We last discussed Sharon Lohr in the context of her article with J. Michael Brick on using the famous 1936 Literary Digest survey as a positive example of statistical adjustment rather than a negative example of non-probability sampling.
In the meantime she came out with a new edition of her classic textbook, Sampling: Design and Analysis. The book also comes with a software supplement in R or Stata; see the webpage for further detail.
One of the challenges of teaching survey sampling or writing a book on the topic is that the design and analysis of sample surveys looks easy but is actually really hard.
Why is it hard? A bunch of reasons. For starters, nonresponse: we very rarely have actual random samples. Then there’s adjusting for nonresponse: even if all adjustment variables are available, it can still be a challenge to fit a model to make the predictions to do the adjustment. Then there’s small-area estimation, time series, and other things that we care about but are tough to estimate. Oh, and did I mention the challenge of adjusting for non-census variables? And then there are those surveys that aren’t even close to random samples, for example obtained by referral sampling. And did I mention that a random sample of a network does not look anything like a baby version of the original? And causal inference from sample data? We could go on and on. Some people have even said that survey weighting is a mess.
But from the outside, survey sampling looks so simple and boring: simple random sampling, stratified sampling, cluster sampling, balls from an urn, etc etc etc. You can fill up a semester with all the simple basics, staying clear of the “there be dragons” areas on the edge of the map.
Lohr’s book has a good balance for a textbook and reference book, in that it presents the basic material clearly, with lots of examples, explanations, and exercises, and in later chapters touches on some of the challenges. She never gets around to Mister P, but it’s good to have something to save for the 4th edition, right? In all seriousness, you can’t cover the world in any one book, and I recommend Lohr’s as a textbook and reference work.
FWIW, I just finished teaching half a semester using this book (the 2nd half is dedicated to observational studies).
One nice feature of the book is that it presents both deign-based and model-based analyses (though the focus is on design-based analysis). The other textbooks I happen to look at at the library only included design-based stuff.
I also rather liked her treatment of non-response, though it was only one chapter–those other textbooks didn’t include any discussion (that I could find, at least)!
Anyway, I share your frustration that the standard treatment focuses on the well-randomized surveys without non-response–something I’ve never actually seen in real life–rather than the more common scenarios (that, to me, are also more interesting)