When classifier predictions route cases to specialized queues, a model update changes both accuracy and the work each team receives. Queue Shift chooses the batch assignment that maximizes candidate-model probability subject to budgets on work moved between queues, counted in cases, hours, or money, and on expected negative flips. At the movement and expected flips of any comparison rule, it has at least that rule's expected accuracy under the candidate's probabilities. Simulations show that holding queue totals fixed costs little when an update mostly reorders cases and a great deal when it moves volume between queues, and that learned queue prices nearly match exact assignment in accuracy but miss the movement budget in most batches.

Standard CART decision trees are unstable—small changes in training data can produce substantially different tree structures and predictions. Stable CART addresses this by trading a small amount of accuracy for lower cross-bootstrap prediction variance through techniques like honest estimation (using separate data subsets for learning tree structure vs. estimating leaf values), lookahead search (considering multiple future splits before committing rather than making greedy single-step decisions), and bootstrap-aware split selection (penalizing or filtering out splits that are unstable across resampled datasets).

Isotonic calibration fits flat regions: runs of scores that receive one probability. A flat region costs a decision only when the cases it pools fall on both sides of the decision's threshold, and a ranking only when its ties are broken at random. Using an exact known-risk simulation and a pre-registered benchmark, the paper finds flat regions cheap for cost-threshold decisions: choosing a calibrator by validation payoff did no better than choosing it by Brier score. For allocation, random tie-breaking costs up to 1.7 successes per 1,000 cases, which breaking ties by the original score recovers exactly.

Standard calibration minimizes average calibration error, which can hide large errors for minority groups. Leximin calibration instead minimizes the worst-off group's calibration error first, then the second-worst, and so on—applying the Rawlsian leximin criterion to the distribution of calibration quality across groups.

Multiclass calibration methods can reorder predicted class probabilities, so the class a calibrated model ranks first may differ from the class the original model ranked first. This is problematic when downstream decisions depend on the ranking, not just the probabilities. We develop calibration methods that guarantee the within-example class ranking is preserved while still improving probability reliability.

Raking needs every row seen so far, so keeping weights current on a stream means storing it and refitting. This paper gives a raking method whose state is a fixed set of running totals: each arriving row gets, in closed form, the weight that balances closeness to its design weight against the running gap to population targets. In simulations it comes close to full-data raking with moderate targets at constant cost per row; it falls short with extreme targets and many margins, and where rows can be stored, periodic raking remains the better choice.

streamcal keeps a deployed model's binary probabilities calibrated as outcomes arrive, without retaining its prediction history. It fits a monotone (isotonic) correction from fixed-size per-bin sufficient statistics, can down-weight stale history, and predicts before observing each outcome so a forecast never uses its own label.

LLM-as-judge systems produce numerical scores, but downstream decisions often require pairwise comparisons: which response is better? Converting scores to win rates requires calibration—the mapping from score differences to win probabilities. We develop calibrated estimators for pairwise win rates from cardinal LLM judge scores, accounting for judge miscalibration and non-transitivity in preferences.

setjoin implements group-constrained record matching: it assigns source groups to target groups, then matches records within the assigned groups. This is useful when both files have trustworthy group membership and the linkage should preserve it. Its current performance example is synthetic; gains on observed panel data have not been established.

A 7-step record linkage pipeline—preprocess, deduplicate, block, score, filter, decide, inspect—with multi-pass support for progressively relaxed thresholds. Implements Hungarian, greedy, and row-sequential decision rules over pairwise string similarity scores.

An R package implementing Bloom filter-based joins for improved performance with large datasets. Bloom filters provide a probabilistic test for set membership that can dramatically reduce the number of expensive exact comparisons needed during a join.

