Questionnaire design

Guide

On this page

"Data are only as good as the instrument that collected them" – this is how Pallant (2020, chapter 1) opens a book on SPSS with a chapter on research design, and she explains why: the file that goes into the software comes from somewhere – a questionnaire, an interview, an observation – and a carelessly designed questionnaire will not yield data that answer the research question. This page follows that chapter and what Field (2024) writes about measurement (sections 1.5.2-1.5.4) and about questionnaire development (chapter 18), and adds what the site needs: how each question type becomes a column in the file, and which calculator will read it. Version 2 added two chapters from the Cambridge handbook of research methods in the social and behavioral sciences (Nichols & Edlund, 2024): Yan (2024) on how people answer a question – the four stages and what goes wrong in each – and Thomander and Krosnick (2024) on what a hundred years of experiments have shown about every design decision: open or closed, how many points, words or numbers, agreement statements, "don't know", the order of options and questions. Where this evidence settles a question Pallant left open, or contradicts accepted advice – the page says so.

1. Plan from the end: the analysis determines the question

Pallant's (2020, chapter 1) first rule is that before writing a questionnaire you need to know how the information will be used – that is, which statistical tests are planned – because the analysis determines the form of the question and the response format. Her example: someone planning a correlation between age and self-esteem needs exact age in years; someone who asked "up to 30 / 31-50 / over 50" got an ordinal variable with three levels, and a correlation is no longer possible. The same logic works for every question: write the research question (Research question and hypotheses), derive from it the variables and the level of measurement each one needs (Variables, measurement scales and research designs), choose the test – and only then word the question that will produce exactly that variable. Table 1 in section 5 takes the reverse route: from the question type to the column in the file and to the calculator.

More from Pallant's list for research design, as far as it concerns the questionnaire: recruit more subjects than needed ("people are unreliable – they don't turn up, get sick, drop out and don't fill in questionnaires properly"; plan pessimistically); think in advance about confounding variables and measure them if they cannot be controlled – that is, include them in the questionnaire; and have someone else go over the design before going into the field (Pallant, 2020, Chapter 1). How many subjects are needed – on the sample and sampling page.

2. An existing scale or your own items

A concept that cannot be measured directly – self-esteem, burnout, community resilience – is measured with a scale: several items that together "represent" the concept (Pallant, 2020, Chapter 1; Field, 2024, Section 18.2). Pallant: there are thousands of validated scales, and the place to start is the literature review in the field – what did other researchers use? Sometimes the items are printed in an appendix of the paper; otherwise, locate the original paper that described the development and validation of the scale. Some scales are copyrighted and an official copy must be purchased; a scale published in full in a paper is considered "public domain" and can be used without payment – but always with full citation of the source. For each scale, collect the reliability and validity data that were reported, because they go into the method section.

Even so, a scale shown to be reliable in one population can be "completely unreliable" in another – Pallant gives English-speaking adults versus children who are not English speakers – and so even an excellent scale goes through a pilot in the intended sample (Pallant, 2020, Chapter 1). Field (2024, section 18.9.1) puts this as a principle: reliability is not a property of the scale but of the scale, the people who respond to it and the context in which it is administered. A scale translated into Hebrew is a special case of the same principle – the translation is a new instrument, and its reliability is computed anew (on the alpha page).

Your own scale. Field's example (2024, section 18.5) of questionnaire development: he interviewed anxious and non-anxious students about SPSS, wrote 23 candidate items from the interviews, each a statement with a five-point Likert scale – and then checked with factor analysis which items cluster into which dimensions (which measure one thing, and which can be dropped) and with reliability analysis how consistent they are. That is the route: more items than needed, a large sample (Field, section 18.5.3: the "10-15 subjects per item" rule has no empirical basis; what matters is the total sample size – 300 is considered good, and with high loadings fewer is enough), factor analysis (the principles on the unsupervised learning page; for a structure not known in advance – exploratory factor analysis; for testing a structure set in advance – confirmatory factor analysis) and Cronbach's alpha.

