A potential voter must incur a number of costs in order to successfully cast an in-person ballot, including the costs associated with identifying and traveling to a polling place. In order to investigate how these costs affect voter turnout, we introduce two quasi-experimental designs that can be used to study how the political participation of registered voters is affected by differences in the relative distance that registrants must travel to their assigned Election Day polling place and whether their polling place remains at the same location as in a previous election. Our designs make comparisons of registrants who live on the same residential block, but are assigned to vote at different polling places. We find that living farther from a polling place and being assigned to a new polling place reduce in-person Election Day voting, but that registrants largely offset for this by casting more early in-person and mail ballots.

We consider the measures of partisan symmetry proposed for practical use in the political science literature, as clarified and developed in Katz, King, and Rosenblatt (2020, American Political Science Review 114, 164–178). Elementary mathematical manipulation shows the symmetry metrics to have surprising properties that call their meaningfulness into question. To accompany the general analysis, we study measures of partisan symmetry with respect to recent voting patterns in Utah, Texas, and North Carolina, flagging problems in each case. Taken together, these observations should raise major concerns about the available techniques for quantitative scores of partisan symmetry—including the mean–median score, the partisan bias score, and the more general “partisan symmetry standard”—with the decennial redistricting underway.

We offer methods to analyze the “differentially private” Facebook URLs Dataset which, at over 40 trillion cell values, is one of the largest social science research datasets ever constructed. The version of differential privacy used in the URLs dataset has specially calibrated random noise added, which provides mathematical guarantees for the privacy of individual research subjects while still making it possible to learn about aggregate patterns of interest to social scientists. Unfortunately, random noise creates measurement error which induces statistical bias—including attenuation, exaggeration, switched signs, or incorrect uncertainty estimates. We adapt methods developed to correct for naturally occurring measurement error, with special attention to computational efficiency for large datasets. The result is statistically valid linear regression estimates and descriptive statistics that can be interpreted as ordinary analyses of nonconfidential data but with appropriately larger standard errors.

Recent advancements in natural language processing (NLP) methods have significantly improved their performance. However, more complex NLP models are more difficult to interpret and computationally expensive. Therefore, we propose an approach to dictionary creation that carefully balances the trade-off between complexity and interpretability. This approach combines a deep neural network architecture with techniques to improve model explainability to automatically build a domain-specific dictionary. As an illustrative use case of our approach, we create an objective dictionary that can infer conflict intensity from text data. We train the neural networks on a corpus of conflict reports and match them with conflict event data. This corpus consists of over 14,000 expert-written International Crisis Group (ICG) CrisisWatch reports between 2003 and 2021. Sensitivity analysis is used to extract the weighted words from the neural network to build the dictionary. In order to evaluate our approach, we compare our results to state-of-the-art deep learning language models, text-scaling methods, as well as standard, nonspecialized, and conflict event dictionary approaches. We are able to show that our approach outperforms other approaches while retaining interpretability.

We develop a new measure of party position based on a scaling of ideology tags supplied in infoboxes on political parties’ Wikipedia pages. Assuming a simple model of tag assignment, we estimate the locations of parties and ideologies in a common space. We find that the recovered scale can be interpreted in familiar terms of “left versus right.” Estimated party positions correlate well with ratings of parties’ positions from extant large-scale expert surveys, most strongly with ratings of general left–right ideology. Party position estimates also show high stability in a test–retest scenario. Our results demonstrate that a Wikipedia-based approach yields valid and reliable left–right scores comparable to scores obtained via conventional expert coding methods. It thus provides a measure with potentially unlimited party coverage. Our measurement strategy is also applicable beyond Wikipedia.

