Skip to main content

Question bank — Item Analysis (psychometric metrics)

Goal: Review psychometric quality metrics for questions that have real candidate response data, and prioritise which questions to keep, monitor, or retire.

Who can do this: Platform administrators and Directory administrators only. The Item Analysis tab is not visible to any other role — MCA users, company admins, staff with browse access, and Staff viewers can open the Question Bank but only see the Question List. They cannot see Item Analysis.

Before you start: Real candidate response data must exist for live assessments (minimum 250 response rows). If the tab is empty, select an assessment from the dropdown and click Run Analysis.

/question-bankItem Analysis tab


What this tab is for

Item Analysis shows Classical Test Theory (CTT) quality signals for questions that have enough real attempts—similar to what psychometrics teams review in spreadsheets:

IdeaOn screen
Difficultyp-value (proportion who answered correctly) with a 5-band label (excellent, good, fair, poor, negative)
DiscriminationD-index (how well the item separates stronger vs weaker performers)
Overall recommendationRecommendation chip per row (Bank It, Monitor, Review, Mis-keyed, Remove)

Use it to prioritise review before promoting bank items or after a high-stakes deployment—not as the only approval gate (operators still use Review Queue and status lifecycle).


A. Open Item Analysis

  1. Sidebar → Question Bank.
  2. Select the Item Analysis tab (fourth tab after Question List, Review Queue, Usage Dashboard).
    • The top stat cards (Total Questions, Live Unvalidated, Review Queue) are hidden while this tab is active — they don't apply here.
  3. Wait for the summary cards and table to load.

B. Read the summary cards

Four counters reflect all analysed items across every assessment — they do not change when you filter the table:

CardMeaning
Total AnalysedTotal question-level analytics rows in the system
Bank ItItems the engine recommends keeping as strong candidates
ReviewItems flagged for human review — combines review and review distractors verdicts
RemoveItems recommended for removal — combines remove and mis-keyed verdicts

These help you size a validation sprint before drilling into rows.


C. Select an assessment and run analysis

The Assessment dropdown lists every assessment that has at least 1,000 candidate response rows (from both Build and Match platforms), identified by their human-readable name where available.

  1. Select an assessment from the dropdown.
  2. The table filters to only questions from that assessment.
  3. An Assessment Reliability Card appears below the dropdown.
  4. Difficulty and Discrimination distribution tables appear, showing the band breakdown for that assessment.
  5. Use the Run Analysis button (right of the dropdown) to compute CTT metrics for that assessment.
    • If no results exist yet, the button shows Run Analysis (filled style).
    • Once results exist, it changes to Re-run Analysis (outlined style) — use this to refresh after new response data arrives.
    • The table, stat cards, and reliability card all update automatically when the run completes.
  6. To see all assessments together, select All assessments from the dropdown (hides the reliability card and distribution tables).

:::info Match platform assessments Assessments imported from the Match platform appear in the dropdown alongside Build assessments. They are labelled with their assessment name if available, or "Match Assessment #<id>" as a fallback. :::


D. Assessment Reliability Card

Visible when a single assessment is selected. It shows three sections:

Items count, test-taker count, max/min score, and banked item count, plus when the metrics were last computed.

Reliability table

MetricMeaning
Cronbach's AlphaInternal consistency reliability for all item types. Scale: excellent (≥ 0.90) · good (≥ 0.80) · acceptable (≥ 0.70) · questionable (≥ 0.60) · poor (≥ 0.50) · unacceptable (< 0.50).
KR-20Kuder-Richardson formula 20 — reliability for MCQ-only items, with the MCQ item count shown. Same 6-level scale.

If either value is ≥ 1.0, a ⚠ Degenerate badge appears instead of an interpretation — this signals a pathological item set (e.g. all items perfectly correlated or sample too small). Do not act on interpretation until response data is larger.

Central tendency

MetricMeaning
MeanAverage candidate total score
MedianMiddle score — less sensitive to outliers than mean
VarianceSpread of scores
Std DevStandard deviation — typical distance from the mean

Verdict distribution

Five chips show how questions in the assessment are distributed: Bank It, Monitor, Review (combines review + review distractors), Mis-keyed, Remove.


E. Difficulty and Discrimination distribution tables

Visible when a single assessment is selected. Each table shows how many items fall into each quality band.