3. Reliability and validity – what is reported about the scale

The two properties every scale needs, according to both sources: reliability – how free it is from random error, or: does it give the same result under the same conditions; and validity – does it measure what it is supposed to measure (Pallant, 2020, Chapter 1; Field, 2024, Section 1.5.4). Field adds the order: validity is a necessary but not sufficient condition, and to be valid the instrument must first be reliable. And why this is pressing precisely in a questionnaire: self-report (self-report) suffers from greater measurement error than direct measurement, because the answer is also influenced by things that are not the concept being measured – would you admit in a questionnaire that you shoplifted? (Field, 2024, Section 1.5.3).

Reliability – the two common measures (Pallant, 2020, Chapter 1): test-retest reliability (stability over time: the same people, two time points, the correlation between them; for a mood scale a low correlation is expected, for a personality trait – a high one) and internal consistency – the extent to which the items "hang together", usually Cronbach's alpha, which expresses the average correlation among all the items. The accepted cutoff is .7 (Nunnally, 1978, as cited in Pallant), but alpha depends on the number of items: in a short scale (fewer than ten items) it can be low even when the items are consistent, and then the mean inter-item correlation is reported – optimal range .2 to .4 (Briggs & Cheek, 1986, as cited in Pallant). Field (2024, section 18.9.4) adds a caveat: .7-.8 is "bandied around" everywhere as acceptable, but alpha is affected by the number of items and by things that are not reliability, so it is interpreted in the context of the measurement and the study and not by a blanket cutoff. Validity – there is no single measure; the literature speaks of content validity (the items cover the domain of the concept), criterion validity (a relationship with a measurable criterion – concurrent or predictive) and construct validity (relationships with other concepts, close ones – convergent – and distant ones – discriminant), and validation is the gathering of empirical evidence, not a formula (Pallant, 2020, Chapter 1; Field, 2024, Section 1.5.4). In more depth – on the concepts page; the computation – in Cronbach's alpha and scale score.

4. Question types

A closed question offers a list of defined answers – yes/no, or several options of which one is marked. Pallant (2020, chapter 1) illustrates with the education question (six options numbered 1-6) and notes the advantage: coding into numbers is almost automatic – no = 0, yes = 1; male = 1, female = 2; for education, type the number of the option marked. The small piece of advice that saves the most: number the options on the questionnaire itself, and set a convention for the numbering order (for example across and then down) and stick to it throughout the questionnaire.

An open question – when it is impossible to guess all the answers in advance ("What is the main source of stress in your life at the moment?"). Respondents are free, but the answers are not a number: after collection, go through them, identify recurring categories (work, money, relationships, health), give each category a code – and "other" for everything else – and type in the code (Pallant, 2020, Chapters 1-2). A combination – a closed list plus "other, please specify" with a line to write on – is especially useful in the early stages of research, because it shows whether the categories offered cover what people want to answer (Pallant, 2020, Chapter 1).

What the evidence says about the choice. Thomander and Krosnick (2024) summarize a hundred years of experiments, and two conclusions run counter to intuition. The first: for quantities – an open question. To choose a range ("2-5 times") the respondent first answers with a number in their head anyway, and the set of ranges itself pushes toward the middle, because respondents assume that the middle range is the typical one (Schwarz et al., 1985, as cited in Thomander & Krosnick). The second: for a categorical question whose answer space is unknown – also open; "other" does not solve it, because respondents limit themselves to the options offered to them, and a closed list is suitable only when it is exhaustive – which usually requires a broad pilot with the open version. Examples inside the question ("for example: German, Polish…") focus attention on the examples, and it is better without them. And "check all that apply": a series of separate, balanced questions ("Have you read this newspaper, or not?") gives higher reliability and deeper processing (Dolnicar et al., 2012; Smyth et al., 2008, as cited in Thomander & Krosnick) – and in the file it is exactly the same thing: a column for each option (Table 1). Rating or ranking. When life forces a choice – one candidate, one value to raise children toward – asking to rank the items (ranking) gives better data than rating each item separately, because in a battery of ratings some respondents mark the same point for all; when one can like two things equally – rate, and do whatever is possible to make the questions easy and the respondents engaged (Thomander & Krosnick, 2024).

