-
A bigger local language model knows religious groups differ, then answers No to everything
Qwen2.5-3B on a CPU ranks GSS group gaps better than 1.5B (Spearman 0.56) but saturates to No, so its gap on the probability scale shrinks to 2% of the truth.
-
Recounting 107 pre-registered predictions by script found a total that added to 63, not 62
A script that re-reads every logged prediction and verdict for the GSS imputation paper: 59 held, 22 partly, 26 refuted, and one arithmetic slip fixed.
-
A small local language model knows which way religious groups differ, but not by how much
Qwen2.5-1.5B on a CPU, given only a GSS respondent's background, ranks unasked items at AUC 0.57 and predicts group gaps at a tenth of their size.
-
Six paper figures that refuse to draw if they disagree with the text
The figures are drawn from committed CSVs after the script recomputes the 24 numbers the text quotes. All 24 agree; two figures add readings.
-
Shrinkage survives the survey design; its interval needs one more variance term
Against the population share under the GSS's strata and PSUs, shrinkage keeps its gain, but its interval covers 93% until the survey's variance is added.
-
Survey design effects on a rotated item: small clustering, but a 4 to 9% interval gap
With GSS strata and PSUs, design effects on pooled subgroup shares are only 1.06, but the survey's own error leaves calibration intervals 4-9% too narrow
-
Pick the shrinkage prior by kind of subgroup, then stop re-choosing it
Choosing each subgroup kind's shrinkage prior from the other survey items agrees with the rule 37 of 39 times, but the two misses erase most of the gain
-
Rotating a survey item off a subgroup's ballot: the paper draft
Seven pre-registered GSS experiments, one argument: keep a calibration subsample, shrink toward the transferred model, borrow only shared drift
-
A shrinkage prior rule, tested on survey items it never saw: half the gain survives
Fixed in advance, the by-kind prior rule cuts subgroup error to 0.85 of the zero prior's on 7 fresh GSS items, against 0.65 on the items it came from
-
Which subgroup is the model getting wrong? Ask the other subgroups
Shrinking toward the other traditions' miss cuts 2016–24 subgroup error to 0.26 of a zero prior's, matching an oracle; it fails for immigrant generations
-
A shrunk subgroup estimate is more accurate. Can you still trust its 95% interval?
On 227 GSS religion subgroups, the robust empirical Bayes interval covers 96.8% on average at 50 respondents per group and is 17% narrower than the unshrunk one
-
One calibration budget, several subgroups: the split barely matters, the estimator does
With one budget for several subgroups asked a rotated survey item, Neyman allocation moves the error under 1%. Shrinkage halves it.
-
If you stop asking a survey question, how many people still need to answer it?
A model trained on another group saves 17% of the calibration sample a rotated GSS religion item needs; ±3 points still takes 700 to 1,400 respondents.
-
One extra variable, the decade, takes a quarter off a mixture imputer's error
Adding the survey decade to the Si-Reiter mixture cuts its error on held-out GSS Catholic years from 4.8 to 3.55 points and same-sex marriage from 20 to 11.5.
-
Why a survey model trained on one Christian group gets another group's level wrong
A transferred GSS imputation model misses a group's share by 5.6 points. Covariate shift explains under a point; the group's other answers cannot fix the rest.
-
Two opposite imputation models miss an unseen survey year by the same amount
On held-out GSS years of Catholic opinion, the Si-Reiter mixture misses by 4.8 points, MICE by 4.7, interpolation by 3.4: the barrier is period, not model.
-
Does fitting a survey model on nearby years predict an unseen ballot better than copying its neighbours?
Fitting only on the nearest 5 years cuts a boosted model's held-out-year error to 3.44 points, level with linear interpolation at 3.43.
-
Can a trend term predict a whole unseen GSS survey year of Catholic opinion?
A level shift fitted on neighbouring survey years brings a boosted imputation model to parity with simple interpolation, no better, on 13 Catholic GSS items.
-
Does a survey model trained on US-born Catholics work for Hispanic immigrant Catholics?
On GSS 1972-2024, a model fitted on third-generation Catholics ranks Hispanic immigrants about as well as the pooled model but misses levels by up to 13 points.