Once again, I did it in the exact same way you did it, and included the teams that you said you wanted to create a bigger sample, and then when it didn't show what you expected/wanted, you tossed it aside as completely worthless.
2019:
All teams between 40-45% taxes: 100.4%
All teams between 49-53% taxes: 105.2%
2018:
All teams between 40-45% taxes: 101.7%
All teams between 49-53% taxes: 109.3%
There is still a significant differences between low tax and high tax areas.
No, that is not remotely what happened.
I said that doing things this way would be BETTER than just averaging averages as you initially did. However, it is still not the best way to go about this, because you are dichotomizing two continuous variables. Do you not understand what that means? Do you not understand why that is a bad idea? Let me explain.
Let's create a hypothetical scenario in which teams in the 40% range spend 105%, teams in the 45% range spend 95%, teams in the 47% range spend 120%, teams in the 49% range spend 110%, and teams in the 53% spend 100%. Here is what that would look like on a chart:
You actually get a (minuscule) negative correlation with those numbers. However, if you repeat the same exercise that you just performed with these hypothetical numbers, you would get:
Teams in the 40-45% range: 100%
Teams in the 49-53% range: 105%
The actual correlation between the numbers would be slightly negative, and essentially non-existent. But you would come to the same exact conclusion as you did; that there is a significant difference between low tax and high tax areas.
The reason that you would come to this conclusion in this hypothetical scenario is due to a few factors. For one, you are assigning the same significance to a 40% tax rate as a 45% tax rate, and the same significance to a 49% tax rate as a 53% tax rate; both of which are fundamental errors, for obvious reasons. Second, you would be completely excluding all data in the 47% range, which is also a fundamental error for obvious reasons.
This is very similar to the issue that is present with this study regarding concussion rate and altitude. Look at the actual data, plotted on a chart:
The correlation is slightly negative between their x and y variable. However, by separating the variables into two bins that were selected by a completely arbitrary method, they were able to come to a conclusion that was completely out of line with the actual correlation between the two variables. And unlike your analysis, they did not completely exclude some variables. What you are doing is slightly worse than what they did in this flawed study.
Now, the conclusion that is drawn from your methodology is not as far from the truth as the conclusion drawn from your methodology in my hypothetical scenario. In fact, the conclusion that you came to is pretty close to the conclusion that is drawn from the two R^2 values in the correlation; which is to say that there probably is a correlation, but it is not extreme, or severe.
But does this not show you the clear flaw with your methodology?