakdown suggests a larger bias for Affs at lower point spreads that evens out as the gap between teams rises.  The data does provide some support for Rashad Evans's hypothesis that bigger upsets can be scored as the negative.  Once the point spread reaches around 300, Neg win percentage starts to outpace the Aff -- though, of course, at this point any chance of upset is very small.
 

The first set of ratings for 2015-16 are now posted.  

Full disclosure: In addition to my work with Concordia, I am also helping Michigan this year.  To avoid possible conflict of interest problems, I have made no changes to the ratings algorithm since summer and will also make no changes over the course of this year.  This is pretty easy with Glicko style ratings because once you set them going all you have to do is enter new results data as they arrive.

A quick refresher on how the ratings work: Glicko style ratings are determined in a self-correcting relational way based on who a team competes against.  If you win, your rating goes up.  If you lose, your rating goes down.  How much it moves is based on the rating of your opponent.  If you beat somebody with a much higher rating, yours will go up a lot.  If you beat somebody with a much lower rating, yours might barely move at all.  And vice versa for losing.  At the beginning of the season, each team starts with the same rating (1500).  As results come in, the ratings begin to separate teams by moving teams up and down as they win and lose debates.  Since there is little data early on, the ratings are much rougher at the beginning.  They gradually become more fine tuned over the course of the season.  They need some time to sort themselves out.  More data = better.  More data also gradually stabilizes a team's rating.  At the beginning of the season the ratings are more unstable and react more quickly to change than they do at the end of the season.  The numeric value of a team's rating is essentially a predictive value.  The difference between two teams' ratings forms the basis of a prediction concerning the outcome of a debate between them.  For example, a team with a 200 point advantage is considered to be a 3:1 favorite over their opponent.  You can find out the predictions for your own debates by using the prediction calculator.

Comments on where the ratings sit now:

1. The relative dearth of data doesn't allow the ratings to be terribly sophisticated yet.  As it stands, the ratings line up pretty directly with win/loss percentage.  This will change as they are able to account more and more for strength of opposition.
   
2. The bifurcation of tournament travel also impacts the ratings.  The algorithm makes no assumptions about the quality of a tournament overall.  Instead, one's rating is determined in relation to the "pond" that one swims in.  This can have a big impact if we are trying to compare teams from relatively discrete ponds that don't interact much.  In other words, 1700 for a team that went to UMKC/Weber does not necessarily mean the same thing as it does for a team that went to GSU/UK.  We will only be able to get a better picture as there begins to be more circulation between the two pools of debaters.
   
3. Neither the Kentucky nor Weber round robins were included in the first period of data.  I discussed the reason here.  In short, including the round robins in the first data period has a large distorting effect because the ratings have not yet been able to distinguish quality of opposition strongly enough.  Since at a round robin you have a large number of rounds against almost exclusively high quality opponents, the ratings can unduly punish teams that underperform and inadequately reward teams that do well.  The solution is to move the inclusion of the results back one time period (in effect, pretend that the round robins happen a week later than they do).  This allows the ratings to settle in and more properly evaluate round outcomes.  For now, this means that at least Michigan State ST (as well as maybe Iowa KL and one or two others like Binghamton SS) are mostly likely underrated in the standings.  Inclusion of the round robins will also likely pull Michigan KM and possibly Harvard HS down a little.
   
4. To make sure that there's a minimum of data available for each team, there is a two tournament minimum for inclusion into the ratings.  It will go up to three later in the season.  Consequently, you will not see your name in the ratings if you have only attended one tournament, even if you did well.