Austin Jang

Kevin Munger argues that, when an agnostic approach is applied to social scientific inquiry, the goal of prediction to new settings is generically impossible. We aim to situate Munger's critique in a broader scientific and philosophical literature and to point to ways in which gnosis can and, in some circumstances, must be used to facilitate the accumulation of knowledge. We question some of the premises of Munger's arguments, such as the definition of statistical agnosticism and the characterization of knowledge. We further emphasize the important role of microfoundations and particularism in the social sciences. We assert that Munger's conclusions may be overly pessimistic as they relate to practice in the field.

Leave-k-out (LKO) robustness checks are widely used to assess whether excluding specific observations meaningfully changes the results of an empirical investigation. However, this practice often lacks clear theoretical justification, and existing explanations are often unsatisfactory. This work addresses this gap by demonstrating that LKO diagnostics can function as sensitivity analysis for unobserved confounding through a novel conceptual tool called the "corrosive covariate." The key insight is that omitted variables can negate the influence of specific observations without literal exclusion from the sample. Unlike existing sensitivity analyses that provide only aggregate bias assessments, this framework identifies particular observations for targeted investigation, creating a principled bridge between statistical diagnostics and domain expertise. An application to divided government studies demonstrates the approach.

The design-based paradigm may be adopted in causal inference and survey sampling when we assume Rubin's stable unit treatment value assumption (SUTVA) or impose similar frameworks. While often taken for granted, such assumptions entail strong claims about the data generating process. We develop an alternative design-based approach: we first invoke a generalized, non-parametric model that allows for unrestricted forms of interference, such as spillover. We define a new set of inferential targets and discuss their interpretation under SUTVA and a weaker assumption that we call the No Unmodeled Revealable Variation Assumption (NURVA). We then reconstruct the standard paradigm, reconsidering SUTVA at the end rather than assuming it at the beginning. Despite its similarity to SUTVA, we demonstrate the practical insufficiency of NURVA for identifying substantively interesting quantities. In so doing, we provide clarity on the nature and importance of SUTVA for applied research.

We present a large-scale, multi-country assessment of the most frequently employed survey measures of support for democracy. We contrast such conventional, attitudinal measures to revealed-preference measures. The latter are based on voting for candidates in experimentally manipulated scenarios that mimic real-world electoral trade-offs between democratic principles and competing political considerations, such as partisan loyalty or policy preferences. Only a subset of attitudinal measures are both monotonic and discriminating in experimentally revealed commitment to democracy, two criteria that we develop to assess their performance. Two popular measures of support for democracy exhibit either no or reverse relationship to experimentally revealed commitment to democracy, resulting in potentially misleading findings. These findings are robust to a range of statistical techniques, including conventional and machine learning approaches for detecting treatment effect heterogeneity, and hold across a number of countries with diverse levels and histories of democracy. We propose a number of recommendations for practice, including improved formulations of established measures of support for democracy as well as set of alternative measures.

Decentralized autonomous organizations (DAOs) are a novel, blockchain-based governance systems where decision-making authority is distributed through the ownership of digital governance tokens. However, DAOs have recently come under scrutiny given the highly concentrated empirical distribution of voting power challenging the promise of decentralization. In this paper, analyzing over 10 million votes across 126 DAOs, we document three novel empirical regularities: high-powered token holders frequently abstain from voting, the highest participation rates occur among low-powered voters, and proposals typically pass with broad consensus across power levels. These findings pose a challenge to both conventional participation theory and prevailing critiques of DAO governance. These organizations may achieve de facto decentralization through strategic restraint by powerful actors. We develop three hypotheses to explain these patterns: strategic abstention as bargaining for community engagement, selective participation to safeguard against malicious proposals, and participation as filtering for governance alignment. Our findings demonstrate that democratic governance can emerge even when formal power structures appear undemocratic, offering new insights into digital governance and theories of political participation.

Researchers often rely on simulation studies to evaluate statistical methods, but the design of these simulations—particularly the choice of parameter values—is typically ad hoc. This paper proposes a principled framework for adversarial parameter selection, treating the simulation design process as an optimization problem. We introduce a general algorithm based on Bayesian optimization with deep ensembles to identify parameterizations that expose the limitations of a candidate method. Alongside this, we outline simpler diagnostic strategies that work in well-behaved settings. Applications to polarization measurement, multi-arm bandits, and heterogeneous treatment effect estimation illustrate both the flexibility and diagnostic clarity of our approach. This framework provides researchers with tools to stress-test methods more systematically, improving transparency and robustness in simulation-based evidence.

Influence functions provide a mathematical framework for understanding how individual observations contribute to statistical estimates, but their diagnostic potential remains underutilized in applied research. This paper demonstrates how the key insight that many estimators can be expressed as sample means of observation-specific influence values enables practical diagnostic applications. We present three such applications: influence function-based estimation and inference that simplifies standard error calculation for machine learning methods; systematic identification of influential subsets for assessing empirical robustness; and observation-level decomposition that allows direct comparison of how different estimators weight empirical evidence. These techniques require minimal theoretical background while providing principled tools for understanding estimator behavior. Applications to studies of political violence illustrate how these diagnostics can inform both methodological choices and case selection for mixed-methods research, offering applied researchers greater analytical control over increasingly sophisticated statistical methods.

The Negative Income Tax experiments were large-scale randomized policy studies carried out by the United States government in the 1960s through 1980s. Levels of attrition were high across the programs, which relied on extensive surveys for data collection; this disincentivized participation for households that did not receive substantial program benefits, possibly introducing correlation between missingness and potential outcomes. With application to the Gary Negative Income Tax experiment, we exploit the longitudinal nature of the data to test the robustness of treatment effect estimates to relaxed identifying assumptions. We apply several approaches to estimation that are compatible with these assumptions, including nonparametric and semiparametric estimation paired with machine learning methods to flexibly model the censoring mechanism and the response surface. Using these methods, we find no evidence of program effect on full-time employment. These methods may be more generally applied in longitudinal experiments with differential attrition.