Political scientists have increasingly deployed conjoint survey experiments to understand multidimensional choices in various settings. In this paper, we show that the average marginal component effect (AMCE) constitutes an aggregation of individual-level preferences that is meaningful both theoretically and empirically. First, extending previous results to allow for arbitrary randomization distributions, we show how the AMCE represents a summary of voters’ multidimensional preferences that combines directionality and intensity according to a probabilistic generalization of the Borda rule. We demonstrate why incorporating both the directionality and intensity of multi-attribute preferences is essential for analyzing real-world elections, in which ceteris paribus comparisons almost never occur. Second, and in further empirical support of this point, we show how this aggregation translates directly into a primary quantity of interest to election scholars: the effect of a change in an attribute on a candidate’s or party’s expected vote share. These properties hold irrespective of the heterogeneity, strength, or interactivity of voters’ preferences and regardless of how votes are aggregated into seats. Finally, we propose, formalize, and evaluate the feasibility of using conjoint data to estimate alternative quantities of interest to electoral studies, including the effect of an attribute on the probability of winning.

Katz, King, and Rosenblatt (2020, American Political Science Review 114, 164–178) introduces a theoretical framework for understanding redistricting and electoral systems, built on basic statistical and social science principles of inference. DeFord et al. (2021, Political Analysis, this issue) instead focuses solely on descriptive measures, which lead to the problems identified in our article. In this article, we illustrate the essential role of these basic principles and then offer statistical, mathematical, and substantive corrections required to apply DeFord et al.’s calculations to social science questions of interest, while also showing how to easily resolve all claimed paradoxes and problems. We are grateful to the authors for their interest in our work and for this opportunity to clarify these principles and our theoretical framework.

A core part of political research is to identify how political preferences are shaped. The nature of these questions is such that robust causal identification is often difficult to achieve, and we are not seldom stuck with observational methods that we know have limited causal validity. The purpose of this paper is to measure the magnitude of bias stemming from both measurable and unmeasurable confounders across three broad domains of individual determinants of political preferences: socio-economic factors, moral values, and psychological constructs. We leverage a unique combination of rich Swedish registry data for a large sample of identical twins, with a comprehensive battery of 34 political preference measures, and build a meta-analytical model comparing our most conservative observational (naive) estimates with discordant twin estimates. This allows us to infer the amount of bias from unobserved genetic and shared environmental factors that remains in the naive models for our predictors, while avoiding precision issues common in family-based designs. The results are sobering: in most cases, substantial bias remains in naive models. A rough heuristic is that about half of the effect size even in conservative observational estimates is composed of confounding.

While a difference-in-differences (DID) design was originally developed with one pre- and one posttreatment period, data from additional pretreatment periods are often available. How can researchers improve the DID design with such multiple pretreatment periods under what conditions? We first use potential outcomes to clarify three benefits of multiple pretreatment periods: (1) assessing the parallel trends assumption, (2) improving estimation accuracy, and (3) allowing for a more flexible parallel trends assumption. We then propose a new estimator, double DID, which combines all the benefits through the generalized method of moments and contains the two-way fixed effects regression as a special case. We show that the double DID requires a weaker assumption about outcome trends and is more efficient than existing DID estimators. We also generalize the double DID to the staggered adoption design where different units can receive the treatment in different time periods. We illustrate the proposed method with two empirical applications, covering both the basic DID and staggered adoption designs. We offer an open-source R package that implements the proposed methodologies.

Katz, King, and Rosenblatt recently wrote a broad survey developing and extending the theory of partisan symmetry. Our paper reviewed the implementability of the theory, focusing on simplified scores of symmetry—seemingly compatible with their formulation—that are in wide use. We analyzed these simplified scores and concluded that they are not suited for redistricting reform. By our reading of their response, Katz, King, and Rosenblatt agree.

Some scholars build models to classify documents into chosen categories. Others, especially social scientists who tend to focus on population characteristics, instead usually estimate the proportion of documents in each category—using either parametric “classify-and-count” methods or “direct” nonparametric estimation of proportions without individual classification. Unfortunately, classify-and-count methods can be highly model-dependent or generate more bias in the proportions even as the percent of documents correctly classified increases. Direct estimation avoids these problems, but can suffer when the meaning of language changes between training and test sets or is too similar across categories. We develop an improved direct estimation approach without these issues by including and optimizing continuous text features, along with a form of matching adapted from the causal inference literature. Our approach substantially improves performance in a diverse collection of 73 datasets. We also offer easy-to-use software that implements all ideas discussed herein.

