We next explain how to fit multilevel models in Bugs, as called from R. We illustrate with several examples and discuss some general issues in model fitting and tricks that can help us estimate multilevel models using less computer time. We also present the basics of Bayesian inference (as a generalization of the least squares and maximum likelihood methods used for classical regression), which is the approach used in problems such as multilevel models with potentially large numbers of parameters.

Appendix C discusses some software that is available to quickly and approximately fit multilevel models. We recommend using Bugs for its flexibility in modeling; however, these simpler approaches can be useful to get started, explore models quickly, and check results.

Introduction

Generalized linear modeling is a framework for statistical analysis that includes linear and logistic regression as special cases. Linear regression directly predicts continuous data y from a linear predictor Xβ = β0 + X1β1 + ⋯ + Xkβk. Logistic regression predicts Pr(y = 1) for binary data from a linear predictor with an inverselogit transformation. A generalized linear model involves:

1. A data vector y = (y1, …, yn)

2. Predictors X and coefficients β, forming a linear predictor Xβ

3. A link function g, yielding a vector of transformed data ŷ = g−1(Xβ) that are used to model the data

4. A data distribution, p(y|ŷ)

5. Possibly other parameters, such as variances, overdispersions, and cutpoints, involved in the predictors, link function, and data distribution.

The options in a generalized linear model are the transformation g and the data distribution p.

• In linear regression, the transformation is the identity (that is, g(u) ≡ u) and the data distribution is normal, with standard deviation σ estimated from data.

• In logistic regression, the transformation is the inverse-logit, g−1(u) = logit−1(u) (see Figure 5.2a on page 80) and the data distribution is defined by the probability for binary data: Pr(y = 1) = ŷ.

We now go through the steps of understanding and working with multilevel regressions, including designing studies, summarizing inferences, checking the fit of models to data, and imputing missing data.

Now that we can fit multilevel models, we should consider how to understand and summarize the parameters (and important transformations of these parameters) thus estimated.

Inferences from classical regression are typically summarized by a table of coefficient estimates and standard errors, sometimes with additional information on residuals and statistical significance (see, for example, the R output on page 39). With multilevel models, however, the sheer number of parameters adds a challenge to interpretation. The coefficient list in a multilevel model can be arbitrarily long (for example, the radon analysis has 85 county-level coefficients for the varying-intercept model, or 170 coefficients if the slope is allowed to vary also), and it is unrealistic to expect even the person who fit the model to be able to interpret each number separately. We prefer graphical displays such as the generic plot of a Bugs object or plots of fitted multilevel models such as displayed in the examples in Part 2A of this book.

Our general plan is to follow the same structures when plotting as when modeling. Thus, we plot data with data-level regressions (as in Figure 12.5 on page 266), and estimated group coefficients with group-level regressions (as in Figure 12.6). More complicated plots can be appropriate for non-nested models (for example, Figure 13.10 on page 291 and Figure 13.12 on page 293). More conventional plots of parameter estimates and standard errors (such as Figure 14.1 on page 306) can be helpful in multilevel models too.

Once data and a model have been set up, we face the challenge of debugging or, more generally, building confidence in the model and estimation. The steps of Bugs and R as we have described them are straightforward, but cumulatively they require a bit of effort, both in setting up the model and checking it—adding many lines of code produces many opportunities for typos and confusion. In Section 19.1 we discuss some specific issues in Bugs and general strategies for debugging and confidence building. Another problem that often arises is computational speed, and in Sections 19.2–19.5 we discuss several specific methods to get reliable inferences faster when fitting multilevel models. The chapter concludes with Section 19.6, which is not about computation at all, but rather is a discussion of prior distributions for variance parameters. The section is included here because it discusses models that were inspired by the computational idea described in Section 19.5. It thus illustrates the interplay between computation and modeling which has often been so helpful in multilevel data analysis.

Debugging and confidence building

Our general approach to finding problems in statistical modeling software is to get various crude models (for example, complete pooling and no pooling, or models with no predictors) to work and then to gradually build up to the model we want to fit.

Causal inference using regression has an inherent multilevel structure—the data give comparisons between units, but the desired causal inferences are within units. Experimental designs such as pairing and blocking assign different treatments to different units within a group. Observational analyses such as pairing or panel study attempt to capture groups of similar observations with variation in treatment assignment within groups.

Multilevel aspects of data collection

Hierarchical analysis of a paired design

Section 9.3 describes an experiment applied to school classrooms with a paired design: within each grade, two classes were chosen within each of several schools, and each pair was randomized, with the treatment assigned to one class and the control assigned to the other. The appropriate analysis then controls for grade and pair.

Including pair indicators in the Electric Company experiment. As in Section 9.3, we perform a separate analysis for each grade, which could be thought of as a model including interactions of treatment with grade indicators. Within any grade, let n be the number of classes (recall that the treatment and measurements are at the classroom, not the student, level) and J be the number of pairs, which is n/2 in this case. (We use the general notation n, J rather than simply “hard-coding” J = n/2 so that our analysis can also be used for more general randomized block designs with arbitrary numbers of units within each block.)

Aim of this book

This book originated as lecture notes for a course in regression and multilevel modeling, offered by the statistics department at Columbia University and attended by graduate students and postdoctoral researchers in social sciences (political science, economics, psychology, education, business, social work, and public health) and statistics. The prerequisite is statistics up to and including an introduction to multiple regression.

