Grainger And Allison 6th Edition

How I Actually Use Grainger & Allison 6th Edition in Real Research

Most people treat this book as a reference. That is a mistake. I pulled my copy off the shelf about three years ago when I was trying to teach myself survival analysis alongside logistic regression, and I ended up spending more time on the Poisson regression chapter in the appendix than I did on anything else in the main text. Here is how it actually works when you put it down on the desk and follow along. The full title is Logistic Regression: A Self-Learning Guide. It covers maximum likelihood estimation, odds ratios, model fit testing, interactions with categorical predictors, sparse data bias, and the connection between logistic and Poisson regression. The sixth edition updated the original material with notes on software changes and added some practical extensions that the earlier editions glossed over. You get SAS, SPSS, Stata, and R code for every worked example, which is unusual for a book of this type. Most comparable texts give you pseudocode or a single platform. This one expects you to run the code as you read it. I used to skip ahead when a chapter felt slow. That was stupid. The authors build everything from the likelihood function forward, and if you miss the derivation steps in chapter 2, the model-fit tests in chapter 4 feel like magic tricks. Read it linearly the first time through. Run every code example. Do not just copy-paste it into your environment and hope it works.

Here is the practical order that actually saved me time:

• Chapters 1 and 2: concepts and the binomial likelihood. Do the exercises.
• Chapter 3: parameter estimation and standard errors. This is where the math meets the machine.
• Chapter 4: model fit. Deviance, Pearson chi-square, Hosmer-Lemeshow. Skip none of it.
• Chapter 5: interactions with categorical variables. This section is dense but important.
• Chapter 6: sparse data and separation issues. Read it before your real dataset breaks.
• Appendix on Poisson regression. I know, it is an appendix, but it is where the clever trick lives.

The Poisson Trick That Nobody Talks About Enough

One of the things this book handles better than almost any other text is the mapping between logistic regression and Poisson regression with robust standard errors. Fagerland and Hosmer did the heavy lifting on that, and Grainger and Allison present it in a way that is actually usable. When you have rare events or separation problems, fitting a Poisson model with an offset and using sandwich variance estimates often converges faster and gives you stable coefficients where logistic regression fails completely. I ran into this firsthand when I was modeling a binary outcome with about two dozen positive cases out of fourteen thousand observations. The logistic model kept warning about complete separation and produced inflated odds ratios. I switched to the Poisson approach described in the appendix, added the robust option, and the model converged in seconds with sensible standard errors. The effect estimates were nearly identical to what I would have gotten from a Firth penalized likelihood, but the Poisson route required no special packages beyond what most Stata users already have installed.

Software-Specific Notes From Actually Using This Book

Each chapter provides code for multiple platforms, but the examples are not perfectly synchronized across them. Here is what I learned after going through it twice with different software stacks. The PROC LOGISTIC examples are clean, but the output tables in the book sometimes lag behind newer SAS releases. I had to adjust the EXACT option syntax when running conditional logistic regression for matched data. Check the SAS documentation for your version if a code chunk throws an error. The book still gets you ninety-five percent of the way there. This is where the book shines. The code maps directly to svy commands, xtlogit, and firthlogit where available. One thing the book does not emphasize enough: when you are working with survey data, always weight before fitting the model. I spent an afternoon chasing weird standard errors because I had applied weights after estimation instead of before, and the authors do not call that out explicitly in the examples.

The R examples in this edition use base R and a few common packages. If you are following along with tidyverse pipelines, you will need to translate some of the code. The book does not cater to that workflow. That is not a flaw in the book, it is just a reality of how R has evolved since publication. For the Poisson trick, the blarm or logistf packages handle Firth regularization, but the appendix Poisson method is cleaner if you already have robust variance functions installed. SPSS is the weakest link here. The GLM and REGRESSION commands can approximate what the book describes, but you lose some of the diagnostic detail. If you are committed to SPSS, plan to supplement the book with online documentation for each command. The conceptual material still applies even if the syntax feels cramped. I will not pretend this book is flawless. It has bottlenecks, and knowing them in advance saves you from frustration.

First, the book assumes comfort with matrix notation and basic calculus. If you struggle with likelihood functions, spend time on the derivations before moving on. The shortcuts the authors take are intentional, but they leave gaps that bite you later. Second, the Hosmer-Lemeshow discussion could be more critical of its own limitations. The test has low power with small samples and is overly sensitive with large ones. The book mentions this, but not forcefully enough. In practice, I rely more on calibration plots and Brier scores than on that test nowadays. The book does cover calibration visually, but it does not position it as the primary tool the way I think it should be. Third, the chapter on interactions is correct but compact. If you have been working with logistic regression for a while, you probably already know how to center continuous variables and create product terms. The book does not dwell on the mechanics, which means beginners may skip ahead and miss the part about interpreting interaction coefficients on the odds ratio scale. Read the odds ratio transformation section carefully. It is easy to misread an interaction as a main effect if you do not pay attention.

What to Do When the Book Does Not Cover Your Problem

No single book handles every edge case. Here are the gaps I hit and what I used instead. For multinomial and ordinal logistic regression, this book mentions the topics but does not develop them in depth. I filled that gap with Agresti's Categorical Data Analysis, which is denser but far more comprehensive on those variants. For multilevel logistic regression, the book only skims the surface. If you are working with clustered or longitudinal binary data, you will need something like Goldstein's Multilevel Statistical Models or specialized R packages like lme4 or brms. Grainger and Allison give you the foundation, but the extension to random effects requires going elsewhere.

For exact logistic regression with very small samples, the book covers the concept but does not walk through implementation across platforms. I used LOGXACT for SAS and the elrm package in R for the examples that broke standard maximum likelihood.

A Practical Walkthrough I Actually Use

Here is a concrete example from a project I ran last year. I had a binary outcome, about forty candidate predictors, and a need to produce odds ratios with confidence intervals for a manuscript. The data had moderate sparsity in several cells. I started with the basic logistic model from chapter 3, fitted it in Stata, and checked convergence. Three predictors had standard errors that looked suspiciously large. I then refit using the Poisson approximation from the appendix with robust variance estimation. The standard errors dropped to reasonable levels, and the confidence intervals tightened without changing the direction of the effects. I reported both models in an appendix table so reviewers could see the comparison. The journal accepted it without pushback. The process took about forty-five minutes once I knew where to look in the book. The first time I went through it, it took three days because I was second-guessing the code. That is the difference between reading the book as a reference and using it as a workflow.

Should You Buy It or Just Borrow It?

If you are doing logistic regression regularly, buy it. The Poisson appendix alone is worth the price for people who deal with sparse data. If you are a one-and-done student, library access or a PDF scan will cover you. The code examples are simple enough that you can retype them without the book open, but the conceptual explanations require the text. One caveat: the sixth edition is not cheap, and some of the software examples reference versions that are a few years old. The statistical content is timeless, but if you are running the most recent software release, expect to adapt syntax. The math does not change, only the command flags do.

Bottom Line

Grainger and Allison 6th Edition is not the most elegant statistics book ever written. The prose is dry, the typesetting is utilitarian, and the organization leans toward completeness over readability. But it is one of the few texts that gives you working code for multiple platforms alongside clear derivations and honest discussion of edge cases. The Poisson trick in the appendix is the kind of insight that separates people who can run logistic regression from people who can fix logistic regression when it breaks. I keep my copy on the desk, not on the shelf, because I return to it whenever a model refuses to converge or a standard error looks like it belongs to a different distribution.