While Amazon's Mechanical Turk (MTurk) has reduced the cost of collecting original data, in 2018, researchers noted the potential existence of a large number of bad actors on the platform. To evaluate data quality on MTurk, we fielded three surveys between 2018 and 2020. While find no evidence of a "bot epidemic," significant portions of the data—between 25%-35%—are of dubious quality. While the number of IP addresses that completed the survey multiple times or circumvented location requirements fell almost 50% over time, suspicious IP addresses are more prevalent on MTurk than on other platforms. Furthermore, many respondents appear to respond humorously or insincerely, and this behavior increased over 200% from 2018-2020. Importantly, these low-quality responses attenuate observed treatment effects by magnitudes ranging from approximately 10-30%.

When the population is divided into known strata with known variances, the optimal allocation of a fixed sample budget across strata depends on stratum size, variance, and sampling cost. We derive the optimal allocation and show how much efficiency is lost by common rules of thumb like proportional allocation.

Estimating quantities like average potholes per kilometer or pedestrian density requires randomly sampling street locations within a region. Geo-sampling addresses this by downloading street network data from OpenStreetMap for a specified administrative region, splitting each street into 0.5km segments (recording the lat/long of segment endpoints), building a database of all segments, and then drawing a random sample that can be exported as a CSV or visualized on a map for field data collection.

Given a set of spatially distributed tasks (e.g., field survey locations, audit sites), Allocator computes optimal itineraries that minimize total travel time or distance, assigning tasks to enumerators and sequencing visits within each assignment.

Randomly sample GitHub repositories, optionally filtered by language, creation date, or star count. Useful for constructing representative samples of open-source projects for empirical software engineering research.

When running OLS on per-row trait shares (e.g., fraction of items in a category), labeling every item is expensive. Random labeling wastes budget on items that barely affect the regression. fewlab identifies the items with the highest statistical leverage—those that most influence the coefficient estimates—and prioritizes them for labeling, achieving unbiased OLS with a fraction of the labeling cost.

StatQA is a modern Python framework for automatically extracting structured facts, statistical insights, and multimodal Q/A pairs from tabular datasets. It converts raw columns and values into clear, human-readable statements paired with rich visualizations, enabling rapid knowledge discovery, CLIP-style multimodal RAG corpus construction, and LLM training.

In randomized experiments with imperfect compliance, the LATE requires observing treatment receipt, and the Wald estimator requires monotonicity. When receipt is unobserved but the experiment has a pre-treatment time series and a distinct delivery window, a structural break test applied to the delivery window can classify treated units into compliers, never-takers, and defiers. This yields an inferred compliance rate with closed-form bias correction, an empirical test of monotonicity via defier detection, and a characterization of the complier subpopulation that can be projected onto the control group.

Researchers increasingly use machine-learning predictions as covariates, including controls for confounding. Prediction error can then change the estimated effect of interest. Regression calibration offers an established baseline: use a labeled sample to predict the true covariate from its machine-generated surrogate and the other regression variables. We explain when this procedure identifies a linear model and what can go wrong when calibration must itself be learned. A population projection can suffice without a correct conditional-mean model, but estimated calibration must preserve the residual relationships that identify the downstream coefficients. We connect departures from those relationships to coefficient error, distinguish many calibration predictors from many mismeasured covariates, and separate estimation error from failures of support and transport. Exact examples show how omitting treatment or imposing shrinkage changes confounding adjustment; simulations illustrate these mechanisms. The resulting framework makes regression calibration a transparent benchmark for evaluating methods that use machine-learned covariates.

Correcting errors in machine-learned covariates can change a treatment-effect estimate in ways that separate corrections do not predict. Each correction changes the regression's adjustment for other variables, so its effect can depend on which errors have already been corrected. Existing methods use verified measurements to correct inference and guide validation. We study the joint effects of finite corrections on a fitted coefficient. Our diagnostic replaces sets of proxy covariates with verified values and compares the resulting refits with the sum of isolated correction effects. Exact least-squares identities locate the source of nonadditivity in changes to the regression's cross-product matrix. Correction paths describe how the coefficient moves as more measurements are corrected, and their dispersion describes how much the choice of correction set matters. Established Shapley methods attribute the endpoint change. Simulations illustrate these distinctions. Applications to housing and text-coded incivility compare refits with three additive approximations: isolated corrections track small random correction sets closely but misstate the full correction. The resulting diagnostic measures both the consequences of individual errors and the additional changes produced by correcting them together.

