blog post: wanting workflow We review a covid survey case study from the new Bayesian Workflow book. We model: - measurement: test sensitivity and specificity - representation: differences between sample and population (MRP) How can workflow help us here ?

blog post: structured MRP to smooth survey weights Adjusting for lots of variables can lead to very large weights. So @yajuansi.bsky.social, @trangucc.bsky.social, Jonah Sol Gabry, and Andrew Gelman turned to a structured MRP and its equivalent weights.

blog post: equivalent models, equivalent weights (locally) with survey-weighting methods, we can ask: under which outcome models do they do well ? with outcome-model methods (like MRP), we can ask: what are the (locally) equivalent weights ?

blog post: poststratification without population level information Poststratification uses population data on X to help estimate a population mean E(Y). But sometimes population data on X isn't available: In 2016 Andrew asked pollsters to poststratify on party ID, but how ?

blog post: quantifying uncertainty in ranked choice voting polls RCV uses rankings to get a winner by instant runoff. Polls estimate rank probabilities with uncertainty. Unlike with non-RCV, even in random samples a plurality of uncertainty mass can get an incorrect winner.

blog post: toy example for energy balancing weights How do energy balancing weights (used now by the NYT/Siena Poll) handle unsampled population groups ? Let's work thru a toy example and compare to Poststratification, Raking, and MRP.

blog post: Big Changes in the Times/Siena Poll 2 changes to their survey weights: 1. new weighting variable: support score 2. new weighting method: energy balancing

blog post: perfect collinearity in the sample but not in the population Two variables are perfectly collinear in your sample, so you drop one. You use your model to predict in the population. What can go wrong ? Let's talk thru a Census Bureau toy example from BDA2.

blog post: using MRP in later analyses (pride edition) happy pride ! 🌈 @jeffreylax.bsky.social & Phillips 2009 used MRP to estimate state-level public opinion about policies affecting gays and lesbians. They then use this as a predictor of whether the state adopts the policies.

blog post: should MRP workflow include LOCO-CV ? Individual-level loss orders models differently than the population-level loss we want judging MRP. To get population-level loss, use out-of-sample classical poststratification to compare with MRP. How to split data ? LOCO = leave one cell out.

blog post: it is (still) the people Survey Statistics blog series' 1st birthday 🥳 Andrew Gelman's 60-ish Birthday 🥳 and NYT weights with synthetic past vote 🗳️

blog post: double-plus robustness Meng (2022): GREG is not only “double robust” (consistent if either the outcome model or response model are correct), but “double-plus robust” (consistent if what is left of the outcome model and response model are uncorrelated).

blog post: GREG GREG is Generalized REGression estimator. We can think of it either as: 1. Adjusting an estimate based on the model with a Horvitz-Thompson estimate of the error, or 2. On the flip side, adjusting the Horvitz-Thompson estimate with the model.

blog post: relevant alternatives ? We saw that the multinomial logit model implies independence from irrelevant alternatives (IIA). Let’s expand the model to include choice set C within the logits f(X_ic,C), allowing for non-IIA.

blog post: exploded logit ! a common choice model is multinomial logit. this model implies that rankings follow an exploded logit !

blog post: dCV for MRP Splitting a cluster between training and test fits models with more information than we should ("data leakage"). But not splitting a stratum between training and test fits models with LESS information than we should. bsky.app/profile/stat...

blog post: irrelevant alternatives ? a common choice model is multinomial logit. this model implies Independence of Irrelevant Alternatives (IIA), e.g. the ratio of Left-vs-Right preference is the same in round 1 as in the runoff.

blog post: improving with structure We’ve met Mr. P (Multilevel Regression and Poststratification). We’ve met Mrs. P (Multilevel Regression with Synthetic Poststratification). Now let’s meet Ms. P (Multilevel Structured regression with Poststratification).

blog post: design-based cross validation how to split train and test sets to respect survey design ? what lessons carry over to nonprobability samples ?

blog post: Individualism and the CV Noise Problem Politically meaningful differences among models can be swamped by cross-validation noise.

blog post: individualism doesn't work (even when weighted) individual-level loss (even weighted to the population) orders models differently than the population-level loss of interest to folks using MRP

blog post: work with us at Blue Rose ! use cutting edge statistics, machine learning, and engineering to study public opinion, forecast elections, and advise Democrats.

blog post: sampling-weighted loss we use sampling weights to estimate a population mean E(Y). what about to estimate a conditional mean E(Y|X) ? the best-fit model in the sample may not be the best-fit model in the population.

blog post: sampling to assess data quality @bhedtgauthier.bsky.social et al. (2012) used sampling to assess and improve data quality in Malawi

blog post: Gallup's Presidential Approval Ratings Gallup will no longer track presidential approval after 88 years Let's look at their sampling, mode, and weighting (still used for other survey questions)

blog post: more on recalled vote we've talked about measurement error in recalled vote in the US. how does this change in multiparty states ?

blog post: 5 flavors of calibration 2 from survey statistics 1 from machine learning 2 from Gelman et al.'s workflow article

blog post: Total Margin of Error (Part II) For election polls from 1998 to 2014 Shirani-Mehr et al. found: margin of error = 2 x (reported margin of error) Let's revisit Meng's “Statistical Paradises and Paradoxes” to understand this more generally.

blog post: Total Margin of Error margin of error = 2 x (reported margin of error) and how much of this error is "bias" vs "variance" ?

blog post: Margin of Error how can we get a poll's margin of error ? let's start with MRP and some simplifying assumptions.