5. Response format – and the Likert scale

The response format is the decision on which the analysis depends most (Pallant, 2020, Chapter 1). Three of Pallant's rules: ask for an actual value when there is one (age in years, not an age category – correlation requires continuous scores with a wide range); give as wide a range of answers as possible, because "you can always group later" – but not the reverse; and instead of "agree / disagree" – a Likert-type scale, for example from "strongly disagree" (1) to "strongly agree" (6), which gives a wider range of scores and expands the possible analyses. The number of points has to be decided (Pallant refers to DeVellis, 2012, for a discussion of the advantages and disadvantages); Jamieson (2004) notes that five points are the norm, and there are arguments for seven and for an even number of points (no midpoint) – and this question is settled by the evidence later in this section. What is not in dispute: clear instructions – tick a box? circle a number? – with an example, because for some respondents this is the first questionnaire of their lives (Pallant, 2020, Chapter 1).

What one question tells and what it misses. Pallant's example: "Do you smoke? Yes / No". In the pilot someone will ask whether it means cigarettes, cigars or cannabis; and then – is it enough to know whether, or do you also need how much (two or three cigarettes or two or three packs) and how often (every day or on occasions)? For every question, ask what information it will give and what information will be missing (Pallant, 2020, Chapter 1).

How many points, and how to label them – the evidence. Thomander and Krosnick (2024): reliability and validity increase with the number of points, but the gain keeps shrinking, and the conclusion from the literature is five points for a unipolar scale (zero at the end: from "not at all" to "extremely") and seven for a bipolar scale (zero in the middle: from "like a great deal" to "dislike a great deal", with "neither like nor dislike" in the middle). Do not remove the midpoint: a middle option does not encourage lazy answers; some respondents genuinely belong in the middle, and without it they choose a neighboring point at random; omitting it does not improve a unipolar scale and harms a bipolar one. A "slider" scale (slider) – avoid it: the rating is pulled toward the slider's starting position, and more respondents break off. And verbal labels on all the points, without numbers: higher reliability and validity than with words at the ends only, respondents prefer it, and it is also faster; numbers next to words do harm when the meanings conflict. Labels found to be roughly equally spaced – for unipolar "not at all / slightly / moderately / very / extremely"; for bipolar "extremely bad / moderately bad / slightly bad / neither good nor bad / slightly good / moderately good / extremely good"; for quality in five "very poor / poor / fair / good / excellent" (the site's renderings; in the original – extremely, very, moderately, slightly, not at all). And a bipolar question is answered better in two steps – "for or against?" and then "strongly, moderately or slightly?" – seven categories in two short questions instead of seven options at once (Thomander & Krosnick, 2024). The implication for the file: the number (1-5, 1-7) is the code in the codebook, not what the respondent sees – in the questionnaire, printed or online, show the words; and an existing validated scale is administered in the format in which it was validated (section 2). "Agree / disagree" statements are a bipolar scale – and on their problem, section 7.

Frequency and quantity – not on an "often" scale. Dimensions without a natural metric – importance, liking, friendliness – require a rating scale; dimensions with a natural metric – frequency, quantity, time – are not measured with "vague quantifiers" ("quite a lot", "fairly often") that everyone understands differently, but with an open question in the natural unit: 5 days a week, 3 hours (Thomander & Krosnick, 2024) – Pallant's rule about age in years, extended. Yan (2024) shows why a list of ranges is not a solution either: respondents assume that the questionnaire's author chose the ranges carefully, and so read from them what is "normal" and place themselves relative to it. When people were asked how much television they watch a day, a scale ending with "more than two and a half hours" yielded lower reports than a scale beginning with "up to two and a half hours"; and those who marked two and a half hours at the top end of the low scale also reported afterwards that television takes up a large part of their leisure – and those who marked the same value at the bottom of the high scale, the opposite (Schwarz et al., 1985, as cited in Yan). And more from Yan: "How many times…" encourages counting episodes (and tends to underestimate); "How often…" encourages generalizing from a rate ("twice a year", and tends to overestimate) – when a count is wanted, ask for a time frame, and mention records if there are any.