Event study designs estimate period-specific treatment effects with known standard errors from two-way fixed effects regressions. We treat the coefficient sequence as observations from a local linear trend state-space model and apply the Rauch–Tung–Striebel (Kalman) smoother, using the known heteroskedastic regression standard errors as observation noise. The smoother adapts to local precision—trusting the trend model when a period's estimate is noisy and trusting the data when it is tight—reducing level MSE by 80% and derivative MSE by 98% versus raw estimates, while also providing a derivative-based parallel trends test with correct size.

Spurious associations in training data arise because models learn correlations rather than causal relations. We propose a composite loss that incorporates causal and behavioral priors: an invariance penalty for label-preserving perturbations (e.g., swapping gendered pronouns should not change predictions), a directional penalty for perturbations that should change the label in a known direction, and a falsification penalty inspired by epidemiological negative controls that discourages reliance on features with no causal relationship to the outcome. The framework unifies ideas from CheckList-style behavioral testing, causal inference, and adversarial robustness into a single training objective.

Assignment-based anticlustering splits units into equal groups that are as alike as possible and scales to millions of units, but it is deterministic. Jittering the sort that defines its batches and shuffling the group labels make it an experimental design with exact randomization inference. Each of its steps turns out to be the greedy allocation of the multi-arm adaptive design ARMM, found by an assignment problem instead of a search over all orderings of the arms. In simulations with 4 to 50 arms, it balanced covariate means better than every other design whose cost is linear in the number of units, by a margin that grew with the number of arms; run on the covariates and their squares, it gave the smallest error and the most power for a nonlinear outcome in nearly every setting.

Social scientists increasingly wish to reason causally about high-dimensional treatments (texts, images, prompts) while isolating the effect of a single latent concept. We treat a pretrained model as a measurement device and use minimal edit pairs to estimate a concept-tangent subspace in activation space, with diagnostics that can explicitly reject the existence of a clean subspace for a given encoder and concept. When diagnostics are favorable, activation-steering maps constrained to the subspace generate approximate minimal edits, and concept coordinates serve as treatments in a double/debiased ML estimator. The resulting estimand is a local, representation-dependent linear effect—not a universal causal effect of the underlying human concept.

One predicted item label can enter many records and several constructed variables. This paper traces the consequences for coefficients and standard errors when aggregates are regression outcomes, covariates, or both. It connects generated-variable and shared-source inference, gives joint moment corrections and covariance estimates under item validation, and states the additional conditions required for normal coefficient intervals. Simulations distinguish coefficient bias, empirical variability, estimated standard errors, and coverage, including failures with influential or few items. Browsing-data illustrations show the roles of reuse, visit concentration, and observed classifier disagreement.

Estimating the trend (derivative) at a specific point in a noisy time series is difficult because naive approaches like computing differences between consecutive observations amplify noise rather than reveal the underlying signal. Incline addresses this by first smoothing the time series using either Savitzky-Golay filters (local polynomial fitting) or smoothing splines, then estimating the first or second derivative of the smoothed function at chosen points in time. The difference between naive and smoothed estimates can be substantial: in the provided example, the correlation between them is -0.47, making the choice of method consequential for applications like detecting sudden cost increases or identifying rapidly changing patient health trajectories.

Bandwidth selection for Nadaraya-Watson kernel regression and kernel density estimation typically relies on cross-validation, which is computationally expensive. This package implements an analytic bandwidth selector based on the Hessian of the cross-validation objective, giving a closed-form approximation that avoids the grid search.

Generalized least squares requires estimating and inverting the error covariance matrix, which is expensive when the matrix is large or unstructured. This package uses alternating least squares with a low-rank factor-analytic decomposition of the covariance, making iterative GLS feasible for problems where the full covariance is too large to invert directly.

