Book review of “Causality in a Social World” by Guanglei Hong

Observational Studies 2 (2016) 86-89 Submitted 10/16; Published 10/16 Book review of “Causality in a Social World” by Guanglei Hong Kenneth A. Frank kenfrank@msu.edu Department of Counseling, Educational Psychology and Special Education, College of Education Department of Fisheries and Wildlife, College of Agriculture & Natural Resources Michigan State University Lansing, MI, U.S.A. Guan Kung Saw guan.saw@utsa.edu Department of Educational Psychology College of Education and Human Development The University of Texas at San Antonio San Antonio, TX, U.S.A. Ran Xu ranxu@msu.edu Department of Counseling, Educational Psychology and Special Education College of Education Michigan State University Lansing, MI, U.S.A. As the introduction of Guanglei Hong’s Causality in a Social World makes clear, this book would not be necessary if all treatments we wished to study had constant effects through simple mechanisms on independent individuals who were randomly assigned to treatments. While, such conditions may hold in some idealized agricultural settings, this is not the phenomenon we encounter in a social policy oriented world with human agency. In response, Hong presents a coherent theoretical and empirical framework for estimating causality when people choose their own treatments, when they encounter mediating and moderating effects of treatments and when they influence others’ choices and outcomes. The book is presented in four large sections: overview, moderation, mediation and spillover , with a chapter introducing the core ideas in each section (chapters 4, 7, 11 and 14 respectively). Beyond merely consolidating her own foundational work, the book is steeped in deep and historical statistical principles of sampling, propensity score analysis, mediation and moderation, and spill-over mechanisms. Ultimately, the book will mark a passageway from underlying statistical principles to a framework that may endure and expand beyond even what Hong anticipates. The core of the framework is to conceptualize causal inference as a sampling issue, building upon a long tradition of statistical adjustment through weighting in survey sampling. In particular, Chapter 4 lays a foundation in conventional statistical principles, in this case of sampling, stratified sampling, and weights to correct for purposeful sampling that create a sample disproportionate to population treatment assignments. The turn then is to apply the sampling conceptualization to strata defined by the propensity to have been assigned to the treatment. c ⃝2016 Kenneth A. Frank, Guan Kung Saw and Ran Xu. Book review of “Causality in a Social World” When subjects are placed in strata based on their propensity to receive the treatment the proportion of treated and control subjects within each stratum will typically not represent their respective proportions in the full sample. This creates challenges in synthesizing results across strata. The insight is that the data within each stratum can be weighted to reflect the proportions of treated and control subjects in the overall sample (or in the population if people are randomly sampled). This is known as marginal mean weighting through stratification (MMWS). Drawing on her own seminal work in this area, Hong shows that for binary treatment estimates based on MMWS are equivalent to estimation through stratification without weights. Furthermore, both outperform (in terms of root mean square error – page 94) conventional inverse proportional treatment weighting (IPTW). This is in part because assignment to discrete strata is robust to misspecification of the propensity model relative to continuous weighting approaches (see page 98), while the weights insure that the treatment and control cases in each stratum contribute proportionally to the overall estimate. In addition to its estimation advantages, MMWS is more flexible for studying binary and multivalued treatments than conventional propensity score matching and stratification that are mostly limited to examinations of dichotomous treatments. Extensions to mediation, moderation and spill-over are based on a conceptualization of multiple treatments, either constructing a sequence as in mediation, separable regimes as in moderation, or a two-stage clustered randomized design as in spill-over. Just as chapter 4 presents MMWS within traditional sampling ideas, so do each of the parts show how new ideas emerge out of existing techniques. For example, chapter 6 presents moderation in the context of randomized experiments and factorial designs, chapter 9 builds mediation out of Baron and Kennys (1986) path...