How Equipercentile Equating Works | Match Scores by Percentile Position When the Relationship Between Forms Is Not a Straight Line

eduKateSG Learning Node Series · 0195

Two test forms can be similar overall and still differ in a way that no single straight-line adjustment captures.

Suppose Form B is slightly harder than Form A around the middle of the score range but almost identical near the top. Mean equating would shift every score by the same amount. Linear equating would stretch and shift the entire scale. Neither method can reproduce a curved relationship that changes across score regions.

Equipercentile equating takes a different route. Instead of assuming one algebraic shape in advance, it finds scores on the two forms that occupy the same percentile position in their appropriately comparable score distributions. If a score of 42 on Form A and 39 on Form B each correspond to the same percentile rank, the method treats them as equipercentile equivalents.

Equipercentile equating works by matching scores that have the same percentile standing in comparable score distributions, allowing the score conversion to bend when form differences are nonlinear across the score range.

The 50-Second Read

Equipercentile equating matches scores by cumulative distribution position rather than by a fixed linear formula.
It can represent nonlinear relationships between test forms.
The forms still need an appropriate equating design: random groups, single group, counterbalanced or a valid anchor-based design.
The score distributions must be made comparable before percentile matching is meaningful.
Discrete raw scores create jagged empirical percentile functions.
Smoothing is commonly used so sampling noise is not mistaken for real nonlinear structure.
Presmoothing smooths score distributions before equating; postsmoothing smooths the resulting equating function.
Too much smoothing can erase genuine curvature.
Too little smoothing can chase random irregularity.
Tail regions are especially unstable because few candidates occupy extreme scores.
Equipercentile methods often need larger samples than simple mean or linear equating.
The method is flexible, but flexibility increases the need for diagnostic judgement.

Canonical Owner Boundary

This node owns the observed-score equating method that matches percentile positions and permits nonlinear score transformations. How Random-Groups Equating Works and How Single-Group Equating Works own data-collection designs. How Common-Item Nonequivalent-Groups Linking Works owns the anchor-based design for unequal groups. The next node, How Linear Equating Works, owns the straight-line alternative. This article asks: how do we equate forms when their score relationship bends rather than staying linear?

1. Percentile Position Is the Core Idea

Take the cumulative score distribution for Form A. For each raw score, determine the proportion of the relevant population scoring at or below that point. Do the same for Form B.

If a score on A and a score on B occupy the same cumulative position, equipercentile logic treats them as equivalent. The mapping is based on rank within the distribution rather than on equal raw points.

2. The Design Must Make the Distributions Comparable First

Percentiles only mean the same thing if the underlying distributions belong to populations that can legitimately be compared. In a random-groups design, randomisation supplies that comparability. In a single-group design, the same people provide both distributions. In a nonequivalent-groups anchor design, the anchor is used to adjust or link the populations before the equipercentile relationship is interpreted.

Equipercentile is therefore a method, not a substitute for equating design.

3. Why Percentiles Permit Curvature

Suppose the median score is 30 on Form A and 27 on Form B, but the 90th percentile is 45 on A and 44 on B. The form difference is larger in the middle than at the top. A constant shift cannot reproduce both relationships.

Equipercentile equating can map 30 to 27 and 45 to 44 because each percentile is matched independently before a smooth function connects the points.

4. Raw Scores Are Discrete

A 50-item test has only 51 possible integer raw scores. Percentile distributions therefore jump in steps rather than forming a perfectly smooth curve.

Direct percentile matching needs a convention for handling these jumps. Continuization methods conceptually spread each discrete score category over a small interval so a smoother inverse percentile mapping can be defined.

5. Sampling Noise Makes Empirical Curves Jagged

Even if the population relationship between forms is smooth, a finite sample can produce irregular frequency spikes and dips. Direct unsmoothed equipercentile equating can chase those sample accidents.

That is why smoothing is not cosmetic. It is an attempt to separate likely population structure from random sample irregularity.

6. Presmoothing Acts on the Score Distributions

Presmoothing fits a smoother model to the observed score frequencies before the equipercentile relationship is calculated. Log-linear models are commonly used because they can preserve selected moments and structural features while reducing erratic local variation.

The danger is obvious: a smoothing model that is too restrictive can erase genuine irregularity or curvature in the population distribution.

7. Postsmoothing Acts on the Equating Function

Instead of smoothing the distributions first, analysts can calculate an empirical equipercentile conversion and then smooth the resulting transformation. Postsmoothing works directly on the score relationship.

Both strategies trade variance against bias in different ways. The correct choice depends on the score distributions, sample size and form relationship.

8. Smoothing Is a Model Choice

Moses and Liu’s ETS report Smoothing and Equating Methods Applied to Different Types of Test Score Distributions compared presmoothing, equipercentile, kernel and postsmoothing approaches under different distribution shapes. Their work shows why no one smoothing recipe should be treated as universally correct.

When population score distributions contain systematic irregularities, aggressive smoothing can make the equating function look elegant while becoming less faithful to the underlying structure.

9. Small Samples Create a Bias–Variance Fight

With a small sample, unsmoothed percentile curves are unstable. Smoothing can substantially reduce variance, but it imposes structure that may bias the relationship if chosen badly.

Livingston’s ETS work on small-sample equating with log-linear smoothing found large gains in accuracy in the study conditions, while also showing that overly simple smoothing could miss genuine curvilinearity.

10. The Tails Are Where Flexibility Becomes Dangerous

Very high and very low raw scores are often rare. Their empirical percentile locations can therefore change sharply when only a few candidates are added or removed.

A nonlinear method can amplify this instability into large tail conversions. Tail diagnostics, smoothing choices and minimum sample expectations are therefore crucial.

11. Extrapolation Beyond Observed Scores Is Especially Weak

If nobody in the sample scores 49 or 50 on one form, the empirical distribution contains little direct evidence about the corresponding percentile mapping. Any conversion there becomes extrapolation from the smoothing model or boundary rule.

Publishing a complete conversion table does not mean every row has equal empirical support.

12. Linear Equating Is Nested Inside the Larger Question

If the two score distributions differ only in mean and standard deviation under a location-scale relationship, a linear transformation can be sufficient. Equipercentile methods become valuable when evidence suggests shape differences that matter.

Flexibility should therefore be earned by diagnostics, not selected because “nonlinear” sounds more sophisticated.

13. Equipercentile Equating Preserves Rank Position, Not Item Meaning

Matching percentile positions does not prove the forms measure the same construct. Two unrelated tests can always have percentile distributions, but that would not make their scores meaningfully equatable.

Content comparability and construct equivalence are upstream requirements. Statistical rank matching only operates after those requirements are reasonably satisfied.

14. Chained Equipercentile Uses an Anchor as Intermediate Currency

In the common-item nonequivalent-groups design, one popular approach is chained equipercentile equating. Form A is related to the anchor in one group, and the anchor is related to Form B in the other. The two percentile mappings are chained.

This makes anchor quality central. A narrow or unstable anchor contaminates the nonlinear bridge.

15. Poststratification Equipercentile Builds Synthetic Populations

Another anchor-based approach uses the common items to create weighted or synthetic score distributions representing a common population before percentile matching.

Chained and poststratification equipercentile methods therefore share the equipercentile principle but handle population nonequivalence differently.

16. Kernel Equating Extends the Same Family of Ideas

Kernel equating uses smoothing and continuization techniques within a more general framework for score equating. It can produce smooth nonlinear functions and standard errors while making the smoothing choices explicit.

The existence of multiple nonlinear methods reinforces the main point: flexibility requires assumptions about how much irregularity is signal and how much is noise.

17. Identity Is Always a Useful Baseline

If two forms were built very similarly, the identity transformation—score x equals score x—can be surprisingly competitive. Every more complex method should justify the extra adjustment it introduces.

Equating should correct real form differences, not manufacture them from sampling noise.

18. Score Rounding Creates a Final Discrete Decision

Equipercentile transformations can produce fractional equivalents such as 37.4. Operational systems may then round to integer scaled scores or another reporting metric.

Rounding can create small discontinuities, especially near cut scores. The complete decision pipeline should therefore be checked after rounding, not only at the continuous-function stage.

19. Standard Error of Equating Varies Across Scores

The conversion is not equally precise everywhere. Middle score regions with abundant observations often have smaller uncertainty than sparse tails.

Scale linking error therefore belongs beside the nonlinear conversion rather than being hidden behind a polished lookup table.