Classifiers produce continuous scores; deployment requires a threshold. The optimal threshold depends on the metric you care about—F1, balanced accuracy, or a custom cost function—and the score distribution. This script computes the exact threshold that maximizes a given stepwise metric, avoiding the grid-search approximation common in practice.

pyppur is a Python package that implements projection pursuit methods for dimensionality reduction. Unlike traditional PP objectives geared toward finding 'interesting' projections (non-gaussian), pyppur focuses on finding non-linear projections by minimizing either reconstruction loss or distance distortion.

How should a machine learning startup serve more queries, faster and of more kinds, without costs running away? We split costs into variable costs, of serving models and keeping them performant, and fixed costs, of developing and training new ones, and suggest ways to cut each. Since the largest fixed cost is often finding the root causes of failures, we also describe how we approach that.

Counts of research software use allocate credit, and the available ones are cheap to produce: a download is one fetch of a file. I propose validated use, a package loaded in replication code deposited at a journal that checks it, as a costly signal, and measure it in 13,245 deposits from 75 economics and political science journal collections, which load 4,565 R, Python and Stata packages. Among packages the corpus uses, validated use and downloads agree loosely, and least where automatic installation is commonest: the rank correlation is 0.61 for Stata, 0.53 for R and 0.36 for Python. Downloads track a package's position in the dependency graph more closely than its use. The most used software prepares exhibits: estout is in 43% of Stata deposits. Stata appears in 1.4 times as many deposits as R and is missing from studies of research software because its packages are scattered across SSC, two journals' archives and authors' sites with no index across them. I build that index, 10,147 commands in 5,581 packages, and release it with the mention-level data.

This project estimates Python package use by sampling public GitHub repositories and counting import statements. It provides data and a dashboard as an alternative to download counts, which can reflect automated installation rather than adoption.

Open-source software is widely used in commercial applications. Pair that with the fact that when choosing open-source software for a new problem, developers often use social proof as a cue. These two facts raise concern that bad actors can game social proof metrics to induce the use of malign software. We study the question using two field experiments. On the largest developer platform, GitHub, we buy ‘stars’ for a random set of GitHub repositories of new Python packages and estimate their impact on package downloads. We find no discernible impact. In another field experiment, we manipulate the number of human downloads for Python packages. Again, we find little effect.

Finding an error can improve later research only if researchers respond to it. We examine citations after methodological audits in neuroscience, animal experiments, and political science. OpenAlex histories through 2025 show continued citation to flagged neuroscience papers: their median annual citations rose from 6.5 in 2010 to 14–19 during 2012–2015 and were 7 in 2025. Comparisons within the same journal and publication year suggest somewhat slower proportional citation growth after the critique. A separate sample of citing passages rarely acknowledged a concern. The paper distinguishes persistent attention from the effect of publicity on citation growth.

Using more than 3,000 retracted articles and over 74,000 citations, we study how problematic research continues to circulate. At least 31% of citations occurred a year or more after retraction, and about 91% of post-retraction citations did not acknowledge a concern with the cited article.

A GitHub Action that checks BibTeX references against Retraction Watch data and opens an issue when it finds a potentially retracted citation. It matches by DOI or bibliographic information and can run on a schedule or after bibliography changes.

An archived web application for finding citations to retracted articles in pasted APA references. It compares parsed citations with a database of retractions and highlights possible matches for review.

AutoSum collects passages surrounding citations to a publication in openly accessible papers. Collating these passages provides a view of how other scholars summarize the work and helps identify possible miscitations.

This note proposes a framework for constructing interpretable machine-learning benchmarks. It emphasizes defining the task and target population before choosing data, and assessing label quality and whether performance reflects the intended capability.

A review of Noise: A Flaw in Human Judgment that examines variation in judgments made from the same information. We consider what disagreement reveals about uncertainty, skill, effort, preferences, and the institutions in which decisions are made.

We compare how often empirical articles in the American Political Science Review and American Economic Review include precise numerical results in their abstracts. The project provides coded data and analysis of differences in quantitative reporting.

