Customer Reviews

Causality: Models, Reasoning, and Inference

5 star:		(7)
4 star:		(1)
3 star:		(1)
2 star:		(1)
1 star:		(2)

 

 

Judea Pearl and his colleagues at UCLA (and elsewhere) have published a large number of papers and written unpublished reports over the past 15 years, in which they have developed a modern, analytical approach to causation. Many of these are in somewhat obscure publications, and so it is especially helpful to have the most important of them collected together in this volume. Pearl has edited, written new chapters and connecting prose, to weave this summary of a substantial amount of research.

Although the dust-jacket suggests that only modest mathematics is needed, and although this is technically true, it is misleading, because the whole area requires a sophistication of thought that goes well beyond the simplicity of the tools. Nonetheless, there is currently no other volume that is as easy to read as this, and summarizes so much material so compactly.

It is possible that the new vision of causal analysis developed by Spirtes, Scheines, Glymour, Pearl, Robins, Verma, Heckerman, Meek, and others, will have profound effect on how we analyze research data. If so, this book will be necessary reading for decades to come.

Judea Pearl is one of the leading researchers in the topic of causality. What is causality? In the exploration of statistical data we are often able to find relationships or correlations between two variables. We are often tempted to attribute the results of one variable, say A as an outcome (being high or low)that is due to the result (high or low) of the other, say B. We want to say that B is the cause of the outcome of A. Significant correlation by itself only suggests relationships. It cannot tell you whether A causes B or B causes A or neither. Causality is the study of designing experiments to allow you to determine if a relationship has a cause and effect. The subject matter is very philosophical and somewhat controversial. But a lot of research effort has gone into providing mathematical rigor to the concept. Pearl is one of those rare scientists who can contribute to such theory and explain it. But as Aickin suggests in his amazon review this is not a subject for a novice. Previous exposure to statistical methods such as correlation and regression is important to a clear understanding of this book.

A great text, if for no other reason than the fact that it fills an important niche. Pearl does an excellent job of delineating causal models as both philosophical and statistical problems. I found the coverage of latent variable models particularly useful.

My only complaint is Pearl often makes assumptions without justifying them sufficiently. Usually, the assumptions made are reasonable or of negligible consequence, but at other times, the veracity of the assumptions is arguably core matter of the discussion. The net effect is a feeling of reading a brilliant, detailed exposition of what causal models imply observationally, undermined by doubts about the appropriateness of causality as a concept at all.

Overall, however, this a wonderful text that should be useful to anyone interested in causality or statistical modeling.

The scientific research community has adopted rigorous methods to eliminate the need for subjective judgments about many things, but when it comes to testing whether X causes Y, they revert to intuition and hand-waving. This book makes a strong argument that we shouldn't accept that. It demonstrates that it is possible to turn intuitions about causation into hypotheses that are unambiguous and testable.
But the style is sufficiently dense and dry we will need some additional books with more practical styles before these ideas become widely understood. The style is fairly good by the standards of books whose main goal is rigorous proof, but it's still hard work to learn a large number of new concepts that are mostly referred to by terse symbols whose meaning can't be found via a glossary or index. Pearl occasionally introduces a memorable word, such as do(x), the way a software engineer who wants readable code would, but mostly sticks to single-character symbols that seem unreasonably hard (at least for us programmers who are used to descriptive names) to remember.
If you're uncertain whether reading this book is worth the effort, I strongly recommend reading the afterword first. It ought to have been used as the introduction, and without it many readers will be left wondering why they should believe they will be rewarded for slogging through so much dry material.

I take issue with the previous reviewer. Pearl does not assume that the modeller is able, a priori, to determine what the correct model is. Instead, Pearl asks what conclusions can be drawn if the modeller is able to substantiate only parts of the model. By systematically changing those parts, he then obtains a full picture of what modeling assumptions "must" be substantiated before causal inferences can be derived from nonexperimental data. An anslysis of assumptions is not a license to abuse them.

