Identification for Prediction and Decision

by Charles F. Manski

This book is a full-scale exposition of Charles Manski's new methodology for analyzing empirical questions in the social sciences. He recommends that researchers first ask what can be learned from data alone, and then ask what can be learned when data are combined with credible weak assumptions. Inferences predicated on weak assumptions, he argues, can achieve wide consensus, while ones that require strong assumptions almost inevitably are subject to sharp disagreements.

Building on the foundation laid in the author's Identification Problems in the Social Sciences (Harvard, 1995), the book's fifteen chapters are organized in three parts. Part I studies prediction with missing or otherwise incomplete data. Part II concerns the analysis of treatment response, which aims to predict outcomes when alternative treatment rules are applied to a population. Part III studies prediction of choice behavior.

Each chapter juxtaposes developments of methodology with empirical or numerical illustrations. The book employs a simple notation and mathematical apparatus, using only basic elements of probability theory.

Genres: Economics

368 pages, Hardcover

First published December 14, 2007

Reviews

[Daeus]

Solid book/writing, great intuition, I just dont have the time to read through it right now (read first 2 chapters for a work bookclub).


Quotes
- "Studies of identification ask to characterize the conclusions that could be drawn if one could use the sampling process to obtain an unlimited number of observations. Studies of statistical inference seek to characterize the generally weaker conclusions that can be drawn from a finite number of observations."
- "I will stress that researchers and decision makers need to cope with ambiguity."
- "If one changes from square loss to absolute loss, the best predictor changes from the mean to the median."
- "A jurisdiction contemplating using the approach to make selective incapacitation decisions does not need to ask why the association between background and criminality found in the RAND study was what it was. It needs only to ask whether it is reasonabke to assume that this association holds in its own circumstances."
- "Theories are testable where they are least needed, and are not testable where they're most needed." I suppose studies that show limits of certain premises help reduce overall assumptions/bias, rather than meriting some deterministic rule to implement. I'm see the value of slowing peoples assumptions with this work more than proving some theory about behavior.
- "inference with missing outcome data is a matter of contemplating all the values that the missing data might take."
- "Although bounds using the empirical evidence alone sometimes are 'distressingly wide,' .....they establish a domain of consensus among researchers who may disagree about the precise value of the population parameter... they make explicit the fundamental role that credible assumptions play in inferential methods that yield tighter findings."