Large language models are a powerful tool for conducting text analysis in political science, but using them to annotate text has several drawbacks, including high cost, limited reproducibility, and poor explainability. Traditional supervised text classifiers are fast and reproducible, but require expensive hand annotation, which is especially difficult for rare classes. This article proposes using LLMs to generate synthetic training data for training smaller, traditional supervised text models. Synthetic data can augment limited hand annotated data or be used on its own to train a classifier with good performance and greatly reduced cost. I provide a conceptual overview of text generation, guidance on when researchers should prefer different techniques for generating synthetic text, a discussion of ethics, a simple technique for improving the quality of synthetic text, and an illustration of its limitations. I demonstrate the usefulness of synthetic training through three validations: synthetic news articles describing police responses to communal violence in India for training an event detection system, a multilingual corpus of synthetic populist manifesto statements for training a sentence-level populism classifier, and generating synthetic tweets describing the fighting in Ukraine to improve a named entity system.

We develop a formal framework for accumulating evidence across studies and apply it to develop theoretical foundations for replication. Our primary contribution is to characterize the relationship between replication and distinct formulations of external validity. Whereas conventional wisdom holds that replication facilitates learning about external validity, we show that this is not, in general, the case. Our results show how comparisons of the magnitude or sign of empirical findings link to distinct concepts of external validity. However, without careful attention to the research design of constituent studies, replication can mislead efforts to assess external validity. We show that two studies must have essentially the same research designs, i.e., be harmonized, in order for their estimates to provide information about any kind of external validity. This result shows that even minor differences in research design between a study and its replication can introduce a discrepancy that is typically overlooked, a problem that becomes more pronounced as the number of studies increases. We conclude by outlining a design-driven approach to replication, which responds to the issues our framework identifies and details how a research agenda can manage them productively.

Substantive research in the Social Sciences regularly investigates signed networks, where edges between actors are positive or negative. One often-studied example within International Relations for this type of network consists of countries that can cooperate with or fight against each other. These analyses often build on structural balance theory, one of the earliest and most prominent network theories. While the theorization and description of signed networks have made significant progress, the inferential study of link formation within them remains limited in the absence of appropriate statistical models. We fill this gap by proposing the Signed Exponential Random Graph Model (SERGM), extending the well-known Exponential Random Graph Model (ERGM) to networks where ties are not binary but positive or negative if a tie exists. Since most networks are dynamically evolving systems, we specify the model for both cross-sectional and dynamic networks. Based on hypotheses derived from structural balance theory, we formulate interpretable signed network statistics, capturing dynamics such as “the enemy of my enemy is my friend”. In our empirical application, we use the SERGM to analyze cooperation and conflict between countries within the international state system. We find evidence for structural balance in International Relations.

Past work on closed-ended survey responses demonstrates that inferring stable political attitudes requires separating signal from noise in “top of the head” answers to researchers’ questions. We outline a corresponding theory of the open-ended response, in which respondents make narrow, stand-in statements to convey more abstract, general attitudes. We then present a method designed to infer those attitudes. Our approach leverages co-variation with words used relatively frequently across respondents to infer what else they could have said without substantively changing what they meant—linking narrow themes to each other through associations with contextually prevalent words. This reflects the intuition that a respondent may use different specific statements at different points in time to convey similar meaning. We validate this approach using panel data in which respondents answer the same open-ended questions (concerning healthcare policy, most important problems, and evaluations of political parties) at multiple points in time, showing that our method’s output consistently exhibits higher within-subject correlations than hand-coding of narrow response categories, topic modeling, and large language model output. Finally, we show how large language models can be used to complement—but not, at present, substitute—our “implied word” method.

Confidence intervals are ubiquitous in the presentation of social science models, data, and effects. When several intervals are plotted together, one natural inclination is to ask whether the estimates represented by those intervals are significantly different from each other. Unfortunately, there is no general rule or procedure that would allow us to answer this question from the confidence intervals alone. It is well known that using the overlaps in 95% confidence intervals to perform significance tests at the 0.05 level does not work. Recent scholarship has developed and refined a set of tools for inferential confidence intervals that permit inference on confidence intervals with the appropriate type I error rate in many different bivariate contexts. These are all based on the same underlying idea of identifying the multiple of the standard error (i.e., a new confidence level) such that the overlap in confidence intervals matches the desired type I error rate. These procedures remain stymied by multiple simultaneous comparisons. We propose an entirely new procedure for developing inferential confidence intervals that decouples the testing and visualization that can overcome many of these problems in any visual testing scenario. We provide software in R and Stata to accomplish this goal.

Regression discontinuity designs (RDD) are widely used in the social sciences to estimate causal effects from observational data. Following recent methodological advances, scholars can choose from various RDD estimators for point estimation and inference. This decision is mainly guided by theoretical results on optimality and Monte Carlo simulations because of a paucity of research on the performance of the different estimators in recovering real-world experimental benchmarks. Leveraging exact ties in personal votes in local elections in Colombia and Finland, which are resolved by a random lottery, we assess the performance of various estimators featuring different polynomial degrees, bias-correction methods, optimal bandwidths, and approaches to statistical inference. Using re-running and re-election as outcomes, we document only minor differences in the performance of the various implementation approaches when the conditional expectation function (CEF) of the outcomes in the vicinity of the discontinuity is close to linear. When approximating the curvature of the CEF is more challenging, bias-corrected and robust inference with coverage-error-rate-optimal bandwidths comes closer to the experimental benchmark than more widely used alternative implementations.

Analysts often seek to compare representations in high-dimensional space, e.g., embedding vectors of the same word across groups. We show that the distance measures calculated in such cases can exhibit considerable statistical bias, that stems from uncertainty in the estimation of the elements of those vectors. This problem applies to Euclidean distance, cosine similarity, and other similar measures. After illustrating the severity of this problem for text-as-data applications, we provide and validate a bias correction for the squared Euclidean distance. This same correction also substantially reduces bias in ordinary Euclidean distance and cosine similarity estimates, but corrections for these measures are not quite unbiased and are (non-intuitively) bimodal when distances are close to zero. The estimators require obtaining the variance of the latent positions. We (will) implement the estimator in free software, and we offer recommendations for related work.

We use instruction-tuned large language models (LLMs) like GPT-4, Llama 3, MiXtral, or Aya to position political texts within policy and ideological spaces. We ask an LLM where a tweet or a sentence of a political text stands on the focal dimension and take the average of the LLM responses to position political actors such as US Senators, or longer texts such as UK party manifestos or EU policy speeches given in 10 different languages. The correlations between the position estimates obtained with the best LLMs and benchmarks based on text coding by experts, crowdworkers, or roll call votes exceed .90. This approach is generally more accurate than the positions obtained with supervised classifiers trained on large amounts of research data. Using instruction-tuned LLMs to position texts in policy and ideological spaces is fast, cost-efficient, reliable, and reproducible (in the case of open LLMs) even if the texts are short and written in different languages. We conclude with cautionary notes about the need for empirical validation.