Political Analysis (@polanalysis.bsky.social)

@polanalysis.bsky.social Oct 10, 2026, 5:51 AM

New on FirstView: which cases drive a regression result? Melamed, Wippell, Schoon & Breiger extend Regression Inside Out to hazard models and beyond. Re-analyzing a study of durable democracy, they find South Africa contributes disproportionately to the key effect.
doi.org/10.1017/pan.2026.10056

Dot plot of each country’s contribution to the estimated effect of nonviolent campaign duration on democratic survival, ranked from largest to smallest. South Africa stands far above the rest at about 0.35. Spain, Chile, Argentina and Estonia follow at about 0.1 to 0.2. Most countries make small positive contributions near zero. A few at the bottom pull in the opposite direction, with Guatemala the most negative, followed by Somalia, South Korea and Greece.

@polanalysis.bsky.social Oct 10, 2026, 5:47 AM

New on FirstView: in two field experiments with a Dutch party’s 35,000+ members, Jonne Kamphorst & @hlarreguy.bsky.social find that personal calls raised volunteer sign-ups (+83% and +361%, from low baselines), especially among younger members. Emails had no effect.
doi.org/10.1017/pan.2026.10053

@polanalysis.bsky.social Oct 10, 2026, 5:42 AM

New on FirstView: weighting can fail when nonrespondents differ on the outcome. @lukaslaffers.bsky.social , @mintal.bsky.social & Ivan Sutóris extend a fix to ordinal scales. In the 2024 ANES, it shifts life satisfaction meaningfully but economic views only modestly.
doi.org/10.1017/pan.2026.10055

Three-panel bar chart for the ANES question “How satisfied are you with life?” Top: life satisfaction by how much respondents liked the interview. Those who liked it more report higher satisfaction, and those who disliked it report lower satisfaction. Middle: raw, survey-weighted, and corrected estimates at 50%, 65% and 80% nonresponse. As nonresponse rises, “extremely” and “very satisfied” shrink while “slightly satisfied” and “not satisfied at all” grow. Bottom: at 65% nonresponse, estimated nonrespondents are much more often “slightly” or “not satisfied at all” than respondents.

@polanalysis.bsky.social Oct 10, 2026, 5:29 AM

Letter: Do AI models read ideology from faces? Jeon, Lee, @jacobmontgomery.bsky.social & @calvinklai.bsky.social varied only the face in neutral campaign ads. All three models rated women more liberal than men, beyond survey gaps. Racial effects varied by model.
doi.org/10.1017/pan.2026.10038

Flowchart of the study design in four steps, left to right. 1: A Facebook-style campaign ad showing an AI-generated portrait of a smiling woman, with neutral text asking for an endorsement in exchange for a window sticker. 2: A prompt asking the model to write a caption for the candidate and pick a political view from very conservative to very liberal. 3: The model’s caption, a campaign message about a brighter future that says the candidate leans liberal. 4: A bar chart of ideology labels across five categories, from very conservative to very liberal, with an arrow pointing to the liberal bar.

@polanalysis.bsky.social Oct 10, 2026, 5:26 AM

Letter: Jonathan Colner shows how free computer vision tools can extract text sections from poorly formatted PDFs. Pulling agenda items from city council minutes took about an hour per city vs. about 31 hours by hand, and closely matched two ground-truth datasets.
doi.org/10.1017/pan.2025.10006

Three-step workflow using a Chula Vista city council meeting record from June 10, 2014. Step 1: agenda items on a page are hand-labeled with boxes in Label Studio. Step 2: a trained object detection model finds similarly formatted items on a new page. Step 3: LayoutParser and OCR extract the text from each box.

