Two questionnaires want the same thing. The first asks “How old are you?” and leaves an empty box for the number. The second offers four age bands to tick. Once the responses are in, one of them gives you a mean age to the year, the other only a rough estimate.
The difference is the level of measurement. A level of measurement tells you how much information the values of a variable carry, and it therefore decides which calculations and which analyses are open to you later. Four levels are distinguished: nominal, ordinal, interval and ratio, from the least information to the most. We show you how to place any question in three steps, before the survey goes out.
📌 The key points at a glance
- Four levels: nominal, ordinal, interval and ratio.
- The level is set by the question, not by the analysis.
- Nominal data allows counts, never a mean.
- GCSE grades are ordinal, despite looking like numbers.
- Likert scales are strictly ordinal data.
Create a survey for free
With empirio.ai you can create a modern online survey in minutes — with hosting in the EU.
- AI-built survey
- Adjust by drag & drop
- Real-time analysis
What is a level of measurement?
A level of measurement is the property of a variable that decides which statements its values support: only equal and unequal, additionally a rank order, additionally equal distances, or additionally a natural zero point.
Measurement scale and scale level are used for the same idea. The four-level classification comes from the psychologist Stanley Smith Stevens, who set it out in 1946 in the journal Science. It underpins the way statistical methods are taught today. The University of Leeds library guide to statistical tests sorts variables into categorical data, split into nominal and ordinal, and numerical data, also referred to as interval or ratio.
Two terms get mixed up in day-to-day questionnaire work. A rating scale is the visible answer format, for example the five steps from “very satisfied” to “very dissatisfied”. The level of measurement is a property of the data those steps produce.
Categorical means the answer options carry no usable numeric value, which covers nominal and ordinal. Numerical means the values are real numbers you can calculate with, which covers interval and ratio.
A level of measurement describes not the question, but what ends up in the data table.
The four levels of measurement
The four levels build on one another. Each level does everything the level below it does, plus one thing more. That is why a higher level can always be treated as a lower one, while the reverse is impossible.
For a questionnaire this means the higher the level, the more analysis you have available later. Where it costs nothing, collect at the highest level you can. Where a question would become unanswerable or intrusive as a result, the lower level is the better decision.
| Level of measurement | What the values tell you | Survey example |
|---|---|---|
| Nominal | equal or unequal | degree subject, marital status |
| Ordinal | also greater or smaller | GCSE grade, highest qualification |
| Interval | also equal distances | calendar year, degrees Celsius |
| Ratio | also ratios and a true zero | age in years, income in pounds |
Nominal data: categories without an order
Nominal data sorts values into categories that carry no ranking. Single, married and divorced sit side by side as equals, and their order on the questionnaire can be swapped at will. Numbers you assign to such categories in the dataset are pure labels: whether single becomes a 1 or a 7 changes nothing. More in the article on the nominal scale.
Ordinal data: a ranking without fixed distances
Ordinal data adds a rank order but says nothing about the gaps. GCSEs, A levels and a degree form a clear sequence, yet nobody can say whether the step from the first to the second is the same size as the step from the second to the third. How to build such steps cleanly is covered in the article on the ordinal scale.
Interval data: equal distances, arbitrary zero
Interval data has equal distances but no natural zero point. With temperature in degrees Celsius, the difference between 10 and 15 degrees is the same as the one between 25 and 30 degrees. Because the zero was placed arbitrarily at the freezing point of water, 30 degrees is still not twice as warm as 15. Details in the article on the interval scale.
Ratio data: a true zero and real ratios
Ratio data has equal distances and a natural zero point at which the property genuinely is absent. Zero pounds of income means no income, zero years means no age. Only then does the phrase “twice as much” make sense: £3,000 really is double £1,500. Further cases appear in the article on the ratio scale.
Two further names come up regularly. The cardinal scale is not a fifth level but the umbrella term for interval and ratio, in other words for everything numerical. The absolute scale sits above the ratio scale in some textbooks, because the unit of measurement is fixed as well, as with a count of terms studied.
