Understand the four levels of measurement with survey examples, suitable statistics and tests, plus guidance on Likert scales and "don't know" answers.
The level of measurement tells you what information your response data contains: nominal data only distinguish categories, ordinal data also express an order, interval data have equal distances between values, and ratio data also have a true zero point. Each level supports additional calculations: frequencies for nominal data, the median for ordinal data, and the mean for interval data. Ratio data also support statements such as "twice as much" (Malhotra and Birks 2007, pp. 336-341).
This article answers five questions in order:
- What are the four levels of measurement, and what do they look like in a survey?
- Which statistics and which tests fit which level?
- How do you determine the level of measurement of a question?
- Is a Likert scale ordinal or interval?
- How do you design a rating scale with a midpoint, "don't know" and "not applicable"?
I keep seeing why this matters in studies, theses and company reports: someone calculates the mean of department numbers, reports an average of 3.4 on a satisfaction scale without showing the distribution, or counts "don't know" as the midpoint. None of these mistakes gets flagged by the software, because a numeric code doesn't protect you from a meaningless calculation. Before you calculate anything, ask what level of measurement your data have.
The four levels of measurement with survey examples
In survey research, measuring means assigning numbers to characteristics of people or objects according to a fixed rule. The rule has to stay the same for all respondents and over time (Marx 2026, slide 20). Which calculations make sense with these numbers later depends on what the rule represents. Textbooks usually distinguish four levels. By the way, the division into four levels goes back to the psychologist S. S. Stevens (cited in Malhotra and Birks 2007, p. 336 and p. 365, endnote 2). Interval and ratio scales are often grouped together as "metric" (Schumann 2019, p. 21).
Nominal scale: the number is just a label
On a nominal scale, the numbers only distinguish groups. If you code "Sales = 1, Production = 2, Administration = 3" in an employee survey, administration isn't "more" than sales. You could swap the codes at will without changing any statement. The only meaningful operation is counting (Malhotra and Birks 2007, p. 338).
Typical nominal questions in surveys: department, degree program, state or region, how someone heard about your product or service, yes/no questions. There's no such thing as an average department, just as there's no "average religion" (Schumann 2019, p. 27). For yes/no questions coded 0 and 1, the mean equals the share of yes answers. That doesn't make the question metric. It's simply a convenient way to calculate the proportion.
Ordinal scale: ordered categories without equal distances
On an ordinal scale, the numbers express an order. Whether rank 1 stands for more or less of a characteristic depends on how the question defines it. The scale doesn't tell you how large the difference between two ranks is (Malhotra and Birks 2007, p. 338). The example in Jacob et al. (2019, p. 34) is the school-leaving qualification: in Germany, the university-entrance qualification (Abitur) ranks above the intermediate certificate, but no one would say that people with an Abitur learned "three times as much" as people with the lowest secondary school certificate.
Typical ordinal questions: rankings ("Put these five features in order"), highest level of education, and questions that group an interval or ratio variable into categories, for example income "under 2,000 euros, 2,000 to 3,999 euros, ..." or age groups (Jacob et al. 2019, pp. 35 and 183). Methods in which respondents compare objects directly, such as paired comparisons or rankings, also only produce ordinal data (Marx 2026, slide 28).
Interval scale: equal distances, arbitrary zero point
On an interval scale, equal numeric distances also mean equal distances in the characteristic. The zero point, however, is set, not natural. That's why you can compare differences, but you can't form ratios (Malhotra and Birks 2007, p. 340).
The standard example is temperature in degrees Celsius. Water at 30 °C is 15 degrees warmer than water at 15 °C. It isn't twice as warm, because the zero point of the Celsius scale is the freezing point of water and is therefore set by convention. Only the Kelvin scale has an absolute zero point (Jacob et al. 2019, p. 34). Do the math:
15 °C = 288.15 K, 30 °C = 303.15 K, ratio 303.15 : 288.15 = about 1.05, not 2
In the social sciences, "true" interval scales are rare. Jacob et al. (2019, p. 35) give year of birth as an example and describe conventional calendar years as an uncontroversial case (ibid., n. 4). Where rating scales belong is disputed. That's the topic of the section Is a Likert scale ordinal or interval?.
