You write a survey, field it to 1,000 people and open the data file. Some columns hold brand names, some hold 1 to 5 ratings and some hold dollars spent. Each needs different handling, and the mistakes people make here, such as averaging categories or reading a rating scale as a ruler, are easy to avoid once you know the four scales. They also sit underneath the question of whether a multi-item scale is reliable, which we return to below. Getting scales right is also a foundation of survey data quality.
Key takeaways
- The four measurement scales are nominal, ordinal, interval and ratio, and each adds one property to the last: order, then equal distance, then a true zero.
- The scale decides the statistic: percentages and modes for nominal data, medians or top-box for ordinal, means for interval, and any statistic for ratio.
- Strictly, a rating scale is ordinal, but balanced scales with evenly spaced, clearly labeled points are commonly treated as interval so you can report means.
- Design choices such as 5 or 7 points, a deliberate midpoint, full labels and balance decide how close a rating scale behaves to interval data.
- Use the same number of points, direction and labeling across items you combine, because mixed scales give the longer scale more weight in a total.
What are the four measurement scales?
The scales are ordered by how much information the numbers carry. Each level keeps everything the one before it offers and adds one property: order, then equal distance, then a true zero.
| Scale | What the numbers tell you | Survey example | Typical summary |
|---|---|---|---|
| Nominal | Which category, nothing more | Brand last purchased; Yes/No; region | Counts, percentages, mode |
| Ordinal | Which category, and the order | Rank these three concepts; age bands; “How often do you shop here?” (never to weekly) | Median, percentiles, mode |
| Interval | Order and equal distance, but no true zero | A balanced 7-point satisfaction scale, if treated as equal steps; temperature in Celsius | Mean, standard deviation, correlation |
| Ratio | Order, equal distance and a true zero | Dollars spent; purchases in the past 30 days; age in years | All of the above, plus ratios and percent change |
A short way to remember it is “difference, direction, amount, zero”. Nominal says two answers differ. Ordinal says which is higher.
Interval says by how much. Ratio says how many times as much.
What can you calculate with each scale?
Work out what the scale permits before you pick a statistic. A worked example makes the limits concrete.
Nominal. Suppose 1,000 respondents name the brand they bought last: Brand A 420, Brand B 310, Brand C 270. Those are 42%, 31% and 27%, which add to 100%. The mode is Brand A. A “mean brand” does not exist, because coding the brands 1, 2 and 3 is only labeling.
Ordinal. Five respondents rank three concepts. You can say most put Concept A first. You cannot say that the gap between first and second equals the gap between second and third. One person may love A and find B and C equally poor, and a ranking hides that.
Interval and ratio. Five respondents rate satisfaction from 1 to 5 and give 2, 3, 3, 4 and 5.
| Statistic | Calculation | Result |
|---|---|---|
| Mode | Most frequent value | 3 |
| Median | Middle of the ordered values | 3 |
| Mean | (2 + 3 + 3 + 4 + 5) / 5 = 17 / 5 | 3.4 |
All three are legitimate for an interval reading. Only the mode and median are safe for a strictly ordinal reading. Ratio data adds one more move: if one customer spent $20 and another $40, the second spent twice as much, and “twice” is meaningful because $0 means no spend. A 40 degree day is not twice as hot as a 20 degree day, because Celsius has no true zero.
Are rating scales ordinal or interval?
This is the question that matters most in day-to-day survey work. A single satisfaction item is, strictly, ordinal: nothing guarantees that the step from “Dissatisfied” to “Neither” feels the same as the step from “Satisfied” to “Very satisfied”.
The working compromise in the field is to treat well-built rating scales (balanced, with evenly spaced and clearly labeled points) as interval, report means for tracking, and show the percentage at each point so a reader can see the shape. Where the labels are clearly uneven, or the scale is a short ranking, fall back on medians and top-box shares.
How should you design a rating scale?
Scale choices decide how close your data comes to interval behavior.
How many points?
Five and seven points are the common choices for agreement and satisfaction, and 0 to 10 is common for likelihood to recommend. Too few points, such as three, squeeze respondents into coarse buckets. Too many points ask people to make distinctions they cannot make. Pick one length and keep it within a questionnaire.
Should there be a midpoint?
An odd number of points gives a neutral middle. Keep it when neutrality is a real opinion. Remove it, using an even number of points, only when you need respondents to lean one way, and know that some will choose a side without holding a view.
Should every point be labeled?
Labeling every point with words, from “Very dissatisfied” to “Very satisfied”, makes the meaning of each step explicit. Labeling only the endpoints leaves respondents to interpret the middle themselves, and numeric labels alone push the scale towards looking like a ruler. Whichever you choose, use it consistently so scores stay comparable across questions and waves.
Is the scale balanced?
A balanced scale offers the same number of positive and negative options around the neutral point. A scale with three positive choices and one negative choice builds a positive tilt into the results before anyone answers.
How does scale choice affect reliability?
Reliability asks whether a set of items measures the same thing consistently. Cronbach’s alpha summarizes that from the correlations between items, and correlations assume at least roughly interval data. If your items use a coarse two- or three-point response, the variation they can show is limited, and that tends to depress the correlations alpha relies on.
Mixed scales cause a second problem. If one item uses 1 to 5 and another 1 to 7, their variances differ and a simple sum gives the longer scale more weight.
Use the same number of points, the same direction and the same labeling across the items you intend to combine. Also reverse-score negatively worded items before you compute the total. The pillar guide shows how to compute and interpret the result.
How do you apply this in practice?
Once the scale of each question is clear, the analysis follows: percentages and modes for nominal questions, medians or top-box for ordinal ones, means and correlations for rating scales you treat as interval, and any statistic for ratio data. Halo Reports builds crosstabs and banner tables with weighting and significance testing, plus charts and interactive report pages, so each question type can be cut by the groups you care about.
What are the common mistakes?
- Averaging nominal codes. A mean of region codes 1 to 4 has no meaning. Report shares.
- Reading ranks as distances. Knowing A beat B beat C says nothing about how far apart they are.
- Mixing scale lengths in one index. Combining 5-point and 7-point items without rescaling gives the longer scale more influence.
- Changing the scale between waves. Moving from 5 to 7 points in a tracker breaks comparability with earlier results.
- Collapsing ratio data into bands too early. Asking for age or spend in ranges throws away detail you cannot recover. Collect the exact number when respondents can give it, and band it later.
Before your next questionnaire goes to field, write the scale type next to each question and the statistic you plan to report for it. If a question cannot support that statistic, change the question now, while it is cheap. For wide, complex studies and global brand trackers, a governed platform built for insights and research teams is one place to run them.