In practice, we often find that count data is not well modeled by Poisson regression, though Poisson models are often presented as the natural approach for such data. In contrast, the negative binomial regression model is much more flexible and is therefore likely to fit better, if the data are ...

July marks the end of three years of blogging for us. By our count, we've posted 121 examples across the first three years. We aim to be helpful and interesting.As always, it's hard to get a sense of our readership. At the time we wrote this, Feedbur...

One of us recently read a colleague's first draft of a paper, in which she had written: "All analyses were done in R 2.14.0." We assume we're preaching to the converted here, when we say that the enormous amount of work that goes into R needs to be re...

The assumption of equal variances among the groups in analysis of variance is an expression of the assumption of homoscedasticity for linear models more generally. For ANOVA, this assumption can be tested via Levene's test. The test is a function of the residuals and means within each group, though various ...

A colleague asked for help with randomly choosing a kid within a family. This is for a trial in which families are recruited at well-child visits, but in each family only one of the children having a well-child visit that day can be in the study. The idea is that ...

Nick has a paper in the American Statistician warning about bias in multiple imputation arising from rounding data imputed under a normal assumption. One example where you might run afoul of this is if the data are truly dichotomous or count variables, but you model it as normal (either because ...

In examples 9.30 and 9.31 we explored corrections for multiple testing and then extracting p-values adjusted by the Benjamini and Hochberg (or FDR) procedure. In this post we'll develop a simulation to explore the impact of "strong" and "weak" control of the family-wise error rate offered in multiple comparison corrections. Loosely put, ...

In example 9.30 we explored the effects of adjusting for multiple testing using the Bonferroni and Benjamini-Hochberg (or false discovery rate, FDR) procedures. At the time we claimed that it would probably be inappropriate to extract the adjusted p-values from the FDR method from their context. In this entry we attempt ...

In our book, we discuss the important question of how to assign different parameterizations to categorical variables when fitting models (section 3.1.3). We show code in R for use in the lm() function, as follows:lm(y ~ x, contrasts=list(x,"contr.trea...

A colleague recently asked after a slide suitable for explaining proportional hazards. In particular, she was concerned that her audience not focus on the time to event or probability of the event. An initial thought was to display the cumulative haz...