Ratio scale: with a true zero point
A ratio scale has all the properties of an interval scale and, in addition, an absolute zero point. Zero here really means "none of it". That's why ratios make sense: someone who goes to the gym 4 times a month goes twice as often as someone with 2 visits (Malhotra and Birks 2007, p. 341).
Typical ratio questions in surveys: number of orders in the past year, spending in euros, length of service in years, commuting time in minutes, income in euros entered as a number rather than selected from a range. Malhotra and Birks (2007, p. 341) also include age, costs, sales and market share.
| Level of measurement | What the numbers mean | Survey example | Statement that's allowed |
|---|---|---|---|
| Level of measurementNominal | same or different | department, channel, yes/no | "42% work in sales." |
| Level of measurementOrdinal | plus: order | ranking, education, income bracket | "Half are in bracket 3 or below." |
| Level of measurementInterval | plus: equal distances | temperature in °C, year of birth | "The distance from 10 to 20 °C is as large as from 20 to 30 °C." |
| Level of measurementRatio | plus: true zero point | number of orders, spending in euros, minutes | "Group A orders twice as often as group B." |
Each level contains all the properties of the levels below it. That's what the staircase in diagram A shows: each step adds exactly one property.
Staircase with four steps. At the bottom, nominal: distinguish, example department, allowed frequency and mode. Above it, ordinal: order, example income bracket, plus median. Above that, interval: equal distances, example temperature in degrees Celsius, plus mean. At the top, ratio: true zero point, example number of orders, plus ratios such as twice as much.
Nominal
- New: distinguish
- Example: department, channel
- Allowed: frequency, mode
Ordinal
- New: order
- Example: ranking, income bracket
- Allowed: plus median
Interval
- New: equal distances
- Example: temperature in °C
- Allowed: plus mean
Ratio
- New: true zero point
- Example: number of orders, euros
- Allowed: plus ratios ("twice as much")
more information, more allowed calculations
Each level contains all the properties of the steps below it. Adapted from Malhotra and Birks 2007, pp. 336-341, and Marx 2026, slide 22.
What you can calculate at each level
The level of measurement determines which statistics can be meaningfully interpreted. The rule is cumulative: what's allowed at a lower level remains allowed at the higher ones. The following overview follows the table in my course (Marx 2026, slide 23) and in Malhotra and Birks (2007, pp. 338-341).
| Level of measurement | Central tendency | Dispersion and other statistics | Association and tests |
|---|---|---|---|
| Level of measurementNominal | mode (most frequent category), percentages | no measures of dispersion in the usual sense | chi-square test, binomial test |
| Level of measurementOrdinal | plus median | percentiles, quartiles | Spearman rank correlation, rank tests |
| Level of measurementInterval | plus arithmetic mean | range, standard deviation | Pearson correlation, t-test, regression |
| Level of measurementRatio | plus geometric and harmonic mean (for positive values only) | coefficient of variation | methods from the levels below, if their other assumptions are met |
Two sentences are worth remembering. First: the arithmetic mean can only be interpreted from the interval level upward (Schumann 2019, p. 27). Second: at the ordinal level, the mean is strictly speaking off limits, because the distances between the points can't be interpreted (Jacob et al. 2019, p. 34). The "strictly speaking" matters. How researchers handle this in practice is covered in the section on the Likert debate.
Worked example: similar means, different distributions
Let's look at an invented example from an employee survey. Two teams of 20 people each rate the statement "The new working-hours policy is an improvement for my team." The scale is balanced with a midpoint:
Strongly disagree (1) · Disagree (2) · Neither agree nor disagree (3) · Agree (4) · Strongly agree (5)
| Answer | Team A | Team B |
|---|---|---|
| Answer1 Strongly disagree | 6 | 0 |
| Answer2 Disagree | 2 | 2 |
| Answer3 Neither agree nor disagree | 0 | 10 |
| Answer4 Agree | 4 | 6 |
| Answer5 Strongly agree | 8 | 2 |
| AnswerMean | 3.30 | 3.40 |
| AnswerMedian | 4 | 3 |
| AnswerAgreement (4 and 5) | 60% | 40% |
| AnswerDisagreement (1 and 2) | 40% | 10% |
By the mean, team B is slightly ahead. By the median, team A is ahead. Both numbers are calculated correctly, and both miss the most important point: team A is split. Eight people reject the policy, twelve welcome it, and no one is undecided. In team B, half choose the midpoint. For management, these are two different situations. This question alone doesn't tell you why someone rejects the policy or chooses the midpoint. For that, you need a conversation or an open follow-up question.
The lesson: for a single rating question, report the distribution first, for example as a stacked bar or as the share of agreement and disagreement. The median and mean can supplement the distribution, but they can't replace it.
How to determine the level of measurement: three questions
The level of measurement isn't in the number, but in the question: what matters is what the answers represent in reality. The same characteristic can end up at different levels depending on how you ask. Three questions get you there in almost every case:
- Can the answers be ordered meaningfully? If not: nominal. Department, channel and favorite color have no natural order.
- Are the distances between the answers equal and meaningful? If not: ordinal. The distance between "lower secondary certificate" and "intermediate certificate" isn't the same as between "university-entrance qualification" and "master's degree".
- Is there a true zero point that means "none of it"? If not: interval. If so: ratio scale.
Decision tree. First question: can the answers be ordered? No gives nominal. Yes leads to the second question: are the distances equal and meaningful? No gives ordinal. Yes leads to the third question: is there a true zero point? No gives interval, yes gives ratio. Side note: single rating questions are strictly ordinal, but are often treated as interval data in practice.
The level of measurement depends on the question, not on the number in the dataset. If you group a quantity into categories, you end up with ordinal.
I keep seeing three traps.
Grouping values into categories lowers the level of measurement. If respondents enter their income in euros as a number, the data are ratio-level. If they select an income bracket, the data are ordinal, because the exact values are no longer known (Jacob et al. 2019, p. 35). That's often still the better choice, because people prefer giving a bracket to stating their exact salary. Decide deliberately, not out of habit.
Codes aren't values. Every survey tool stores answers as numbers, even for nominal questions. The fact that the export says "Sales = 1" doesn't make the department metric. Before every analysis, check whether a mean actually means something.
The indicator isn't the construct. Age recorded in years is measured on a ratio scale. But if you use it as an indicator of life experience or generation, what you actually want to measure is at best ordinal (Jacob et al. 2019, p. 36). So the question isn't only "How did I measure?", but also "What does the number stand for in my research question?".
Is a Likert scale ordinal or interval?
In short: a single Likert question is strictly speaking ordinal. Researchers have argued about this for decades, but in practice most people calculate the mean anyway. You can do that too, as long as you honestly show the median and the distribution alongside it. Both camps have good reasons.
The strict view. No one can show that the distance between "Disagree" and "Neither agree nor disagree" is exactly as large as between "Agree" and "Strongly agree". Jacob et al. (2019, p. 35) put it clearly: rating scales of all kinds don't reach the interval level, even with 5, 7, 10 or more points, and their level of measurement remains ordinal. If you follow this view, you report frequencies and the median and use rank-based methods.
The pragmatic view. In practice, the mean of rating questions is calculated very often anyway. It's a tradition, in company reports as well as in textbooks. In marketing research, as Malhotra and Birks note, "attitudinal data obtained from rating scales are often treated as interval data" (Malhotra and Birks 2007, p. 340). My course, too, lists attitudes, opinions and customer satisfaction in the row for the interval scale, with the note "quasi-metric" (Marx 2026, slide 22 and slide 23). That's the marketing research view. Simulation studies support it: Allerbeck (1978), Diehl and Kohr (2004) and Labovitz (1970) found that methods such as regression or analysis of variance don't lead to notable distortions with such data (cited in Jacob et al. 2019, p. 36). Norman (2010) summarizes studies going back to the 1930s and considers parametric methods robust to this violation of their assumptions, including with Likert data.
Reasons to use the mean. It sums up a question in one number that's easy to compare across groups and points in time. It takes all numerically coded answers into account, whereas the median of a five-point scale can only take a few values and often doesn't show small shifts. And it's the basis for widely used methods such as the t-test, regression or analysis of variance.
Reasons for caution. It assumes equal distances between the points, although the answer labels alone don't prove them. It hides the distribution: in the example above, team B is ahead by the mean, even though team A is split and has more agreement. It invites statements like "twice as satisfied" that ordinal data can't support. And it's sensitive to incorrectly coded opt-out answers, more on that in the section Rating scale.
Where the two views agree. Jacob et al. (2019, p. 36) also consider the pragmatic route acceptable, with one clear condition: you may calculate with the data pragmatically, as if they were quasi-interval, but you should only interpret the results ordinally:
A mean of 4.0 compared with 2.0 means "more agreement", not "twice as much agreement".
Single question or scale of several items?
Much of the dispute dissolves once you keep two things apart.
- A single rating question ("I'm satisfied with my work") has five or seven points. Here, the distribution is the most honest presentation. You can add the mean if you want to compare groups, but not as the only number.
- A Likert scale in the proper sense consists of several statements on the same topic, whose values are added up or averaged for each person. In the words of Malhotra and Birks, "the Likert scale is also referred to as a summated scale" (Malhotra and Birks 2007, p. 349). Such a sum or mean score has more possible values and is usually treated as interval data in practice. More possible values alone don't establish the interval level, though. The items must address the same underlying dimension, and reliability alone doesn't establish that. How to combine several items correctly is shown in the article Aggregating multi-item (Likert) scales, and how to check the scale in Reliability and validity.
My recommendation
You can calculate the mean, as most people do. But it only becomes honest once you show what it hides. For a study, a thesis or a report that has to stand up to critical readers, this is how I proceed:
- For single questions, always show the median and the distribution next to the mean, for example as a bar for each point or as the share of agreement. One sentence in the methods section justifies treating the data as metric.
- For multi-item scales, calculate a mean score for each respondent when combining the items is justified both conceptually and methodologically. Check the assumptions of the statistical methods before using them.
- Always phrase results ordinally: "higher", "lower", "more agreement", never "twice as satisfied".
- If you want or need to use only methods appropriate for ordinal data, for example because your supervisor, a reviewer or a journal takes the strict view, use methods for ordinal data: the median, rank correlation or a rank test that fits your design, see Rank tests with your data. If you also calculate the mean and both routes lead to the same conclusion, that strengthens your result. If they differ, report and explain the difference.
This gives you a defensible approach whichever position your readers take. More on designing single Likert questions can be found in A list of common Likert scales and in the article Likert scales explained.
What the level of measurement means for choosing a test
The level of measurement narrows down which methods are possible. The study design and the assumptions of each method also have to fit. The t-test compares means and therefore requires data treated as metric. The chi-square test checks frequencies in a crosstab and already works with nominal data (Malhotra and Birks 2007, p. 338). The following table assigns typical survey questions to methods. It's based on the overview in my course (Marx 2026, slide 23).
| Your question | Level of the variables | Suitable method |
|---|---|---|
| Your questionIs the preferred contact method related to the age group? | nominal × ordinal or nominal | Chi-square test with a crosstab |
| Your questionAre the ranks of two ratings related? | ordinal × ordinal | Spearman rank correlation |
| Your questionDo two independent groups differ in their responses to a single item when it is analyzed as ordinal data? | nominal (group) × ordinal | rank test, for example the Mann-Whitney U test |
| Your questionDo two independent groups differ in the mean of a scale? | nominal (group) × interval | t-test for independent samples |
| Your questionAre two metric quantities related? | interval × interval | Pearson correlation, regression |
Two notes on this. The chi-square test ignores the order of ordinal categories and treats them as nominal. And if you analyze a single Likert question with a t-test, the guidance in the Likert debate section applies: using the test is defensible, but interpret the results only in ordinal terms.
How to run rank tests with your data
QUESTIONSTAR calculates Spearman's rank correlation directly. Through the MCP interface, QUESTIONSTAR also calculates Kendall's tau. QUESTIONSTAR doesn't calculate rank tests such as the Mann-Whitney U test. For those, you have two options.