@polanalysis.bsky.social Oct 10, 2026, 5:23 AM

How strong is Gamson’s Law? Using the isometric log-ratio transformation, new to political science, Martin & Vanberg find seat–portfolio proportionality closer to 4:3 than 5:4. It also varies: 148 of 308 governments fall significantly below the OLS estimate.
doi.org/10.1017/pan.2026.10036

Chart of estimated seat-share effects for 308 coalition governments, each with a 95% confidence interval, ordered from smallest to largest. A band shows the single estimate from a standard OLS model for comparison. Many governments fall well below that band, and a few fall above it.

@polanalysis.bsky.social Oct 10, 2026, 5:20 AM

Cho, Kim & Hazlett: standard methods pick one best-fitting model and miss the uncertainty from extrapolation. Gaussian processes keep the range of plausible fits, so intervals widen past the edge of the data. Shown for poor overlap, interrupted time series, and RD.
doi.org/10.1017/pan.2026.10032

Chart comparing three models fitted on data where X runs from -3 to 1, then extended out to X = 3. Blue is a Gaussian process, red is KRLS, gray is a quadratic polynomial. The solid line is the true function and the shaded bands are 95% intervals. Past the data’s edge, the Gaussian process band widens while the KRLS band narrows.

@polanalysis.bsky.social Oct 10, 2026, 5:12 AM

Does a college degree help every immigrant equally? In “Adaptive Randomization in Conjoint Survey Experiments,” @jennahgosciak.bsky.social , @dmolitor.bsky.social , and @ianlundberg.bsky.social adaptively search for where an attribute’s effect is strongest and weakest. doi.org/10.1017/pan.2026.10040

Chart of the estimated probability that respondents chose the college-educated immigrant over one with no formal education, shown for each of 16 contexts defined by the immigrants' other traits, with 95% credible intervals. Every estimate is above 50%, ranging from below 65% to above 75%. The contexts with the highest and lowest estimates are highlighted in red.

@polanalysis.bsky.social Oct 10, 2026, 5:05 AM

Why do some coalitions collapse sooner? In “A Multilevel Model for Coalition Governments,” Benjamin Rosche accounts for parties that join many governments and finds coalitions whose parties rely more on member funding end sooner, driven most by less cohesive parties. doi.org/10.1017/pan.2026.10039

Diagram of the multilevel structure of coalition governments in Israel, 2001 to 2011, linking three governments to the parties that served in them. Parties appear in more than one government and each government contains several parties, creating a crisscrossing pattern between parties and governments.

@polanalysis.bsky.social Oct 10, 2026, 5:00 AM

Do Americans’ issue opinions hold steady? In “A Dynamic Discrete Choice Approach to Attitude Stability and Constraint,” @alecia-nepaul.bsky.social and Steven Stern find party ID is the most influential attitude, while views on spending and immigration are unstable. doi.org/10.1017/pan.2026.10034

Heatmap from the U.S. General Social Survey showing how strongly shifts in six attitudes carry forward over time: ideological identification, military spending, welfare, immigrants, party identification, and environment. Rows show where a shift starts. The darkest cells are mostly on the diagonal, each attitude affecting itself. Most cells between different attitudes are pale or yellow (not statistically significant), with party identification to ideological identification among the stronger links.

@polanalysis.bsky.social Oct 10, 2026, 4:46 AM

Why do standard models rank the Squad as moderate Democrats? In “Accounting for Protest Voting in the U.S. Congress,” Anthony Fowler and Jeffrey B. Lewis show that once protest votes are accounted for, the Squad are among the most liberal House members. doi.org/10.1017/pan.2026.10037

Scatterplot of ideology scores for members of the 117th U.S. House, comparing a standard model (horizontal axis) with the authors' protest-adjusted model (vertical axis). Republicans are red, Democrats blue, and members of the Squad green. The Squad move from moderate Democrats under the standard model to the most liberal members under the adjusted one.

@polanalysis.bsky.social Oct 10, 2026, 4:32 AM

The October 2026 issue of Political Analysis (Volume 34, Issue 4) is out, and every article is open access. Eight papers on protest voting in Congress, coalition governments, conjoint designs, Gaussian processes, vision-language models, and more. Highlights in the thread below.