Method: from studies to O*NET-wide estimates
The imputation model spreads observed effects from the occupations and activities that were studied to those that weren’t, following how strongly each occupation relies on each work activity. This page shows the inputs to that model at the canonical setting (β = 2.5) and explains every parameter.
This is a read-only view. To launch runs with other settings, run the review app locally
(
python scripts/review_app.py) and open /run.
Metric
A run targets one metric at a time. Switching changes which activities and
SOC groups are flagged as observed in the charts below.
Stage B: how occupations load on work activities
Mean Stage-B weight per SOC major group × O*NET work activity. Weights come from applying a
row-wise softmax with temperature β to each occupation's (Importance × Level / 5) profile, then
averaging within each SOC-major group. Red markers and bold labels flag activities and groups with an observed
effect for this metric; shaded rows are excluded. Click a row label or checkbox to toggle its inclusion.
Activities whose summed weight (over kept occupations) is below this are dropped.
Observed activities are always kept.
What each parameter does
metric
Which effect-size column feeds the observations. speed = mean log-ratio;
quality = mean Hedges' g. Two independent runs.
β (specificity)
Softmax temperature for Stage B. Higher β means each occupation concentrates on
fewer activities. Default 2.5 yields roughly 7 effective activities/occupation.
aggregation level
occupation (default, ~894 O*NET occupations), soc_minor (~92 3-digit groups
like
13-2000 Financial Specialists), or soc_major (22 2-digit majors). In aggregated
modes, the composite weight matrix is the per-group row-mean; observations are IV-weighted (or simple-mean
fallback) within the group; the AIOE baseline is averaged within the group.excluded_soc_majors
SOC 2-digit prefixes to drop entirely. Default excludes the physical majors
(37, 45, 47, 49, 51, 53) plus 11 & anything the user has toggled off.
activity_weight_threshold
See the bar chart above — activities whose summed weight across the kept
occupations falls below this are pruned. Observed activities are always retained regardless.
Ωref
Stage D observation-trust precision. Higher pulls estimates harder toward the observed
values. Default 100.
use_baseline / Ωbase
speed only. When on, mixes in the Felten et al. AIOE score
(moment-matched to the observed metric's mean and SD) as a per-occupation soft anchor. Ωbase
controls the strength: 0.1 ≈ tiebreaker, 0.5 ≈ balanced, 1.0 ≈ AIOE-led,
5.0 ≈ AIOE-replaces-graph.
σref
Stage D SE floor / scale (default 0.1). Effectively converts a paper's SE into
an inverse-variance weight relative to Ωref.
ε (regularizer)
Ridge added to the diagonal of the linear system to keep it invertible (default
10-6).