Likert: ordinal or interval? Here the literature is divided, and the site takes a position. Field (2024, Box 1.2) holds that any subjective rating – preference, confidence, understanding – should be treated as an ordinal variable, "although many scientists do not do so": one cannot be sure that the distance between 1 and 2 equals the distance between 4 and 5. Jamieson (2004) is stricter: a Likert scale is an ordinal level of measurement, and therefore median or mode and not mean, and nonparametric tests (chi-square, Spearman, Mann-Whitney) and not parametric ones – and she points out that the distributions of Likert items are often skewed or polarized, and that this question must be decided already at the design stage. Norman (2010) answers her with evidence on robustness: in a sample of 93 respondents on 64 pairs of rating items, the correlation between Pearson and Spearman was .99, and even when he collapsed the scale to four highly skewed points – .987; parametric tests are robust to non-normality and to ordinality; and he also brings the argument of Carifio and Perla (2008, as cited in Norman) – a single Likert item is perhaps ordinal, but a Likert scale, a sum of many items, is interval, just like a score on a multiple-choice test that is a sum of correct answers – though in his view the decision rests not on this argument but on the robustness evidence. The site's convention, as set on the variables and measurement scales page, is "both": a single rating item is an ordinal variable (frequencies, median, Spearman), a multi-item scale score is an interval variable (mean, Pearson, t, regression) – and in both cases report what was done. The implication for design: anyone who wants means, t-tests and regression on a concept builds a scale of several items for it, not a single question; and anyone who has only one question reports it as frequencies and a median.

6. How people answer a question – and why this determines the wording

The two chapters of the Cambridge handbook start from the same model: to answer a question the respondent goes through four stages – comprehension (what is being asked, and what is meant), retrieval from memory, judgment (combining what was retrieved into one answer, or an estimate) and mapping the answer onto one of the options (Yan, 2024; Thomander & Krosnick, 2024). Each stage requires effort, and not every respondent is willing to invest it. One who does all of them properly is "optimizing" (optimizing); one who cuts corners – mildly (all the stages, carelessly) or strongly (skipping retrieval and judgment and choosing an answer that seems reasonable, or just one that is easy to defend) – is "satisficing" (satisficing; Krosnick, 1991, as cited in Thomander & Krosnick). Three things determine how many will satisfice: the difficulty of the task, the respondent's ability and motivation – and motivation declines with fatigue, as the questionnaire grows longer. The symptoms are familiar to anyone who has analyzed a questionnaire: "don't know" instead of thinking; straightlining (straightlining) – the same point for a whole battery of items; agreeing with every statement; the extremes; the middle (Yan, 2024). And respondents also read from the context: they assume that the questionnaire's author is a cooperative conversational partner – that every question is relevant, that the options were chosen carefully, that the previous question is related to the next one – and answer according to what they infer was meant (Yan, 2024, on Grice's conversational maxims). Hence the working rule of both chapters: every design decision is tested by whether it reduces difficulty and increases motivation – and whether it respects the conversational maxims.

What goes wrong at each stage. In comprehension: long questions, with several clauses or in the negative, take longer and are burdensome; words with many meanings; and even a simple word – when asked "Have you smoked at least 100 cigarettes in your entire life?", 54% understood "smoking" as any puff and 46% only as inhaling (Suessbrick et al., 2000, as cited in Yan) – and a definition inside the question dramatically improves accuracy for vague concepts. In retrieval: people forget – especially routine and frequent events – and shift things in time (telescoping: an event that occurred outside the period is reported as if within it); and what was not encoded in memory under the same category (walking the dog is not "physical activity") will not be retrieved. In judgment: counting up from the bottom underestimates, generalizing from the top overestimates, and an impression ("I never have dental problems") – the worst of all; time, records, an anchoring event marking the start of the period, and a diary help. In mapping: the respondent edits the answer – in sensitive questions (section 8), or to get rid of follow-up questions: someone who answers "no" to a filter question spares themselves everything after it, so ask all the filter questions together and only then the follow-up questions; and the order of the options has an effect – in a written list people choose the first ones more (primacy), in a list read aloud – the last ones (Yan, 2024; Thomander & Krosnick, 2024). In a categorical list with no natural order, in an online questionnaire, the options can be presented in random order to each respondent – unless the order itself carries meaning (Thomander & Krosnick, 2024). Yan's first recommendation comes even before all this: to know exactly which concept is being measured – "unemployed" is both someone who has no job and someone who wants one, and a question on "Are you working?" alone measures only half the concept – and the research question page starts there.

