Response bias: how answers get tilted, and what to change

Your results came back cleaner than the thing they describe. This is the vocabulary for what happened, the size it reaches when somebody measures it, and the handful of changes that shrink it.

Michael Hodge Survey methodologist Updated: 11 September 2026

The short answer

Response bias is a systematic tilt in how people answer. They do answer, which is what separates it from nonresponse bias. But the answers lean somewhere predictable: too positive, too agreeable, too close to the middle, or toward whatever seems the acceptable thing to say out loud.

The tilt survives a bigger sample. More replies shrink random error and leave the lean exactly where it was, so you finish with a precise estimate of the wrong number.

It is measurable. Pew Research Center put the same 60 questions to 3,003 US adults, randomly assigned half to telephone and half to the web. On the phone, 62 per cent said they were very satisfied with their family life. On the web, 44 per cent.⁵

What you are seeing, and what to change first
What you see	Likely mechanism	First action
Nearly everything scores 4 or 5	Acquiescence: agree-or-disagree statements invite agreement.	Ask about the thing instead of asking for agreement. Rewrite the items.
The middle option does most of the work	Neutral response bias: an unclear item, no stake in the answer, or a question that does not apply.	Split "does not apply" out before touching the midpoint. Which to change.
Complaints are milder in person than by email	Courtesy bias: the person or place doing the asking.	Ask away from the premises and without an interviewer. How big the gap gets.
A trend moved after the questionnaire was reordered	Question order bias: the previous question changed what came to mind.	Fix the order and randomise only inside blocks. Set the order.
Sensitive items skipped or implausibly clean	Social desirability: the survey feels identifiable.	Say exactly what anonymous means, then drop the fields that break it. The six changes.
Identical answers down a whole grid	Habituation: a long run of statements that all read alike.	Break up the grid next wave. Check the rows now, do not delete them. The checks.

Response bias, response set and lie scale

Four words get used for overlapping things, and most of the confusion on this topic lives in the gaps between them. Here is what each one actually denotes.

Four terms, kept apart
Term	What it means	The tell
Response bias	The umbrella. Any systematic tendency for answers to depart from the truth in a predictable direction, whatever causes it.	The whole distribution has moved, not just a few rows.
Response set	Cronbach's 1946 term for the subset driven by the shape of the item rather than its content: "a habit or temporary disposition that causes a person to respond to test items differently than he would if the same content were presented in a different form".¹	Rewrite the item in a different format and the same person answers differently.
Social desirability	Answering to look acceptable rather than to be accurate. A cause of response bias, not a synonym for it.	The gap widens on sensitive items and narrows on neutral ones.
Deliberate deception	A conscious decision to misreport. Rare in ordinary business surveys, common in high-stakes assessment.	Answers are internally consistent but contradict a record.

A response set is a kind of response bias. Social desirability is a motive that produces one. Deliberate deception is a choice, and the only one of the four that requires intent.

What a lie scale is, and what a validity scale is

A lie scale is a small set of items scattered through a questionnaire, describing behaviour that is creditable but very rare, or discreditable but close to universal. Agree with enough of the first kind and deny enough of the second and you have described a person too clean to exist. The best-known standalone version is the Marlowe-Crowne social desirability scale of 1960.² A validity scale is the wider family the lie scale belongs to: any indicator built into a test to say something about how the test was taken rather than about the trait it measures.

Two things follow, and the second one is uncomfortable. A lie scale is a detector, so it leaves the questions alone and adds a check. Rewriting or dropping the items that invite a slanted answer is the opposite move: it changes the instrument and adds no check.

What a lie scale cannot do is tell you that a particular person lied. A high score says the answer pattern resembles the pattern of people presenting themselves favourably. Some people genuinely are that tidy, some read the items literally, and some were careless. Treat the score as one quality indicator among several, read it alongside completion time, straightlining and any record you hold, and never drop an individual on the lie score alone. Check, do not automatically delete.

Where the evidence is thin

Whether bias indicators earn their place in applied settings is disputed. McGrath and colleagues reviewed the literature and found only 41 studies that had tested whether bias indicators moderate or suppress the validity of other measures. In personality assessment and workplace measurement there were enough to judge, and the support was weak. They concluded that after roughly a century of research "a sufficient justification for the use of bias indicators in applied settings remains elusive".³

Seven authors replied in the same journal under the title "A misleading review of response bias", arguing the inclusion criteria were too narrow and that the evidence in clinical neuropsychology is strong.⁴ Both are worth knowing about before anyone sells you a lie scale for an employee survey.

The types of response bias, with examples

Nine patterns cover almost everything you will meet. Each one has a different trigger, so each one has a different fix.

Nine patterns and their triggers
Type	What it looks like	What sets it off
Social desirability bias	Creditable behaviour over-reported, awkward behaviour under-reported.	Sensitive topics, an identifiable survey, a person listening.
Courtesy bias	Complaints softened to avoid giving offence, especially to whoever is standing there.	Being asked on the provider's premises, or by the provider.
Acquiescence bias (yea-saying)	Agreement across items that contradict one another.	Agree-or-disagree grids, low effort, an authority asking.
Extreme responding (and positive response bias)	Only the endpoints get used and the distribution piles on 1 and 5. When the pile sits at the favourable end it gets called positive response bias.	Vague items, strong feelings, some cultural response styles.
Neutral response bias (central tendency)	The middle box does most of the work and variance collapses.	An unclear question, no stake in the answer, fear of consequences.
Habituation (straightlining)	The same option down a grid because the items all read alike.	Long batteries of similarly worded statements.
Demand characteristics	The respondent works out the wanted answer and supplies it.	A leading introduction, or the person who ran the thing handing out the form.
Question order bias	The same question gets a different answer depending on what came before it.	A preceding question that makes one consideration salient.
Recall bias	Estimates drift, and round numbers appear far too often.	Long recall windows, and events nobody counts as they happen.

Acquiescence, extreme responding and neutral responding are the three classic response sets: all three are driven by the form of the item rather than by its subject.

Response bias examples: what it did to real results

Response bias is easy to assert and hard to size, because you rarely have the truth to compare against. Three studies had something to compare against.

The same question, asked two ways, 18 points apart

Share answering "very satisfied". Panelists were randomly assigned to answer by telephone with an interviewer (1,494 people) or on the web with nobody listening (1,509). Across all 60 questions the average gap was 5.5 points and the largest was 18.

Satisfaction with family life and with social life, by mode of interview. Source: Pew Research Center (2015), From Telephone to the Web: The Challenge of Mode of Interview Effects in Public Opinion Polls. 3,003 US adults randomly assigned to phone or web, 60 questions, 7 July to 4 August 2014.

Pew Research Center took a panel of people who normally answer on the web and randomly split them: 1,494 answered by telephone with an interviewer, 1,509 answered online. Same 60 questions, same four weeks of July and August 2014. Across the whole set the average difference was 5.5 percentage points.⁵

The differences were not spread evenly. They clustered on the questions where an answer might embarrass you in front of a stranger. Satisfaction with family life moved 18 points and social life 14. Saying that gay and lesbian people face a lot of discrimination ran 62 per cent by phone against 48 online; for Hispanic people 54 against 42; for black people 54 against 44. Describing your own neighbourhood as very safe after dark ran 55 against 43. On the question of whether women face a lot of discrimination there was no significant difference at all, which is the detail that makes the rest credible: an interviewer does not move every answer, only the ones with a socially preferred side.

Courtesy bias

Courtesy bias is the softening of a complaint to avoid giving offence to whoever is asking. It is the mechanism behind a pattern most teams have seen: the form handed over at the counter comes back kinder than the same form emailed a week later. Here it has been measured.

A study in a large referral hospital in Dar es Salaam asked women about disrespect and abuse during childbirth. Interviewed at the hospital just before discharge, 1,914 women were asked: 15 per cent reported at least one instance. A separate group of 64 women was interviewed in their communities four to six weeks after delivery: 70 per cent did. The research team also ran 197 direct observations of labour and delivery, which confirmed high rates of some of the behaviours being asked about.⁶

The two groups are very different sizes and the authors say plainly that more work is needed on how to measure this. Take the comparison as an order of magnitude, not a coefficient. As an order of magnitude it is brutal: where you ask changes what you are told, and asking inside the building being evaluated is close to the worst available option.

Response bias that reversed a result

Adida and colleagues had something almost nobody gets: official election results alongside panel survey data, village by village, from a field experiment in Benin. Respondents overreported both turning out to vote and voting for the incumbent, and the overreporting was worse where the question was more sensitive. That much is a measurement problem. What happened next is a decision problem. Read through the survey, the experimental treatment looked like it had worked. Read through the administrative records, it had done nothing.⁷

Nobody in that team did anything careless. The survey was well run. Response bias does not announce itself by making the data look messy; it makes the data look supportive.

Question order bias

Question order bias is a change in the answer caused only by where the question sits. Nothing about the wording changes. The measurement still moves.

Move a question and the answer moves with it

Two split-ballot experiments. October 2003: 45 per cent backed legal agreements for same-sex couples when that question followed one about marriage, 37 per cent when it did not. December 2008: 88 per cent said they were dissatisfied with the way things were going when the question came straight after a presidential approval question, 78 per cent when it came first.

Share giving each answer, by where the question sat in the questionnaire. Source: Pew Research Center, Writing Survey Questions. Split-ballot order experiments of October 2003 and December 2008.

Pew has run this as a split ballot repeatedly, giving half the sample one order and half the other. In October 2003, support for legal agreements giving same-sex couples the same rights as married couples came out at 45 per cent when the question followed one about marriage itself, and 37 per cent when it did not. Answers to the marriage question itself were not significantly affected by where it sat, so the effect ran one way.⁸

In December 2008, asking whether people were satisfied with the way things were going in the country produced 88 per cent dissatisfied when it came straight after a presidential approval question, against 78 per cent when it came first. A similar experiment in December 2004, when both satisfaction and approval were much higher, produced 57 per cent against 51. Order effects do not need a strong political climate to appear.

They also run in the other direction. In November 2008, 81 per cent said Republican leaders should work with the incoming president when that question followed one about Democratic leaders working with Republicans, against 66 per cent when it came first. Ask about Republicans first and support for Democratic cooperation fell from 82 per cent to 71. People answer the second question in a way that stays consistent with the first.

What to do about it

Randomising everything is not the fix, because it destroys comparability between waves. Group items by topic. Randomise only within a block of interchangeable items, such as a set of feature ratings. Keep screeners and anything that drives branching in a fixed position. Put a definition immediately before the section it applies to rather than only in the introduction. Then leave the order alone across waves, because a trend line that moves when the questionnaire is reshuffled is measuring the questionnaire.

Acquiescence bias

The agreement scale is the format that causes it, and two thirds of the rating questions in our own template library use one.

Acquiescence is the tendency to agree with a statement whatever it says. It is not a personality flaw, it is the path of least resistance through a grid of statements that all point the same way. And it is specifically a risk of the agree-or-disagree format, because that format asks people to endorse a proposition rather than to describe a thing.

Which brings us to our own library. We parsed every question in it on 10 September 2026: 109 live templates, 1,422 questions, of which 644 are rating items. Almost all of them, 98.9 per cent, use five points. There are 34 distinct end-anchor pairs in use. One pair accounts for 443 of the 644.

One agreement pair does two thirds of the work

Share of the 644 rating items in our own template library carrying each end-anchor pair. There are 34 distinct pairs; the five largest are shown. The top row is 443 items.

End-anchor pairs by share of rating items, SuperSurvey template library. Source: SuperSurvey Question Corpus, 109 live templates and 1,422 questions, read 11 September 2026.

So a little over two thirds of the rating questions we publish use the exact format most exposed to the bias this page is about. That is worth saying out loud, because the alternative was to write a guide about acquiescence and quietly not look.

The comparison has been made properly. Saris, Revilla, Krosnick and Shaeffer ran a between-subjects experiment inside representative sample surveys, combining random assignment with a multitrait-multimethod design, and found that responses to agree-or-disagree items "had much lower quality than responses to comparable questions offering item specific response options".⁹

Item-specific means the answer options describe the thing being asked about instead of describing how much you agree. "How easy was it?" with a scale running from very difficult to very easy, rather than "This was easy" with a scale running from strongly disagree to strongly agree. The multi-item index case, where agreement batteries still have a job, is covered under 5-point agreement scales.

Five of our own agreement items, rewritten

How easy was it to do what you came to the app to do?
Very difficultDifficultNeither easy nor difficultEasyVery easy
How welcome did you feel on your last visit?
Not at all welcomeSlightly welcomeModerately welcomeVery welcomeExtremely welcome
How good a use of your time was this session?
Very poor use of my timePoor useFair useGood useVery good use of my time
In the last month, how often were you treated with less respect than others on your team?
NeverOnceA few timesMost weeksMost days
Was the price you paid the same as the price you were shown?
Lower than shownThe sameSlightly higherMuch higherI did not see a price

Each one replaces an agreement item published, live, in our own templates today: 1 replaces "This app is easy to use", 2 "I feel welcome here", 3 "This session was worth the time it took", 4 "I am treated with the same respect as everyone else on my team", 5 "The price I paid was the price I had been shown". Source: SuperSurvey Question Corpus, 109 live templates and 1,422 questions, read 11 September 2026.

What makes a question biased

A biased question is one whose wording or answer options make one response easier to give than the others, before the respondent has thought about it. Five patterns produce most of them.

Five patterns, with the rewrite
Pattern	Biased	Neutral
A judgement built into the stem	How helpful was our friendly support team?	How helpful was our support team? (Not at all helpful to Extremely helpful)
Two questions in one (double-barrelled)	Was the training clear and useful?	Ask twice: How clear was the training? Then: How useful was it?
Unbalanced answer options	Excellent / Very good / Good / Poor	Excellent / Good / Neither good nor poor / Poor / Very poor
A recall window nobody can meet	How many times did you contact support last year?	In the last 30 days, how many times did you contact support? (None / 1 / 2 to 3 / 4 to 6 / 7 or more)
An assumption the respondent may not share	How long have you been using the reporting dashboard?	Have you used the reporting dashboard? Ask the how-long question only of the people who say yes.

The unbalanced row is the common one and the easiest to miss: three favourable options against one unfavourable option pulls the mean upward with no loaded word anywhere in sight.

Wording is where most preventable bias starts, and it is a large enough topic to have its own guide: how to write survey questions covers the stem, the options and the scale together.

Response bias is not the same as nonresponse bias

Response bias is about how people answer. Nonresponse bias is about who answers. They travel together and they need different remedies, so it is worth being able to tell which one you have.

An unusually clean measurement of the second one: a trial of a lifestyle intervention for people at high risk of cardiovascular disease invited 8,902 patients from 60 general practices. 1,489 responded, a response rate of 16.7 per cent. Because the invitations came from primary care records, the researchers knew a good deal about the 7,413 people who did not reply.¹⁰

The probability of responding fell as cardiovascular risk rose: an adjusted odds ratio of 0.82 for every five percentage points of QRisk2 score, with a 95 per cent interval of 0.77 to 0.88. It fell with deprivation too, to 0.52 (0.40 to 0.68) in the most deprived fifth of neighbourhoods against the least deprived. Black African or Caribbean patients were less likely to respond, 0.67 (0.45 to 0.98), while South Asian patients were no different from white patients at 1.08 (0.84 to 1.38). A trial designed for the people at highest risk was answered least by the people at highest risk.

