Handbook of matching and weighting adjustments for causal inference by José R. Zubizarreta, Elizabeth A. Stuart, Dylan S. Small, and Paul R. Rosenbaum, Chapman and Hall/CRC, 2023, ISBN: 9781003102670, https://www.routledge.com/Handbook-of-matching-and-weighting-adjustments-for-causal-inference/Zubizarreta-Stuart-Small-Rosenbaum/p/book/9781003102670
Youjin Lee · Biometrics · 2024
Uncovering the causal relationship between a treatment and an outcome is the ultimate goal in most research in the social sciences, public health, and medicine, whether explicitly stated or not. In observational studies, unlike experiments whereby treatments are randomly assigned to units, units that receive the treatment and those that do not can easily differ on multiple factors. This could lead to differences in outcomes between the 2 groups that are not due to the treatment, which will bias the causal effect of interest. This book introduces diverse causal inference methods used in observational studies, each of which aims to adjust for these baseline differences between the treated and control groups using their pre-treatment information. The book is structured into 3 main parts following an initial conceptual overview of the causal inference problem: II. Matching, III. Weighting, and IV. Machine Learning and Bayesian approaches. Then it concludes with additional topics relevant to observational studies beyond covariate adjustment. The idea of matching for causal analysis is relatively straightforward and intuitive: identifying a set of the treated and control units that are similar with respect to pre-treatment covariates. Part II of this book introduces various algorithms and their applications to different types of data and research designs. Chapters 4-6 cover several optimal matching methods that use pre-treatment covariates and explain their connections to the network flow approach. The book also illustrates the use of matching in other causal inference methods, such as instrumental variable methods (Chapter 7) and regression discontinuity designs (Chapter 8), where covariate adjustment can strengthen the instrumental variable and enhance the efficiency and power of regression discontinuity treatment effects. Following chapters introduce matching methods proposed for specific data types and research questions: treatments assigned at different time points (Chapter 9), multilevel data (Chapter 10), effect modification and sensitivity analysis (Chapter 11), and multiple (more than 2) treatment groups (Chapter 12). Chapter 13 addresses recent matching methods designed for large observational studies, which often pose computational challenges in matching. The concept of weighting is to assign weights to each unit to generate a pseudo-population in which the treated and control groups are similar with respect to pre-treatment covariates. Part III of the book discusses various weighting methods beyond the well-known inverse probability weighting (IPW). First, Chapter 14 introduces overlap weighting, which offers improved efficiency compared to IPW, and illustrates its extensions to different settings. Chapters 15 and 16 present balancing approaches that aim to find weights that can balance the covariates between the 2 treatment groups without requiring further steps after weighting. The book also covers advanced weighting methods, including principal score methods in the presence of post-treatment variables that define the target estimand (Chapter 17), incremental propensity scores for estimating incremental treatment effect (Chapter 18), and weighting methods for analyzing mediational effects (Chapter 19). Parts II and III illustrate weighting and matching methods, focusing on their distinctive features for achieving balance between treatment groups (often 2) rather than on statistical modeling. Part IV explores recent advancements in machine learning and Bayesian methods for causal inference, focusing on their statistical advantages. Chapter 20 focuses on the use of machine learning methods for estimating treatment heterogeneity with survival outcomes. Chapter 21 provides a theoretical overview of targeted maximum likelihood estimators combined with the highly adaptive LASSO minimum loss estimator, demonstrating their desirable statistical properties in causal effect estimation. Chapter 23 discusses several Bayesian approaches to propensity score estimation, illustrating the distinct advantages of Bayesian methods. The final part of the book features 3 chapters designed to offer readers critical and diverse perspectives on observational studies, covering topics such as sensitivity analysis (Chapter 25) and evidence factors (Chapter 26). The book benefits from a comprehensive collection of recent causal inference methods, offering a wide range of perspectives on weighting and matching techniques. While all the methods share the common goal of unbiased causal effect estimation in observational studies, each chapter clearly demonstrates its focus (eg, balancing covariates or using survival outcomes). In particular, each chapter includes data application examples at the end or incorporates application studies throughout. Some chapters also provide R code for analysis. As a result, readers can easily pinpoint the focus of each method and find useful resources to reproduce the methodology tailored to specific approaches. As a textbook, the book is most suitable for PhD students in statistics, biostatistics, or other related fields involving quantitative methods. However, (under)graduate students and researchers in other disciplines may find some chapters useful, provided that they are already familiar with R code, basic concepts of probability, and regression models. There are several points that require caution for readers. Because this book is a compilation of monographs, there are some inconsistencies in notation, even for key concepts like the treatment variable (eg, denoted by A, Z, T, etc.) and propensity scores (eg, denoted by e, |$\lambda $|, |$\pi $|, etc.). These inconsistencies could be beneficial for readers specialized in causal inference, as they reflect the widely used notations for each concept in the field. However, this variability may distract readers when using the entire book for educational purposes, especially for those who are relatively new to causal inference. Additionally, the depth of contents varies across chapters, with some emphasizing methodological aspects and others focusing more on practical application. For PhD students in (bio)statistics, I recommend reading the references cited by the authors in each chapter to gain a deeper understanding. For students and researchers who prefer to avoid theoretical complexities, I suggest focusing on the application studies provided in each chapter to better understand the rationale and use of the methods discussed. There has been a notable lack of resources that educators and researchers can rely on for PhD-level causal inference methods covering recent advancements under diverse study designs. Some of the modern methods have been considered niche, accessible, and understood only by a small group of statisticians and biostatisticians specializing in the field. In my view, this is not necessarily due to the complexity of the methods but rather because it often takes considerable effort to track down recent references and stay up-to-date with the work of specific authors. As a result, researchers in other disciplines may struggle to move beyond 1:1 matching and IPW, regardless of their study design or research questions. I am grateful that this book contributes to expanding the accessibility of modern causal inference tools, bringing them together in a cohesive manner for researchers and educators who wish to learn, teach, and apply these methods to obtain unbiased causal evidence from—potentially messy and unkind—observational studies.