With a statistics program. Export the responses as an SPSS, CSV or XLSX file and run the test in SPSS, R or the free JASP. Check beforehand how opt-out answers are coded in the export, so that they don't enter the calculation as a scale value.
With an AI assistant. If QUESTIONSTAR is connected to an AI assistant through the MCP interface, for example in ChatGPT or Claude, the assistant can calculate the rank test from the responses to your survey. The assistant does the calculation, not QUESTIONSTAR. Questions you've marked as personal data in the survey aren't passed on to the assistant when the setting is switched on. Responses without this marking, especially free text, can still contain personal data. So check which questions the assistant will see, and verify the result in a statistics program.
Designing a rating scale: midpoint, "don't know" and "not applicable"
A rating scale lets respondents express an assessment by choosing from a predefined set of response options. The Likert agreement scale is the best known. To analyze the responses correctly, you have to keep three things apart: the midpoint of the scale, the answer "don't know" and the answer "not applicable". They look similar on screen, but they mean different things.
The midpoint: a position on the scale
Depending on its label, the midpoint ("Neither agree nor disagree", "Neutral") marks a mixed or neutral position on the scale. Why someone chooses this answer can't be seen from the answer alone. An odd number of points is appropriate if a neutral attitude is possible for at least some respondents. Use an even number without a midpoint if you want to force a decision or there's no neutral position (Malhotra and Birks 2007, p. 352). The midpoint, however, also attracts undecided respondents and people who are reluctant to reveal their opinion (Marx 2026, slide 60). I recommend balanced scales, that is, as many positive as negative points (Marx 2026, slide 59), with a neutral midpoint in between if a neutral attitude is possible. The article Likert scales explained covers the pros and cons of five- and seven-point scales and of scales without a midpoint. For an overview of the five design decisions, see Likert scales: five aspects to consider.
"Don't know": no opinion because knowledge is missing
"Don't know" or "can't judge" doesn't belong on the scale, but next to it. The answer means: I don't have a basis for a judgment. Without this option, people without an opinion often choose the midpoint, which distorts the mean and dispersion (Malhotra and Birks 2007, p. 353). Even more dangerous are random answers from people who have no attitude at all. They go unnoticed and weaken associations in the data (Schnell 2019, p. 47).
The flip side: a visible "don't know" option is also an easy way out for anyone who wants to avoid the effort of thinking (Jacob et al. 2019, p. 151). That's why it doesn't belong on every question.
What matters is whether someone can have the knowledge for a judgment. In my course, the rule is: use "don't know" for factual questions and knowledge questions, not when measuring attitudes and opinions (Marx 2026, slide 61). Factual questions also include ratings of processes that not everyone has experienced: only someone who has reported a problem knows whether the IT department solves problems quickly. Asked about their own attitude or experience, on the other hand, everyone can answer. For knowledge questions, "don't know" is even a substantive answer that you mustn't leave out (Jacob et al. 2019, p. 151).
| Question | Offer "can't judge"? | Why |
|---|---|---|
| Question"The IT department solves reported problems quickly." | yes | Someone who has never reported a problem can't know. |
| Question"How satisfied are you with how your complaint was handled?" (to all customers) | better: a filter question first | Only customers who have made a complaint should be asked this question. |
| Question"I'm satisfied with my work." | no | Everyone has an opinion about their own work. |
| Question"I feel treated fairly in my team." | no | It's about their own experience. |
Here's a question that keeps the opt-out answer clearly separate. The scale stays balanced with a midpoint, and "can't judge" is set apart next to it, only for this kind of question:
The IT department solves reported problems quickly.
Strongly disagree · Disagree · Neither agree nor disagree · Agree · Strongly agree
(set apart, only for questions that not everyone can judge) Can't judge
"Not applicable": the question doesn't concern me
"Not applicable" means: the question doesn't fit my situation. Someone who has never been to the repair shop can't rate it. Online, skip logic is usually the better solution: you ask first whether someone has used the repair shop and show the follow-up question only to those it concerns (Marx 2026, slide 61). How to plan filters is shown in the article Questionnaire design.