How to determine the level of measurement in three questions
The level of measurement of a variable follows from three yes-or-no questions asked in this order. As soon as one of them gets a no, the level is settled and you can stop.
The test works for any variable, whether it comes from your own questionnaire or from somebody else’s dataset. The only condition is that you keep the exact wording of the question in view rather than the topic in general.
- Do the values carry an order? If no, the data is nominal. If yes, move to question 2.
- Are the distances between neighbouring values equal? If no, the data is ordinal. If yes, move to question 3.
- Is there a zero point at which the property genuinely is absent? If no, the data is interval. If yes, it is ratio.
Question 1 is where most mistakes start. It does not ask whether the values can be sorted somehow, because car makes can be sorted alphabetically too. It asks whether the order means something, whether one value is more, higher or better than the one before it.
Question 3 has a useful cross-check: does the phrase “twice as much” make sense? Twice as much income is a clear statement, twice as much calendar year is not.
The test stops at four levels. Where a variable counts whole units whose size is not freely chosen, such as terms studied or children, some textbooks add an absolute scale on top. For analysis, the same rules apply there as for ratio data.
Which level of measurement does each variable have?
The table below places the variables that questionnaires ask about most often. Each row assumes the obvious answer format, which the third column keeps in mind.
One warning applies to the whole table: the level of measurement follows the answer format, not the topic. Age is ratio data as long as you collect the years as a number. Ask for age bands instead and the data is ordinal, even though the subject is still age.
| Variable | Level of measurement | Why |
|---|---|---|
| Gender | Nominal | categories without a ranking |
| Marital status | Nominal | single and married rank equally |
| Postcode | Nominal | the characters are only a label |
| Degree subject | Nominal | no meaningful order |
| GCSE grade from 9 to 1 | Ordinal | ranked, distances not guaranteed |
| Highest qualification | Ordinal | ordered, distances not measurable |
| Satisfaction on five steps | Ordinal | equal spacing is not guaranteed |
| Year of birth | Interval | the zero point is set arbitrarily |
| Temperature in degrees Celsius | Interval | zero degrees is not an absence of heat |
| Age in years | Ratio | zero years means no age |
| Income in pounds | Ratio | zero pounds means no income |
| Number of terms studied | Absolute | zero point and unit are both fixed |
For the standard questions about the respondent, the article on demographic questions in a questionnaire is worth a look. It also covers which of them you actually need.
Create a survey for free
With empirio.ai you can create a modern online survey in minutes — with hosting in the EU.
- AI-built survey
- Adjust by drag & drop
- Real-time analysis
The level of measurement is set by the question, not by the analysis
The level of measurement of your data is fixed the moment you choose the answer format, not later at the computer. Offer age bands to tick and you have ordinal data, and no analysis software in the world will change that afterwards.
The reason lies in the direction information can travel. A year of birth can be turned into an age band at any point, while an age band never yields a year of birth again. Because the levels build on one another and the information content rises with each step, you can only move down afterwards, never up.
What the open number field gives you
- Mean, median and spread are all available.
- Bands can be redrawn afterwards as you like.
- Relationships with other numbers can be calculated.
What age bands cost you
- The mean can only be estimated roughly.
- The band boundaries are locked in for good.
- Studies with different boundaries barely compare.
⚠️ Careful
Bands still have their place. For income, many surveys choose them deliberately, because the question about an exact figure counts as sensitive. If you expect a lot of drop-off, bands are the better call: ordinal data from everyone beats numerical data from half the sample.
A second reason for bands is comparability with official figures. If you plan to set your results alongside Office for National Statistics data later, adopt their band boundaries while you are still building the questionnaire. The article how to create a questionnaire walks through the order of steps.
Is a Likert scale ordinal or interval?
A Likert scale strictly produces ordinal data, because nobody can demonstrate that respondents perceive the gap between “agree” and “strongly agree” as being the same width as the gap between “neither” and “agree”.
