Abstract
When I started graduate school in 2003, I knew that I liked history and had very little interest in statistics. Inspired by the work of Andrew Abbott, I was committed to the idea that the problem with regression was not so much that it produced bad results as it was that the ubiquity of traditional forms of variable-oriented thinking undermined our ability to think creatively about how patterns of social action are informed by the spatial and temporal context in which they unfold. From this perspective, the fundamental challenge associated with the development of quantitative history in the social sciences was finding an approach to regression that would allow statistical relationships to vary over space and time.
As I slowly warmed up to the possibility of integrating regression into my research, I became particularly interested in methods associated with the study of spatial heterogeneity. This put me at odds with both my more qualitatively oriented colleagues in historical sociology as well as more conventional quantitative researchers and reviewers who saw little utility in the added complexity that came with allowing one or more variables to exert multiple effects depending on where a given case was located. Part of what made this experience so difficult was that I had very few resources to help explain why this type of approach is so central to the prospect of quantitative history. Until now.
Drawing on a combination of intellectual history, replication, and formal statistical critique, Gregory Wawro and Ira Katznelson's Time Counts: Quantitative Analysis for Historical Social Science makes a compelling case for the claim that quantitative history can be made more historical by introducing methods that allow for various forms of parameter heterogeneity. More specifically, the authors argue for the use of Bayesian semi-parametric models in conjunction with what they refer to as “historically relevant priors” (pp. 52–53). In this respect, their work, which draws extensively on examples from political science, can be understood as a realization of work begun in sociology by scholars including Larry Isaac, Larry Griffin, Bruce Western, and Meredith Kleykamp.
The book is divided into two parts, the first of which sets the intellectual and methodological stage for the various applications that follow. This discussion plays an important role in the context of a book that is intended to help scholars bridge the qualitative-quantitative divide. With this goal in mind, the authors are deeply attentive to the challenges traditionally associated with trying to make historical research more quantitative. While historians themselves have never been wholly opposed to the prospect of using numbers in their work, the push for a quantitative turn in historical research that began in the postwar period was ultimately dismissed as a form of methodological imperialism insofar as the use of historical materials was increasingly divorced from historical analysis, leaving questions of space and time to fall by the wayside.
Reflecting on the legacy of this debate, Wawro and Katznelson push back on the notion that qualitative and quantitative methods can be forced together under a common inferential framework, while simultaneously eschewing the claim that the two approaches constitute wholly distinct cultures. The authors suggest a middle path predicated on a rejection of the standard unit homogeneity assumption, which allows quantitative researchers to not only assign a global effect to each of our covariates, but to increase the precision of the resulting parameter estimates by pooling large numbers of ostensibly comparable observations. Unfortunately, this rather heroic assumption is difficult to maintain in the context of quantitative history, where we are often forced to pool observations over disparate times and/or locales, increasing the likelihood that a conventional global model will fundamentally mischaracterize the varied data-generating processes in play.
So rather than resorting to simplifying assumptions in the vain hope that the world will magically align with our desire for linear generalization, Wawro and Katznelson propose that we adopt methods that allow us to capture the type of contextual complexity traditionally embraced by historians. This approach is made possible by innovations in statistical methodology such as the advent of changepoint models, structured additive regression models, and Markov switching models with time-varying transition probabilities. The utility of these methods is demonstrated at length in the second half of the book, which revisits a number of prominent examples from political science with an eye toward questions regarding temporal heterogeneity, path dependence, and historical persistence.
For example, Wawro and Katznelson use a structured additive regression model with a Markov random field prior to capture spatial and temporal variation in the effect of unionization on pro-labor voting in the United States Senate between 1933 and 1948. As the authors describe, the advantage of using a Markov random field prior is that it allows them to capture variation in relationships over both space and time, while borrowing information from adjacent regions and time points in a historically informed way. The analysis reveals an unexpected divide among southern senators, with senators from the Border South becoming increasingly responsive to unionization efforts in their home states, unlike their more staunchly anti-labor colleagues in the Deep South.
The use of Markov switching models to capture path dependence in the ebb and flow of partisan polarization in the United States House of Representatives is especially novel, if somewhat speculative. By the authors’ own account, the results of this analysis are too noisy to be sure about the location of critical junctures in the observed data, but there is good reason to believe that the relationship between income inequality and polarization varied over time, with higher levels of inequality increasing the probability of remaining in a polarizing state. The discussion of parameter heterogeneity is arguably less pronounced in the book's penultimate chapter, which takes up the question of causal inference in historical research. Nonetheless, the authors show that careful thinking about time and temporality is integral to the viability of popular methods such as instrumental variable models, which depend on assumptions regarding the nature of historical persistence in order to estimate the effect of past conditions on contemporary outcomes.
It is important to note that while this is very much a book about the use of statistics in historical research, it is not a statistics book in the usual sense. In other words, while Wawro and Katznelson provide a clear description of a range of different procedures, interested readers will almost certainly need to consult other resources if they intend to implement these methods on their own. This is going to be especially true for the Bayesian techniques, which pair well with historical research but are, unfortunately, still relatively unfamiliar to the average social scientist. That being said, it isn't reasonable to expect one book to do everything. Time Counts is a thought-provoking take on the possible future of quantitative history that warrants serious consideration by quantitative and qualitative scholars alike. The authors provide proof of concept of what is possible when you allow quantitative methods to bend to the complexity of historical reality. I look forward to seeing others carry on this project, with more sustained empirical treatments sure to follow.