7. Wording

"There is a real art to writing clear items", writes Pallant (2020, chapter 1) – there are no unequivocal rules, but there is a list of things to avoid, and every item on it is familiar to anyone who has ever analyzed someone else's questionnaire:

Agreeing with everything. Thomander and Krosnick's (2024) best-supported recommendation is also the hardest to apply: "agree / disagree", "true / false" and "yes / no" questions suffer from acquiescence bias (acquiescence) – the tendency to endorse statements regardless of their content. More people agree with "I like pineapple" than disagree with "I don't like pineapple", although the questions are supposed to be mirror images; the reasons – satisficing, politeness, avoiding confrontation, deference to the person in authority. Instead of a statement – a question that mentions the concept being measured explicitly, with the two options balanced: not "I am satisfied with the service – agree / disagree", but "How satisfied or dissatisfied are you with the service?" with a scale from "not at all satisfied" to "very satisfied". The site holds two positions here: a validated scale built from agreement statements – use it as it was validated, and its reversed items (section 9) are the familiar protection against response bias (Field, 2024, Section 18.9.4); new items of your own – word them as a question about the concept, not as a statement.

"Don't know" and "not applicable" – when yes and when no. Pallant (2020, chapter 1) recommends adding such a category when called for – otherwise someone who does not know will choose the middle of the scale, and that will be read as a position. Thomander and Krosnick (2024) show that in attitude questions choosing "don't know" usually reflects not an absence of opinion but satisficing, ambiguity of the question, ambivalence or protecting one's self-image – and therefore recommend omitting it from attitude questions, and asking afterwards about the strength of the attitude. Both: "not applicable" for a factual question that does not apply to everyone (with a filter question before it) – yes; "don't know" in an attitude question – usually not. And in the file, such a code is not a missing value but an answer: document it in the codebook, and exclude it before a mean or correlation (see missing values and the standard template). Question length. Not the shortest – the shortest that conveys the idea uniformly to everyone: unambiguous, familiar, short words; as few sentences as possible; and a definition inside the question when the concept is vague (Thomander & Krosnick, 2024; Yan, 2024).

8. Question order, questionnaire length and sensitive questions

Order. The good news: in analyses of hundreds of experiments, question-order effects are rare – 11-12 out of more than 500 variables in Smith (1988), 8 out of 113 experiments in Schuman and Presser (1981; both as cited in Thomander & Krosnick, 2024). When they occur, it is because of the conversational maxims: the previous question explains the next one. Yan's (2024) example: satisfaction with marriage and then with life – the correlation is high, because the marriage is "fresh" in memory; with a joint introduction ("two questions about your life – the first about your marriage, the second about life in general") the correlation drops, because the respondent understands that the general question is about "the rest of life" (Strack et al., 1988; Schwarz et al., 1991, as cited in Yan). The cumulative recommendations: general before specific on the same topic; questions on the same topic together; open with easy, pleasant questions; filter questions so as not to ask what does not apply; and sensitive and demographic questions – at the end (Thomander & Krosnick, 2024; Yan, 2024). Length. Fatigue accumulates, and satisficing with it – so as few questions as possible, only what serves a research question (section 1). Thomander and Krosnick mention a growing literature showing that single-item measurement can work well even for concepts customarily measured with a battery; the site's convention (section 5) remains: a multi-item scale score – interval; a single item – ordinal, reported as frequencies and median.

