HEW-NODE-0130 · How Education Works · test equating, score linking, scaling, alternate forms, anchor items, common-item design, vertical scales, mode comparability, population invariance, psychometrics, transformed scores, uncertainty and assessment comparability
Two students can answer different examination papers, receive the same reported score and still require a serious technical question before anyone says those scores mean the same thing.
Different test forms contain different items. Difficulty drifts. Administration changes from paper to screen. Cohorts differ. Some programmes build new forms every year. Others connect grades across age levels. A score can look stable while the measurement underneath it has changed.
Score comparability is not created by printing the same number on two reports. It is created by evidence that justifies treating those numbers as meaningfully related.
This node sits beside the How Education Works hub, Educational Measurement, National Learning Assessment Systems, Assessment Item Banks & Test Form Assembly, Examination Marking, Standard Setting & Results Processing, School-Based Assessment Moderation & Standardisation and Student Results Reporting & Report Cards.
Those pages keep their jobs. Educational Measurement owns validity, reliability, scales and interpretation broadly. Item Banks owns item pools and test-form assembly. Examination Marking owns scripts, marks, standard setting and result processing. National Assessment owns country-level assessment systems and trend interpretation. This node owns the comparability bridge: how scores from different forms, modes, years or related assessments are placed on scales that support carefully bounded comparisons.
The 60-Second Read
- Raw scores are tied to the specific set of items administered.
- Scaling converts raw performance into a reporting scale; it does not automatically make different tests comparable.
- Linking is the broad family of methods used to connect score scales.
- Equating is a stronger form of linking used when forms are intended to be interchangeable measures of the same construct under stringent conditions.
- Two tests can be linked without being equated.
- Common-item or anchor designs provide a bridge across forms or cohorts.
- Equivalent-group designs use comparable groups rather than common items.
- Item-response models can support linking but do not make weak designs strong by themselves.
- Anchor items must represent the relevant construct and behave stably enough to support the link.
- Population invariance matters: a conversion that works differently for subgroups can undermine fairness.
- Mode changes from paper to computer may require separate evidence.
- Vertical scaling links achievement across levels or ages, but strong interpretation requires caution because the construct itself can change with development.
- Standard setting and equating are different. Equating addresses form difficulty; standard setting interprets performance against grade or achievement standards.
- Score transformations can create precision-looking numbers without increasing measurement precision.
- Linking error and ordinary measurement error both matter.
- Extreme scores are often less stable under transformations than central scores.
- Comparability should be stated as a degree and purpose, not as a magical yes/no property.
- Changing curriculum or test specifications can break the historical link.
- Technical documentation should state design, assumptions, diagnostics and limitations.
- The goal is continuity of meaning, not continuity of numerals.
One-Sentence Definition
Assessment score linking is the family of statistical and design procedures used to place scores from different tests or forms into a defined relationship, while equating is the strongest form of linking intended to support interchangeable score meaning across sufficiently similar forms.
The First Distinction: Raw Score Is Not Scale Score
A raw score of 42 means forty-two units of awarded credit on that particular paper. It does not by itself tell us whether another paper with a raw score of 42 was equally difficult, measured exactly the same content mix or produced the same degree of precision.
A reporting scale can transform raw scores into a common numerical frame. The scale is useful only if the transformation is supported by a defensible design.
The Second Distinction: Scaling Is Not Equating
You can convert every raw score to a 100–800 scale using a formula. That is scaling. Nothing about the conversion proves that a 650 from Test A is equivalent to a 650 from Test B.
Equating requires evidence that the forms are similar enough in construct and intended use, and a data-collection design that can estimate unintended difficulty differences.
The Third Distinction: Linking Is a Hierarchy
The National Council on Measurement in Education defines score comparability as varying along a continuum depending on the form of linking, and points to distinctions among equating, calibration, moderation, projection, vertical scaling and other relationships.
That hierarchy matters. The stronger the claim — especially interchangeability — the stronger the design and assumptions required.
Current Measurement Reference: Equating Is the Strongest Linking Claim
The 2026 fifth edition of Educational Measurement distinguishes equating from other forms of linking and describes equating as the strongest relationship in the linking hierarchy: alternate forms of the same test are adjusted for unintended difficulty differences so the resulting reported scores can be treated as interchangeable for intended uses.
That word — interchangeable — is demanding. It should not be borrowed casually for assessments that merely correlate.
Why New Test Forms Need Equating
High-stakes testing programmes often need new forms because secure items cannot be reused indefinitely, content specifications change slightly, or multiple forms are administered in the same window. Even carefully assembled parallel forms will not have exactly equal difficulty.
If one form happens to be harder, candidates should not receive lower reported scores merely because of form assignment. Equating attempts to remove this unintended difficulty difference.
Common-Item Designs Build a Bridge
One common design places a set of anchor items on both forms. Differences on the anchors help estimate how much of the overall score difference reflects group ability and how much reflects form difficulty.
The anchor is not just any repeated subset. It should cover enough of the intended construct and behave stably across administrations to carry the linking job.
Weak Anchors Create Strong-Looking Errors
If anchor items overrepresent one topic, are exposed before the second administration, or function differently across groups, the linking function can inherit those distortions while still producing smooth conversion tables.
Technical elegance cannot repair an anchor that does not connect the intended construct appropriately.
Equivalent-Group Designs Use People as the Bridge
If genuinely comparable groups take different forms, score differences can be attributed to form difficulty under the design assumptions. Random assignment can strengthen the case when feasible.
The practical problem is that large operational programmes rarely receive perfectly equivalent groups by chance unless the design creates them.
Non-Equivalent Groups Need More Structure
When cohorts differ — perhaps one year’s students are stronger than another — common-item designs can help separate cohort difference from form difference. This is one reason year-to-year trend programmes need more than comparison of raw means.
Item Response Theory Can Place Items and Learners on a Common Metric
Item-response models estimate relationships between latent proficiency and item response probability. With suitable common items or common persons, item parameters and proficiency estimates can be placed onto a common scale.
The model does not eliminate design requirements. If the construct changes, anchors drift, or model fit is poor, the common metric can be technically produced while interpretive comparability weakens.
Classical Equating Methods Still Matter
Mean equating, linear equating and equipercentile equating represent different relationships between score distributions. No method is universally best. Choice depends on test design, sample size, score properties, intended interpretation and operational constraints.
ETS’s Principles and Practices of Test Score Equating emphasises that equating is a practical quality-control system as well as a statistical calculation because errors in the conversion can affect every candidate whose score is reported.
Equipercentile Equating Matches Relative Positions
At a high level, equipercentile methods identify scores on two forms that correspond to the same percentile rank in a target population. The relationship can be nonlinear.
Because observed score distributions contain sampling noise, smoothing choices matter. A jagged conversion table can reflect random irregularity rather than meaningful measurement structure.
Linear Equating Matches Key Distribution Features
Linear approaches align means and standard deviations under their assumptions. They are simpler and can be stable, but they may miss nonlinear differences between forms.
The correct question is not which method sounds more sophisticated. It is which method is defensible for the programme and whether conclusions are robust to reasonable alternatives.
Population Invariance Is a Fairness Question
If the linking relationship differs substantially for language groups, regions, genders or accommodation conditions, one conversion may not preserve score meaning equally across populations.
NCME instructional work on population invariance has long treated this as a central linking concern because a conversion that depends strongly on subgroup composition can threaten validity and fairness.
Subgroup Difference Does Not Automatically Mean Bias
A difference in linking functions can arise because the assessments measure somewhat different content that subgroups have genuinely learned differently. The response is diagnostic: inspect construct representation, item functioning, sample size, administration and intended interpretation before assigning a cause.
Mode Changes Need Separate Evidence
Moving from paper to computer can change scrolling, navigation, typing demands, timing, access to tools, reading behaviour and item formats. If the construct is intended to remain the same, the programme should investigate whether mode changes scores in ways that require adjustment or qualification.
ETS has published technical work on linking scores across computer- and paper-based administration, illustrating that mode comparability is a linking problem rather than an interface footnote.
Adaptive Tests Complicate the Idea of a Common Form
Computer-adaptive testing means candidates may receive different item sequences by design. Comparability then depends on calibrated item pools, model fit, test information, content constraints and scoring algorithms rather than identical forms.
The common scale is built into the measurement system rather than added only after administration.