We collect article metadata from five political science journals to examine changes in authorship, article length, and readership. The data document the rise of coauthorship and the highly uneven distribution of article and abstract views.

SuperDF extends R data frames with metadata such as version, author, and notes. The metadata is designed to persist through common data operations and input/output, keeping documentation attached to the data.

We apply static analysis to R scripts in replication archives for American Journal of Political Science articles. The project documents style issues, warnings, and errors detected by lintr and provides the scripts used for the audit.

This note organizes explanations for inter-group prejudice, including group identity, personality, motivated beliefs, and differences in reasoning. It distinguishes possible mechanisms behind aversive beliefs and feelings rather than treating prejudice as a single process.

We examine polarization as social distance and hostility between partisans, rather than only disagreement over policy. Evidence from several sources shows growing dislike of opposing partisans, with policy attitudes providing an inconsistent explanation and campaign messages offering a plausible alternative.

Americans greatly overestimate how many party supporters belong to stereotypical groups. Survey experiments suggest that these misperceptions are not simply expressive responding or ignorance of population shares; correcting beliefs about the opposing party reduces perceived extremity and social distance.

We test whether the representativeness heuristic contributes to party stereotypes. Experiments involving conjunction judgments, information about group partisanship, and cognitive load suggest that this mental shortcut can distort perceptions of who belongs to each party.

Why do moderate partisans remain warm toward more extreme leaders of their own party? We find both distorted perceptions of leaders' positions and asymmetric treatment of policy disagreement: even when informed about positions, partisans do not necessarily penalize more distant co-partisan candidates.

Candidate evaluations diverge over the 2000, 2004, and 2008 presidential campaigns, with differences across elections and measures. Ratings of both one’s own candidate and the opponent contribute. Trait-rating gaps widen faster in battleground states, while overall favorability shows a less consistent geographic pattern. These survey comparisons describe change without identifying the effects of campaigning or negative advertising.

Two experiments and a survey test whether partisan cues bias simple visual evaluations, such as counting errors or assessing a scene. Across these tasks, the estimated partisan effects are generally small.

We use variation in state right-of-way regulations to study the effect of broadband access on partisan hostility. Combining broadband availability with survey data from 2004 and 2008, we find greater hostility and greater consumption of partisan media where broadband access expands.

Observed browsing data show that online pornography consumption is concentrated among a small share of users. Republicans consume somewhat more than Democrats in the unadjusted data, but the difference disappears after accounting for age and gender.

An average endorsement effect does not identify how many people favor the endorser. Reanalysis of a 2009 Pakistan survey with an ordinal measurement model yields a 43% probability of a positive latent reaction, averaged across four militant groups in the unweighted sample (95% credible interval: 32–53%). Small negative mean reactions coexist with substantial dispersion.

We compare White respondents' racial stereotype reports in confidential computer-assisted and oral interviews in the 2008 American National Election Study. Confidential responses are slightly more negative about both Black and White people and have no greater predictive validity, casting doubt on a large social-desirability distortion in the oral measures.

Is partisan affect principled? In a 2013 MTurk experiment, reports of declining rather than sustained support for a leader after controversy lowered estimated warmth toward fellow partisans, suggesting a possible preference for loyalty. In a 2018 YouGov experiment, portraying opponents as fair-minded modestly improved estimated trait ratings, with little estimated movement in warmth. Portraying allies as biased produced no clear penalty, but many readers also reported that their party was about as biased as expected. The evidence is suggestive: uncertain estimates, missing warmth ratings, bundled article features, and incomplete comparisons prevent firm conclusions about effect sizes, mechanisms, or a partisan double standard.

If people prefer more for both sides to less for both, everyone should select the plan giving both sides more. In a 2018 CCES highway choice, only 35.0% of Democrats and 27.1% of Republicans did so; the rest chose smaller allocations that left their side ahead. In a post-election experiment, raising opponents’ income gain from 3% to 7%, while keeping the own-party gain at 5%, changed policy support by -20.0 points on a 0–100 scale (95% CI [-23.9, -16.1]). Relative partisan gains affect support, but the exercises do not distinguish spite from dislike of falling behind; the highway comparison also changes total spending.

