Thread by @rbkeeney

How should you use breakout age in your rookie WR analysis?

A thread on when it matters and... when it doesn't.

Breakout age can be challenging to evaluate at times. First, there are two primary methods: 20%, and 30% - which is better?

Second, both of these methods also leave us with some interesting questions such as: How different is 18 vs. 20 vs. no breakout age?

We can run a simple linear regression on the variables, but that typically removing ALL non-breakout players.

For example, looking at a 20% dominator driven breakout age, Hill would be excluded from our analysis!

So instead, we can treat each value as a category - and create dummy variables for each.

For example: Assign a "1" when a player has a breakout age of 18 and "0" for the rest, and so on.

This means that we're building a function that looks like this:

Fantasy points PPG avg = [b0, x1 ~ No breakout age] + [b1]*[x1 ~BOA = 18] + [b2]*[x2 ~BOA = 19].... and so on.

Put another way, if a player didn't have a breakout age, our equation is: FF PPG = b0

Next, we can run a linear regression on our model to find the coefficients:

Next, let's walk through the results:

1/ if a player has no breakout age they should score the b0 (or intercept) = 3.83

2/ if a player BOA = 19, simply add the intercept, b0, to the b2... or 3.83 + 2.58 = 6.41

What does this mean?

Well, since our null hypothesis would be BOA no impact (or the means would be the same ~ be equal to "no BOA"), we can regret the null hypothesis for ages 18, 19, and 20.

A breakout age of 18, 19, and 20 with a 20% CD threshold is statistically significant.

Going further, we CANNOT reject the null hypothesis that a breakout age of 21, 22, or 23 is the same as a player without a breakout age.

However, we cannot say that a BOA=18 is different than BOA=19 or 20... yet.

To do that, we need to adjust our formula so that a BOA =18 is our new intercept, or b0, and then compare.

like so:
AVG.PPG.YR.1.3 ~ BOA_none + BOA19 + BOA20 + BOA21 + BOA22 + BOA23

And, the results tell us that we should reject the null hypothesis that a BOA=18 is the same any other BOA.

Note, our model r-squared didn't change, and a prediction of FF PPG for a BOA of 19 is 8.41 - 1.98 ~ 6.43... the same as before!

How about using a threshold CD of 30% to calculate BOA?

Well, ages 18, 19, and 20 are still statistically significant, that is, we reject that they are the same as no breakout age.

Additionally, a BOA calculated from 30% CD method has a slightly higher r-squared value, perhaps adding to the argument that it should be preferred.

But wait! If we wanted to input this into another model... how could we?

One method is to replace data with these outputs (here, for 30% CD BOA):

No breakout age -> 4.22
18 -> 4.22 + 5.96
19 -> 4.22+ 2.72
20 -> 4.22+ 2.59
21 -> 4.22 + 0.41
22 -> 4.22 + 0.54
23 -> 4.22 - 0.07

The neat thing about this common trick (inputting dummy variables for non-numerical values) with a linear regression is that you can do it for all sorts of stuff, like conferences, coaches, teammates, etc.

@threadreaderapp unroll