Aligned OLS sensitivity configs and notebook design

2 min read

Aligned OLS sensitivity configs and notebook design

  • Day: 2026-06-15
  • Time: 11:52 to 12:20
  • Project: Dev
  • Workspace: WP 2: Operational
  • Status: Completed
  • Priority: HIGH
  • Assignee: Matías Nehuen Iglesias
  • Tags: Yaml, Ols, Fixed-Effects, Makefile, Sensitivity-Analysis, Config-Drift

Description

Session Goal

Refine the OLS experiment stack for specification-sensitivity analysis, with emphasis on keeping YAML configs, Makefile targets, and notebook design aligned to a single experimental contract.

Key Activities

  • Separated mechanical YAML validation from analytical design so structural consistency can be checked independently from the interpretation of temporal, geographic, and geo-time effects.
  • Audited the current OLS feature contract and fixed-effect variants, identifying the canonical baseline and the intended sensitivity variants.
  • Reframed the work for question 17 / notebook 17 as a specification-sensitivity workflow, not a search for the “best model”.
  • Proposed a cleaner experimental naming scheme centered on ols_core as the baseline, with controlled variants for dropping sanitacion_nivel, dropping Max_Nivel_Educativo, and adding household_pyramid.
  • Detected configuration drift in YAML and Makefile orchestration: some aglo_fe_no_* configs had inherited stale experiment.id values and swapped drop-columns, creating a risk of inconsistent results.
  • Recommended a production-style contract check so each Makefile target matches an existing YAML exactly in experiment ID, feature view, blocks, drops, and feature engineering settings.

Achievements

  • Clarified the analytical intent: the session is about sensitivity and ablation analysis rather than model selection.
  • Established a cleaner baseline convention for the experiment grid, reducing naming ambiguity and avoiding redundant no_pyramid terminology.
  • Identified concrete config mismatches that could invalidate comparisons if left uncorrected.
  • Defined the notebook structure around interpretable comparisons of geographic, temporal, and compositional sensitivity, including coefficient-shift analysis and comparison tables.

Pending Tasks

  • Correct the misaligned YAML files so each config’s experiment.id and dropped columns match the intended variant.
  • Update Makefile targets to mirror the cleaned YAML contract exactly.
  • Add or run an audit script to detect future naming drift between orchestration and config files.
  • Finalize notebook 17 outputs/tables for baseline vs. ablation vs. geo-rank sensitivity comparisons.

Evidence

  • source_file=2026-06-15.sessions.jsonl, line_number=2, event_count=0, session_id=e35b0064104eb2a83b6df53f93aae726a1799d80eb43128436f10d009a714674
  • event_ids: []

Graph View

Backlinks