... the nitty-gritty of the math ...
A little bit less nitty-gritty and more conceptual.
There's a lot of techniques you can use to get a number. A simple example is to take the average total for Celtics games with Thomas, and the average total for Celtics games without Thomas and take the difference. An even simpler (but stupid) option is to just say the number is 0. Instinctively we want to say that the difference of averages is better, but why is that better, and is that the best option?
(If you just want a way to calculate a number, you can use the difference of averages and stop here. Heck, you could just use 0, but we all know that's silly.)
So, really, the question is really either, what's a way to predict the impact of Thomas sitting that's "good enough," or what's the "best" way to predict the impact of Thomas sitting. Without more insight, "good enough" and "best" are pretty vague notions here.
So let's take a step back. When you ask a question like "what's the numerical impact of X", you probably already have some kind of formula in mind that you want to use to predict point totals going forward. That formula is going to include things that you can measure directly or see in advance (for example, whether Thomas sits) and things that you can only get at indirectly (like how much it matters whether Thomas sits).
So what we can do - at least in principle - is we go back to every game in the past and write out the formula for it, filling in the things we know directly, and leaving the things we can only get at indirectly as variables. This gives us a "prediction" (in terms of those variables) for the point total in each game.
Then we start working out values for the variables so that the predictions in total for those past outcomes are as good as possible. (That means that we need to make up a 'goodness formula' too.)
It turns out that this is particularly easy, or works particularly well for certain kinds of 'prediction formulas' and 'goodness formulas'. For one of the simplest examples you can find youtube videos about 'least squares regression'. One part of the math knowledge is knowing the kind of prediction formulas that work well.
A convenient thing is that you can use the 'goodness formula' to estimate how well your prediction formula works. However, that's dangerous because we made up the prediction formula and the goodness formula so we can't really be sure that they'll work well for new data points, and it's also dangerous because there might be something important that's different between the data points we have, and the situation we're trying to predict.
So, in addition to working out how to deal with particular kinds of prediction formulas, people also work on identifying overly naive predictions. This part of the math is more subtle, but can help you avoid overconfidence.