Sensitive questions and social desirability. A question is sensitive in three senses (Yan, 2024): private – beyond the bounds of ordinary conversation (income, sexuality); risky – identifying information whose disclosure to a third party is dangerous; or one that has a socially "correct" answer – and then people overreport the desirable (voting) and underreport the undesirable (heavy drinking). Thomander and Krosnick (2024) add a caveat: a considerable part of what looks like false reporting is actually nonresponse – those who did not vote also respond less to surveys – and most of the "sophisticated" techniques have not been proven. What is documented: self-administration without an interviewer reduces the bias, so an online or printed questionnaire is better than an interview for sensitive questions (Yan, 2024); anonymity increases reporting of the undesirable but also lowers effort and accuracy; "forgiving" wording ("many people did not manage to vote…") increases reporting, but it has not been shown to increase validity (Thomander & Krosnick, 2024). And what is not a technique: informed consent that states in advance what will be asked, and whom to turn to if the questions caused distress – on the research ethics page.

9. Reversed items

In a scale of statements it is customary to word some of the items in the opposite direction – in Field's SPSS anxiety questionnaire (2024, section 18.9.4), 22 items for which agreement means anxiety ("Statistics makes me cry") and one reversed item ("Standard deviations excite me"). Reversed items matter because they reduce response bias – respondents must read in order to answer – but they have a price that must be paid before the analysis: a reversed item correlates negatively with the other items, lowers Cronbach's alpha and can even make it negative. So its score is reversed before the reliability analysis: the maximum of the scale plus one, minus the value (on a 1-5 scale: 6 − the value) – 5 becomes 1, 1 becomes 5, 3 stays 3 (Field, 2024, Section 18.9.4). On the site: enter the item as answered, write "reverse" in the notes column of the codebook, and the alpha page reverses it by itself; a negative item-scale correlation in the output is almost always a reversed item that was not marked.

10. Pilot

Before distribution, Pallant (2020, chapter 1) requires a pilot – and on people of the same kind who will answer in the study itself (adolescents, the unemployed, prisoners): to make sure that the instructions, the questions and the scale items are clear, that respondents understand and answer properly, and to identify questions that might offend. This is also where it turns out that "Do you smoke?" is not one question (section 5), and that a scale that worked in another culture does not work here (section 2) – alpha can already be computed on the pilot data on the alpha page. And in an experiment – a full dress rehearsal of the manipulation and the measures, interviewer training, and an interrater reliability check when several people code behavior (Pallant, 2020, chapter 1; the calculator: intercoder reliability). "Iron out the sources of confusion before distributing hundreds of questionnaires."

Yan (2024) extends the toolbox by stage: focus groups at the conceptualization stage – how people talk about the topic, and in what words; expert review and cognitive interviews – the respondent thinks aloud while answering, and the researcher hears what they understood and how they chose – before the field; and during data collection – response times (an online questionnaire records them). The methods do not always point to the same problems, but several methods are better than one, and "any evaluation method is better than none" (Yan, 2024); Thomander and Krosnick (2024) call this "cognitive pretesting", and it is their way of measuring whether the question conveys the idea uniformly (section 7).

11. From the questionnaire to the file

The codebook (codebook) is prepared before data entry, not after it: for each variable a short, unique name, a description, how the answers are coded and the level of measurement; and for each questionnaire – an identification number written on it, so that it can be returned to when an error is found (Pallant, 2020, Chapter 2). On the site the codebook is the "legend" sheet in the Excel file, and the file structure – a row per subject, a column per variable – is described in the standard data template, with an empty template to download and a checker that reports what each calculator will recognize. Two shortcuts: anyone whose data are already in SPSS generates the template from the file in From SPSS to Excel – and if the questionnaire is written in Word in tables whose first cell holds the question code (Q1, Q2, D3), the page also fills in the full question wording and the filter instructions from the questionnaire; and anyone entering data by hand – numbers the options on the questionnaire, and types the number.

Common mistakes

Asking in categories what can be asked as a number. Age, seniority, income, hours – in categories they are ordinal variables, and the correlation and the mean are lost (Pallant, 2020, Chapter 1).

