Likert Scale vs Radial Scale: Discrete Categories or Continuous Measurement
Discrete Likert categories versus continuous radial and slider formats: what each measures, what the evidence says about break-off, and when to use which.
TL;DR
- A Likert scale gives a small number of labelled, discrete categories. A radial scale asks the respondent to place a position on a circular continuum, producing a near-continuous value.
- The developer of the Sama Radial Indicator claims three anchors and 180 response options, on the argument that continuous measurement reduces response error and normalises the resulting distribution.
- The independent evidence on continuous input formats is mixed and format-specific: sliders performed badly in a controlled web experiment, while visual analogue scales did not.
- For CX tracking, discrete Likert remains the safer default. Radial formats are worth testing in controlled research settings, not in a live tracker you need to keep comparable.
The structural difference
A Likert item offers a fixed set of labelled options and asks the respondent to pick one. Five or seven categories, each with a verbal meaning attached, and nothing in between. The output is ordinal: the categories are ordered, but the distance between them is not guaranteed to be equal.
A radial scale replaces the row of boxes with a circular or arc-shaped control. The respondent positions a marker along the arc, and the position is read off as a value. The output is effectively continuous, or fine-grained enough to be treated as such.
The Sama Radial Indicator (SRI), developed by W.G.S. Konarasinghe and described in a paper available on SSRN comparing SRI with the Likert scale for measuring psychometric variables, is the most concrete published example. Its design is built on the properties of a circle. The stated claim: the scale carries only three anchors but offers 180 response options, which its author argues reduces response error, and it converts psychometric ratings into a continuous scale so that latent variables become normally distributed.
That is the claim as published. It should be read as the developer's argument for his own instrument, not as an independently replicated finding. We have not found peer-reviewed replication of the error-reduction or normality claims by researchers outside the SRI work, and the earlier version of this page presented them with more confidence than the evidence base supports.
The genuine argument for continuous measurement
Set aside the specific instrument. The underlying critique of Likert scales is real and worth stating properly.
The problem is category ambiguity at the inner edges. On a 7-point agreement scale, the distinction between "moderately agree" and "slightly agree" is not something most respondents hold a stable opinion about. Konarasinghe makes exactly this point: respondents can hardly distinguish those categories, so the answer becomes partly a guess, and guessing is measurement error. A continuous control sidesteps the problem by never asking the respondent to name their position, only to place it.
The second argument is statistical. Likert data is ordinal but routinely analysed with methods that assume interval or continuous data. A continuous instrument, if it works, removes that mismatch rather than papering over it.
Both arguments are legitimate. Neither is settled by the existence of an instrument that claims to solve them.
Extra resolution is worthless if the respondent cannot supply it. A control offering 180 positions assumes a person holds an opinion precise to fractions of a degree. Most do not. What you gain in recorded decimal places, you can lose in placement noise, and the analysis then treats that noise as if it were signal. Precision in the instrument is not the same thing as precision in the respondent.
What the independent evidence actually shows
The closest well-controlled evidence on continuous input formats in web surveys is Frederik Funke's 2016 experiment in Social Science Computer Review. It compared slider scales and visual analogue scales against standard HTML radio buttons, with three, five or seven response options.
The findings matter here because they cut both ways:
- Slider scales performed badly. Significantly higher break-off, with an odds ratio of 6.9, and substantially longer response times. The recommendation in that work is to avoid slider scales.
- Visual analogue scales did not. VAS and radio buttons could be used without those negative side effects, including on touchscreen devices.
The distinguishing factor was the interaction mechanic, not continuity as such. Drag-and-drop imposed a cost that point-and-click did not. A radial control that requires dragging a marker around an arc on a phone is closer to the slider case than the VAS case, which is a real reason for caution. A radial control that responds to a single tap is closer to VAS.
This is also why the earlier claim on this page, that radial formats produce 8% missing responses against 1% for Likert, should not be used. We could not locate a source for those figures, and the best available experimental evidence points the opposite way for the drag-based formats radial controls most resemble.
