How Circle-Arc Equating Works | Add Just Enough Curvature for Small Samples Without Letting Noise Draw the Score Curve

eduKateSG Learning Node Series · 0198

When a straight line is too rigid and full equipercentile equating is too unstable, a controlled curve can be the better compromise.

Small-sample equating creates a recurring problem. Mean and linear methods are stable because they estimate little, but they can miss genuine curvature. Equipercentile equating can follow curvature, but with only 25, 50 or 100 candidates it may follow sampling noise just as enthusiastically.

The circle-arc method was developed as a middle route. It constrains the equating relationship to a smooth circular arc passing through carefully chosen end points and a middle point estimated from the data. The method is not trying to reconstruct every wiggle in the empirical score distributions. It asks whether one restrained bend can capture the important departure from linearity.

Circle-arc equating works by constraining the score conversion to a smooth arc through predefined end points and a data-determined middle point, giving small samples limited nonlinear flexibility without allowing every random irregularity to become part of the equating function.

The 50-Second Read

Circle-arc equating was developed for small-sample score equating.
It permits curvature but far less freedom than full equipercentile equating.
The curve is constrained by two end points and one middle point.
The middle point carries much of the sample information about form difference.
End points can be chosen from theoretical score limits or other defensible boundary information.
The method can outperform linear and equipercentile alternatives when samples are small and the true relationship is moderately nonlinear.
Its advantage is a bias–variance compromise, not a universal superiority claim.
Poorly chosen end points can force the wrong curvature.
The method can behave badly if the real equating relationship has more complex shape than one arc can represent.
It still depends on a valid random-groups, single-group or common-item design.
Small-sample uncertainty remains and should be reported.
Circle-arc equating is best understood as constrained nonlinearity.

Canonical Owner Boundary

This node owns the constrained curved equating method developed to balance nonlinear form differences against small-sample instability. How Mean Equating Works owns constant-shift adjustment. How Linear Equating Works owns straight-line shift-and-stretch adjustment. How Equipercentile Equating Works owns full percentile-based nonlinear mapping. This article asks: what if we need some curvature, but our sample is too small to estimate a free nonlinear curve safely?

1. Small Samples Force a Different Kind of Humility

With tens of thousands of examinees, empirical score distributions can support detailed nonlinear equating. With 50 candidates per form, each individual can visibly change percentile ranks, tail frequencies and local curve shape.

The correct response is not necessarily to pretend the true relationship is linear. It is to reduce the number of nonlinear features we ask the data to estimate.

2. Circle-Arc Equating Uses Three Anchoring Conditions

Livingston and Kim’s original ETS work describes a family of methods in which the equating function is constrained to pass through two specified end points and one middle point determined from sample data.

Three points are enough to define a circular arc under the method’s geometry. Instead of estimating a conversion independently at many score points, the entire relationship is disciplined by those three conditions.

3. The End Points Are Prior Structure

The minimum and maximum possible scores provide natural places to think about boundary relationships, but the precise end-point specification depends on the version and context of the method.

The crucial idea is that the small sample is not asked to estimate everything. Some boundary information comes from test structure or defensible prior constraints, reducing variance.

4. The Middle Point Carries the Empirical Difference

The middle point is estimated from the observed data and determines how the arc bows between the end points. If the middle relationship sits close to the straight line connecting the ends, the resulting arc is nearly linear. If it departs, the curve captures moderate nonlinearity.

This gives the sample one major nonlinear degree of freedom rather than dozens.

5. Why an Arc Rather Than a Polynomial?

A flexible polynomial can swing wildly outside densely observed regions, especially with small data. High-order curves can produce implausible reversals or extreme tail behaviour.

A circular arc imposes smooth monotonic geometry and limited curvature. The restriction is the point: it prevents a small dataset from inventing complicated score structure.

6. The Method Is a Bias–Variance Compromise Made Visible

Mean equating has low variance but can have high bias if the relationship bends. Full equipercentile equating can have lower structural bias but high sampling variance. Circle-arc equating deliberately sits between them.

Its success therefore depends on the true relationship being complex enough to need curvature but simple enough for one restrained arc to approximate.

7. ETS Small-Sample Research Motivated the Method

Livingston and Kim’s 2008 ETS report, Small-Sample Equating by the Circle-Arc Method, introduced related arc methods intended to estimate nonlinear score relationships from small groups. Their preliminary results showed promising performance relative to mean equating and other small-sample alternatives.

