Working Through Hogg's Mathematical Statistics
I've used this textbook for years when I need to actually understand why a statistical result works instead of just knowing the name. The Hogg book — Introduction To Mathematical Statistics Hogg — sits somewhere between an undergraduate bridge course and a graduate-level proof-based text. It's rigorous enough to force you to derive things but accessible enough that you don't need measure theory as a prerequisite on day one. The book's approach is deliberately deductive. It starts with probability foundations and builds straight into point estimation, confidence intervals, hypothesis testing, and regression from first principles. You'll encounter the factorization theorem early, Rao-Blackwell later, and Lehmann-Scheffé after that. The proofs are presented cleanly, which is either exactly what you want or slightly frustrating if you prefer intuition-first material. I prefer it, but I know people who switch to Casella and Berger mid-course because they need more motivation before the rigor lands.
Where Most People Get Stuck
The chapter on sufficient statistics and exponential families is where the book earns its reputation. The factorization theorem seems simple until you're asked to recognize a sufficient statistic for a non-standard distribution, and that's when most students stall. I spent an afternoon once trying to verify sufficiency for a curved exponential family where the parameter space restricted the natural parameter to a parabola. The textbook examples don't cover this. What worked was going back to the definition directly and showing that the conditional distribution of the sample given the proposed statistic didn't depend on the parameter. It took longer, but it made the factorization theorem's boundary conditions clear in a way the proof in the book doesn't explicitly highlight. UMVUE derivation is another area that trips people up. The Lehmann-Scheffé theorem gives you a clean path: find a complete sufficient statistic, then find any unbiased estimator and condition on it. The theorem itself is straightforward. The difficulty is knowing when completeness actually holds and recognizing it quickly enough during an exam. For the normal distribution with unknown mean and variance, completeness of (X-bar, S^2) is standard material. For certain non-i.i.d. setups or truncated distributions, you have to verify it yourself. I've seen students lose points not because they misapplied the theorem but because they assumed completeness without checking the support condition. The book mentions this but doesn't drill it hard enough for a first pass. The hypothesis testing chapters cover Neyman-Pearson lemmas and UMP tests thoroughly. The counter-intuitive part that beginners miss: UMP tests rarely exist beyond one-parameter exponential families with monotone likelihood ratios. Once you move to two-sided alternatives or composite hypotheses in multiple dimensions, you're usually looking at UMPU tests or admissible procedures, not uniformly most powerful ones. The book does discuss this but the implication isn't always clear on first reading. You won't find a UMP test for H0: mu = 0 versus H1: mu != 0 in a normal variance-unknown setting. You'll find a UMPU test based on the t-statistic. That distinction matters when you're actually doing inference and someone asks whether your test is optimal.
How I Actually Use This Book
I don't read it cover to cover. I use it as a reference when I need to verify a derivation or understand the conditions behind a standard result. The asymptotic theory section is genuinely useful — the treatment of consistency, asymptotic normality, and the delta method is concise and correct. I also return to the Chapter on likelihood inference because the book handles profile likelihoods and invariant tests better than most comparable texts at this level. For homework or self-study, work through the examples before the exercises. The worked examples in Hogg are carefully chosen and often contain the exact technique needed for the problems. The exercise set ranges from mechanical computation to proof-based questions. If you can do about 70% of the exercises without looking at solutions, you've internalized the material. If you're below 40%, you need to rework the probability foundations chapters first. There's a freely available PDF floating around on various academic repositories. I don't have a current link to verify, and sharing direct download URLs for copyrighted material isn't something I'd do. The book is widely available through university libraries, and the sixth edition is the current standard version with updated exercise sets and some reorganized chapters on multivariate methods.
Get the Full Details
Limitations and When to Look Elsewhere
The book has real weaknesses. It doesn't cover bootstrap methods in any depth. The treatment of Bayesian statistics is minimal — basically a single chapter that introduces conjugate priors and posterior inference without connecting to modern computational approaches. If your work involves MCMC or decision-theoretic frameworks, you'll need supplementary material regardless of how thoroughly you master Hogg. The regression chapter is adequate but thin compared to a dedicated text like Kutner, Nachtsheim, Neter, and Li. For applied work with real datasets, the theoretical focus here will leave gaps. The pacing assumes you can handle proofs at a reasonable speed. Students who are stronger computationally than theoretically sometimes struggle through the first third of the book. That's fine — it's a mathematical statistics text, not an applied one. If your goal is to run analyses rather than derive properties, consider pairing this with a more hands-on resource or using it selectively for the theory chapters while skipping or skimming the heavier proof sequences. The sixth edition fixed several errata from earlier printings but introduced new ones. Always cross-reference tricky results with published corrigenda or discussion forums when a derivation seems off. I caught a typo in the sufficient statistic example for the Weibull distribution in my copy — the exponent on x was wrong in one place, which propagated through a subsequent exercise. These errors are minor but annoying when you're three hours into a problem set.
If you're working through this book seriously, expect to spend roughly two to three hours per chapter on a first read if you're doing the derivations yourself. The exercises in the later chapters on sequential analysis and nonparametric methods can take significantly longer. That's normal. The material rewards the time investment, but it's not efficient reading. You learn it by doing it, not by absorbing it passively.
