|Links||Annotated MM Literature||About the Author|
Multilevel modeling (MM) is a family of statistical procedures that try to come to terms with influences that are located on different, well, levels. So naturally the question arises what is meant by "level".
One way to think about it is as follows: People do not live entirely on their own, but rather embedded in social units. Even though today, in a globalized world, we may say that people have relationships with other people all over the world, most people have some relationships that are more special than others. People who are linked together via special relationships frequently communicate among each other, and thus the possibility rises that the people you are linked to influence your views. So we may think about the individuals as one (the lowest) level and their network (whether it consists of people that are met in person or of people communication with whom may take place only via artifical media) as a next (higher) level.
(Note that "low" and "high" are just labels; we may well think about things the other way round. "High" just means something like "aggregate"; that is, several individuals – entities on the "low" level – are seen as agglutinated).
A second way: Opportunities structure the behaviour of individuals, and as many people select their opportunities by local proximity, the region in which a person lives may enhance or restrict his or her opportunity. For instance, if a person lives in a region with high unemployment, this may influence his or her behaviour about acceptable wage levels when looking for a new job.
A third way: Often people, be it voluntarily or not, are subject to common external influences. Take, for instance, a university professor. All the students that come to him or her are subject to her or his way of teaching. Could be that this way of teaching influences these students (even though this certainly – if sometimes fortunately – happens less frequently than we professors might desire). Again, we may think of a multitude of professors as the "higher" level units and of their many students as the "lower" level units.
For a variety of reasons, data referring to more than one level often cannot be analyzed by conventional statistical models. For instance, classical OLS regression analysis requires that residuals from individual observations are not correlated. This requirement becomes doubtful if these individual observations are subject to the same influences or are related to each other in other ways.
After so many words, this page – for the time being – does very little: It provides a few links to MM related pages, and it also provides selected references to the literature, with short comments.
Regrettably, updates of this page are quite limited:
Dear reader, the following remarks are intended to serve as a guide to the growing literature on multilevel modelling. They cover most of the textbooks and many of the "classical" papers by those statisticians who did most of the pathbreaking work.
Regrettably my overview covers only the years up to 1999 (even though I will try to stay up to date as far as textbooks are concerned)! Of course, since that time many useful new papers have appeared, but even though I take an occasional glance at some of them, I do not find the time to include them here. Please accept my apologies!
Note that the Centre for Multilevel Modeling now has a similar website that introduces you particularly to newer, and also to more specialised, books. (However, it does not treat the texts in German that are covered below).
The subtitle runs through the book from the beginning. Bickel of course is right. But in my view the book sometimes is cumbersome to read, as it is rather wordy, which Bickel admits himself. On the other hand, if you like long and wordy explanations, this may be the book of your choice. What's really nice is the short overwiew of useful resources at the end of each section. Another great feat, even though I think this is not the effect intended by Bickel: The examples of how to do multilevel analysis via SPSS's menu system that are scattered over the entire book make it clear how any understanding of what you're doing is utterly obfuscated by doing statistics this way (i.e. using menus instead of commands written in plain text).
Some like it French: An introductory paper on multilevel analysis of educational achievement, with good data. No maths.
Very good for those who wish to arrive at an advanced understanding.
A very careful introduction that demonstrates how multilevel models can be estimated and explored with Stata (version 11). Material used in this book (datasets, do files) can be downloaded here as a zip file.
Almost everything you always wanted to know about multilevel modeling.
A good overview of the basic ideas and of applications, with most emphasis on random-coefficient models.
This book is specifically useful for those who want to analyze school data (and have to resort to books in German). However, readers of chapter 3 (dealing largely with centering) should consult chapter 5.2 in Kreft/de Leeuw 1998 and the paper by Kreft/de Leeuw/Aiken 1995.
The wide coverage of this book has much to recommend it - for users who already have acquired an elementary understanding.
Not surprisingly, after an introduction to conventional multilevel modeling (via R) this book demonstrates the application of Bayesian methods to this field using BUGS. It is a pity that doing it the Bayesian way is so self-evident for the authors that they forget to convince the readers of the advantages of Bayesian analysis.
A more advanced introduction by one of the " fathers" of multilevel modeling. The first edition appeared in 1987. The manuscript of the second edition can be downloaded here for free. Some material concerning the third edition, not least some errata, may be found here.
This book is the expanded and updated version of an earlier book, Multilevel Analysis, Amsterdam: TT Publishers, 1995, which is downloadable for free (attention, this is a PDF file with a size of several MB).
This introduction to a special issue of SMR gives a brief overview of problems resulting from using traditional approaches for multilevel data, of the basic structure of the random coefficient model, of available software, and of current problems and possible future developments.
A solid review of the state of the art.
A brief introduction that should be found in most German sociology libraries, but (naturally) cannot replace a textbook.
This book, now downloadable, is certainly the most helpful textbook for those with a strong dislike of maths, formal derivations and the like. The reader is carefully guided through a number of examples. However, one should be aware that there are many advanced topics that are not dealt with in this book.
This should definitely be the source for beginners who need a text in German. What I like particularly is the author's approach to start with detailed exploratory analyses and only then to proceed to more advanced modeling strategies. Complex mathematical detail (for instance, concerning estimation methods) is omitted, but the practical implications are dealt with in a fairly comprehensive manner. Examples use MLA, a free software for estimating multilevel models.
A website provides data sets and command files used in this book as well as links to software and other sources.
A concise introduction for German readers.
In this book, statistical reasoning is paramount, but it is also applied to several datasets with helpful discussions of the results.