The obvious inference from that is the wrong one. Groves and Peytcheva assembled 59 methodological studies that had each estimated nonresponse bias directly, and the relationship between a survey's nonresponse rate and the size of its bias turned out to be weak.¹¹ A high response rate is not a certificate and a low one is not a verdict. Estimate the bias against something you know about the people who did not answer, rather than reading it off the rate. Who ends up in the frame at all is a separate decision, covered under sampling methods.

How to reduce response bias before you send the survey

Six changes, in the order they pay off.

Ask about the thing, not about agreement

Replace "This was clear" with "How clear was it?" and label every point of the scale for clarity, from not at all clear to completely clear. There is no proposition left to endorse, so the acquiescence route closes entirely. It is also the only change on this list that alters what the question measures rather than the conditions it gets answered under. The label sets for 1-to-5 and 0-to-10 formats are under rating scale wording.
Take the human out of the room

Answers given to an interviewer and answers given alone are different answers, and the gaps concentrate on exactly the questions where a reply could embarrass someone: 5.5 percentage points on average across all 60 Pew questions, 18 on the most personal one, and no significant gap at all on a question with no socially preferred side. If the topic is sensitive and you have a choice of mode, take the interviewer out.
Say exactly what anonymous means, and then be it

Anonymous means you cannot link a response back to a person. Confidential means you can but you will not. They are different promises and respondents can tell when the wrong one is being made: an "anonymous" employee survey that asks for team, tenure and job title has identified everyone in a team of six. Pick the promise you can keep, describe it in one sentence, and stop collecting the fields that break it.
Put the sensitive questions where trust has been earned

Later, in almost every case. Our own library does this without being asked: demographic questions make up 2.9 per cent of all 1,422 questions, appear in only 26 of the 109 templates, and sit at a median of 83 per cent of the way through. The phrasing and the opt-out options are covered under demographic questions.
Balance the answer options and stop the grid repeating

Count the favourable and unfavourable options and make them match, and make sure the options do not overlap and leave nobody without a box: the rules for answer option design cover both. Then break up any run of statements that all read alike, because a long uniform battery is what produces habituation in the first place.

Do not add reverse-worded items just to catch people out. A negatively worded statement inside an otherwise positive block confuses a slice of honest respondents, and the confusion shows up as noise in the block, so a trap item costs you data from the people who were paying attention. If you are using a validated instrument, keep the reversed items it specifies, in the positions it specifies, and reverse-code them exactly as its scoring instructions say. Judge attention from several indicators together, such as completion time, straightlining and an answer you can check against a record, rather than from any single item.
Fix the order, then leave it alone

Decide the sequence once, randomise only inside blocks of interchangeable items, and keep it identical across waves. Every change you make to the order becomes a change in the trend line you cannot separate from a change in opinion.

How to spot response bias in your results

Prevention is cheaper, but the survey is usually already out. Five checks, in ascending order of effort. Every one of them flags rows to look at, not rows to remove: check, do not automatically delete. A respondent who is genuinely satisfied with everything gives uniform answers, and uniform answers on their own are not evidence of deception and not grounds for exclusion.

Look at the shape, not the mean. A mean of 4.2 built from a pile on 5 and a mean of 4.2 built from a spread are different results with the same summary. Publish the distribution beside every average.
Straightlining. Identical answers down a grid are a pattern worth opening, and nothing more until you have opened it. If the instrument carries a reversed item and the row agrees with that too, the row contradicts itself and deserves a closer look; if the grid is all one direction, a straight line can be an accurate report. Check, do not delete on sight.
Speed. Completions far below the median for the same questionnaire. Take a look before you do anything: fast is evidence, not proof, and a familiar questionnaire is answered quickly by people who mean every answer.
A record to check against. Attendance logs, ticket counts, incident reports, transaction data. Anything you already hold that overlaps with something you asked about. A gap between the record and the answers is the one check on this list that sizes the bias rather than hinting at it.
A condition to compare across. If some responses arrived anonymously and some identified, or some by phone and some online, compare the two groups as a diagnostic. Treat it as exactly that. People chose their own mode, so the groups can differ in who they are as well as in how they answered, and a gap between them cannot be attributed to the mode. Attributing cause needs random assignment, which is what Pew did with its 3,003 panelists, or an assignment mechanism outside the respondent's control that you can defend. Without one, report the gap as a warning sign and nothing stronger.