Vertical Scaling Connects Levels Across Development
A vertical scale attempts to place achievement from different grades or age levels onto one continuum so growth can be interpreted. This is attractive because a single scale appears to show progress directly.
The interpretation is harder than it looks. As learners develop, the content domain and cognitive demands can change. A Grade 3 reading assessment and a Grade 9 reading assessment may share a broad construct while requiring very different tasks.
A Common Number Does Not Guarantee a Common Construct
If the test specification changes materially — new curriculum, new item types, different calculator policy, different time limit — historical comparability may weaken. Sometimes the honest choice is to break the series rather than force a technical link that overstates continuity.
Standard Setting and Equating Solve Different Problems
Equating asks how scores from different forms should correspond after adjusting for unintended difficulty difference. Standard setting asks what performance level should define a grade, proficiency category or cut score.
They can interact operationally, but a grade boundary is not an equating function and an equated score does not decide what counts as proficient.
Moderation Is Yet Another Relationship
School-based assessment moderation may adjust or review locally generated marks to support consistency across centres. It does not necessarily make the underlying assessments interchangeable. The separate School-Based Assessment Moderation & Standardisation node owns that process.
Scale Transformations Can Create False Precision
A reporting scale might produce 612 instead of raw score 37. The three digits can look more precise even though measurement uncertainty remains. Transformation changes the location and units of the score, not the amount of information in the candidate’s responses.
Report precision should therefore match evidence, not aesthetic preference.
Linking Error Should Be Monitored
An equating function is estimated from data. Different samples would produce slightly different functions. Methods therefore have standard errors or other measures of uncertainty.
Programmes should know whether observed year-to-year shifts are large relative to linking uncertainty rather than treating the conversion as exact.
Extremes Are Often More Fragile
Very high and very low scores may have fewer observations and less anchor information, making transformations less stable. Smoothing and extrapolation choices can matter more near the ends of a scale.
Security Can Affect Linking Design
Common items are useful statistically but can create exposure risk if repeatedly reused. Operational programmes must balance psychometric linking value against security. Embedded anchors, rotating anchor sets and item-bank calibration are possible responses depending on the programme.
Anchor Drift Needs Surveillance
An item can change in effective difficulty over time because curriculum emphasis shifts, language changes, technology makes knowledge common, or the item becomes familiar through exposure. Stable wording does not guarantee stable measurement behaviour.
Anchor diagnostics should therefore test functioning over time rather than assume old items remain neutral forever.
Trend Reporting Requires a Version Ledger
- test specification version;
- item-bank version;
- administration mode;
- timing rules;
- accommodation policy;
- scoring model;
- anchor set;
- equating method;
- population definition;
- reporting scale;
- standard-setting process;
- known breaks in comparability.
Without this ledger, a long trend line can conceal several changes in what the test meant.
Comparability Should Be Purpose-Specific
A link may be adequate for group-level trend monitoring but not for treating individual scores as interchangeable in admissions. Another link may support broad concordance but not fine-grained classification near a cut score.
Ask what decision the comparison will support before deciding what evidence is enough.
Concordance Tables Are Not Automatically Equating Tables
Two distinct admission tests can be statistically linked so users can estimate corresponding score ranges. If the tests differ in construct, reliability or intended use, the resulting concordance should not be described as making the tests interchangeable.
Correlation can be high while individual-level differences remain meaningful.
Classification Consistency Can Matter More Than Score Difference
If the practical decision is whether a learner is above a proficiency threshold, investigate whether linking changes classifications near the boundary. A one-point conversion difference can matter greatly if it crosses a cut score and hardly at all elsewhere.
Equating Should Be Reproducible
Another qualified analyst should be able to reconstruct the conversion from documented data definitions, exclusions, anchor rules, methods and software versions. High-stakes transformations should not depend on undocumented manual steps.
Independent Checks Protect Against Systematic Error
A wrong key affects one item. A wrong equating table can affect an entire cohort. Independent replication or automated reconciliation of score conversions is therefore proportionate to the scale of consequence.
Case Study: The Harder Form
Invented example: two mathematics forms share the same blueprint and administration conditions. Form B turns out slightly harder. Candidates taking B score three raw marks lower on average even after group differences are accounted for.