Advanced mathematics is not assumed—it is important to understand the linear model in regression, but it is not necessary to follow the matrix algebra in the derivation of least squares computations. It is useful to be familiar with exponents and logarithms, especially when working with generalized linear models.

After completing Part 1 of this book, you should be able to fit classical linear and generalized linear regression models—and do more with these models than simply look at their coefficients and their statistical significance. Applied goals include causal inference, prediction, comparison, and data description. After completing Part 2, you should be able to fit regression models for multilevel data. Part 3 takes you from data collection, through model understanding (looking at a table of estimated coefficients is usually not enough), to model checking and missing data. The appendixes include some reference materials on key tips, statistical graphics, and software for model fitting.

We now introduce multilevel linear and generalized linear models, including issues such as varying intercepts and slopes and non-nested models. We view multilevel models either as regressions with potentially large numbers of coefficients that are themselves modeled, or as regressions with coefficients that can vary by group.

This chapter describes a variety of ways in which probabilistic simulation can be used to better understand statistical procedures in general, and the fit of models to data in particular. In Sections 8.1–8.2, we discuss fake-data simulation, that is, controlled experiments in which the parameters of a statistical model are set to fixed “true” values, and then simulations are used to study the properties of statistical methods. Sections 8.3–8.4 consider the related but different method of predictive simulation, where a model is fit to data, then replicated datasets are simulated from this estimated model, and then the replicated data are compared to the actual data.

The difference between these two general approaches is that, in fake-data simulation, estimated parameters are compared to true parameters, to check that a statistical method performs as advertised. In predictive simulation, replicated datasets are compared to an actual dataset, to check the fit of a particular model.

Fake-data simulation

Simulation of fake data can be used to validate statistical algorithms and to check the properties of estimation procedures. We illustrate with a simple regression model, where we simulate fake data from the model, y = α + βx + ∊, refit the model to the simulated data, and check the coverage of the 68% and 95% intervals for the coefficent β.

There are generally many options available when modeling a data structure, and once we have successfully fit a model, it is important to check its fit to data. It is also often necessary to compare the fits of different models.

Our basic approach for checking model fit is—as we have described in Sections 8.3–8.4 for simple regression models—to simulate replicated datasets from the fitted model and compare these to the observed data. We discuss the general approach in Section 24.1 and illustrate in Section 24.2 with an extended example of a set of models fit to an experiment in animal learning. The methods we demonstrate are not specific to multilevel models but become particularly important as models become more complicated.

Although the methods described here are quite simple, we believe that they are not used as often as they could be, possibly because standard statistical techniques were developed before the use of computer simulation. In addition, fitting multilevel models is a challenge, and users are often so relieved to have successfully fit a model with convergence that there is a temptation to stop and rest rather than check the model fit. Section 24.3 discusses some tools for comparing different models fit to the same data.

Posterior predictive checking is a useful direct way of assessing the fit of the model to various aspects of the data. Our goal here is not to compare or choose among models but rather to explore the ways in which any of the models being considered might be lacking.

Simple methods from introductory statistics have three important roles in regression and multilevel modeling. First, simple probability distributions are the building blocks for elaborate models. Second, multilevel models are generalizations of classical complete-pooling and no-pooling estimates, and so it is important to understand where these classical estimates come from. Third, it is often useful in practice to construct quick confidence intervals and hypothesis tests for small parts of a problem—before fitting an elaborate model, or in understanding the output from such a model.

This chapter provides a quick review of some of these methods.

Probability distributions

A probability distribution corresponds to an urn with a potentially infinite number of balls inside. When a ball is drawn at random, the “random variable” is what is written on this ball.

Areas of application of probability distributions include:

• Distributions of data (for example, heights of men, heights of women, heights of adults), for which we use the notation yi, i = 1, …, n.

• Distributions of parameter values, for which we use the notation θj, j = 1, …, J, or other Greek letters such as α, β, γ. We shall see many of these with the multilevel models in Part 2 of the book. For now, consider a regression model (for example, predicting students' grades from pre-test scores) fit separately in each of several schools. The coefficients of the separate regressions can be modeled as following a distribution, which can be estimated from data.

• […]

We start with an overview of classical linear regression and generalized linear models, focusing on practical issues of fitting, understanding, and graphical display. We also use this as an opportunity to introduce the statistical package R.

Chapter 9 discussed situations in which it is dangerous to use a standard linear regression of outcome on predictors and an indicator variable for estimating causal effects: when there is imbalance or lack of complete overlap or when ignorability is in doubt. This chapter discusses these issues in more detail and provides potential solutions for each.

Imbalance and lack of complete overlap

In a study comparing two treatments (which we typically label “treatment” and “control”), causal inferences are cleanest if the units receiving the treatment are comparable to those receiving the control. Until Section 10.5, we shall restrict ourselves to ignorable models, which means that we only need to consider observed pre-treatment predictors when considering comparability.

For ignorable models, we consider two sorts of departures from comparability—imbalance and lack of complete overlap. Imbalance occurs if the distributions of relevant pre-treatment variables differ for the treatment and control groups. Lack of complete overlap occurs if there are regions in the space of relevant pre-treatment variables where there are treated units but no controls, or controls but no treated units.

Imbalance and lack of complete overlap are issues for causal inference largely because they force us to rely more heavily on model specification and less on direct support from the data.