Counterfactuals and Causal Inference: Methods and Principles for Social Research

Citation

Morgan, S. L., & Winship, C. (2015). Counterfactuals and Causal Inference: Methods and Principles for Social Research (2nd ed.). Cambridge University Press.

Chapter Summary

The second edition is a revised and expanded version of the 2007 first edition. It has thirteen chapters in six parts, and the parts follow a clear logic: first the two formal languages of causality, then estimators that work when the relevant variables are observed, then estimators for when they are not, and finally a part on what to do when neither strategy delivers a point estimate.

Part I, Chapter 1: Introduction

The opening chapter poses the book's basic question, whether X causes Y and how large the effect is, using examples such as charter schools, college degrees and earnings, and the butterfly ballot in the 2000 Florida election. It outlines the potential outcome model, offers a selective history of causal analysis in observational social science (including the role of experimental language and of survey research), introduces the examples used throughout the book, and ends with causal graphs.

Part II, Chapters 2–3: Counterfactuals and the Potential Outcome Model; Causal Graphs

Chapter 2 builds the potential outcome framework step by step: defining causal states, individual-level effects, treatment groups and observed outcomes, the average treatment effect, the stable unit treatment value assumption, naive estimation, over-time potential outcomes and many-valued treatments. Chapter 3 introduces the second language, causal graphs: identification, the basic elements of graphs, their link to structural equations, and their relation to the potential outcome model. Each chapter closes with an appendix.

Part III, Chapters 4–7: Conditioning on Observed Variables

Chapter 4, Models of Causal Exposure and Identification Criteria for Conditioning Estimators, presents the back-door criterion and the distinction between conditioning to balance and conditioning to adjust. Chapter 5 treats matching estimators (stratification, weighting and algorithmic matching, with practical issues). Chapter 6 treats regression estimators, including regression as conditional-variance-weighted matching. Chapter 7, Weighted Regression Estimators of Causal Effects, covers estimators of the ATE, ATT and ATC, doubly robust estimators and an extended example.

Part IV, Chapters 8–11: When Back-Door Conditioning Is Ineffective

Chapter 8, Self-Selection, Heterogeneity, and Causal Graphs, takes up selection on unobservables and complex patterns of self-selection. Chapter 9 covers instrumental variable estimators, including those that allow individual-level heterogeneity. Chapter 10, Mechanisms and Causal Explanation, discusses the front-door criterion and the appeal for generative mechanisms. Chapter 11, Repeated Observations and the Estimation of Causal Effects, covers interrupted time series, regression discontinuity designs and panel data, with an appendix on time-varying treatment regimes.

Part V, Chapter 12: Distributional Assumptions, Set Identification, and Sensitivity Analysis

The only chapter of Part V considers estimation when effects are not point-identified: latent variable selection-bias models, set identification with minimal assumptions, and sensitivity analysis for provisional estimates.

Part VI, Chapter 13: Counterfactuals and the Future of Empirical Research in Observational Social Science

The closing chapter answers objections to adopting the counterfactual approach and distinguishes modes of causal inquiry in the social sciences.

Key Concepts

Chapters 1–3: The two frameworks

• Potential outcomes: each individual has an outcome under each treatment state, but only one can be observed, so individual causal effects cannot be directly calculated.
• Average treatment effect and SUTVA: Chapter 2 defines average effects and the stable unit treatment value assumption on which they rest.
• Causal graphs: directed graphs that represent causal relationships and support identification arguments.

Chapters 4–7: Conditioning

• Back-door criterion: a graph-based rule for choosing variables whose conditioning blocks non-causal paths.
• Matching and regression: two families of conditioning estimators, treated from a potential outcomes perspective.
• Weighted regression and doubly robust estimators: Chapter 7's estimators of the ATE, ATT and ATC.

Chapters 8–12: Beyond conditioning

• Selection on the unobservables: the problem that motivates instrumental variables, mechanisms and longitudinal designs.
• Front-door criterion and mechanisms: identification through intermediate variables, linked to demands for deeper explanation.
• Set identification and sensitivity analysis: reporting bounds or testing how conclusions change when assumptions are relaxed.

Critical Analysis

Strengths

• Two languages, kept side by side. The book introduces potential outcomes (Chapter 2) and causal graphs (Chapter 3) before any estimator, and its chapter appendices relate them, so later chapters can use either.
• Organized by identification problem. The parts follow a decision logic: if the determinants of exposure are observed, condition (Part III); if not, look for instruments, mechanisms or repeated observations (Part IV); if nothing is point-identified, bound or probe sensitivity (Part V). A reader can find the right chapter from the research situation.
• Heterogeneity stressed. The publisher's blurb states that causal effect heterogeneity is stressed throughout, and the contents reflect this, for instance in Chapter 8 and in the heterogeneity-aware instrumental variable section of Chapter 9.
• Worked material. Chapter 1 sets out examples used throughout the book, and Chapter 7 includes an extended example.

Limitations

• A demanding text. At roughly 500 pages with formal notation and technical chapter appendices on graphs and identification, it is a graduate-level treatise rather than a first course in methods.
• Focus on observational data. Experiments appear mainly as the benchmark against which observational designs are judged; there is no chapter on experimental design.
• Statistical background assumed. Chapters 5–7 build on regression and related statistical ideas, and the book is written for readers who already work with quantitative data.

Where It Fits in the Course

This is the specialist causal-inference text in a methods course list that otherwise holds general introductions. Babbie's The Practice of Social Research, Walliman's Research Methods: The Basics and Bryman's Social Research Methods are broad introductions to the research process, and Creswell's Research Design is a guide to planning a study and writing a proposal; all of these are sensible earlier reading. Charmaz's Constructing Grounded Theory covers a qualitative tradition and is complementary rather than a competitor.

Among the political science books with the same topic, King, Keohane and Verba's Designing Social Inquiry discusses inference in qualitative research and Gerring's Social Science Methodology is a unified framework for method generally. Morgan and Winship is the one to turn to when the question is specifically how to estimate causal effects from non-experimental data.

How to Use This Book

The book suits graduate students and advanced undergraduates in sociology, demography, education and related fields who already know regression and want a rigorous foundation for causal claims; the publisher's blurb names the social, demographic and health sciences.

A practical plan:

• Core: Chapters 1–3 in order, then Chapters 4–6. Do not skip Chapter 4's treatment of identification, since the later chapters apply it.
• Applied estimation: Chapter 7 and then Chapter 9 for those who will run weighted regressions or instrumental variable analyses.
• Depth and design: Chapters 8, 10 and 11 for selection, mechanisms and longitudinal designs; Chapter 12 for sensitivity analysis.
• Optional on first pass: the chapter appendices (2, 3, 4, 11), which supply technical detail, and Chapter 13, which is best read once the methods are familiar.

Read Chapter 1 and 2.1–2.4 first and try restating a research question of your own as a contrast between potential outcomes; then draw it as a causal graph. If the two representations disagree about what must be controlled, work out why before moving on.