The method was later published in the Journal of Educational Measurement as The Circle-Arc Method for Equating in Small Samples.

8. Random-Groups Studies Show the Trade-Off Clearly

In their 2010 ETS resampling study of randomly equivalent groups of 50 to 400 examinees per form, Livingston and Kim compared circle-arc, linear, mean and smoothed equipercentile methods against a large-sample criterion.

Circle-arc equating produced the smallest overall errors across the studied sample sizes, particularly in the upper half of the score distribution. The result should not be universalised, but it demonstrates why constrained curvature can be valuable when samples are sparse.

9. Common-Item Designs Add Another Layer

Kim and Livingston also studied small-sample common-item linking with new-form samples as small as 10 candidates and reference samples three times larger. The competing methods included chained equipercentile, chained linear, chained mean and circle-arc approaches.

Their 2009 report found chained mean equating strongest for some low-score regions while circle-arc performed especially well in the upper half. The key lesson is not that one method always wins; it is that score region and sample size can change the ranking.

10. One Overall Error Number Can Hide Regional Failure

A method can have low average root-mean-squared error while performing poorly near a consequential cut score. Small-sample equating should therefore be evaluated across score regions, not only through one global summary.

If a certification decision sits in the lower tail, a method that performs best in the upper half may not be the right operational choice.

11. The Arc Does Not Remove the Need for Comparable Forms

Circle-arc equating can flex between score scales. It cannot make different constructs equivalent. Alternate forms still need aligned blueprints, administration conditions and score meaning.

The more the forms differ substantively, the more dangerous it becomes to interpret geometric curve fit as measurement equivalence.

12. The Design Remains Upstream

A circle-arc transformation can be estimated within a random-groups or common-item framework, but the curve itself does not solve population nonequivalence. The data must first support the form comparison.

Use random groups, single groups or common-item nonequivalent groups appropriately before choosing the shape of the transformation.

13. Endpoint Choice Can Dominate the Curve

Because the arc must pass through its specified end points, poor boundary assumptions pull the entire conversion. If the forms have different effective floors or ceilings, forcing common raw-score endpoints can create misleading curvature.

Endpoint logic therefore deserves substantive review rather than being accepted as a software default.

14. The Middle Point Can Be Noisy Too

The method reduces the number of empirical degrees of freedom, but the middle point still comes from finite data. If the sample mean or chosen middle relationship is unstable, the entire arc moves.

Bootstrap or repeated-sampling analyses can show how sensitive the resulting curve is to that single empirical anchor.

15. One Arc Cannot Represent Every Nonlinear Relationship

A form relationship could bend one way in the lower scores and another way near the top. One circular arc has limited shape. It cannot reproduce arbitrary inflection patterns.

When large samples reveal more complex population curvature, equipercentile or other nonlinear methods can be more appropriate.

16. This Is Not “Approximate Equipercentile” by Default

Circle-arc equating does not simply smooth an empirical equipercentile curve. It imposes its own constrained geometric form. The resulting conversion should be evaluated as a model in its own right.

Its success depends on whether that restricted family of curves approximates the population relationship well enough.

17. Prior Information Is the General Principle Beneath the Method

Circle-arc equating works partly because it refuses to ask a tiny sample to estimate every feature. Structural information about endpoints and smoothness acts like prior information.

Livingston and Lewis explored a different small-sample route in their ETS report Small-Sample Equating With Prior Information, using empirical Bayes information from previous equatings. Both approaches express the same systems insight: when new data are scarce, trustworthy prior structure can reduce variance.

18. But Prior Structure Can Be Wrong

Constrained methods reduce variance by refusing some possible curves. If the true form relationship lies outside the allowed family, the constraint becomes systematic bias.

The discipline is therefore two-sided: do not let small data invent complexity, and do not let prior geometry suppress real complexity.

19. Standard Error Still Belongs Beside the Curve

A smooth arc can create an illusion of certainty because it lacks the visible jaggedness of an empirical equipercentile function. But the middle point and underlying group statistics are still estimated.

Report linking uncertainty and, where possible, resampling stability across score points.

20. Cut Scores Change the Method Evaluation

Suppose circle-arc equating has slightly lower overall error than mean equating, but mean equating is more accurate around the pass mark. If the test’s primary use is pass/fail classification, local performance around the cut can outweigh global average performance.

