Working With Casella and Berger's Statistical Inference
The book is the standard graduate-level text for mathematical statistics. It covers point estimation, hypothesis testing, and interval estimation with a rigorous measure-theoretic foundation. Most programs require it for qualifying exams. You will find it referenced constantly in PhD comps and qualifying exam study guides. Students often search for the PDF because the hardcover runs $120 and the electronic license from the publisher costs similarly. The library reserves at many universities have a copy, and interlibrary loan works if you are patient. I have seen students use scanner apps on campus copies to produce personal study PDFs. That is fine for personal use. Distributing the file or posting it publicly violates copyright. The publisher monitors those channels. The first edition came out in 1990. The second edition followed in 2002. The third edition arrived in 2021. The second edition is the most widely used version in course syllabi. The third edition adds exercises on bootstrap theory and updates notation. You can compare the table of contents online to decide which edition matches your course. If your professor lists chapter numbers and problem ranges, the edition matters less than you think. The core material on unbiasedness, sufficiency, completeness, and the Neyman-Pearson lemma is identical across editions.
I used the second edition while preparing for my comprehensive exam. The difficulty spikes around Chapter 8, where uniformly most powerful tests meet the Karlin-Rubin theorem. Students who skip the monotone likelihood ratio condition end up stuck on problems that require recognizing a family of distributions has MLR. I spent three hours on one problem because I treated the statistic $T(X)$ as arbitrary instead of verifying the ratio property. The workaround was to rewrite the joint density and factor it into $g(T(x)|\theta)h(x)$. Once I identified that factorization, the test followed mechanically. Another edge case appears in Chapter 7 with the Lehmann-Scheffé theorem. The theorem states that a complete sufficient statistic yields a unique minimum variance unbiased estimator. Beginners often assume any unbiased estimator based on a sufficient statistic is automatically best. That assumption fails when completeness is missing. I ran into this on a practice problem involving a uniform distribution on $(0,\theta)$. The sufficient statistic is the maximum order statistic, but it is not complete in the usual parametrization unless you restrict the parameter space. The fix is to verify completeness by checking whether $E_\theta[g(T)]=0$ for all $\theta$ implies $g(T)=0$ almost surely. I wrote a short R script to simulate bias across candidate estimators and confirmed that the textbook estimator based on $(n+1)/n \cdot X_{(n)}$ dominates alternatives. The book assumes familiarity with real analysis. You should know $\epsilon$-$\delta$ arguments, convergence modes, and Fubini-type theorems before opening Chapter 5. If you have only seen probability through combinatorics or simulation, the proof of the Cramér-Rao lower bound will feel like word salad. I recommend skimming a measure-theoretic probability reference like Billingsley before attempting the regularity conditions section. Even a cursory review of almost sure convergence versus convergence in probability saves hours of confusion.
Problem sets are where the book earns its reputation. The exercises range from computational derivations to theoretical proofs. Many students treat the problems as optional. That is a mistake. The exam questions are usually modeled on the end-of-chapter exercises. I learned this the hard way during my first attempt at the qual. The question asked for the asymptotic distribution of a method-of-moments estimator under a nonstandard parameter constraint. I recognized the pattern from Problem 7.32, which required applying the delta method after verifying differentiability of the transformation. The solution took six lines. Without that prior exposure, I would have spent twenty minutes reinventing the expansion. The book has limitations. It does not cover Bayesian inference beyond Chapter 6, and even that chapter is brief. If your program expects Bayesian methods, you will need supplementary readings like Gelman or Robertson. The treatment of empirical likelihood is also thin. For modern resampling techniques, you might pair this text with Efron and Tibshirani. The exposition on rank tests is adequate but not exhaustive. If you need depth on nonparametrics, turn to Lehmann and Datar. Another practical note: the index is useful but not exhaustive. Cross-referencing terms like "Pitman efficiency" or "Rao-Blackwellization" can require searching multiple chapters. I kept a personal glossary mapping key definitions to page numbers. This saved time during exam preparation. You can create a similar document by extracting terms from the notation list and linking them to proof sketches in the margins.
Get the Full Details
If you are using the book for self-study, budget two semesters for full coverage. The pacing assumes a year-long graduate sequence. Attempting to compress it into eight weeks leads to superficial understanding. I advised a student who tried to skim Chapters 4 through 7 in three weeks. She could replicate solutions but could not derive the Cramér-Rao bound from first principles. We restarted with slower pacing, and she gained fluency by week ten. The PDF format is convenient for annotation. I used a tablet stylus to mark proof steps and note alternative approaches. Some students highlight entire pages. That practice obscures the text. I prefer marginal notes that reference related results, such as linking Lehmann-Scheffé to Basu's theorem in Problem 6.17. The habit strengthens retention more than passive highlighting. For instructors, the solutions manual exists but is distributed under embargo. Departments typically obtain it through publisher channels. Individual instructors cannot purchase it without verification of teaching appointment. If you are a student without access, consider forming a study group. Collaboration helps clarify tricky points like the difference between uniformly most powerful and uniformly most powerful unbiased tests. The distinction hinges on the unbiasedness constraint, which appears explicitly in the definition of $\mathcal{U}_\alpha$ in Chapter 8.
The book remains the canonical reference. Its clarity outweighs its gaps. Use it alongside lecture notes that fill in missing context, and you will build a solid foundation for further study in semiparametric methods or high-dimensional inference.