Equating adjusts the conversion so reported scores preserve intended form interchangeability. The system does not lower standards; it corrects an unintended form-difficulty difference.
Case Study: The Paper-to-Screen Shift
Invented example: a reading assessment moves online. Overall means remain similar, but long passage items show stronger mode effects for students using smaller screens. A simple total-score link hides the pattern.
The programme investigates item-level mode functioning, screen design and subgroup effects before claiming full comparability.
Case Study: The Curriculum Break
Invented example: a science curriculum shifts from factual recall toward data interpretation and modelling. Assessment content follows. Technical staff can still fit a statistical link to the old scale.
The policy team decides the trend should be marked with a series break because the construct emphasis changed materially. The technically possible link is not the educationally honest interpretation.
Case Study: The Anchor That Changed Meaning
Invented example: a digital-literacy anchor asks learners to identify a browser security icon. Five years later the interface has become familiar and the item is much easier even at similar underlying proficiency.
The anchor no longer behaves as a stable bridge. It is retired from linking rather than preserved because it is historically convenient.
Failure Modes and Repairs
- Same scale, assumed same meaning: repair by documenting the empirical linking design.
- Scaling called equating: repair by using the correct relationship and claim strength.
- Weak anchors: repair by testing representativeness, security and stability.
- Model worship: repair by checking construct and design before fitting sophisticated models.
- Subgroup blindness: repair with population-invariance and differential-function diagnostics.
- Mode blindness: repair by testing paper, computer and device effects where relevant.
- Curriculum break hidden by mathematics: repair by allowing honest series breaks.
- False precision: repair by matching reported precision to measurement evidence.
- No linking uncertainty: repair by estimating and monitoring transformation error.
- Undocumented conversion: repair through reproducible code, versioning and independent checks.
The Score-Linking Operating Chain
- Define the comparison purpose.
- Define the construct for each assessment.
- Confirm whether forms are intended to be interchangeable.
- Choose the appropriate linking strength.
- Design common-item, common-person or equivalent-group evidence.
- Select and review anchors where used.
- Collect representative data.
- Check administration comparability.
- Check mode and accommodation effects.
- Fit candidate linking or equating methods.
- Inspect model fit and score distributions.
- Estimate linking uncertainty.
- Test population invariance.
- Inspect subgroup and item-level anomalies.
- Compare alternative methods.
- Evaluate score conversion at boundaries and extremes.
- Check implications for classifications.
- Document transformation and scale.
- Independently reproduce the conversion.
- Apply the approved link.
- Validate reported-score generation.
- Monitor anchor drift over time.
- Record specification and model versions.
- Re-evaluate after curriculum or mode change.
- Break the trend series when comparability is no longer defensible.
A Score Comparability Dashboard
- forms linked;
- linking purpose;
- construct similarity;
- anchor coverage;
- anchor exposure risk;
- sample sizes;
- method used;
- alternative-method comparison;
- linking standard error;
- population-invariance diagnostics;
- subgroup anomalies;
- mode-effect evidence;
- classification changes near cut scores;
- extreme-score stability;
- conversion-table validation;
- specification version;
- administration version;
- known comparability breaks.
Canonical Owner Boundaries
- Educational Measurement owns validity, reliability, measurement error and score interpretation broadly.
- Assessment Item Banks & Test Form Assembly owns item pools, blueprints and construction of forms.
- Examination Marking, Standard Setting & Results Processing owns marking, grade boundaries and result processing.
- National Learning Assessment Systems owns national sampling, administration and trend-use architecture.
This node owns the comparability bridge itself: scale construction, score linking, equating designs, anchor behaviour, mode linking, population invariance and the boundary between defensible common-scale interpretation and overclaim.
The Return Path
Return to two students holding the same reported score from different forms.
The number looks simple because the difficult work happened before it appeared: test design, common evidence, statistical linking, subgroup checks, conversion validation and decisions about what kind of comparability the evidence can support.
That is the point of good measurement infrastructure. Complexity should live inside the system so the public score can remain interpretable without becoming misleading.
The purpose of equating is not to make different tests look the same. It is to prevent irrelevant differences between sufficiently similar tests from changing the meaning of the score.
Return to the How Education Works hub.