"impression that more advanced statistics is technical elaboration that doesn't offer major additional insights"

Why did you have this impression?

Sorry for the off-topic, but I see this a lot in LessWrong (as a casual reader). People seem to focus on textual, deep-sounding, wow-inducing expositions, but often dislike the technicalities, getting hands dirty with actually understanding calculations, equations, formulas, details of algorithms etc (calculations that don't tickle those wow-receptors that we all have). As if these were merely some minor additions over the really important big picture view. As I see it this movement seems to try to build up a new backbone of knowledge from scratch. But doing this they repeat the mistakes of the past philosophers. For example going for the "deep", outlook-transforming texts that often give a delusional feeling of "oh now I understand the whole world". It's easy to have wow-moments without actually having understood something new.

So yes, PCA is useful and most statistics and maths and computer science is useful for understanding stuff. But then you swing to the other extreme and say "ideas from advanced statistics are essential for reasoning about the world, even on a day-to-day level". Tell me how exactly you're planning to use PCA day-to-day? I think you may mean you want to use some "insight" that you gained from it. But I'm not sure what that would be. It seems to be a cartoonish distortion that makes it fit into an ideology.

Anyway, mainstream machine learning is very useful. And it's usually much more intricate and complicated than to be able to produce a deep everyday insight out of it. I think the sooner you lose the need for everything to resonate deeply or have a concise insightful summary, the better.

I don't believe you can obtain an understanding of the idea that "correlation does not imply causation" from even a very deep appreciation of the material in Statistics 101. These courses usually make no attempt to define confounding, comparability etc. If they try to define confounding, they tend to use incoherent criteria based on changes in the estimate. Any understanding is almost certainly going to have to originate from outside of Statistics 101; unless you take a course on causal inference based on directed acyclic graphs it will be very challenging to get beyond memorizing the teacher's password

PCA and other dimensionality reduction techniques are great, but there's another very useful technique that most people (even statisticians) are unaware of: dimensional analysis, and in particular, the Buckingham pi theorem. For some reason, this technique is used primarily by engineers in fluid dynamics and heat transfer despite its broad applicability. This is the technique that allows scale models like wind tunnels to work, but it's more useful than just allowing for scaling. I find it very useful to reduce the number of variables when developing models and conducting experiments.

Dimensional analysis recognizes a few basic axioms about models with dimensions and sees what they imply. You can use these to construct new variables from the old variables. The model is usually complete in a smaller number of these new variables. The technique does not tell you which variables are "correct", just how many independent ones are needed. Identifying "correct" variables requires data, domain knowledge, or both. (And sometimes, there's no clear "best" variable; multiple work equivalently well.)

Dimensional analysis does not help with categorical variables, or numbers which are already dimensionless (though by luck, sometimes combinations of dimensionless variables are actually what's "correct"). This is the main restriction that applies. And you can expect at best a reduction in the number of variables of about 3. Dimensional analysis is most useful for physical problems with maybe 3 to 10 variables.

The basic idea is this: Dimensions are some sort of metadata which can tell you something about the structure of the problem. You can always rewrite a dimensional equation, for example, to be dimensionless on both sides. You should notice that some terms become constants when this is done, and that simplifies the equation.

Here's a physical example: Let's say you want to measure the drag on a sphere (units: N). You know this depends on the air speed (units: m/s), viscosity (units: m^2/s), air density (units: kg/m^3), and the diameter of the sphere (units: m). So, you have 5 variables in total. Let's say you want to do a factorial design with 4 levels in each variable, with no replications. You'd have to do 4^4 = 256 experiments. This is clearly too complicated.

What fluid dynamicists have recognized is that you can rewrite the relationship in terms of different variables, and nothing is missing. The Buckingham pi theorem mentioned previously says that we only need 2 dimensionless variables given our 5 dimensional variables. So, instead of the drag force, you use the drag coefficient, and instead of the speed, viscosity, etc., you use the Reynolds number. Now, you only need to do 4 experiments to get the same level of representation.

As it turns out, you can use techniques like PCA on top of dimensional analysis to determine that certain dimensionless parameters are unimportant (there are other ways too). This further simplifies models.

There's a lot more on this topic than what I have covered and mentioned here. I would recommend reading the book Dimensional analysis and the theory of models for more details and the proof of the pi theorem.

(Another advantage of dimensional analysis: If you discover a useful dimensionless variable, you can get it named after yourself.)

What resources would you recommend for learning advanced statistics?