I propose a new model, ordered Beta regression, for continuous distributions with both lower and upper bounds, such as data arising from survey slider scales, visual analog scales, and dose–response relationships. This model employs the cut point technique popularized by ordered logit to fit a single linear model to both continuous (0,1) and degenerate [0,1] responses. The model can be estimated with or without observations at the bounds, and as such is a general solution for these types of data. Employing a Monte Carlo simulation, I show that the model is noticeably more efficient than ordinary least squares regression, zero-and-one-inflated Beta regression, rescaled Beta regression, and fractional logit while fully capturing nuances in the outcome. I apply the model to a replication of the Aidt and Jensen (2014, European Economic Review 72, 52–75) study of suffrage extensions in Europe. The model can be fit with the R package ordbetareg to facilitate hierarchical, dynamic, and multivariate modeling.

In recent decades, political science literature has experienced significant growth in the popularity of nonlinear models with multiplicative interaction terms. When one or more constitutive variables are not binary, most studies report the marginal effect of the variable of interest at its sample mean while allowing the other constitutive variable/s to vary along its range and holding all other covariates constant at their means, modes, or medians. In this article, we argue that this conventional approach is not always the most suitable since the marginal effect of a variable at its sample mean might not be sufficiently representative of its prevalent effect at a specific value of the conditioning variable and might produce excessively model-dependent predictions. We propose two procedures to help researchers gain a better understanding of how the typical effect of the variable of interest varies as a function of the conditioning variable: (1) computing and plotting the marginal effects at all in-sample combinations of the values of the constitutive variables and (2) computing and plotting what we call the “Distribution-Weighted Average Marginal Effect” over the values of the conditioning variable.

We introduce and assess the use of supervised learning in cross-domain topic classification. In this approach, an algorithm learns to classify topics in a labeled source corpus and then extrapolates topics in an unlabeled target corpus from another domain. The ability to use existing training data makes this method significantly more efficient than within-domain supervised learning. It also has three advantages over unsupervised topic models: the method can be more specifically targeted to a research question and the resulting topics are easier to validate and interpret. We demonstrate the method using the case of labeled party platforms (source corpus) and unlabeled parliamentary speeches (target corpus). In addition to the standard within-domain error metrics, we further validate the cross-domain performance by labeling a subset of target-corpus documents. We find that the classifier accurately assigns topics in the parliamentary speeches, although accuracy varies substantially by topic. We also propose tools diagnosing cross-domain classification. To illustrate the usefulness of the method, we present two case studies on how electoral rules and the gender of parliamentarians influence the choice of speech topics.

Theory has long suggested that swing voting is a response to cross-pressures arising from a mix of individual attributes and contextual factors. Unfortunately, existing regression-based approaches are ill-suited to explore the complex combinations of demographic, policy, and political factors that produce swing voters in American elections. This gap between theory and practice motivates our use of an ensemble of supervised machine learning methods to predict swing voters in the 2012, 2016, and 2020 U.S. presidential elections. The results from the learning ensemble substantiate the existence of swing voters in contemporary American elections. Specifically, we demonstrate that the learning ensemble produces well-calibrated and externally valid predictions of swing voter propensity in later elections and for related behaviors such as split-ticket voting. Although interpreting black-box models is more challenging, they can nonetheless provide meaningful substantive insights meriting further exploration. Here, we use flexible model-agnostic tools to perturb the ensemble and demonstrate that cross-pressures (particularly those involving ideological and policy-related considerations) are essential to accurately predict swing voters.

