P36 k=1: The Pre-Registered Prediction Fails, and theta Is Not an Identifiable Parameter at k=1 — A Composition Audit of Weighting in Direction-Matched Roll-Call Pairs (P36 k=1 的方向内空白与第二种「方向」定义:预测被打脸,且 θ 在 k=1 不是一个可识别的参数)
Year: 2026
Type: conference-abstract
Abstract: Zhang San, validation layer, pre-registered execution plus exploratory estimator audit (independent author; public data voteview HSall rollcalls/votes/members, CC0; congress>=90; 55,958 roll calls, 15,563,447 same-member pairs at k=1). This paper executes item 1 of P35's next-step list: measure the direction-matched effect at k=1, and add a second definition of direction (vote-semantics direction: passage / amendment / procedural) in place of the party-position definition used in P34/P35. Pre-registration was written at 01:46, before the script (01:51) and every result file (01:53), and is deposited with this record as a separate file. Findings. (1) Calibration passes bit-level on both grids: QCTL k=1 theta_full = +0.06353594799671548 and CCTL k=1 = +0.1278431713372692, both identical to P32 with difference 0.00e+00. (2) The pre-registered prediction H2 fails, in the opposite direction: H2 said the direction-matched theta would land in +0.03 to +0.10; the measured value is +0.172986, above the upper bound and above P35's k=2 value. (3) The second direction definition B is degenerate as a test: the passage/passage layer has zero estimable cells, so B collapses onto the amendment/amendment layer and reproduces it to the last digit (+0.245602506211); it filters vote types rather than varying direction within a vote type. (4) Exploratory audit (not pre-registered): at k=1 the full-sample theta is a weighted mixture, not a parameter. The same cells give +0.0635 under the standard weight nS*nX/(nS+nX), +0.4696 under an S weight, -0.1129 under an X weight and +0.1031 under equal weights, a span of 0.58 that crosses zero. (5) The sign can be traced to cells: two procedural vote types hold 0.36 percent of the same-bill pairs but 37.4 percent of the weight and contribute -0.1816, while all other cells contribute +0.1203; without those two cells the same-day full-sample theta is +0.192. Coverage is reported on S pairs while weights are driven by X pairs, so 1-3 day cells hold 2.5 percent of S pairs yet 36.1 percent of the weight. (6) Direction-stratified signs are stable across all four weightings (same-direction +0.143 to +0.246, opposite-direction -0.171 to -0.309); what is unstable is the undirected full sample. Honest boundaries: the passage/passage layer is not estimable at all; B is a degenerate design; the audit is exploratory and no weight is declared correct, so neither +0.4696 nor -0.1129 is quoted as an estimate; the direction label has weak endogeneity (the member's own vote enters the majority-party yea rate); 97.5 percent of S pairs are same-day, so k=1 conclusions concern consecutive same-day votes only. The paper therefore does not upgrade the literal pass to a claim that within-direction stickiness holds, and sets the next step as a method audit of every headline theta in P30-P35 rather than another data domain.
Author: San, Zhang
Language: English
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7222970807