What not to do

Do not remove respondents until the answer looks right, and do not remove them on one indicator. Decide the exclusion rule before you look at what it does to the result, require at least two independent signals before a row goes, then report the figure with and without. Weighting is a fix for composition, not for tilt: it can correct a sample that is short of night-shift staff and it cannot correct night-shift staff who felt unable to answer honestly.

There is real work on statistical correction. Grimmond, Brown and Hawkins published a Thurstonian model in 2025 that separates response bias from the latent state it is contaminating.¹² That is a modelling exercise, not a filter you drop into a spreadsheet. For most teams the honest move is to report the limitation and fix the instrument for the next wave.

Frequently asked questions

Are answer bias, respondent bias and response bias the same thing?

Yes, in practice. "Answer bias", "respondent bias" and "responder bias" all get used for the same phenomenon, and none of them is a distinct technical term. The methodology literature settled on "response bias" decades ago, so that is the phrase to search with and the phrase to put in a write-up. The trap is the other way round: published papers sometimes use "response bias" in the title to mean who agreed to take part at all, which is a different thing entirely. Reference 10 below is one of them. Read the method, not the title.

Does a bigger sample reduce response bias?

No, and the arithmetic is worth seeing. Sampling error falls with the square root of the number of responses, so going from 400 replies to 1,600 halves your margin of error. Bias is a constant added to the estimate, so it is entirely unaffected by that. Quadruple the sample on a survey with a five-point tilt and you get a much tighter interval centred five points from the truth, which is worse than the small survey because it looks authoritative.

Does an anonymous survey remove social desirability bias?

It closes one channel and leaves another open. Anonymity removes the fear that a named answer will be traced back, and taking the interviewer away removes the wish to look reasonable to a stranger. What is left is the wish to look reasonable to yourself, and survey respondents overstate church attendance, charitable giving and their own likelihood of voting whoever is or is not listening.⁸

Anonymity is also a promise the respondent has to believe. An employee survey that collects team, tenure and job title is not anonymous, whatever the introduction claims, and people work that out faster than the people who wrote it expect.

What is neutral response bias, and should I drop the midpoint?

Neutral response bias is overuse of the middle option. Before removing it, work out which of three quite different things your midpoint is being used for: a genuine middling opinion, no opinion at all, or the question not applying to this person. Those need different treatment and one answer option is currently absorbing all three. Splitting "does not apply" out as its own option is the change to try first, because it takes out a group that was never neutral. Deleting the midpoint instead forces that group to pick a side they do not hold, which manufactures data rather than removing bias.

How do I report response bias?

Name the specific mechanism you think was at work, say which direction it pushes, and give the reader a bound if you can. "Respondents were invited by their line manager, so the satisfaction figures are more likely to be too high than too low" is useful. "Results may be subject to bias" is not, because it applies to every survey ever run and therefore tells nobody anything. If you have a comparison, an early-wave against late-wave split, or a record to check against, put the size of the gap in the limitations paragraph rather than the adjective.

How we counted

The library figures on this page come from parsing every question in the SuperSurvey template library on 10 September 2026: 109 live templates and 1,422 questions, of which 644 are rating items, 516 single-select, 221 open text and 41 multi-select. Only live template rows were counted. Drafts and retired rows were excluded. A question was classified as a rating item when both of its end anchors were labelled, which is how the 34 anchor pairs were counted; the 443 figure is items whose anchors read exactly "Strongly disagree" and "Strongly agree".

The demographic count is questions asking age, gender, tenure, location, income or similar, identified by matching the question text and its options. Position is the index of the question as a share of the way through its own template, so 83 means a median demographic question sits 83 per cent of the way down.

