The first thing I want to say is this: If you plan on doing regression analysis in your research, stop what you are doing, and read this book first. I think this book represents THE current statement on how we should use regression. For Angrist and Pischke, regression is a technology for summarizing data. If regression is to be used for causal inference, then there is nothing in the specification of the model or the choice of estimator that can ultimately make the causal story persuasive. That is, you don't identify causal effects simply by including "control" variables in your regression. The identification comes from elsewhere---either a real or "quasi" experiment---and the regression is what you use to clean up the imperfections of the experiment and measure effects. Angrist and Pischke have done an enormous service to social science by writing a regression textbook that nonetheless emphasizes the primacy of design. This is a terrific corrective for the "101 flavors of regression" approach of textbooks to date.
Even with this emphasis on design, Angrist and Pischke show us that are a lot of nuances to the way that regressions measure such effects---e.g., in the presence of effect heterogeneity---and that's what this book explores in exquisite detail. It's a hugely important book and a very serious and rigorous treatment, despite it's apparently causal style. They make some claims that may strike some as outrageous---e.g., always using OLS, even for limited dependent variables---but the rigor of their presentation means that the onus is on those who disagree to think harder about why, exactly, they would prefer, say, a more parametric approach.
Nonetheless, it isn't a "5 star" book. It often feels a bit rough-draft-like. The presentation of technical material skips important steps rather haphazardly. I wonder if this was due to bad editing? Hopefully there will be a second edition that cleans up these rough edges, in which case it would be the ideal textbook on regression analysis.