Like all other volumes in this series, this is a nice and accessible introduction which, due to the usual limitations of space, cannot discuss all the complications that may arise during a particular analysis.
Part I gives an overview of multilevel modeling, with special chapters for growth curve models, cross-classified random effects models and models for categorical dependent variables. Part II deals with assessment of model fit and discusses issues of power, sample size and design. Part III discusses extensions of the multilevel framework to issues like meta-analysis and measurement modeling. Part IV, finally, presents chapters about how to report results, about available software and about estimation procedures. All chapters are written in an accessible style.
A fine brief introduction, but I feel that it is somewhat infortunate that MM is introduced via an example where the gains from using a MM approach are almost nil (statistically speaking, i.e. the MM analysis barely differs from OLS regression results).
Although focussing on one specific application of multilevel modelling, i.e. within-individual change, this book certainly is worthwhile reading. – There is also a website for this book with data sets, errata and other stuff.
This is a very thorough introduction that requires quite some effort on part of the student - but it pays. Especially chapters 8 and 9 should be studied carefully, as I have found no comparable discussion of heteroscedascity and the basic assumptions (and how to check them) of multilevel modeling. There is a website with some material pertaining to this book, not least some additional remarks and some corrections.
The second edition appeared in 2011, with material for this edition to be found here.
This paper introduces a variance component (or "random intercept" and "random slope") model based on a Maximum Likelihood algorithm. Results of this model on a school data set are contrasted to some individual level models (including one with context effects) and to an aggregate level analysis.
See statisticians work. This book is largely about how to assess school effectiveness. Some chapters give you a good idea of what Empirical Bayes estimation is about.
A fine paper with a very clear structure, demonstrating the possibilities of multilevel modeling in the analysis of longitudinal data.
Discusses nicely some of the limitations of traditional contextual analysis and then proceeds to elaborate at some length the statistical properties of different estimation methods for RC models. These methods eventually are applied to a Dutch school example.
Illuminating discussion of some implications of multilevel modeling.
Technical discussion (but without any maths) of inferential issues in multilevel modelling; pleas for more frequent use of Markov-Chain Monte Carlo techniques.
This paper develops a penalized quasi-likelihood (PQL) method for the estimation of multilevel models with a dichotomous dependent variable. It demonstrates that this method is superior to the earlier marginal quasi-likelihood approach, especially in the case of extreme data constellations such as a very skewed dependent variable.
This paper highlights some of the problems of comparing schools; see also Goldstein & Spiegelhalter (1996) for more extensive discussion of these topics
Despite its title, this paper is basically a rather formal exposition of the estimation theory in multilevel models, accessible only to readers with substantial statistical training.
A newer paper on centering.
The see-saw effect is a difference -- perhaps even an (apparent) contradiction -- between individual and aggregate level effects, as for instance when a variable (say, sex) has a positive slope on the individual level, but the same attribute measured on the aggregate level (the proportion of girls in a class) has a zero or even negative slope. This paper discusses the conditions under which such effects may occur and in addition propagates random coefficient models as procedures to get appropriate estimates of standard errors, appropriate specifically when compared to the "Cronbach procedure" that is discussed as a contrast.
Mandatory reading for all who want to know more about centering of variables.
A sociological example with a binary dependent variable. Focus is not on statistical basics, but on substantive meaning of multilevel vs. single-level modelling.
A useful substantive application, especially for those who are tired of school studies ....
This is an applied paper dealing mainly with the substantive results of multilevel (random coefficient) modeling.
Another substantive application that does not focus on school data.
Focusses on substantive implications; no maths.
A short presentation of a random coefficient model with applications on a school data set to demonstrate the different variance components and cross-level interaction.
Main focus here is on substantive interpretation and discussion of theoretical implications.
This paper first discusses the problems of drawing inferences from aggregate level analysis at individual level relationships (cross level inference). Next, discussions about the appropriate level of analysis (individual or aggregate) are summed up. Another section discusses cross-level effects, that is, possible effects of aggregate level on individual level characteristics. Finally, some statistical procedures for a multilevel analysis are discussed, such as the "slope-as-outcomes" strategy developped by Burstein et al. in other papers. This paper is quite thorough, but probably of interest mainly to those with a strong interest in the developments that occurred during the 1970s.
Compares three strategies of analysis, (1) a Cronbach analysis (see Cronbach & Webb 1976), (2) a Slopes-as-Outcomes analysis, and (3) a Keesling-Wiley procedure. Concludes that all three procedures can give misleading estimates of class effects on mean class outcomes, but that they may also indicate sources of model mis-specification.
This is sort of a "transitional" document. It discusses and criticizes at great length the "context analysis" of the 1970s and earlier 1980s, and then proceeds to present the more recent approaches by Goldstein, Bryk/Raudenbush, or Longford, which at that time were in an early stage, especially computer-wise.
This small volume is a careful and elementary introduction to modeling individual and aggregate level effects, but only in the final chapter random coefficient models are mentioned briefly and the problems inherent in the older approach are not really discussed. Iversen seems to imply (even though I am not sure whether he says so explicitly) that smaller standard errors for estimated coefficients are superior to larger standard errors (see p. 58 et seq.), but of course this is only true if the smaller errors are indeed valid, which in this case is very doubtful.
This page is a process initiated and maintained by
Prof. Dr. Wolfgang Ludwig-Mayerhofer
Universität Siegen / University of Siegen
Philosophische Fakultät – Soziologie /Faculty of Arts – Sociology
Homepage at the University of Siegen
This e-mail address is a graphics file in order to protect me from spam mails. Sorry for the inconvenience – you have to type it into your email client.
Top of page | Last update: 25 May 2017