{"site":{"name":"Koji","description":"AI-native customer research platform that helps teams conduct, analyze, and synthesize customer interviews at scale.","url":"https://www.koji.so","contentTypes":["blog","documentation"],"lastUpdated":"2026-07-14T00:55:24.758Z"},"content":[{"type":"documentation","id":"92f86df3-a379-465a-9b26-05783fca3108","slug":"extreme-response-bias","title":"Extreme Response Bias: Why Some Respondents Always Pick the Extremes","url":"https://www.koji.so/docs/extreme-response-bias","summary":"Extreme response bias (ERS) is a trait-like tendency for some respondents to over-use the endpoints of a rating scale regardless of item content. Because it is systematic, not random, it does not average out with larger samples: it inflates top-box scores, distorts correlations, and manufactures false differences between demographic and cultural groups (documented from Marin et al. 1992 through De Jong et al. 2008 across 26 countries). Fix it with item-specific scales, within-respondent standardization, and by probing the reasoning behind extreme ratings.","content":"# Extreme Response Bias: Why Some Respondents Always Pick the Extremes\n\n**Extreme response bias — also called extreme response style, or ERS — is the tendency of some respondents to disproportionately choose the endpoints of a rating scale (\"strongly agree,\" \"strongly disagree,\" a 1 or a 10) regardless of what the question actually asks.** It is a *style*, not a considered reaction: a stable, person-level habit that shows up consistently across unrelated items. And because it is systematic rather than random, it does not wash out in a larger sample. It quietly widens your variance, inflates top-box scores, and can manufacture \"differences\" between groups that do not really exist.\n\nIf you run rating scales, Likert batteries, NPS, or CSAT, ERS is silently shaping your numbers. This guide explains what it is, how much damage it does, and how to design questions — and follow-ups — that capture genuine opinion instead of response style.\n\n## What extreme response bias actually is\n\nThe defining word is *style*. Extreme responding is treated in the methodology literature as **trait-like**: a disposition that persists across topics, question wording, and even across time. Longitudinal work has found that response-style indices show individual differences and cross-time stabilities comparable to commonly used personality measures. In other words, an extreme responder tends to stay an extreme responder — it is a property of the *person*, not of any single question.\n\nThat is what separates a *style* from a one-off error. Extreme responders are not necessarily more opinionated; they simply express themselves by reaching for the ends of whatever scale you give them. Two people who feel identically about your product can land on \"7 out of 10\" and \"10 out of 10\" purely because of how they use scales.\n\n## How ERS differs from acquiescence and central-tendency bias\n\nERS belongs to a family of \"response styles\" that also includes its close relatives:\n\n- **Acquiescence bias** (\"yea-saying\") is the tendency to *agree* with statements regardless of content. It is directional — it pushes answers toward the agree side. ERS is *non-directional*: an extreme responder is just as happy to slam \"strongly disagree.\" A person can have both tendencies, but they are distinct.\n- **Central-tendency (midpoint) bias** is the mirror image of ERS: respondents huddle on the neutral middle and avoid committing. Where ERS *widens* the apparent spread of your data, midpoint responding *compresses* it. (See [central tendency bias](/docs/central-tendency-bias) for the flip side of this coin.)\n\nBaumgartner and Steenkamp's influential taxonomy (*Journal of Marketing Research*, 2001) treats these as members of a single family, identifying five stylistic tendencies: acquiescence, disacquiescence, extreme response style, midpoint responding, and noncontingent responding. Recognizing which one is at work matters, because the fixes differ.\n\n## How much it distorts your data — the evidence\n\nExtreme responding is not a fringe concern; it is one of the most heavily documented biases in cross-cultural measurement.\n\n- **It differs sharply across cultural and demographic groups.** In four large pooled datasets, Marin, Gamba and Marin (1992) found that Hispanic respondents chose extreme responses more than non-Hispanic white respondents; their paper concludes, verbatim, that \"Hispanics prefer extreme responses to a greater extent than non-Hispanic Whites.\" Acculturation and education both reduced the effect.\n- **The gap can be large on wide scales.** A later study in *Survey Practice* (Cerda and Basar, 2010) found that combining four or five extreme answers across five questions, about **30% of Hispanic respondents** used the extremes versus **18% of non-Hispanics** — but the difference appeared only on 10-point scales, not on 5-point ones. Scale width itself changes how much ERS you get.\n- **The pattern repeats across nations.** Van Herk, Poortinga and Verhallen (2004) found extreme responding and acquiescence were both more pronounced in Mediterranean than in Northwestern Europe, with Greek respondents highest and British, German and French lowest. Crucially, the pattern did **not** match actual consumer behavior in those countries — strong evidence that the \"differences\" were measurement artifact, not real attitude gaps.\n- **It is measurable at global scale.** De Jong, Steenkamp, Fox and Baumgartner (2008) built an item-response-theory model of ERS on **12,506 consumers across 26 countries**, showing that different questions measure extreme responding better or worse, and that this varies by country.\n\nOne caveat worth stating honestly: the popular claim that ERS rises with age and falls with education is **contested**. The direction is often reported, but effect sizes are modest and some studies find little demographic correlation at all. Treat demographics as a weak predictor, not a rule.\n\n## Why a \"style\" is more dangerous than random noise\n\nRandom error shrinks as your sample grows. Systematic style does not — it biases every respondent who has it, in the same direction, every time. That produces three concrete harms:\n\n1. **It distorts correlations.** When two unrelated measures both carry a respondent's extreme-response signal, that shared style variance can manufacture a spurious correlation between them — or deflate a real one on reverse-worded items. Relationships you \"discover\" may be style, not substance.