In this post, the effect of cold weather on home field advantage will be measured more precisely. By using logistic regression, relative team strength is accounted for. In addition, the statistical significance of the observed weather effect indicates if the effect is real and systematic, or just a result of luck and small sample size.
The last post left off with this table. The left most column describes the visiting team's climate, and the top row describes the home team's climate. For each combination of visiting and home climate, the change in home team winning percentage from early to late in the season is listed. Positive numbers indicate that cold weather may favor the home team. For example, when dome teams play at warm cities, the home team winning percentage was 20% higher late in the season than early in the season. But when dome teams play at moderate cities, the home team winning percentage appeared to be 9% lower late in the season.
| Difference | Dome | Warm | Mod | Cold |
| Dome | 0% | +20 | -9 | +35 |
| Warm | -5 | +14 | -7 | -3 |
| Mod | -8 | -24 | -6 | -13 |
| Cold | -8 | +2 | -15 | +8 |
Several match-ups indicate that the cold and wind of late-season outdoor football has an effect on HFA. Consistent with other research, it appears that dome teams suffer when playing in cold climates.
Another remarkable combination is moderate weather teams playing in warm cities (-24%). But it's not clear why moderate teams would have an easier time playing in the balmy breezes of Florida or San Diego in December. Other notable match-ups are dome teams playing at warm cities (+20%), and cold teams playing at moderate cities (-15%).
To determine if the observed differences in HFA between early and late season games is really due to the change of weather, and not due to relative team strength or luck, several logistic regression models were run. The models were based on every regular season game from the 2002 through 2006 seasons (n=1280). For each game, each team was designated either Team A or Team B. The general model specification was the following:
Dependent variable:
Team A won
Independent variables:
Team A season efficiency stats
Team B season efficiency stats
AHome
[Weather dummy variable]
The team efficiency stats include offensive and defensive passing and running efficiency, turnover rates, and penalty rates. AHome is a dummy variable that is 0 when Team A is away and 1 when Team A is home. The [Weather dummy variable] is 1 when the particular climate match-up of interest is present for the game, and 0 when otherwise.
A general HFA variable (AHome) was included to isolate the effect of weather from the general home field advantage due to travel, psychology, officiating, or other effects.
Several models were run for each climate match-up of interest. The table below lists the statistical significance of the 'weather variable' and the resulting home field advantage calculated from the regression results. Note that the overall HFA rate is 57.5%.
| Visitor | Home | p-value | HFA (%) |
| Dome | Cold | 0.03* | 88 |
| Dome | Warm | 0.28 | 75 |
| Warm | Warm | 0.58 | 71 |
| Warm | Cold | 0.74 | 76 |
| Mod | Warm | 0.11 | 40 |
| Mod | Cold | 0.95 | 62 |
| Cold | Mod | 0.13 | 49 |
Accounting for relative team strength, the only truly significant result is for dome teams in cold weather (p=0.03), with an expected home team winning percentage of 88%. This result is confirmed by the actual 86% home team winning percentage when dome teams play at cold cities late in the year. The regression's estimate of 88% suggests that the cold city home teams may have been slightly weaker compared to their dome opponents over the past 5 years.
Two other types of match-ups might be considered marginally significant--moderate teams at warm cities (p=0.11), and cold teams at moderate cities (p=0.13). Both results indicate a reduced HFA later in the season for those match-up types. But since there were 16 combinations of climate match-ups, chances were that we could see one or two type-II errors, i.e. see significance when none is truly there. Because there is no a priori theoretical reason why to expect those results, we shouldn't deem them significant.
Finally, one last model specification was run to make sure no other climate match-up types were significant. The final model included all late-season match-ups types together. No additional types were significant.
It's clear that situation of a dome team playing playing in a cold city late in the season creates a much stronger HFA than normal. But a larger data set is needed to conclusively analyze other weather match-ups. Dividing 1280 games into 16 types of match-ups and 2 weather periods creates small sub-samples.
A prediction model's accuracy may benefit from enhancing the weight of HFA in certain situations. Looking ahead, there are only three 'dome at cold' match-ups in 2007. STL travels to CIN in week 14, and NO visits CHI and DET visits GB in week 17.
Posts
4 comments: