IfD director Dr. Jozef Michal Mintal with IfD faculty affiliate Dr. Lukas Laffers and Dr. Ivan Sutoris from tne National Bank of Slovakia, tackle in their new paper in Political Analysis one of the biggest current issues in survey research.
The problem: widely used adjustments such as post-stratification or raking get you to the demographic margins, but they assume that once you account for demographics, the people who responded look like the people who didn’t. But when respondents differ from nonrespondents on the outcome itself (nonignorable nonresponse), these adjustments can fail, introducing bias even into basic descriptives. The people you fail to reach can be the ones who feel differently about the question itself, and then a “perfectly weighted” sample still gives you a very biased estimate.
Mintal, Laffers, and Sutóris, now have a practical way to correct for this in ordinal (and binary) outcomes.
This began as a very practial problem for Mintal. While working on reweighting the hashtag#VolebnýKompas data, this issue kept coming up. Michael Peress had a clever answer back in 2010: take something you observe among respondents that tracks how willing they were to cooperate (interviewer-coded cooperativeness, how they rated the interview, and so on) and use it to extrapolate from the reluctant respondents toward the people who never responded, while keeping the benefits of post-stratification. The catch was that it was set up for binary outcomes, and a lot of what we actually care about in political science, sociology and neighboring fields sits on an ordinal scale. Collapsing a 5-point scale into binary throws away information and can change the answer.
Mintal, Laffers, and Sutóris generalized it to ordinal outcomes. The key parameter is ρ, the correlation between the latent errors in the outcome and response equations, which captures selection on the unobserved outcome after conditioning on the covariates. Near zero means there is nothing to correct beyond what your covariates already handle. If it isn’t, you get corrected shares, and the sign of ρ tells you which way the more reluctant respondents and by extrapolation the nonrespondents lean.
They ran it on the 2024 ANES. Life satisfaction moved a lot (ρ ≈ 0.47; the corrected distribution is noticeably less satisfied than the weighted one). Retrospective economic evaluations moved a little (ρ ≈ 0.14). Death penalty, trust in media: ρ statistically indistinguishable from zero, corrected and weighted shares nearly identical. Whether nonresponse bias matters depends on the question, and it is now something you can explore for both binary and ordinal outcomes rather than argue about.
The whole thing is a compact R routine and runs on a normal laptop, our ANES case took about ten minutes. The paper is open access and the code is on Dataverse + a reproducible capsule on Code Ocean: Paper: https://doi.org/10.1017/pan.2026.10055