We are grateful to DeFord et al. for the continued attention to our work and the crucial issues of fair representation in democratic electoral systems. Our response (Katz, King, and Rosenblatt Forthcoming) was designed to help readers avoid being misled by mistaken claims in DeFord et al. (Forthcoming-a), and does not address other literature or uses of our prior work. As it happens, none of our corrections were addressed (or contradicted) in the most recent submission (DeFord et al. Forthcoming-b).

Video advertisements, either through television or the Internet, play an essential role in modern political campaigns. For over two decades, researchers have studied television video ads by analyzing the hand-coded data from the Wisconsin Advertising Project and its successor, the Wesleyan Media Project (WMP). Unfortunately, manually coding more than a hundred of variables, such as issue mentions, opponent appearance, and negativity, for many videos is a laborious and expensive process. We propose to automatically code campaign advertisement videos. Applying state-of-the-art machine learning methods, we extract various audio and image features from each video file. We show that our machine coding is comparable to human coding for many variables of the WMP datasets. Since many candidates make their advertisement videos available on the Internet, automated coding can dramatically improve the efficiency and scope of campaign advertisement research. Open-source software package is available for implementing the proposed methodology.

Political event data are widely used in studies of political violence. Recent years have seen notable advances in the automated coding of political event data from international news sources. Yet, the validity of machine-coded event data remains disputed, especially in the context of event geolocation. We analyze the frequencies of human- and machine-geocoded event data agreement in relation to an independent (ground truth) source. The events are human rights violations in Colombia. We perform our evaluation for a key, 8-year period of the Colombian conflict and in three 2-year subperiods as well as for a selected set of (non)journalistically remote municipalities. As a complement to this analysis, we estimate spatial probit models based on the three datasets. These models assume Gaussian Markov Random Field error processes; they are constructed using a stochastic partial differential equation and estimated with integrated nested Laplacian approximation. The estimated models tell us whether the three datasets produce comparable predictions, underreport events in relation to the same covariates, and have similar patterns of prediction error. Together the two analyses show that, for this subnational conflict, the machine- and human-geocoded datasets are comparable in terms of external validity but, according to the geostatistical models, produce prediction errors that differ in important respects.

Sentiment analysis techniques have a long history in natural language processing and have become a standard tool in the analysis of political texts, promising a conceptually straightforward automated method of extracting meaning from textual data by scoring documents on a scale from positive to negative. However, while these kinds of sentiment scores can capture the overall tone of a document, the underlying concept of interest for political analysis is often actually the document’s stance with respect to a given target—how positively or negatively it frames a specific idea, individual, or group—as this reflects the author’s underlying political attitudes. In this paper, we question the validity of approximating author stance through sentiment scoring in the analysis of political texts, and advocate for greater attention to be paid to the conceptual distinction between a document’s sentiment and its stance. Using examples from open-ended survey responses and from political discussions on social media, we demonstrate that in many political text analysis applications, sentiment and stance do not necessarily align, and therefore sentiment analysis methods fail to reliably capture ground-truth document stance, amplifying noise in the data and leading to faulty conclusions.

We propose and explore the possibility that language models can be studied as effective proxies for specific human subpopulations in social science research. Practical and research applications of artificial intelligence tools have sometimes been limited by problematic biases (such as racism or sexism), which are often treated as uniform properties of the models. We show that the “algorithmic bias” within one such tool—the GPT-3 language model—is instead both fine-grained and demographically correlated, meaning that proper conditioning will cause it to accurately emulate response distributions from a wide variety of human subgroups. We term this property algorithmic fidelity and explore its extent in GPT-3. We create “silicon samples” by conditioning the model on thousands of sociodemographic backstories from real human participants in multiple large surveys conducted in the United States. We then compare the silicon and human samples to demonstrate that the information contained in GPT-3 goes far beyond surface similarity. It is nuanced, multifaceted, and reflects the complex interplay between ideas, attitudes, and sociocultural context that characterize human attitudes. We suggest that language models with sufficient algorithmic fidelity thus constitute a novel and powerful tool to advance understanding of humans and society across a variety of disciplines.