els-2002-conventions

ELS:2002 Conventions

The one thing to get right (opposite of HSLS)

This CSV stores numeric codes, not labeled strings. pd.to_numeric is safe everywhere. But NEGATIVE INTEGERS ARE MISSING-DATA SENTINELS and must be mapped to NaN BEFORE any imputation, encoding, or statistics:

• Integer-coded variables: every negative value (−1 don't know, −2 refused, −3 legitimate skip, −4 nonrespondent, −8 component missing, −9 missing) is a sentinel.
• Continuous composites (BYSES1, BYMATHSE) have VALID negative scale values (BYSES1 range −2.11..1.82). For these, ONLY exact codes ≤ −3 are sentinels. Blanket v < 0 → NaN silently deletes the lower half of the SES distribution (registry pitfall continuous_composite_sentinels, severity critical).

Waves and temporal ordering

BY (2002, grade 10) → F1 (2004, grade 12) → F2 (2006) → F3 (2012). Prefix-based wave inference: a predictor's prefix must come strictly before the outcome's prefix in that order. BY predictors for F1/F2/F3 outcomes are safe; F1 predictors for F2/F3 outcomes are safe with the usual later-wave caveats.

Tier-3 exclusions

Never model: STU_ID, SCH_ID, STRAT_ID, PSU, F1SCH_ID, F*UNIV* universe flags, any ^BYSTUWT/W[0-9]/BRR/*WT weight, and *IM imputation-flag twins (e.g. BYSEXIM next to BYSEX).

Encoding

One-hot categorical CODES with meaningful names: map codes to the registry's codebook_codes labels when constructing dummy names (BYRACE_White not BYRACE_7) so downstream SHAP/subgroup outputs are readable. Continuous variables pass through as single numeric columns per D1.

School clustering on ELS (F1SCH_ID — use directly, no reconstruction)

Base-year SCH_ID is 100% suppressed (−5), but F1SCH_ID carries 752 real school IDs covering ~76% of students. Use it as the cluster ID for the school-aware split, clustered bootstrap CIs, and ICC — NEVER as a predictor (tier-3). Do NOT run fingerprint reconstruction on ELS.

sch = pd.to_numeric(df["F1SCH_ID"], errors="coerce")
school_ids = sch.where(sch >= 0)
# students with no F1 school (~24%): unique singleton clusters
school_ids = school_ids.fillna(-(pd.Series(range(len(df)), index=df.index) + 1))

Report in the paper: clustering is PARTIAL (F1 school, one wave after baseline; ~24% singletons) — better than ignoring the design, weaker than true base-year clustering. This replaces the old blanket "no multilevel structure available" limitation.

HSLS analogues (for cross-cohort work)

ELS	HSLS	Construct
BYTXMSTD	X1TXMTSCOR	baseline std math score
BYSES1	X1SES	SES composite
BYMATHSE	X1MTHEFF	math self-efficacy
BYSEX / BYRACE	X1SEX / X1RACE	protected attributes
F2EVRATT	X4EVRATNDCLG	ever attended postsecondary

Source provenance

Authored in Arc G Phase A (2026-07-03) from the EDAT export codebook + sentinel-aware profiling of the delivered CSV.