We assess deliberative polls and other mini-publics using Robert Dahl's criteria for a democratic process: inclusion, effective participation, enlightened understanding, voting equality, and control of the agenda. The chapter examines where these institutions meet the criteria and where their democratic claims require qualification.

Using open-ended responses from a Deliberative Poll on schooling in Northern Ireland, we compare the reasons articulated by returning participants and contemporaneous controls. Participants gave somewhat more reasons, but the estimate is imprecise and participation was self-selected. The samples differed little in directional balance. Among returnees, repertoires moved toward balance chiefly because supporting reasons declined, and the estimate is sensitive to coder choice.

Across 27 Deliberative Polls, factual knowledge increased by an average of 14.3 percentage points, a relative gain of 39.3%. In two polls with control groups, attendees gained an average of 11.2 percentage points more than controls. Measurements before arrival, on arrival, and after deliberation in two studies show substantial gains before participants arrive. The gender gap narrows but remains. Across 20 polls with poll-specific education cutoffs, the gap between more and less educated attendees widens.

Across 14 Deliberative Polls, participants became more accurate about facts that challenged their initial positions in 13 studies. Gains on these facts exceeded gains on congenial facts in 11 studies, but only two conditional intervals excluded zero. The comparisons describe learning across initial positions; they do not identify a causal effect of deliberation or distinguish motivated learning from initial knowledge, response behavior, and room to improve.

Across 19 misinformation-type items in seven Deliberative Polls, incorrect answers fell from 23% to 10%, and participants who began wrong were as likely to end right as those who began not knowing. In polls with control groups, attending cut overestimates of the number of undocumented immigrants by 15 points, but reduced confident denial of human-caused warming only slightly and not durably.

Across 2,601 group-issue pairs in 21 Deliberative Polls, we test whether discussion routinely homogenizes or polarizes attitudes, or shifts them toward those of socially advantaged participants. We find no strong or routine pattern of these distortions; the modest patterns include some homogenization and moderation.

Observed within-group variance mixes true convergence with changing measurement error. Lower error at T2 can manufacture homogenization and raise the intraclass correlation, but classical independent error cannot increase the absolute variance of group means. The paper proves these boundaries and specifies a latent partially nested model that separates convergence, between-group divergence, and crystallization. The full empirical model has not yet been fitted.

A survey and deliberative field experiment examine young, educated Kirkuk residents' preferences for governing their divided society. Participants support an equal say for ethnic groups, and informed deliberation broadens support for regional autonomy without substantially changing support for equal say.

We classify roughly 5.4 million webpages from 276 British news outlets to estimate how much coverage concerns topics outside public affairs. A classifier trained using article URL categories and checked against hand coding estimates that about 39% of articles are apolitical news.

We combine text-based and audience-based measures of British news outlets' ideology. Parliamentary speeches, party manifestos, and Twitter following patterns provide complementary information about ideological position and differences between measurement approaches.

We distinguish which issues news outlets cover from the positions they take within those issues. A supervised topic model and congressional-speech-based slant measures reveal systematic differences across 255 news programs on 50 television channels.

A Python library for classifying news by topic and distinguishing hard from soft news. It offers URL-pattern rules, trained US and UK text models, and optional language-model classification through Python and command-line interfaces.

We compare readability across New York Times articles and CNN, NPR, and MSNBC transcripts. The analysis examines variation across outlets and over time as one possible barrier to accessing political information.

We estimate the ideological positions of more than 2,300 media sources using the following patterns of politically interested social-media users. Comparisons with content-based measures support the approach, and applications examine journalists' ideology and ideological diversity within outlets.

We study presidential campaign horse-race headlines across six online news outlets in 2012 and 2016. Headlines tend to favor their outlets' typical readers, but within-outlet engagement provides little evidence that readers prefer congenial headlines and somewhat more evidence of demand for uncongenial ones.