20. Cross-Domain Comparison: Mapping Two Uneven Thermometers

Imagine two sensors where one compresses differences near the top of its range. A straight correction can align the middle but still miss the upper end. A nonlinear calibration curve is needed because the relationship changes with level.

Equipercentile equating is the score-distribution analogue: it allows different parts of the scale to receive different adjustments.

21. Cross-Domain Comparison: Currency Purchasing Power

A single market exchange rate can hide the fact that the relative price of food, housing and services differs across countries. A more detailed purchasing-power comparison effectively uses different parts of the consumption distribution.

Equipercentile equating likewise refuses to assume one constant score adjustment when the relationship changes across the distribution.

22. Failure Mode: Use Unsmoothed Equipercentile With Tiny Samples

A jagged sample distribution produces a jagged conversion table that is treated as real form structure.

Repair: evaluate smoothing, bootstrap or resampling stability, and simpler methods. Flexibility without enough data becomes noise fitting.

23. Failure Mode: Smooth Until the Curve Looks Nice

The analyst chooses the smoothest transformation because it appears professional.

Repair: choose smoothing through statistical diagnostics and substantive expectations. A beautiful curve can erase genuine population structure.

24. Failure Mode: Ignore Tail Uncertainty

The top two raw-score conversions are based on only a handful of candidates but are reported with the same confidence as middle scores.

Repair: inspect score frequencies, standard errors and extrapolation dependence before using tail conversions for high-stakes decisions.

25. Failure Mode: Use Nonlinearity to Equate Unlike Constructs

Two tests measure different content emphases, so a flexible percentile mapping is used to force comparability.

Repair: stop. Equating assumes strong construct similarity. Nonlinear statistics cannot repair a construct mismatch.

26. A Practical Equipercentile Workflow

Confirm the forms are suitable for equating.
Use a defensible random-groups, single-group, counterbalanced or anchor-based design.
Construct comparable score distributions.
Inspect sample size and score-frequency coverage.
Plot empirical cumulative distributions.
Evaluate presmoothing options.
Compute the equipercentile transformation.
Inspect curvature and tail behaviour.
Compare with identity, mean and linear alternatives.
Estimate standard error and resampling stability.
Check the effect of rounding and cut scores.
Document smoothing assumptions and weak score regions.

27. Classroom Translation

A teacher comparing two test versions may notice that one form is three marks harder for middle-performing students but nearly identical for top performers. A single “add three marks” rule would overcorrect the top.

Formal equipercentile equating requires much more data than an ordinary classroom provides, but the conceptual lesson remains useful: form difficulty can vary across the ability range, so one constant adjustment may be too crude.

28. Missing-Node Scan

The missing node may be equipercentile equating when linear transformations leave systematic score-region differences; when two forms have similar means but different skew or tail behaviour; when empirical percentile mappings are extremely jagged; when smoothing choices materially change score conversions; when chained equating is used with a common-item anchor but the nonlinear mechanism is poorly understood; when extreme-score conversions rest on very few observations; or when a sophisticated nonlinear method is being used despite no evidence that the form relationship is meaningfully nonlinear.

29. Evidence and Limits

Equipercentile equating is a foundational observed-score method described in ETS references including Livingston’s Equating Test Scores (without IRT), Second Edition and the ETS chapter Equating Test Scores. ETS research by Moses and Liu examines smoothing and equating under different score-distribution shapes, while Livingston’s small-sample research demonstrates both the value and the risks of log-linear smoothing.

The limit is flexibility. Equipercentile methods can model nonlinear relationships that simple transformations miss, but they can also reproduce sampling noise when data are thin. The stronger the curve is allowed to bend, the stronger the evidence must be that the bend belongs to the population rather than the sample.

30. The Return Path

Return to the forms that differ by three marks in the middle and one mark at the top.

Equipercentile equating can follow that changing relationship because it asks where each score sits in the comparable distribution rather than imposing one global line. The price is that every bend must be supported by enough evidence and enough smoothing discipline to separate structure from noise.

Equipercentile equating works when the score relationship genuinely bends. Its strength is flexibility; its risk is believing every wiggle in a finite sample.

Research and Further Reading

eduKateSG Learning Node Series · 0195 · Previous: 0194 — How Single-Group Equating Works.