These counts describe what one survey vendor publishes as good practice. They are not a sample of the world's surveys, and nothing on this page depends on them being one. The published figures from Pew, Sando, Bayley, Adida, Saris and Groves are reproduced as their authors reported them and have not been recalculated here, with one exception: the 16.7 per cent response rate is 1,489 divided by 8,902 and is ours.

References

Cronbach, L. J. (1946). Response sets and test validity. Educational and Psychological Measurement, 6(4), 475-494. doi.org/10.1177/001316444600600405
Crowne, D. P., and Marlowe, D. (1960). A new scale of social desirability independent of psychopathology. Journal of Consulting Psychology, 24(4), 349-354. doi.org/10.1037/h0047358
McGrath, R. E., Mitchell, M., Kim, B. H., and Hough, L. (2010). Evidence for response bias as a source of error variance in applied assessment. Psychological Bulletin, 136(3), 450-470. doi.org/10.1037/a0019216
Rohling, M. L., Larrabee, G. J., Greiffenstein, M. F., Ben-Porath, Y. S., Lees-Haley, P., Green, P., and Greve, K. W. (2011). A misleading review of response bias: comment on McGrath, Mitchell, Kim, and Hough (2010). Psychological Bulletin, 137(4), 708-712. doi.org/10.1037/a0023327
Keeter, S. (2015). From Telephone to the Web: The Challenge of Mode of Interview Effects in Public Opinion Polls. Pew Research Center. 3,003 respondents, 60 questions, 7 July to 4 August 2014. pewresearch.org
Sando, D., Ratcliffe, H., McDonald, K., Spiegelman, D., Lyatuu, G., Mwanyika-Sando, M., Emil, F., Wegner, M. N., Chalamilla, G., and Langer, A. (2016). The prevalence of disrespect and abuse during facility-based childbirth in urban Tanzania. BMC Pregnancy and Childbirth, 16, 236. doi.org/10.1186/s12884-016-1019-4
Adida, C., Gottlieb, J., Kramon, E., and McClendon, G. (2019). Response bias in survey measures of voter behavior: implications for measurement and inference. Journal of Experimental Political Science, 6(3), 192-198. doi.org/10.1017/xps.2019.9
Pew Research Center. Writing Survey Questions. Methods primer, question order section. pewresearch.org/writing-survey-questions
Saris, W. E., Revilla, M., Krosnick, J. A., and Shaeffer, E. M. (2010). Comparing questions with agree/disagree response options to questions with item-specific response options. Survey Research Methods, 4(1), 61-79.
Bayley, A., Stahl, D., Ashworth, M., Cook, D. G., Whincup, P. H., Treasure, J., Greenough, A., Ridge, K., Winkley, K., and Ismail, K. (2018). Response bias to a randomised controlled trial of a lifestyle intervention in people at high risk of cardiovascular disease: a cross-sectional analysis. BMC Public Health, 18, 1092. doi.org/10.1186/s12889-018-5939-y
Groves, R. M., and Peytcheva, E. (2008). The impact of nonresponse rates on nonresponse bias: a meta-analysis. Public Opinion Quarterly, 72(2), 167-189. doi.org/10.1093/poq/nfn011
Grimmond, J., Brown, S. D., and Hawkins, G. E. (2025). A solution to the pervasive problem of response bias in self-reports. Proceedings of the National Academy of Sciences, 122(3), e2412807122. doi.org/10.1073/pnas.2412807122
SuperSurvey Question Corpus, 109 live templates and 1,422 questions, read 11 September 2026.

Michael Hodge: Survey methodologist and editor, SuperSurvey. Bachelor of Science (Psychology), University of Wollongong, with coursework in psychometrics and research methods. Designing surveys since 2003. About the author and how these guides are reviewed

What to read next

The rest of the survey learning centre covers the decisions on either side of this one.

Reduce response bias in your next survey

Item-specific wording, balanced options, a fixed order and demographics at the end. None of it removes bias entirely, and all of it is faster to do in the editor than in a document.

Create a survey with clearer questions

Free to send, no card. Or open one of the ready-made survey templates and edit the wording from there.