Choice (November 00) calls both Pearl's Causality (and Juarrero's Dynamics in Action, which Choice reviews together with Pearl), a "radically new perspective on causation and human behavior... Pearl critically reviews the major literature on causation, both in philosohy and in applied statistics in the social sciences. His formal models, nicely illustrated by practical examples, show the power of positing objectdively real causation connetions, counter to Hume's skepticism, which has dominated discussions of causality in both analytic philosophy and statistical analysis. Probabilities, Pearl argues, reflect subjective degrees of belief, whereas causal relations describe objective physical constraints. He reveals the role of substantive causes in statistical analyses in examples from medicine, economics, and policy decisions. "Both works are highly ambitious in rejecting traditional views. Although the arguments ar meticulous and represent intensive research, their criticisms of mainstream traditions are destined to arouse controversy... Juarrero and Pearl's books will greatly interest philosophers and scientists who are concerned with causality and the explanation of human behavior."

This is a pioneering book dealing exhaustively with the subject of causation. Its contribution to the field of "Uncertainty in AI" is unmeasureable. It dealt with graphical models for reasoning in depth. For computer scientists looking for an encyclopedia of algorithms and applications on causation, there can not be a better book. I highly recommend this book for researchers in UAI. A word of caution: This is not a book for starters and those who do not have a well developed concept of uncertainty.

This is a very interesing book that Judea Pearl worte. The topic is currently of general interest for diverse fields as economics, social sciences and biology, however, this book is not intended for practitoners from these field who face a special problem and search for a possible solution. If you want to buy this book for this reason you will not be able to extract this information for this book. The reason therefor is that important technics like Bayesian Networks or Structural Equations are treated in 3 pages in each case. Judea Pearl assumes that the reader is already familiar with such methods beforehand. (Readers interested in the later subject are strongly refered to Bollen's book "Structural Equations with latent variables".)

Moreover, I do not think that this book presents state of the art information about our current knowledge of this subject. For example, the important problem to extract a network structure (structure learning) from data rather than estimating the parameters of a given networks structure is completely missing.

Nevertheless, this is a good book, because it might give you in the long run (you can not read it in one piece) insights you did not have before. Of course not to all topics causality is involved (see, e.g., above) but the given topics are thorough explained albeit on an advanced level.

Update: I add one star (total three) to my evaluation, because in the meanwhile I appreciate the historical development described in the book including references to the literature.

First off, the rating of three stars is relative to my expectations that this book would provide me with some insights in how to use graphical models for purposes of making inferences from statistical data and, in general, to facilitate the process of (machine) learning from data. And although Pearl and his colleagues have made great progress in this area, this book seems more targeted for researchers in areas outside of AI, such as economics, statistics, and medical research. Although the author gives a number of rigorous definitions to help support his notions of causality, the book is written in a somewhat abstract manner with few if any nontrivial examples (although enough trivial ones to satisfy a more general audience) to support the definitions and concepts. References to the literature are favored over mathematical proofs. Thus, aside from the references, I found this book of little use, but on the other hand, I do recommend it for its intended audience, for I do believe that graphical models can be of great use in these other areas.

Finally given the controversy and general misunderstanding about "causality", I wonder why Pearl would even use definitions like "causal model" and "...variable X is a causal influence of variable Y". His justification seems that researchers still think in terms of cause and effect, and thus it would serve them well if they had a mathematical foundation to fall back on.
Even if I did not have issue with some of the techniques and algorithms endorsed in this book, it would still seem much more appropriate to supply fresh, distinguished definitions (devoid of the "cause" word and its synonyms) and thus when future researchers use and make reference to Pearl's structural methods, they will call them as such and hopefully avoid confusion and controversy.

Pearl supposes that the modeller is able, a priori, to determine, *exactly* what the correct model is. One must be able to specify the model correctly, knowing what the possible confounding variables are, what moderators are important, etc., in advance. How reasonable is this? Isolation and pseudo-isolation are 'swept under the rug,' with an inadequate interpretation of the error term. This is dangerous work that will lead to situations where a researcher will calculate a path model using nonexperimental, cross-sectional data and conclude that they have found a 'causal' model with their X's having a cause-and-effect relationship with their endogenous variables. It is easy to imagine a saturated path model with cross-sectional survey data that produces 'structural' coefficients that are large relative to their standard errors by virtue of a large sample size, yet 'causes effects' that are relatively small. Without proper consideration of the effect sizes or alternative models, the researcher will cite Pearl for the value of their work (no puns about pearls and swine:). I doubt that Pearl has thoroughly thought through the garbage that will be printed citing him as justification; at least, I hope he hasn't.