Schnell (2019, p. 46) advises separating four forms of missing substantive answers in the data: question not asked, not applicable, don't know and answer refused. Distinguishing these categories during data collection helps prevent artifacts in the analysis. A classic mistake is reading "no answer" to a frequency question as "zero times".
Worked example: what an incorrectly coded opt-out answer does
An invented example. 40 employees rate "The IT department solves reported problems quickly." 8 choose "can't judge". The 32 valid answers are distributed as follows: 2 × Strongly disagree, 4 × Disagree, 6 × Neither agree nor disagree, 12 × Agree, 8 × Strongly agree.
| How "can't judge" is treated | Mean | Share of agreement (4 and 5) |
|---|---|---|
| How "can't judge" is treatedcorrect: removed from the calculation (32 valid answers) | 3.6 | 62.5% (20 of 32) |
| How "can't judge" is treatedincorrect: counted as the midpoint (3) | 3.5 | 50.0% (20 of 40) |
| How "can't judge" is treatedincorrect: counted as the next code (6) | 4.1 | 50.0% (20 of 40) |
Depending on the coding, the same answers give a mean between 3.5 and 4.1. The third case happens easily when the opt-out answer simply gets the next numeric code in the export and no one removes it before the calculation. If you then also filter agreement as "4 and higher", the 8 opt-out answers count as agreement: 70% (28 of 40). So always report two things: the result from the valid answers and how many people chose "can't judge". Here, that's 8 of 40, or 20%. This number is a finding in itself: a fifth of respondents apparently had no basis for a judgment, for example because they've never reported a problem.
Levels of measurement and opt-out answers in QUESTIONSTAR
QUESTIONSTAR doesn't decide the level of measurement for you. The software stores what you set up as a question. These points help you put the decisions from this article into practice:
- Common scales: In the question editor, under "Common Scale Labels", you'll find ready-made balanced answer scales that you can apply with one click. The labels are listed in A list of common Likert scales.
- Opt-out answer: Set up "can't judge" as an opt-out answer, not as a point on the scale. In QUESTIONSTAR, an answer like that doesn't count in the mean. Check with a test response that it isn't in the denominator of your share of agreement either.
- Separate analysis: "Don't know" and "no answer" are analyzed separately in QUESTIONSTAR. That way, you can see later whether people couldn't judge or didn't want to answer.
- Skip logic: Instead of "not applicable", show follow-up questions only to those they concern.
- Analysis: Use the tables above to decide which statistic fits each question. For rank tests, export the data, see Rank tests with your data.
Checklist: clarify the level of measurement before the analysis
- Record the level of measurement for every question: nominal, ordinal, interval or ratio.
- Decide deliberately whether to collect age and income as numeric responses or as ordered categories.
- For single rating questions, always report the distribution and the median next to any mean.
- Explain in the methods section why you use means for rating questions, and describe the results in ordinal terms ("higher", not "twice as high").
- Use balanced rating scales, with a midpoint if a neutral attitude is possible.
- Offer "can't judge" only for questions that not everyone can judge, and set it apart from the scale.
- Replace "not applicable" with a filter question wherever possible.
- Exclude opt-out answers before the calculation and state their number in the report.
- Choose a test that fits the level of measurement and the study design, and add a rank test where a metric analysis needs justifying.
Frequently asked questions about levels of measurement
What is a level of measurement?
The level of measurement (also called scale level) describes which relationships between recorded values are meaningful: only equality (nominal), plus order (ordinal), plus equal distances (interval) or plus a true zero point (ratio). This determines which statistics and tests make sense (Malhotra and Birks 2007, pp. 336-341).
Is a Likert scale ordinal or interval?
A single Likert question is strictly speaking ordinal (Jacob et al. 2019, p. 35). Researchers have long disputed this, but in practice and in marketing research it's often analyzed like an interval scale anyway (Malhotra and Birks 2007, p. 340). This is defensible if you report the median and response distribution alongside the mean and interpret the results only in ordinal terms. A mean across several items of a scale is mostly treated as metric in practice.
Can I calculate the mean for ordinal data?
