The Attendance-Replication Cannot Use Committee Data: Roll-Call Voting Is the Only Non-Anchored Behavioral Outcome Currently Available, and It Shows a -3.92pp Pre-Election Shift for the R17 Runner Cohort (Descriptive Only, N=9)

Rejected Paths

Before committing to the query below, I considered and rejected: - Run the committee-attendance replication directly from committee_meetings_{17-22}.parquet: rejected because those files contain bill-level committee-review events (BILL_ID, PPSL_DT, JRCMIT_CONF_NM, JRCMIT_CONF_RSLT), not member-level attendance. Audit of all six parquet files (below) returns zero columns matching MBR/MEM/ATTEND/출석/참석/의원. Scout's R21 plan assumed an attendance variable that does not exist in the processed corpus. - Reconstruct attendance from speech counts in the kr-hearings speeches parquet: rejected because the 1.1GB file is not downloaded locally; a speech-count proxy also collapses "attended but silent" into "absent," which is the opposite of the 대리출석 measurement error Scout flagged. Pursuing it would introduce a second validity problem while trying to solve the first. - Expand the 16-member cohort to include 19th-Assembly runners: rejected because roll_calls_all.parquet has only 1,247 rows for term 19 (vs 1.0M for term 20 and 970K for term 21). Coverage gap kills any identification for pre-2016 cohorts; staying on terms 20-21 where roll-call coverage is dense is the only defensible move.

1. The first finding is a data finding: committee-attendance replication is not feasible on the processed corpus

Scout's R21 post (061_literature_scout.md) asks for "pre-resignation committee attendance drop" as the non-anchored replication outcome and even suggests validating 대리출석 contamination against published committee minutes. Both tasks presume a member-level committee-attendance variable. Audit of the six committee_meetings_{17-22}.parquet files (code below) returns zero such columns:

import pandas as pd, os
KBL = os.environ['KBL_DATA']
for asm in [17,18,19,20,21,22]:
    df = pd.read_parquet(f"{KBL}/committee_meetings_{asm}.parquet")
    att = [c for c in df.columns if any(k in c.upper() for k in ['MBR','MEM','ATTEND','MEMBER'])]
    kor = [c for c in df.columns if any(k in c for k in ['출석','참석','의원'])]
    print(asm, list(df.columns)[:6], 'member_cols=',att,'korean=',kor)
# Every assembly: attendance_cols=[], korean=[]  -- bill-level event data only.

Schema is uniform across terms: [ERACO, BILL_ID, BILL_NO, BILL_NM, PPSR, PPSL_DT, JRCMIT_CONF_NM, JRCMIT_CONF_DT, JRCMIT_CONF_RSLT, _BILL_ID]. Each row is one bill × one committee-session event. The Scout-proposed attendance outcome cannot be constructed from this table. The 대리출석 validation target is therefore moot - we have no attendance field to cross-check against the minutes in the first place.

Data gap (flag to Critic): Paper B's attendance specification requires ingesting assembly.go.kr의 nhwxpdvciuxpykbdc attendance-roster API or scraping commissionMemberList pages per meeting. Neither is in master_bills_* / committee_meetings_*. This is a Phase-1 acquisition task that should be scoped before the PAP commits to an attendance section.

2. Roll-call voting participation is the only non-anchored behavioral outcome already in the corpus

roll_calls_all.parquet (2,425,113 rows, 16-22nd Assembly, dense for terms 20-22) records member-level floor votes coded as 찬성 / 반대 / 기권 / 불참 (yes / no / abstain / absent). Because 불참 is an explicit code rather than a missing row, the file supplies the exact denominator needed to compute attendance as (찬성+반대+기권) / (찬성+반대+기권+불참). This is floor-vote attendance, not committee attendance, but it matches Koo, Kim, and Choi's (2018) behavioral outcome and is genuinely non-anchored to bill dates (votes are called on the Speaker's schedule, not the MP's).

I built the attendance outcome for the 16-member local_executive_runner cohort coded in knowledge/hand_coding/round_21.jsonl, restricted to terms 20-21 where roll-call coverage is complete (9 of 16 runners survive the coverage filter). Each runner's election anchor is the corresponding local-election date (2018-06-13 for term 20; 2022-06-01 for term 21). Per 공직선거법 §53, local-executive candidates must resign ~30 days before filing, so the -3 to 0 months-to-election window is the pre-resignation / pre-exit analog. Baseline window: -12 to -4 months.

Per-member attendance (descriptive only, N=9)

Pooled aggregates (N=9 runners, 20-21st Assembly)

Same-assembly non-runner control (N=313 in 20th, N=316 in 21st)

Naive difference-in-differences: runner shift (-3.92) minus pooled control shift (≈ -1.0) = -2.9 pp. This is directionally consistent with Paper B's sponsorship-shirking sign, but the signal is tiny relative to sponsorship (where R19 reported -1.5 bills/month). The per-member table shows the pooled figure masks enormous heterogeneity: some runners (이철우, 박완수) drop to 0% attendance in the exit window while others (양승조, 박남춘) stay at baseline.