Computing the mean of a single question and reporting it as if it were a scale score. A single item is an ordinal variable (Field, 2024, Box 1.2; Jamieson, 2004) – frequencies and median; mean and t-tests – for a multi-item scale score (Norman, 2010).

Writing your own scale when a validated one exists. Literature review first; and if you still do – many candidate items, a large sample, factor analysis and alpha (Pallant, 2020, Chapter 1; Field, 2024, Section 18.5).

Reporting the alpha from the original paper instead of computing it in your sample. Reliability belongs to the scale in this population, in this context (Field, 2024, Section 18.9.1; Pallant, 2020, Chapter 1).

Double-barreled question, double negative, leading question. Table 2.

Distributing without a pilot. Everything not caught in the pilot is caught in the analysis – when it can no longer be fixed (Pallant, 2020, Chapter 1).

Forgetting the reversed items. Low or negative alpha, a negative item-scale correlation (Field, 2024, Section 18.9.4); "reverse" in the codebook solves it.

"Don't know" as a missing value, or as the middle of the scale. A separate, documented code, excluded before a mean (Pallant, 2020, chapter 1 – the recommendation to add the category; the handling in the file – the site's convention) – and in attitude questions, usually without it (Thomander & Krosnick, 2024).

"Agree / disagree" statements for new items. Acquiescence bias; a question about the concept, with balanced options (Thomander & Krosnick, 2024).

Frequency as "often / sometimes / rarely". Everyone understands it differently, and the scale itself teaches what is "normal"; a number in a natural unit (Thomander & Krosnick, 2024; Yan, 2024).

Numbers instead of words on the scale, a scale without a midpoint, or a slider. Words on all the points, 5 or 7 with a midpoint (Thomander & Krosnick, 2024).

Filter questions interspersed among the follow-up questions. Someone who realizes that "no" saves questions – answers "no"; all the filter questions first (Yan, 2024).

Checklist – before distribution

References

Field, A. (2024). Discovering statistics using IBM SPSS statistics (6th ed.). SAGE.

Jamieson, S. (2004). Likert scales: How to (ab)use them. Medical Education, 38(12), 1217-1218. https://doi.org/10.1111/j.1365-2929.2004.02012.x

Norman, G. (2010). Likert scales, levels of measurement and the "laws" of statistics. Advances in Health Sciences Education, 15(5), 625-632. https://doi.org/10.1007/s10459-010-9222-y

Pallant, J. (2020). SPSS survival manual: A step by step guide to data analysis using IBM SPSS (7th ed.). Routledge.

Thomander, S. D., & Krosnick, J. A. (2024). Question and questionnaire design. In A. L. Nichols & J. E. Edlund (Eds.), The Cambridge handbook of research methods and statistics for the social and behavioral sciences: Vol. 2. Performing research (pp. 352-370). Cambridge University Press. https://doi.org/10.1017/9781009000796.017

Yan, T. (2024). Self-report measures. In A. L. Nichols & J. E. Edlund (Eds.), The Cambridge handbook of research methods and statistics for the social and behavioral sciences: Vol. 2. Performing research (pp. 329-351). Cambridge University Press. https://doi.org/10.1017/9781009000796.016

Nunnally (1978), Briggs and Cheek (1986), DeVellis (2012), De Vaus (2014), Kline (1999, 2005) and Streiner and Norman (2015) – as cited in Pallant (2020); Pedhazur and Schmelkin (1991), Comrey and Lee (1992) and Guadagnoli and Velicer (1988) – as cited in Field (2024); Carifio and Perla (2008) – as cited in Norman (2010); Krosnick (1991), Schwarz et al. (1985), Dolnicar et al. (2012), Smyth et al. (2008), Smith (1988) and Schuman and Presser (1981) – as cited in Thomander and Krosnick (2024); Suessbrick et al. (2000), Strack et al. (1988), Schwarz et al. (1985, 1991) and Grice – as cited in Yan (2024). Following the site's rule, the reference list includes only sources read in the original. Field (2024) and Pallant (2020) in a reflowable digital copy – cited by section and chapter; the two papers and the two chapters of the Cambridge handbook – by the whole item, without page numbers in the body text.