So now four people, all I think rather smarter than I am, have asked me what statistics books to read.
I think these folks are projecting something more than is warranted onto my recent point touching on statistics: They seem to think I know better! Haha! Alas, I am as dumb as a sack of wet hammers when it comes to modern statistics, having been crippled in my youth with an incident involving DEC PDP teletypes and Minitab [yes, I am that old.]. I am still wrestling with the effects of second-hand two-tailed p-value abuse, and though I retain substantial use of my elementary probability theory, I have a seriously diminished capacity to fathom any statistical process involving greek symbols or vertical lines (|).
My one scant advantage is that I recognize that I have a problem. And that I like pretty graphs with lots of squiggly lines and colors, which apparently abound in machine learning texts and papers, and always seem to me to be fraught with powerful and convincing implications about both the model and the explanatory acumen of the authors. [Wait; is that last one an advantage?]
Ah, but lacking a blue handicapped hang-tag for my desk. I have, in fact, paid good money to the esteemed Dr. Shalizi to tell me important statistical stuff [he advised me on a consulting gig]. Perhaps if we all ask him pointedly, he can suggest a reasonable course of action?
I do not teach. But I thought of a way that teaching statistics could be changed over, say, the last 25 years. Perhaps one might organize a class around resampling or bootstrap methods with simulated data. I don’t know if this would lead to improved understanding by students, although I guess Julian Simon used to claim it did.
I doubt that is why there is the variation in textbooks that you have seen. But that is one option on the plate of teaching methods that probably wasn’t available 25 years ago, given the computing power of the time.