Used the Association tool to check the correlation. Most of the variances have p-value not low enough to be significant. Below is what I got...
How to decide if Super Bowl is (or isn't) a game? Predictions for Super Bowl are just similar as those for week 13
Could have parsed it in a cleaner way but it did the job! Then I learned about the association analysis tool from @Natasha's workflow and how to then use those chosen variables for regression analysis. Predicted scores were done using only one variable in the end, just to see how close you could get with one variable (Offense - PassY)
Looks like defensive stats might be better at predicting Super Bowl scores than regular season scores. Hopefully this reinforces the idea that a strong defense is a strong offense, and shows how that giving an inch can indeed result in losing a mile.
I've been sitting on this one for a looong time, because I got sidetracked trying to figure out how to use the R tool to generate residual plots (we were on an older version at the time, that didn't appear to include them). I finally figured out how to do it, even though they're part of the standard linear regression tool output now! I took the Intro to Advanced Analytics training at Inspire 2019, and we asked about residual plots. The instructor advised us to calculate the residuals and create the plots ourselves, so I'm guessing this is a pretty recent addition.
Workflow:
Hold out the super bowl weeks, plus one random week from each year:
(I need to find a better way to randomly sample within groups - this works, but isn't reproducible)
Data investigation:
Build the model and score the sample weeks:
(I used stepwise regression to select the model)
In practice, I'd do a little more investigation, as I'm not satisfied with this model. I'd also want to educate myself more on the "business context," since I know very little about football!
Results:
Predictions:
with many correlated features I should have considered extracting principal components, but I can interpret that because I don't know the sport. I manually removed fields from the equation (which doesn't even bother with interactions) starting with the field with the highest P value (thus lowest absolute t-value). I reran the workflow until I noticed that I had significance values of all below .1 so I stuck with that.