GitHub

Small-n checklist

Continuous MTP and mediation with tens to low hundreds of units (conservation, ecology, early trials) need different defaults from large-cohort epi software. Theory still comes from LMTP and targeted learning (Díaz et al., 2023; van der Laan & Rose, 2011); the checklist below is about finite-sample practice. See Methods and References.

Before you estimate

  1. Call recommend_run_options(n; engine, n_mediators) and pass the kwargs into run_lmtp_grid / run_mediation_grid / execute_estimand.
  2. Prefer SMALL_N_SL_LEARNERS or adaptive_learners(n) over RICH_SL_LEARNERS when n < 80 (Super Learner library must be estimable at the given n; van der Laan, Polley & Hubbard, 2007).
  3. Keep parallel=false unless you have measured headroom (δ-grid threading can thrash memory on multi-core laptops).
  4. Run positivity_report (or positivity=true on the grid) and inspect weak_support / unsupported_shift cells before interpreting TE curves (Petersen et al., 2012; Díaz et al., 2023 on MTP-designed positivity).
  5. For mediation, run mediation_n_mc_sweep and mediation_stability_summary; raise n_mc until TE SE and signs stabilise (nested MC is part of the estimand’s approximation error under continuous MTP mediation; cf. Liu et al., 2024).
  6. Attach sensitivity_report (tipping-point / partial-R²) for any claim that is policy-relevant under possible omitted confounding (Cinelli & Hazlett, 2020).
  7. Treat discovery graphs as sensitivity, not oracles: discovery_adjustment_sensitivity / merge_discovery_sensitivity! (Pearl, 2009; Spirtes et al., 2000).

Policy library

Policies implement scientifically interpretable MTPs (Díaz & van der Laan, 2012; Díaz et al., 2023):

  • Additive z-shift: additive_shift_policy (default)
  • Multiplicative: multiplicative_shift_policy
  • Threshold (“raise if below c”): threshold_shift_policy

Longitudinal

Use SequentialPolicy / run_sequential_lmtp for multi-time exposures, with ID from CausalDynamics TemporalEffectQuery when a lag DAG is available (Díaz et al., 2023). Point-treatment LongitudinalPolicy carries temporal ID metadata for single-time pathway→AUC analyses.

Sysimage (optional)

See scripts/build_mtp_sysimage.jl in the CDCS monorepo for a PackageCompiler recipe (community convenience; not required for CI).