The equating method should be judged against the decision system it supports, not only a generic mathematical loss function.

21. Cross-Domain Comparison: A Suspension Bridge Cable

A cable constrained at two towers and shaped by one central sag can form a smooth, physically restricted curve. It cannot zigzag freely between every gust of wind.

Circle-arc equating behaves similarly. The endpoints are the towers, the middle point controls the bow, and the structural constraint prevents random local noise from writing the entire curve.

22. Cross-Domain Comparison: Three-Point Calibration

An instrument engineer may know two boundary values from design and use one mid-range calibration measurement to estimate modest nonlinearity. The result is less flexible than a full empirical lookup table but more realistic than a straight line.

The analogy captures the method’s central economy: spend scarce data where it changes the model most.

23. Failure Mode: Use Circle-Arc Because the Sample Is Small

A programme has 40 candidates, so it automatically adopts circle-arc equating without checking whether the form relationship is actually curved.

Repair: compare identity, mean, linear and arc alternatives. If a constant or straight-line relationship is adequate, extra curvature adds unnecessary estimation risk.

24. Failure Mode: Treat Endpoints as Sacred

The software uses theoretical score minima and maxima even though one form has a severe floor effect and the other does not.

Repair: inspect whether the endpoint assumptions reflect meaningful score equivalence. Geometry should follow measurement logic, not the other way around.

25. Failure Mode: Judge Only Overall RMSE

The arc has the lowest average error but performs poorly near the operational cut.

Repair: evaluate region-specific error, classification consequences and score-use priorities.

26. Failure Mode: Confuse Smoothness With Truth

The circle arc looks elegant, so the result feels more trustworthy than a jagged empirical curve.

Repair: quantify uncertainty. Smoothness is imposed structure; it is not evidence that the underlying relationship is known precisely.

27. A Practical Circle-Arc Workflow

Confirm the forms are substantively suitable for equating.
Use a defensible equating design.
Inspect sample size and score coverage.
Compare identity, mean and linear baselines.
Specify defensible end points.
Estimate the middle point from the data under the chosen circle-arc procedure.
Construct the arc and inspect monotonicity and boundary behaviour.
Compare with smoothed equipercentile or criterion equating where possible.
Bootstrap or resample to examine stability.
Evaluate error near decision-relevant score regions.
Document why limited curvature was preferred to a line or a more flexible nonlinear method.

28. Classroom Translation

A teacher with 25 students should not attempt formal circle-arc equating from one classroom dataset. But the method carries a valuable design lesson: when evidence is sparse, do not let every local irregularity drive the adjustment.

Use simple, constrained explanations unless repeated evidence shows that a more complex pattern is real. This is as true in learning diagnosis as it is in score equating.

29. Missing-Node Scan

The missing node may be circle-arc equating when a programme has too few candidates for stable equipercentile curves but clear evidence that mean or linear equating is too rigid; when tail percentile conversions jump wildly between samples; when prior boundary information is trustworthy but the middle score relationship needs data-driven curvature; when a small-volume testing programme repeatedly chooses between high-bias simple methods and high-variance flexible methods; or when a smooth nonlinear transformation is needed without granting the sample unlimited freedom.

30. Evidence and Limits

The circle-arc method was developed by Samuel Livingston and Sooyeon Kim at ETS specifically for small-sample equating. Their 2008 research report proposed constrained arc methods, their 2009 Journal of Educational Measurement article developed the approach further, and their resampling studies compared its accuracy with mean, linear and smoothed equipercentile methods under both random-groups and common-item designs.

The limitation follows directly from the method’s design. One arc cannot represent arbitrary form relationships, and chosen endpoints can impose bias. Circle-arc equating is therefore not a new default. It is a specialised answer to a specific problem: limited data plus plausible moderate curvature.

31. The Return Path

Return to the testing programme with 50 candidates, a straight line that misses the population relationship and an equipercentile curve that changes shape every time the sample changes.

The circle-arc method refuses both extremes. It allows the relationship to bend, but only in a controlled way. It spends scarce information on one meaningful piece of curvature instead of dozens of noisy local adjustments.

Circle-arc equating works by putting a speed limit on flexibility: enough curvature to escape a wrong straight line, not enough freedom for a small sample to draw fiction.

Research and Further Reading

eduKateSG Learning Node Series · 0198 · Previous: 0197 — How Mean Equating Works.