Difficulty (p-value) bands

Bandp-value rangeAction
Excellent0.30 – 0.70Retain — optimal discrimination zone
Good0.20–0.30 or 0.70–0.85Retain — slightly off-optimal
Fair0.15–0.20 or 0.85–0.90Review — worth monitoring
Poor0.05–0.15 or 0.90–0.95Remove — too extreme
Negative< 0.05 or > 0.95Remove — essentially all wrong or all right

Discrimination (D-index) bands

BandD-index rangeAction
Excellent≥ 0.40Strong discriminator
Good0.30 – 0.40Good discriminator
Fair0.20 – 0.30Marginal — review
Poor< 0.20Weak or negative — remove

F. Distractor summary (MCQ items, visible page)

When MCQ items are in the current page, a summary shows how many items have fully functional distractors versus items with weak, non-functional, or likely mis-keyed options.

:::note Page scope The distractor summary reflects only the items visible on the current page (up to 25). For a full picture, use the detail modal to inspect individual items, or filter to a verdict chip and page through results. :::


G. Filter by recommendation

Below the reliability card, filter chips narrow the table:

  • All — no verdict filter
  • Bank It, Monitor, Review, Mis-keyed, Remove

Review covers both the review and review_distractors verdicts — items with distractor issues appear under Review, and the detail modal shows the specific distractor finding.

Selecting a chip reloads the list and resets pagination to page 1. The assessment filter and recommendation filter work together. The summary cards at the top always show global totals regardless of these filters.


H. Work the table

Columns:

ColumnWhat to look for
QuestionQuestion text (truncated to two lines). Falls back to a short bank ID if text isn't available.
p-valueDifficulty index + coloured band (green ≈ retain band, orange ≈ review/remove band)
D-indexDiscrimination index + label (e.g. excellent, good, fair, poor)
RecommendationAutomated recommendation chip

Click any row to open the Item Analysis Detail modal.


I. Detail modal (per question)

The modal shows:

  1. Header — full question text, assessment name, date computed.
  2. Difficulty (p-value) and Discrimination (D-index) — numeric value, interpretation band, retain/review/remove decision.
  3. Group sizes — total responses, upper 27%, lower 27% (standard CTT slicing).
  4. Distractor analysis (multiple-choice only) — for options A–D: % chosen in upper vs lower groups and a status (functional, weak, likely mis-keyed, non-functional).
  5. Recommendation — overall verdict label with detail text explaining the driver.
  6. Edit Question button — opens the edit modal for that question's bank item so you can correct content or metadata without leaving Item Analysis.

The assessment is shown as its human-readable name. If only an ID is available (e.g. for older Match imports before name storage was added), it falls back to a short truncated ID.

Close the modal to return to the table. Use these details when escalating to psychometrics or content authors.


J. Editing a question from Item Analysis

The detail modal includes an Edit Question button. Clicking it closes the modal and opens the standard Edit Question form for that item. Changes to content, status, or classification take effect immediately. Re-run analysis after edits to see updated metrics.


K. Pagination

When more than 25 rows match your filter, use the pager at the bottom. Recommendation summary cards still reflect global counts from the server.


L. Relationship to other Question Bank tabs

TabHow Item Analysis differs
Question ListCatalogue CRUD, status, filters on metadata
Review QueueHuman approve/reject workflow
Usage DashboardWhere unvalidated / flagged bank statuses cluster
Item AnalysisPerformance psychometrics after real attempts

An item can be live (validated) in the bank but still show Review here if field data says difficulty or discrimination is weak — treat both signals.


Common questions

SituationGuidance
Tab loads but table is emptySelect an assessment and click Run Analysis.
Assessment not in dropdownFewer than 1,000 response rows exist for it — wait for more completions, or check with engineering.
Reliability card shows degenerate (≥ 1.0)Pathological result — typically too few items or near-perfect correlation. Wait for more data before acting.
Reliability card shows 0 for all verdict chipsAnalysis was run before the new verdict columns were added to the schema. Re-run analysis to repopulate.
Very low response countsMetrics are unstable; wait for more completions before retiring items.
"Remove" / "Mis-keyed" verdictOpen detail modal; check distractor table for swapped keys or dead options.
Staff viewerCan open the Question Bank and browse the Question List, but cannot see the Item Analysis tab.

For catalogue operations (create, import, review queue), see Question Bank — operator runbook.