You can calculate it, but strictly speaking it isn't justified by measurement theory, because the distances between the points aren't known (Jacob et al. 2019, p. 34). For rating scales, it's still common in practice, out of tradition. It's easy to compare and is the basis for methods such as the t-test or regression, but it hides the distribution. The simulation studies that Jacob et al. (2019, p. 36) summarize support methods such as regression or analysis of variance with such data, but they aren't a blanket approval for every dataset. So report the median and the distribution alongside it and avoid statements like "twice as satisfied". If you want to use only methods appropriate for ordinal data, use the median and rank-based methods.
What is the difference between an interval and a ratio scale?
Both have equal distances. Only the ratio scale has a true zero point. That's why 30 °C isn't twice as warm as 15 °C (interval), but 4 orders are twice as many as 2 (ratio).
Is age nominal, ordinal, interval or ratio?
Age recorded in years is measured on a ratio scale. Age groups ("18 to 29", "30 to 44", ...) are ordinal, because they indicate an order without preserving exact ages (Jacob et al. 2019, p. 183).
Does "don't know" belong in the middle of the scale?
No. In rating questions, the midpoint marks a mixed or neutral position, whereas "don't know" means there's no basis for a judgment. Offer "don't know" or "can't judge" set apart next to the scale, and only for questions that not everyone can judge. Calculate the mean and the share of agreement from the valid scale answers and report the opt-out answers separately. For knowledge questions, on the other hand, "don't know" is a substantive answer (Jacob et al. 2019, p. 151).
Are there more than four levels of measurement?
Yes, the four primary types don't cover every case. There are intermediate forms, such as nominal scales with partial order (Malhotra and Birks 2007, p. 341). Schumann (2019, p. 21) also names the absolute scale for counts, such as the number of orders: here, the number itself already has a fixed meaning. For analyzing surveys, the four primary types are almost always enough.
Sources
Allerbeck, K. (1978). Cited in Jacob et al. (2019), p. 36.
Diehl, J. M. and Kohr, H.-U. (2004). Cited in Jacob et al. (2019), p. 36.
Jacob, R., Heinz, A. and Décieux, J. P. (2019). Umfrage. Einführung in die Methoden der Umfrageforschung [Surveys: an introduction to the methods of survey research] (4th ed.). Berlin/Boston: De Gruyter Oldenbourg. In German.
Labovitz, S. (1970). Cited in Jacob et al. (2019), p. 36.
Malhotra, N. K. and Birks, D. F. (2007). Marketing Research. An Applied Approach (3rd European edition). Harlow: Pearson Education.
Marx, P. (2026). Principles of Survey Research. Online course, chapter 2, slides 20, 22, 23, 28, 59, 60 and 61. QUESTIONSTAR. Course, chapter 2.
Norman, G. (2010). Likert scales, levels of measurement and the "laws" of statistics. Advances in Health Sciences Education, 15(5), 625-632. DOI 10.1007/s10459-010-9222-y.
Schnell, R. (2019). Survey-Interviews. Methoden standardisierter Befragungen [Survey interviews: methods of standardized surveys] (2nd ed.). Wiesbaden: Springer VS. In German. DOI 10.1007/978-3-531-19901-6.
Schumann, S. (2019). Repräsentative Umfrage. Praxisorientierte Einführung in empirische Methoden und statistische Analyseverfahren [Representative surveys: a practical introduction to empirical methods and statistical analysis] (7th ed.). Berlin/Boston: De Gruyter Oldenbourg. In German.
Stevens, S. S. (1951). Mathematics, measurement and psychophysics. In S. S. Stevens (ed.), Handbook of Experimental Psychology. New York: Wiley. Cited in Malhotra and Birks (2007), p. 336 and p. 365, endnote 2.
I hope every number in your analyses says what it's allowed to say, and not a word more.
Related articles:
- A list of common Likert scales
- Aggregating multi-item (Likert) scales
- The t-test: comparing means the methodologically clean way
Dr. Paul Marx
Author: Dr. Paul Marx
Dr. Paul Marx is the founder and Managing Director of QUESTIONSTAR and was a professor of marketing at the University of Siegen.