\n2. **It fabricates group differences.** If one segment responds more extremely as a *style*, any comparison of their means or top-2-box scores against another segment is confounded. The Van Herk finding — response patterns that did not track real behavior — is the cleanest demonstration that cross-group \"insights\" can be pure artifact.\n3. **It inflates or deflates top-box scores.** In NPS and CSAT reporting, \"% strongly agree\" or top-2-box is exactly the zone extreme responders overpopulate. A market that simply responds more extremely will look more enthusiastic — or more hostile — than it is, corrupting benchmarks and trend lines over time.\n\n## How to reduce extreme response bias\n\n### Design the scale to resist it\n\n- **Prefer item-specific scales over agree/disagree.** Agree/disagree Likert framing is the natural habitat of both acquiescence and ERS. A content-anchored scale (\"How satisfied were you?\" from *Very dissatisfied* to *Very satisfied*) gives respondents a real dimension to reason about rather than an invitation to reach for an endpoint.\n- **Pilot your scale width.** More points is not automatically safer — recall that the Hispanic/non-Hispanic gap appeared on 10-point but not 5-point scales. Test the format on your own audience before standardizing.\n- **Balance positively and negatively keyed items** so a pure endpoint-picker cannot produce a coherent score, which helps expose the pattern.\n- **Consider forced-choice formats** for high-stakes comparisons, which remove the endpoint option by making respondents choose between statements.\n\n### Correct for it in analysis\n\n- **Standardize within respondent** (ipsatization): center and scale each person's answers to their own mean and spread before comparing groups, removing individual-level style.\n- **Model style as a parameter** using latent-variable or IRT approaches (as De Jong et al. did) so you can partial it out of the substantive estimate. Be aware the correction model you choose changes the result — this is not a free lunch.\n\n### Ask why — the qualitative check\n\nThe most direct defense against misreading an extreme rating is simply to **ask the person to explain it.** A \"strongly disagree\" a respondent can back with a concrete reason is signal; one they cannot articulate is likely style. Probing the reasoning behind endpoint answers is something a static survey can never do on its own — and it is where modern tools change the game.\n\n## The modern approach: AI interviews that probe the extremes\n\nTraditional survey tools like SurveyMonkey or Qualtrics capture the *rating* but not the *reasoning* — you are left staring at a top-box number with no way to know whether it is genuine intensity or habitual extremeness. Following up on thousands of extreme ratings by hand is impossible, so teams simply don't.\n\nAn AI-native platform like Koji closes that gap. Koji combines [structured questions](/docs/structured-questions-guide) — including all six question types (open_ended, scale, single_choice, multiple_choice, ranking, and yes_no) — with an AI moderator that automatically **asks \"why\" behind a rating in the same session.** When a respondent gives a 10 or a 1, Koji probes for the concrete reason, so you can tell the person with a real story from the person who simply always picks the ends. Because the AI runs the same protocol with every respondent, it also removes moderator-driven variation, and it can conduct interviews in many languages — useful precisely because ERS varies so much across cultures. Teams using AI-assisted research report far faster time-to-insight than manual follow-up allows, turning a scale number into an explained, trustworthy signal in minutes rather than weeks.\n\nThe result: instead of correcting for extreme response bias after the fact with statistical guesswork, you prevent the misread at the source by capturing the *meaning* behind every extreme answer.\n\n## Related Resources\n\n- [Central Tendency Bias](/docs/central-tendency-bias) — the mirror image, where respondents cluster on the midpoint\n- [Acquiescence Bias](/docs/acquiescence-bias) — the \"yea-saying\" tendency ERS is often confused with\n- [Survey Response Bias](/docs/survey-response-bias) — the full family of respondent-side biases\n- [Likert Scale Research Guide](/docs/likert-scale-research-guide) — designing rating scales that hold up\n- [Cross-Cultural User Research](/docs/cross-cultural-user-research) — why response styles vary across markets\n- [Structured Questions Guide](/docs/structured-questions-guide) — the six Koji question types and when to use each\n- [Research Bias Guide](/docs/research-bias-guide) — the umbrella overview of bias in research\n","category":"Research Methods","lastModified":"2026-07-13T03:23:38.147527+00:00","metaTitle":"Extreme Response Bias (ERS): Why Respondents Pick the Extremes & How to Fix It","metaDescription":"Extreme response bias makes some respondents over-use scale endpoints regardless of the question, distorting averages and group comparisons. Learn the causes, the evidence, and how to fix it.","keywords":["extreme response bias","extreme response style","ERS","response style","rating scale bias","endpoint responding","survey bias"],"aiSummary":"Extreme response bias (ERS) is a trait-like tendency for some respondents to over-use the endpoints of a rating scale regardless of item content. Because it is systematic, not random, it does not average out with larger samples: it inflates top-box scores, distorts correlations, and manufactures false differences between demographic and cultural groups (documented from Marin et al. 1992 through De Jong et al. 2008 across 26 countries). Fix it with item-specific scales, within-respondent standardization, and by probing the reasoning behind extreme ratings.","aiPrerequisites":["Basic understanding of rating and Likert scales","Familiarity with survey design"],"aiLearningOutcomes":["Define extreme response bias and explain why a response style is not random noise","Distinguish ERS from acquiescence and central-tendency bias","Describe how ERS distorts correlations, group comparisons, and top-box scores","Apply scale design, statistical correction, and AI probing to capture real opinion"],"aiDifficulty":"intermediate","aiEstimatedTime":"10 min read"}],"pagination":{"total":1,"returned":1,"offset":0}}