In research practice such items are very often treated as numerical data anyway, means included. That is not sloppiness but a deliberate assumption, and it needs stating. It holds up better the more carefully the scale has been built.
The GESIS Survey Guidelines by Natalja Menold and Kathrin Bogner (2015) set out clear requirements: a rating scale should be balanced, with the same number of positive and negative categories, and the distances between categories should be equal, in other words equidistant. That equidistance is the condition under which numerical treatment becomes defensible at all.
The same guidance recommends five to seven categories and a verbal label on every single step rather than only on the two ends. Fully labelled scales raise reliability and validity and help respondents with less formal education in particular.
💡 Tip
Do not number the steps from minus 2 to plus 2. GESIS reports that respondents avoid the negative numeric range and give systematically more positive answers as a result.
How such a scale is built in detail is covered in the article on the Likert scale.
Which analysis does each level of measurement allow?
The level of measurement decides which summary statistics make sense. With nominal data you can count, with ordinal data you can additionally take the median, and with numerical data the full range is open to you.
Permissibility accumulates upwards: whatever is allowed at a lower level stays allowed at every level above it. The mode therefore fits at every level, the median from ordinal upwards, and the mean only from interval upwards.
| Level of measurement | Statistics that fit | What no longer holds |
|---|---|---|
| Nominal | frequency, mode | median, mean, ranking |
| Ordinal | mode, median, quartiles | mean, standard deviation |
| Interval | mean, standard deviation | claims such as twice as high |
| Ratio | every statistic, ratios included | no restriction |
The same split runs through significance testing. The University of Leeds guide puts it plainly: parametric tests require numerical data, whereas non-parametric tests are the ones for categorical data that is nominal or ordinal.
Statistical software will not check this for you. In SPSS the measurement level is set by hand in the variable view, and nothing stops anyone from averaging nominal codes. Which methods build on which level is set out in the article on evaluation methods in empirical research.
Common mistakes with levels of measurement
Most mistakes happen while reading the data rather than while calculating. Four patterns show up especially often, and all four come from the same source: a number in a dataset always looks the same, whatever it stands for.
All four are avoidable if you note the level of measurement once, while the questionnaire is still in front of you. Later on, the error usually surfaces only when the results have already been written up.
Treating labels as measurements
Postcodes, phone numbers, bus routes and student numbers are made of characters and digits but measure nothing. They are labels and therefore nominal. A mean of postcode digits produces a number, but no information about where your respondents live. Frequencies are the only thing that fits here.
Mixing grading scales from different countries
The GCSE scale from 9 down to 1, with 9 as the highest grade, applies to England. Wales and Northern Ireland use different scales, and in Germany and Austria the best mark is a 1 while a 5 is the worst. Put answers from several systems in one column and you invert the ranking for part of your sample. In cross-border surveys, collect the grade per country or convert it to a shared sequence first.
Averaging school grades
You often read that a grade average is an ordinary mean. Formally it is not: grades are ordinal, because the distance between a 4 and a 5 is not demonstrably the same as between an 8 and a 9. Institutions average them anyway because it is practical. In a dissertation that step belongs in the methods section as a stated assumption.
Collecting age in bands from the start
Age bands are quick to tick and look privacy-friendly. The price only shows up later: a mean age can then only be estimated, and anyone hoping to compare the results with another study runs into mismatched boundaries. Unless something speaks against it, ask for the year of birth or the number of years.
Conclusion
The level of measurement is not a formality for the methods section. It is the decision that fixes your analysis weeks in advance. Writing next to every question which level it will produce saves you from asking later why the one comparison you cared about is no longer possible.
Where to go next
- Looking for the right question format? Question types in a questionnaire
- Building answer options that hold up? Closed questions and their answer lists
- Curious about the level above ratio? The absolute scale with examples
Want to try the answer formats straight away?
With empirio.ai, an online survey tool from Germany, you pick between answer lists, rating scales and open number fields for each question and see immediately what kind of analysis comes out. Create a survey for free