Practical trade-offs
Likert (discrete) | Radial / continuous | |
|---|---|---|
Output | Ordinal categories | Near-continuous values |
Respondent task | Pick a labelled option | Place a position |
Verbal labelling | Every point can be labelled | Anchors only |
Mobile rendering | Well understood, robust | Depends heavily on interaction mechanic |
Benchmarking | Comparable to existing data | No external comparison set |
Analysis | Ordinal methods, or interval by convention | Continuous methods available |
Accessibility | Standard controls, screen-reader friendly | Custom controls need explicit work |
The two that usually decide it in a commercial CX programme are labelling and benchmarking.
Full verbal labelling is a documented reliability advantage: Krosnick and Presser's review found fully labelled scales more reliable than partially labelled ones. A three-anchor radial control cannot label its intermediate positions by construction, which means it gives up the thing that most reliably improves a rating scale in exchange for resolution the respondent may not have.
Benchmarking is the harder constraint. Almost every external CX comparison set, and almost certainly your own tracker history, is built on discrete scales. Switching format breaks the series. Dawes (2008) showed that even changing the number of points on a discrete scale shifts the mean, coefficient of variation, skewness and kurtosis. Moving from discrete to continuous is a much larger change, and there is no defensible conversion.
Amitayu Basu, CEO, NumrI have no objection to radial scales in a research study where you control the sample and you are not comparing to anything. In a client tracker they are a trap. The moment you change the instrument you have started a new series, and six months later somebody presents a trend line that spans the switch as if it were one line. The methodological gain is small. The reporting damage is permanent.
Where each format belongs
Use a Likert scale when you are running a tracker, benchmarking externally or internally, fielding across languages and devices, or reporting to stakeholders who need familiar categories. This covers nearly all commercial CX measurement.
Consider a radial or continuous format when you are running controlled academic or exploratory research, the analysis genuinely requires continuous input, you control the device environment, and you are not comparing to historical data. Test the interaction mechanic on real mobile devices before fielding, and check break-off in a pilot rather than trusting the format's own literature.
Do not mix formats within one questionnaire section. Switching the respondent between a labelled row and a continuous arc mid-battery adds a cognitive switching cost with no measurement gain.
For the related question of how many points a discrete scale should carry, see 5-point vs 7-point Likert scales. For why the recommendation question is a fixed 11-point exception to all of this, see why NPS uses an 11-point scale.
Frequently asked questions
Does a radial scale really offer 180 response options? That is the claim made for the Sama Radial Indicator by its developer, W.G.S. Konarasinghe: three anchors and 180 response options, based on the geometry of a circle. It is the instrument's design specification as published, not an independently verified performance result.
Is a continuous scale more accurate than a Likert scale? Not demonstrably, in general. The argument is that removing ambiguous intermediate labels removes guessing. The counter-evidence is that continuous controls requiring dragging increase break-off and response time. Accuracy depends on the specific implementation, not on continuity as a principle.
Are radial scales bad on mobile? It depends on the interaction. Funke's experiment found drag-based sliders caused significantly higher break-off, an odds ratio of 6.9, while visual analogue scales and radio buttons did not. A radial control that requires dragging is the risky case. Pilot it before committing.
Can I convert historic Likert data to a radial scale? No. There is no defensible mapping between a small set of ordinal categories and a continuous position. Treat a format change as a new baseline and say so in the report.
Is Likert data ordinal or interval? Formally ordinal, because equal spacing between categories is not guaranteed. In practice it is widely analysed as interval. That convention is one of the motivations behind continuous instruments, though it is not by itself a reason to switch.
What about accessibility? Standard radio-button Likert scales work with screen readers and keyboard navigation by default. Custom radial controls need deliberate accessibility work, including keyboard input and announced values, or they will exclude part of your sample.
Should I A/B test the two formats? If you are seriously considering a switch, yes, but run them in parallel on split samples rather than sequentially, and measure break-off and completion time alongside the scores. Sequential testing confounds format with time.
What does Numr use? Discrete scales as standard: labelled Likert batteries for attributes, 0 to 10 for the recommendation question, and the net method for effort, meaning percent easy minus percent difficult. Comparability across waves and markets is the priority, and discrete formats protect it.
Related guides
For related reading, see Why NPS Uses an 11-Point 0 to 10 Scale, How to Design a CX Survey That Customers Actually Complete, NPS vs CSAT vs CES: Which CX Metric Should You Use, and What Is Net Promoter Score? The Metric, and What to Do With It.
