MRM modules

Part of Statistical Methods — MORIE’s statistical-methods reference.

An MRM module pairs a (treatment, outcome, covariates) design choice for a specific OTIS dataset with the full MRM (Multilevel Reconciliation Methodology) ten-estimator framework, applied to that design. See the attribution block at the head of morie.otis_causal for the framework’s lineage.

Vocabulary

  • MRM — Multilevel Reconciliation Methodology: the ten-estimator framework applied to any design. The umbrella name for the full causal-estimator ensemble described below.

  • RF — formulation, one (T, Y, covariates) design choice for a dataset. Used as a code-level abbreviation (rf_* callables and *_ruhela_formulations analyzers) and aliased under MRM-prefixed names (mrm_*).

  • RDF — dual formulation: a formulation paired with a naive-arm sensitivity contrast. Code-level: rdf_* and the matching MRM-prefixed alias.

MRM (10 estimators)

For every formulation that resolves to per-row panel data with a binary T and numeric Y, MRM runs:

  1. IPW (Hájek) — single-robust on propensity, Lunceford-Davidian sandwich SE

  2. AIPW (RRZ doubly-robust) — cross-fitted IF plug-in

  3. g-computation (Robins 1986) — single-robust on outcome model + bootstrap SE

  4. PSM 1:1 NN (Austin 2011) — nearest-neighbour with caliper, returns ATT

  5. PSM subclass (Rosenbaum-Rubin 1983) — 5-strata weighted ATE

  6. IRM-DML (Chernozhukov 2018) — cross-fitted RF nuisance, reports ATE + ATTE + ATC, cluster-robust SE

  7. PSM → IRM-DML (match_first) — the author’s MatchIt-then-DoubleML pipeline

  8. ATC AIPW — E[Y(1)−Y(0) | D=0] via cross-fitted IF reweighted on the D=0 stratum

  9. PLR DML (Chernozhukov 2018) — partially linear regression, FWL-residualised, homogeneous-effect θ

  10. SuperLearner-stacked AIPW — convex stack of RF + ridge + GLM + mean (xgboost optional), NNLS weights

Plus multi-SE comparison on the IRM-DML primary estimate: pooled (iid), cluster on EndFiscalYear, cluster on UniqueIndividual_ID, multi-way (year × id) — same point estimate, four standard errors.

Plus propensity calibration (Platt or isotonic) on the propensity scores in IPW / AIPW / SuperLearner-AIPW. Reports Brier-score diagnostic.

Per-row formulations (panel data)

For a01 and b01 the canonical formulation is T_high_ac → Y_vm_count paired with a naive arm (any-flag → vm-binary).

# MRM-prefixed names (preferred going forward)
from morie.otis_all_analyze import (
    analyze_a01_mrm,
    analyze_b01_mrm,
    analyze_b02_mrm,
    analyze_a01_mrm_alt_gender,
    analyze_a01_mrm_alt_age,
    analyze_a01_mrm_alt_toronto,
    analyze_a01_mrm_per_year,
    analyze_b01_mrm_per_year,
)
# Original names remain as aliases:
#   analyze_a01_ruhela_formulations == analyze_a01_mrm   (etc.)

Aggregate formulations (count outcomes)

For aggregate datasets the analog is Poisson + NB GLM with IRR.

from morie.otis_all_analyze import (
    analyze_b03_mrm_aggregate,
    analyze_b06_mrm_aggregate,
    analyze_b07_mrm_aggregate,
    analyze_c01_mrm_aggregate,
    analyze_c03_mrm_aggregate,
    analyze_c04_mrm_aggregate,
    analyze_c06_mrm_aggregate,
    analyze_c07_mrm_aggregate,
    analyze_c09_mrm_aggregate,
)
# Original `*_ruhela_aggregate` names remain as aliases.

Doob chi-square companion

from morie.otis_all_analyze import analyze_c_doob_chi2, analyze_d_doob_chi2

Constraints

  • OTIS IDs are year-locked; no cross-year individual linkage by design. See methods/otis_linkage.md.

  • Aggregate IRRs are confounded with cell-size; per-row a01/b01 formulations are the canonical individual-level contrast.

  • Federal counterpart: SIU IAP (Doob, Sapers, Sprott et al.). See methods/siuiap.md.