Explain

W07 · 5 h 30 min

Metrics, uncertainty, and repeated observation

Essential question

When is a visibility change repeatable?

What you should be able to do

  1. Specify numerator, denominator, unit, and aggregation for a metric.
  2. Separate within-time variation from drift.
  3. Report intervals, missingness, and sensitivity to query mix.
PrerequisitesW06 query bank.Bridge note on proportions, bootstrap intervals, and clustered observations.
Builds on

W06 query bank. · Bridge note on proportions, bootstrap intervals, and clustered observations.

Investigates

When is a visibility change repeatable?

Feeds

L06 protocol dry run.

Arrive with a prepared artifact

Reading route
Core PAPER-29/PAPER-33; Extend PAPER-38; candidate synthesis R07 and interface case PLAT-04.
Viewing route
Open the complete text-first lecture packageSeven-minute metric-card construction and denominator audit. The current equivalent is notes, slide script, worked case, and no-video transcript; no recording is claimed.
Readiness check
Rewrite a vague “visibility increased” statement as a complete metric specification.
Bring
One metric card with formula, unit, denominator, time window, and uncertainty.

Concepts, assumptions, and boundary

Metric names are not specifications

Mention rate, citation rate, share, rank, and visibility score can hide incompatible denominators. A metric card must state the event counted, eligible unit, repeated-run rule, query weighting, missing-response treatment, aggregation, and uncertainty method. Mention, citation, entailment, absorption, and referral belong to different stages and should remain separate before any justified composite is considered.

Variance and drift occupy two time scales

Repeated runs close together reveal stochastic or session-level variation; observations across time can reveal platform, corpus, or query-population drift. Pooling them without a time model can make intervals too narrow or changes look stable when they are not. Resampling must respect clustering by query and time block. Missing responses are not automatically zero because failure, refusal, outage, and unavailable citation UI may represent different mechanisms.

Inspect the mechanism or evidence structure

Conceptual modelMetric ladder with two time scales
Metric ladder with two time scalesFive separate event metrics occupy distinct rows. Three equal circles represent repeated observations without assigning values. Two labeled brackets distinguish nearby repetitions from comparisons across time blocks. The figure is a denominator and sampling map, not observed data.Mentionevent / eligible unitCitationevent / eligible unitEntailmentevent / eligible unitAbsorptionevent / eligible unitReferralevent / eligible unitwithin-block repeatsbetween-block driftno values encoded
Figure design. Five event/eligible-unit rows separate mention, citation, entailment, absorption, and referral. Brackets distinguish within-block repetitions from between-block drift; no values or confidence intervals are encoded. This is a denominator map, not an estimate. Within-block repetition and between-block drift require different intervals.
Long description

Five separate event metrics occupy distinct rows. Three equal circles represent repeated observations without assigning values. Two labeled brackets distinguish nearby repetitions from comparisons across time blocks. The figure is a denominator and sampling map, not observed data.

One action, one feedback state

Action

Vary repetitions, query weights, time blocks, and missing-data rules on a frozen panel.

Feedback

Intervals and estimates update with a warning whenever independence or denominator assumptions change.

Accessible alternative

A static sensitivity table reports every predefined configuration.

Open the repeated-measurement explorer

Produce a reviewable intermediate file

Task
Analyze a frozen repeated-observation panel and compare two time scales.
Inputs
Panel CSV, data dictionary, metric formulas, analysis script.
Timebox
75 minutes
Intermediate file
metric-table.csv and sensitivity-plot.svg
Stop condition
Do not pool engines, locales, time blocks, or missing states without an explicit estimand and sensitivity check.

Why each controlled source is here

Complete, retrieve, and revise

Checkpoint
L06 protocol dry run.
Reflection
Explain how one missing-data rule changes the conclusion.
Revision
Add cluster structure and query weights to the analysis specification.
Low-compute route
Run the supplied analysis on a small CSV; all charts have precomputed table equivalents.

Five retrieval questions

01Why is a numerator alone insufficient?

Without the eligible denominator, observation unit, weighting, and window, the rate and its uncertainty cannot be interpreted.

02How do repeated-sampling variance and drift differ?

The first is variation among nearby repetitions under nominally similar conditions; the second is change across time blocks or system states.

03Why may a missing response not equal zero visibility?

It can arise from refusal, outage, interface failure, censoring, or collection error, each requiring a declared rule.

04What is the evidence boundary for W07?

Reported metrics inherit each study’s sampling and platform boundary; matching a label does not make estimates comparable.

05What must you submit or revise after this week?

L06 protocol dry run. Add cluster structure and query weights to the analysis specification.

Continue in the practice package