We reconsider claims that greater choice over political information increases knowledge gaps and polarization. The review questions the premise of widening knowledge gaps and finds that the evidence connecting selective exposure to polarization is more limited and nuanced than common accounts suggest.

We use the annotated New York Times corpus from 1987 to 2007 to study the mix of public-affairs and apolitical coverage. The project also develops analyses of geographic focus using article metadata and location fields.

A coded sample of more than 5,000 network television news segments from 1968 to 2019 documents changes in political content and geographic focus. Coverage unrelated to politics increases over time, and local news becomes a substantially larger share in the later years.

We combine scraped PolitiFact ratings with independently coded party affiliations to examine differences in scrutiny and ratings across parties. The analysis distinguishes how often statements are checked from the average rating among the statements selected for checking.

A scheduled collector archives URLs from major news sources for studying news production. The repository contains source-specific URL collections and links to historical full-text releases, whose coverage differs from the current collections.

A corpus and scraper for CNN transcripts spanning 2000 through March 2025, with program names, dates, times, and transcript text. The repository documents collection across versions of CNN's site and links to research-access data on Harvard Dataverse.

We combine television schedules with program metadata to study changes in genres and the race and gender composition of casts and production teams. The repository includes collection scripts, data, and figures describing these trends.

We code employment among fictional female characters in Indian evening television soaps. In data collected in 2015 on 172 characters across 20 shows, 36.6% of working-age female characters are portrayed as working, with substantial variation across shows.

We tally the coded race and gender of victims and criminals across three Law and Order series. Black characters appear at lower shares than in several historical arrest and victimization series, while women appear at higher shares among criminals than in several arrest categories. The comparisons are descriptive because the sources differ in crimes, places, and years.

A practical framework for diagnosing problems: generate specific explanations, assess their plausibility, and design checks that distinguish among them. Examples show how correlations, close examination of failures, and comparisons with successful cases can help identify causes.

Two larger surveys replicate the finding that people sometimes prefer a certain option to a lottery whose worst outcome is better. Clarifying the guarantee reduces the effect, while expressing the lottery as natural frequencies or using monetary choices eliminates it in these experiments.

Five archival survey experiments test whether equivalent gain and loss descriptions change risky choices. All 23 contrasts show the classic framing direction, but explicitly stating the certain outcome removes much of the gap, suggesting that inferences from incomplete descriptions matter alongside gain and loss labels.

In a two-arm comparison in the 2020 Cooperative Election Study, 711 respondents rated one of two tweets. Both said the system rewards bad behavior; one said politics attracts good people, the other bad people. The bad-people tweet scored 6.2 points lower in perceived sophistication on a 0–100 scale (95% CI: −9.4 to −3.0). Covariate adjustment and survey weighting preserve the negative difference. A causal interpretation requires random assignment within the rating sample; the original questionnaire and randomization program are unavailable.

We compare movie ratings across platforms using data on 16,319 American films released between 1950 and 2020. Among platforms with sufficient coverage, average ratings are only moderately correlated, revealing substantial differences in assessments of the same films.

Surveys from 2010 and 2012 find substantial misunderstanding of the Affordable Care Act despite support for many of its provisions. The analysis estimates that fuller understanding could increase support for the law; this is a modeled implication, rather than a demonstrated effect of an information campaign.

We examine whether Americans evaluate the Affordable Care Act through beliefs about its individual provisions. Despite partisan disagreement, both Democrats' and Republicans' evaluations are associated with what they believe the law does and how confident they are in those beliefs.

We revisit evidence that legislators with daughters vote more liberally on women's issues, extending the analysis to eight congresses before and eight after the original study period. We find no daughters effect in either extension, challenging the explanation that later party polarization alone erased the relationship.

Two surveys nearly a decade apart show that people's images of political parties center on a few prominent national figures, with more extreme politicians recalled more often. Television news coverage is associated with which politicians come to mind.