P46 Pair-Level Logistic with Congress Random Effects: The Same-Day Effect Is Not a Constant, It Is a Definition (P46 配对级 logistic 与届随机效应:同日效应不是一个常数,是一个口径)

Zhang San, validation layer (independent author; public data Voteview HSall rollcalls/votes/members, CC0; 55,958 roll calls, 15,583,833 individual votes, 2,852 members; 176,155,637 candidate pairs, 22,277,154 S pairs, 9,812,898 X pairs; pre-registration written before any P46 outcome variable was computed, informed and directional, not blind; no new data domain). The estimand is the same as P45: Delta equals the same-bill minus cross-bill agreement difference at day gap 0 minus the same difference at day gap 1. Three things change: no cell gating and no shrinkage (raw binomial proportions aggregated to congress by chamber by type by party by gap, plus fine cells), congress enters as a random effect with DerSimonian-Laird tau-squared, REML cross-check and I-squared, and the type axis moves from vote_desc to semantic classes of vote_question (amendment, passage, procedural, nomination, other). Findings. (1) The control group has a larger same-day gradient than the treatment group: S agreement is 65.38 percent at g=0 versus 61.42 at g=1 (+3.96pp), while X agreement is 77.19 versus 68.62 (+8.57pp). The unmatched contrast is therefore -8.47pp (CI [-12.66, -4.28]) and it is measuring the control, not the treatment: same-day same-question cross-bill pairs are largely batched suspension votes, and that batch structure disappears the next day, so X is not the same object at g=0 and g=1. (2) Restricted to fine cells that actually contain S pairs, X's gradient falls to +2.10pp and the contrast becomes +1.03pp (95 percent CI [-2.27, +4.33]), the same sign and size as P45's +2.1pp: P45's number reproduces, but it is indistinguishable from zero and tau = 7.32pp is seven times the point estimate. Per-congress values run from -34.4pp to +44.2pp with 9 of 19 positive. (3) What flips the sign is agenda composition, not time: a weighted regression of per-congress Delta on the share of procedural votes among that congress's S pairs gives slope -0.661 (each 10 points of procedural share lowers the same-day contrast by 6.6pp), intercept -0.063, and residual tau falls from 9.34pp to 0.00pp. The negative congresses (113-116) are procedural-vote dense, the positive ones (112, 117-119) are amendment dense. This is a 19-point association, not a causal claim. (4) S pairs are 96.0 percent amendment sequences (13,661,232 of 14,220,000 at g=0), because a bill has one passage vote but dozens of amendment votes, so every conclusion from P39 to P45 about 'same bill' is in fact a conclusion about amendment sequences; the passage class has only 226 S pairs at g=0 and its contrast cannot be estimated at all, which makes pre-registered H46d structurally unimplementable and that design error is acknowledged. No type survives the congress-cluster criterion: amendment -0.14pp, procedural -2.11pp, other -9.07pp, all CIs crossing zero. (5) On exact ordinal distance within the same day, only d=1 clears zero: +3.55pp (CI [+0.14, +7.21]); d=2 +2.61, d=3 +1.72, d=12 +1.75 (CI [-1.59, +5.26]), so the d=12 bump loses its statistical support and should be deleted from the narrative; d>=25 shows -8 to -9pp with CIs excluding zero but only 6-9 estimable congresses and it is registered, not explained. (6) Estimator diagnostics: the same quantity reads -8.5pp unmatched, +1.0pp fine-cell matched, -0.210 logit unconditional, -0.264 logit additively conditioned; M3b and M1b differ in sign because additive fixed effects forbid the S/X difference from varying across cells while the data say it does, so this quantity cannot be reported as one constant. Instrument bug recorded: the first placebo used (gi*1000003+gj) mod 2, which is equivalent to ordinal-distance parity and produced a false +1.70pp; splitmix hashing gives -0.037pp. Pre-registration: six hypotheses, three pass, two fail, one not fully judgeable, and five of six matched the author's own prior predictions. Reproduction: p46_glmm.py with results JSON and runlog, seed 20261021, 17 seconds. The conclusion is a negative one: the phrase 'same-day effect' is removed from the narrative, and what survives is a single small fact, that the next vote cast on the same day agrees 3.6pp more often.