C6 compliance reminder: N=9 at the member level, so every number here is reported as DESCRIPTIVE ONLY. No p-values, no inferential language, no equivalence ranges. If the PAP wants to commit to Titiunik-Feher (2017) equivalence testing on the attendance outcome (per Scout's Recommendation 4), the N=9 floor means the smallest equivalence range the cohort can support is roughly ±15pp - too wide to be useful. This is itself an argument for demoting attendance to a secondary outcome or acquiring committee-level attendance data before the PAP is signed.

3. Response to Scout (061): three concrete consequences for the PAP

On the Høyland-Hobolt-Hix (2017) anchor. Their framework predicts that progressive-ambition legislators participate more as elections approach under closed-list PR (because the party controls renomination), and less under district SMD. Korea's local-executive cohort is almost entirely district SMD under 공직선거법: 12 of 16 runners resigned from district seats, only 4 from 비례대표 (proportional) seats. The framework therefore predicts shirking, matching the sign I find. But the mechanism prediction is sharper than the overall sign - if attendance is the outcome, the PAP should pre-specify the district vs proportional heterogeneity test. Hand-coding dict entries already tag election_type; this is cheap.

On Scout Recommendation 2 (대리출석 validation). The validation target does not exist in the processed data. I recommend the PAP replace this with a different measurement-validity check: compare roll-call attendance (this paper) against member-level speech counts in kr-hearings-data for the same member-months. Discrepancy between the two proxies is itself a measurement-validity statement Paper B can report honestly.

On Scout Recommendation 3 (committee-role 4b). The hand-coded dictionary does not contain committee-chair / 간사 status. Adding this is a 2-hour hand-coding task against National Assembly press-release archives for each of the 16 runners. Pre-committing to Commitment 4b in the PAP is fine; the data enablement is cheap. But with N=9 in the modern subsample, a chair-vs-rank-file split would produce cells of roughly N=2 vs N=7 - firmly below the N=10 guardrail, so 4b must be pre-registered as descriptive-only.

4. What Critic should evaluate for theoretical framing

1. Whether the "attendance" section should survive the PAP at all, given that (a) the processed corpus cannot supply committee attendance, (b) the roll-call proxy pools to a -3.92 pp shift that is smaller than the control-group seasonal decline in the 20th Assembly, and (c) the N=9 modern subsample precludes inferential claims. One defensible posture: drop attendance, keep sponsorship, and reframe Paper B as "sponsorship-specific shirking with attendance null scope-condition."
2. Whether the Høyland-Hobolt-Hix (2017) anchor should be replaced by Koo, Kim, and Choi (2018) doi:10.1017/s1468109918000403 as the Korean-cohort precedent on voting participation. KKC uses roll-call voting participation and position-change in a Korean lame-duck setting; their outcome variable is closer to what the available data supports than Høyland et al.'s EP participation data.
3. Whether the heterogeneity in per-member delta (0 pp to -58 pp) is itself the paper. The pooled headline number obscures that some runners (particularly 이철우, 박완수, 오영훈) collapse to near-zero attendance immediately post-resignation, while others (양승조, 박남춘) maintain baseline. A paper about who shirks under progressive ambition is more interesting than a paper about whether shirking happens on average, and the hand-coded cohort could support qualitative case-coding even with N=9.

5. Data gaps this round surfaced

• No member-level committee attendance in the processed KNA corpus. This blocks Scout's original R21 specification and requires an upstream data-acquisition sprint before any attendance-based PAP can be signed.
• Roll-call coverage is thin for terms 17-19 (≤25K rows each vs 1M+ for 20-21). The attendance replication is effectively restricted to the 2018 and 2022 cycles, collapsing the 16-member cohort to 9 analyzable members.
• No resignation-date field in members_{17-22}.parquet. The election date is a proxy anchor, but the true exit event is 후보등록일 - 30일. Building the true-exit event would need candidate-registration dates from the NEC registry, which Scout's topic-gate commitment 2 already flagged.
• 공직선거법 §53 exit window would benefit from a pre-registered distinction between "statutory-mandated" and "voluntary" exit timing; absent a per-member registration date, all 9 runners are pooled into the same -3 to 0 window.

Reproducibility

All analyses use /Users/kyusik/Desktop/kyusik-github/kna/data/processed/roll_calls_all.parquet with the parser:

def parse_date(s):
    if pd.isna(s): return pd.NaT
    s = str(s).strip()
    if '-' in s[:10]: return pd.to_datetime(s[:10], errors='coerce')
    if len(s)>=8 and s[:8].isdigit(): return pd.to_datetime(s[:8], format='%Y%m%d', errors='coerce')
    return pd.NaT

The mixed-format date column (some YYYY-MM-DD, some YYYYMMDD HHMMSS) silently collapses 98% of roll-call rows under a naive pd.to_datetime call — a reproducibility trap worth recording for future arc-2 roll-call work.

References

Høyland, Bjørn, Sara B. Hobolt, and Simon Hix. 2017. "Career Ambitions and Legislative Participation: The Moderating Effect of Electoral Institutions." British Journal of Political Science 49 (2): 491-512. doi:10.1017/s0007123416000697

Koo, Bon Sang, Junseok Kim, and Jun Young Choi. 2018. "Testing Legislative Shirking in a New Setting: The Case of Lame Duck Sessions in the Korean National Assembly." Japanese Journal of Political Science 19 (4): 608-624. doi:10.1017/s1468109918000403